author stp

1
 Matrices : Theory & Applications
Additional exercises
Denis Serre
École Normale Supérieure
de Lyon

Contents
Topics 3

Themes of the exercises 5

Exercises 23

Index 243

Notation
Unless stated otherwise, a letter k or K denoting a set of scalars, actually denotes a (commu-
tative) field.

Related web pages

See the solutions to the exercises in the book on
http://www.umpa.ens-lyon.fr/ ˜serre/exercises.pdf,
and the errata about the book on
http://www.umpa.ens-lyon.fr/ ˜serre/errata.pdf.
See also a few open problems on
http://www.umpa.ens-lyon.fr/ ˜serre/open.pdf.

2
Topics
Calculus of variations, differential calculus: Exercises 2, 3, 49, 55, 80, 109, 124, 184, 215,
249, 250, 292, 321, 322, 334, 371, 389, 400, 406, 408, 462, 464

Complex analysis: Exercises 7, 74, 82, 89, 104, 180, 245, 354, 391

Commutator: 81, 102, 115, 170, 203, 232, 236, 244, 256, 289, 306, 310, 315, 321, 368, 368,
369, 469, 470

Combinatorics, complexity, P.I.D.s: Exercises 9, 25, 28, 29, 45, 62, 68, 85, 115, 120, 134,
135, 142, 154, 188, 194, 278, 314, 315, 316, 317, 320, 332, 333, 337, 358, 381, 393, 410,
413, 425, 430, 443, 445, 457

Algebraic identities: Exercises 6, 11, 24, 50, 68, 80, 84, 93, 94, 95, 114, 115, 119, 120, 127,
129, 135, 144, 156, 159, 172, 173, 174, 203, 206, 242, 243, 244, 265, 288, 289, 335, 349,
379, 385, 388, 402, 415, 434, 440, 451, 461

Inequalities: Exercises 12, 27, 35, 40, 49, 58, 63, 69, 77, 82, 87, 101, 106, 110, 111, 125, 128,
139, 143, 155, 162, 168, 170, 182, 183, 198, 209, 210, 218, 219, 221, 222, 253, 254, 264,
272, 279, 285, 334, 336, 341, 361, 363, 364, 368, 371, 386, 419, 448, 458, 466, 471

Bistochastic or non-negative matrices: Exercises 9, 22, 25, 28, 76, 96, 101, 117, 138, 139,
145, 147, 148, 150, 152, 154, 155, 164, 165, 192, 193, 212, 230, 231, 251, 322, 323, 332,
333, 341, 347, 350, 365, 380, 383, 384, 389, 392, 394, 407, 421, 422, 433, 445, 468

Determinants, minors and Pfaffian: Exercises 8, 10,11, 24, 25, 50, 56, 84, 93, 94, 95, 112,
113, 114, 119, 120, 127, 129, 143, 144, 146, 151, 159, 163, 172, 181, 188, 190, 195, 206,
207, 208, 213, 214, 216, 221, 222, 228, 234, 246, 270, 271, 272, 273, 274, 278, 279, 285,
286, 290, 291, 292, 308, 335, 336, 349, 357, 372, 374, 383, 384, 385, 386, 387, 393, 400,
414, 424, 427, 430, 434, 446, 451, 453, 454, 457, 461, 462, 463, 467

Hermitian or real symmetric matrices: Exercises 12, 13, 14, 15, 16, 27, 40, 41, 51, 52, 54,
58, 63, 70, 74, 75, 77, 86, 88, 90, 92, 102, 105, 106, 110, 113, 116, 118, 126, 128, 143, 149,
151, 157, 164, 168, 170, 180, 182, 184, 192, 198, 199, 200, 201, 209, 210, 211, 215, 216,
217, 218, 219, 223, 227, 229, 230, 231, 239, 241, 248, 258, 259, 263, 264, 271, 272, 273,
274, 279, 280, 285, 291, 292, 293, 294, 301, 307, 308, 312, 313, 319, 326, 328–330, 334,
336, 347, 351, 352, 353, 355, 360, 361, 364, 371, 376, 382, 383, 386, 389, 396, 400, 404,
405, 406, 408, 409, 412, 413, 423, 424, 426, 428, 441, 446, 447, 448, 450, 461, 464, 465,
466, 467, 469, 474

Orthogonal or unitary matrices: Exercises 21, 64, 72, 73, 81, 82, 91, 101, 116, 158, 226,
236, 255, 257, 276, 277, 281, 297, 302, 314, 343, 344, 348, 357, 358, 366, 395, 417, 426,
437, 442, 444, 450, 471, 475, 476

Norms, convexity: Exercises 7, 17, 19, 20, 21, 40, 65, 69, 70, 71, 77, 81, 82, 87, 91, 96, 97,
100, 103, 104, 105, 106, 107, 108, 109, 110, 111, 117, 125, 128, 131, 136, 137, 138, 140,

3
 141, 153, 155, 162, 170, 183, 199, 223, 227, 242, 243, 261, 297, 299, 300, 301, 302, 303,
313, 323, 361, 366, 368, 380, 408, 411, 420, 455, 460, 464, 472

Eigenvalues, characteristic or minimal polynomial: Exercises 22, 31, 37, 41, 47, 61, 69,
71, 83, 88, 92, 98, 99, 101, 113, 116, 121, 123, 126, 130, 132, 133, 145, 146, 147, 156, 158,
168, 175, 177, 185, 186, 187, 194, 195, 205, 252, 256, 260, 267, 269, 276, 277, 294, 295,
304, 305, 339, 340, 344, 359, 362, 367, 373, 376, 379, 384, 398, 403, 418, 438, 439, 441,
443, 449, 455, 456, 458, 474

Diagonalization, trigonalization, similarity, Jordanization: Exercices 4, 5, 85, 100, 187,

189, 191, 205, 212, 240, 282, 284, 287, 324, 325, 326–330, 338, 339, 340, 345, 359, 360,
378, 395, 426, 431, 432, 444, 450, 456, 469

Singular values, polar decomposition, square root: Exercises 30, 46, 51, 52, 69, 82, 100,
103, 108, 109, 110, 111, 139, 147, 162, 164, 185, 198, 202, 226, 240, 254, 261, 262, 281,
282, 296, 304, 354, 363, 370, 394, 395, 396, 401, 417, 450, 466, 476

Classical groups, exponential: Exercises 30, 32, 48, 53, 57, 66, 72, 73, 78, 79, 106, 123, 148,
153, 176, 179, 196, 201, 202, 241, 246, 268, 295, 316, 321, 436

Numerical analysis: Exercises 33, 42, 43, 65, 67, 140, 194, 197, 344, 353, 355, 392, 401, 409

Matrix equations: Exercises 16, 34, 38, 59, 60, 64, 166, 167, 237, 238, 319, 325, 331, 399,
435, 442

Other real or complex matrices: Exercises 26, 35, 39, 46, 80, 103, 118, 131, 166, 171, 185,
220, 221, 232, 242, 243, 245, 247, 249, 250, 253, 268, 269, 284, 290, 298, 303, 304, 306,
309, 311, 324, 331, 351, 356, 360, 369, 373, 391, 403, 418, 429, 433, 438, 452, 454, 459

Differential equations: Exercises 34, 44, 122, 123, 124, 145, 195, 196, 200, 236, 241, 263,
295, 365, 418

General scalar fields: Exercises 1, 2, 4, 5, 6, 10, 18, 23, 36, 115, 133, 160, 161, 167, 169, 173,
174, 178, 179, 181, 186, 187, 204, 206, 207, 213, 214, 224, 225, 237, 238, 239, 244, 256,
265, 266, 283, 288, 289, 310, 316, 317, 318, 320, 339, 340, 345, 349, 359, 372, 375, 377,
390, 399, 427, 430, 431, 432, 435, 444, 463, 470, 475

Algorithms: 11, 24, 26, 42, 43, 67, 88, 95, 112, 344, 353, 355, 392, 397

Factorizations. 41, 344, 349, 351, 370

Just linear algebra: 233, 235, 375, 382, 388

4
More specific topics
Numerical range/radius: 21, 65, 100, 131, 223, 253, 269, 306, 309, 356, 369, 418, 419, 437,
439, 452

H 7→ det H, over SPDn or HPDn : 58, 75, 209, 218, 219, 271, 292, 308, 408

Contractions: 82, 104, 220, 221, 272, 300, 380

Hadamard inequality: 279, 285, 404, 405, 414, 471

Hadamard product: 250, 279, 285, 322, 335, 336, 342, 343, 371, 383, 389

Functions of matrices: 51, 52, 63, 74, 80, 82, 107, 110, 111, 128, 179, 249, 250, 280, 283,
334, 354

Commutation: 38, 115, 120, 202, 213, 237, 238, 255, 256, 277, 281, 297, 345, 346, 435, 470

Weyl’sWeyl inequalities and improvements: 12, 27, 69, 168, 275, 363

Hessenberg matrices: 26, 113, 305, 443

Themes of the exercises

1. Similarity within various fields. IX

2. Rank-one perturbations. III

3. Minors ; rank-one convexity.

4, 5. Diagonalizability. III

6. 2 × 2 matrices are universaly similar to their matrices of cofactors. III

7. Riesz–ThorinRieszThorin by bare hands. VII

8. Orthogonal polynomials. III

9. Birkhoff’sBirkhoff Theorem ; wedding lemma. VIII

10. Pfaffian ; generalization of Corollary 7.6.1. X

11. Expansion formula for the Pfaffian. Alternate adjoint. IIII

12. Multiplicative Weyl’sWeyl inequalities. VI

13, 14. Semi-simplicity of the null eigenvalue for a product of Hermitian matrices ; a contrac-
tion. VI

5
15, 16. ToeplizToepliz matrices ; the matrix equation X + A∗ X −1 A = H. VI

17. An other proof of Proposition 3.4.2. VI

18. A family of symplectic matrices. X

19, 20. Banach–MazurBanachMazur distance. VII

21. Numerical range.

22. JacobiJacobi matrices ; sign changes in eigenvectors ; a generalization of Perron-FrobeniusPerronFrobenius

for tridiagonal matrices. VIII

23. An other second canonical form of square matrices. IX

24. A recursive algorithm for the computation of the determinant. III

25. Self-avoiding graphs and totally positive matrices. VIII

26. SchurSchur parametrization of upper HessenbergHessenberg matrices. III

27. Weyl’sWeyl inequalities for weakly hyperbolic systems. VI

28. SL+
2 (Z) is a free monoid.

29. Sign-stable matrices. V

30. For a classical group G, G ∩ Un is a maximal compact subgroup of G. X

31. Nearly symmetric matrices. VI

32. Eigenvalues of elements of O(1, m). VI, X

33. A kind of reciprocal of the analysis of the relaxation method. XII

34. The matrix equation A∗ X + XA = 2γX − In . VI, X

35. A characterization of unitary matrices through an equality case. V, VI

36. Elementary divisors and lattices. IX

37. Companion matrices of polynomials having a common root. IX

38. Matrices A ∈ M 3 (k) commuting with the exterior product.

39. Kernel and range of In − P ∗ P and 2In − P − P ∗ when P is a projector. V

40. Multiplicative inequalities for unitarily invariant norms. VII

41. Largest eigenvalue of Hermitian band-matrices. VI

6
42, 43. Preconditioned Conjugate Gradient. XII

44. Controllability, Kalman’sKalman criterion. X

45. NashNash equilibrium. The game “scissors, stone,. . . ”. V

46. Polar decomposition of a companion matrix. X

47. Invariant subspaces and characteristic polynomials. III

48. Eigenvalues of symplectic matrices. X

49. From Compensated-Compactness. V

50. Relations (syzygies) between minors. III

51. The square root map is analytic. VI

52. The square root map is (globally) Hölderian. VI

53. Lorentzian invariants of electromagnetism. X

54. Spectrum of blockwise Hermitian matrices. VI

55. Rank-one connections, again. III

56. The transpose of cofactors. Identities. III

57. A positive definite quadratic form in Lorentzian geometry. X

58. A convex function on HPDn .

59. When elements of N + Hn have a real spectrum. VI

60. When elements of M + Hn are diagonalizable. VI

61. A sufficient condition for a square matrix to have bounded powers. V, VII

62. Euclidean distance matrices. VI

63. A Jensen’sJensen trace inequality. VI

64. A characterization of normal matrices. V

65. Pseudo-spectrum. V

66. Squares in GLn (R) are exponentials. X

67. The Le Verrier–Fadeevlever@Le VerrierFadeev method for computing the characteristic

polynomial. III

7
68. An explicit formula for the resolvent. III

69. Eigenvalues vs singular values (New formulation.) IV, VI

70. HornHorn!Roger & Schur’sSchur theorem. VI

71. A theorem of P. LaxLax. VI

72, 73. The exchange map. III

74. Monotone matrix functions. VI

75. S 7→ log det S is concave, again.

76. An application of Perron–FrobeniusPerronFrobenius. VIII

77. A vector-valued form of the Hahn–BanachHahnBanach theorem, for symmetric operators.

VII

78. Compact subgroups of GLn (C). X

79. The action of U(p, q) over the unit ball of Mq×p (C). X

80. Balian’sBalian formula. VI

81. The “projection” onto normal matrices. VI

82. Von Neumannvonneu@von Neumann inequality.

83. Spectral analysis of the matrix of cofactors. III

84. A determinantal identity. III

85. Flags.

86. A condition for self-adjointness. VI

87. Parrott’sParrott Lemma. VII

88. The signs of the eigenvalues of a Hermitian matrix. VI

89. The necessary part in Pick’sPick Theorem. VI

90. The borderline case for PickPick matrices. VI

91. Normal matrices are far from triangular. V

92. Tridiagonal symmetric companion matrix. III, V

93. An extension of the formula det A = (PfA)2 . III

8
94. The Pfaffian of A + x ∧ y. III

95. A formula for the inverse of an alternate matrix. III

96. Isometries of (Rn ; `p ) when p 6= 2. VII

97. Iteration of non-expansive matrices. VII

98. The eigenvalues of a 4 × 4 alternate matrix. III

99. An orthogonal companion matrix. V

100. The numerical range of a nilpotent matrix. V

101. The trace of AB, when A and B are normal. V

102. Compensating matrix. VI

103. A characterization of singular values. VII

104. The assumption in von Neumannvonneu@von Neumann Inequality cannot be weakened.

V

105. An inequality for Hermitian matrices. VI

106. The LipschitzLipschitz norm of A 7→ eiA over Hermitian matrices. X

107. When (A ∈ Hn+ ) =⇒ (kA2 k = kAk2 ) VI, VII

108. Convexity and singular values. VII

109. Rank-one convexity and singular values. V

110. The square root in operator norm. VI, VII

111. And now, the cubic root. VII

112. A criterion for a polynomial to have only real roots. VI

113. Some non-negative polynomials on Symn (R). VI

114. Invariant factors of the matrix of cofactors. IX

115. About ω-commuting matrices. V

116. How far from normal is a product of Hermitian matrices. VI

117. The extremal elements among symmetric, bistochastic matrices. VIII

118. “Non-negative” linear forms on Mn (C). V, VI

9
119. HilbertHilbert matrices. III

120. When blocks commute, the determinant is recursive. III

121. The characteristic polynomial of some blockwise structured matrix. III

122. The GreenGreen matrix for ODEs. X

123. Stable, unstable and neutral invariant subspaces.

124. The Lopatinskiı̆Lopatinskiı̆ condition in control theory. X

125. A trace inequality. V

126. Can we choose a tridiagonal matrix in Theorem 3.4.2 ? VI

127. 4 × 4 matrices. III

128. The Hölderhold@Hölder inequality for A 7→ Aα . VI

129. A determinantal identity. III

130. A disk “à la Gershgorin”Gershgorin. V

131. Making the diagonal constant. V

132. Connected components of real matrices with simple eigenvalues. V

133. Characteristic and minimal polynomials. IX

134. The converse of Proposition 8.1.3. XI

135. Winograd’s computationWinograd.

136. Unitary invariant norms and adjunction. VII

137, 138. Convex subsets of the unit sphere of Mn (R). VIII

139. An other inequality from von Neumannvonneu@von Neumann. XI

140. Joint spectral radius. VII

141. Bounded semigroups of matrices. VII

142. Commutators and palindromes. III

143. Determinantal inequalities for SDPn ToeplizToepliz matrices. VI

144. Generalized Desnanot–JacobiDesnanotJacobi formula. III

145. An entropy inequality for positive matrices. VIII

10
146. Exterior power of a square matrix.

147. The spectrum of a totally positive matrix. VIII

148. Totally non-negative semi-groups. VIII

149. More about Euclidean distance matrices. VI

150. Farkas’ LemmaFarkas. VIII

151. The hyperbolic polynomial X02 − X12 − · · · − Xr2 . V, VI

152. Another proof of Birkhoff’s Theorem. VIII

153. SO2 (R) and O−

2 (R) are linked inside the unit sphere of M2 (R). X

154. The combinatorics of ∆3 as a polytope. VIII

155. Estimates of the joint spectral radius. VIII

156. The characteristic polynomial of a cyclic matrix. VIII, IX

157. Flat extension of a Hermitian matrix. VI

158. Spectrum of a normal matrix and of its principal submatrices. V

159. The characteristic polynomial of a cyclic matrix (bis). III

160. Schur’s LemmaSchur. III

161. Isomorphic but not conjugated nilpotent sub-algebras of Mn (k). III

162. The unitarily invariant norms. VII

163. The mapping X 7→ AT XA on the space of symmetric matrices. III

164. Spectral gap for tridiagonal symmetric stochastic matrices. VIII

165. Primitive non-negative matrices. VIII

166. Solvability of AXB = C. IX, XI

167. Solvability of AX − XB = C. III

168. The spectrum of A + B when A, B are Hermitian and B is small (a case of the A.
Horn’sHorn!Alfred problem). IV, VI

169. Theorem 6.2.1 is not valid in rings that are not PID. IX

170. HeisenbergHeisenberg Inequality. VI

11
171. Lattices in Cm .

172. Schur’s Pfaffian identity.Schur III

173. Hua IdentityHua, Loo-Keng. III

174. JordanJordan!Pascual and ThedyThedy Identities for symmetric matrices. III

175. Diagonalization in the reals and projectors. V

176. Exponential in a JordanJordan!Pascual algebra. X

177. A characterization of the minimal polynomial. IX

178. The rank of SchurSchur complement. III

179. Group-preserving functions. X

180. Hyperbolic Hermitian pencils. VI

181. The reciprocal of Exercise 127. III

182. Sums of eigenvalues of a Hermitian matrices. VI

183. The p-norms of a matrix and its absolute value. VII

184. Inf-convolution. VI

185. A generalization of the fundamental Lemma of Banach algebras. VII

186. The characteristic polynomial and the invariant factors of submatrices. IX

187. The DunfordDunford decomposition.

188. The SmithSmith determinant. III

189. Similar permutation matrices. III

190. The determinant of M T M . III

191. Matrices that are conjugated to their inverse. III, IX

192. Two kinds of positivity. V

193. A criterion for the spectral radius of non-negative matrices. VII

194. An improvement in Le Verrier’slever@Le Verrier method.

195. Strongly stable matrices. V

196. A formula for exp(tA). X

12
197. Symmetric eigenvalue problem: The order of the JacobiJacobi method when n = 3. XIII

198. The geometric mean of positive semi-definite Hermitian matrices. VI

199. Interpolation of Hermitian spaces. VI

200. Homogenization of elliptic operators. VI

201. Hamiltonian matrices. X

202. The exponential and the index. X

203. A polynomial identity in Mn (k). IV

204. Modular subgroups of GLn (Z) are torsion free (with one exception).

205. Partial similarity and characteristic polynomial. IX

206. An identity in M3 (k).

207. Rank and kernel. IIII

208. A determinantal identity. III

209. H 7→ (det H)1/n is concave over HPDn .

210. A scalar inequality for vectors, which implies an inequality between Hermitian matrices.
VI

211. An extension problem for Hermitian positive semi-definite matrices. VI

212. Conjugation classes in SL2 (Z).

213. Tensor product and determinant (I). III

214. Tensor product and determinant (II). III

215. Differentiation over Symn . VI

216. A converse of Exercise 13 of Chapter 3. VI

217. Diagonalizability of pencils of real quadratic forms.VI

218. H 7→ (det H)1/n is concave over HPDn (bis).

219. H 7→ (det H)1/n is concave over HPDn (ter)

220. Symmetry group of the unit ball of Mp×q (C). X

221. Hua’s InequalityHua, Loo-Keng. V, VI

13
222. Eigenvalues vs singular values (II). IV, XI

223. A convex range for quadratic maps. V

224. Nilpotence. III

225. The dimension of nilpotent subspaces. III

226. The AluthgeAluthge transform. X

227. Computation of a convex envelop. VI

228. The topology of Alt4 (R) ∩ GL4 (R). III, VI

229. SchurSchur complement for positive semi-definite matrices. VI

230, 231. Doubly stochastic n-tuples. VIII

232. A matricial equation. V

233. Search of an adapted orthogonal basis in R3 .

234. The Pfaffian of simple alternate matrices. III

235. Adapted bases.

236. Isospectral flows. X

237, 238. The commutant of a square matrix. III

239. Sums of squares. VI

240. Unitarily similar matrices. V

241. Symmetry of a resolvant. VI, X

242. ZariskiZariski closure of the unit sphere of (M2 (R), k · k2 ). VII

243. The algebraic nature of the unit sphere of (Mn (R), k · k2 ). VII

244. A polynomial identity in M2 (k). III

245. Pencils in GLn (C). V

246. A determinantal identity for orthogonal matrices. V

247. Small subgroups of GLn (C) are finite. VII, X

248. A positive definite Hermitian matrix. VI

249. Differentials of power and exponential. V, X

14
250. Differentiating matrix functions. V

251. The strong PerronPerron–FrobeniusFrobenius property. VIII

252. Strange minimal polynomial ! IX

253. The power inequality for the numerical radius. VII

254. A generalized CauchyCauchy–SchwarzSchwarz inequality. V

255. Fuglede’sFuglede’s theorem and its consequences. V

256. Commutators and nilpotence. III

257. A rational map from Un to Um . X

258. The product of Hermitian matrices of which one is positive definite. VI

259. The spread of a Hermitian matrix. VI

260. The eigenvalues of A−1 A∗ . X

261. Projection onto the unit ball of Mn (C). VII, X

262. An algorithm for the polar decomposition. X

263. Self-adjoint differential equations. X

264. A “Cauchy-Schwarz inequality” for the geometric mean. VI

265. Near generalized inverses. XI

266. Polynomial group action over nilpotent matrices. III

267. Matrices whose roots have integral entries. III

268. A characterization of finite subgroups of Un . X

269. The convex hull of the spectrum vs the numerical range. V

270. Matrices whose roots have integral entries (bis). III

271. A generalization of Hua’s InequalityHua, Loo-Keng. V, X

272. Bellman’sBellman positive matrix. VI

273. Symplectic matrices and SiegelSiegel domain. X

274. Some 3 × 3 matrices, and planar triangles. V

275. The determinant of a sum in HPD2 . VI

15
276. The diagonal and the spectrum of normal matrices. V

277. The property L. III

278. Sylvester’sSylvester Lemma. III

279. Oppenheim’sOppenheim (Sir A.) Inequality. VI

280. Positive unital linear maps. VI

281. Normality of A, B, AB and BA. V

282. Unitary and orthogonal similarity. V, X

283. Matrix version of LagrangeLagrange interpolation. IX

284. Invariant planes for real matrices. V

285. An improvement of the Oppenheim’sOppenheim (Sir A.) Inequality. VI

286. A snail determinant. III

287. Matrices being diagonalizable within Mn (Fp ). III, IX

288. A preparation for the next one.

289. The Amitsur–LevitzkiAmitsurLevitzki Theorem (after S. RossetRosset!Shmuel).

290. The determinant and trace for positive definite matrices. V

291. The theorem of CraigCraig & SakamotoSakamoto. VI

292. The LegendreLegendre transform of H 7→ − log det H. VI

293. Ando’sAndo supremum of Hermitian matrices. VI

294. Stability issues in HamiltonianHamilton systems. X

295. The exponential as a polynomial. X

296. The negative second moment identity. V

297. Von Neumann’svonneu@von Neumann proof of Fuglede’sFuglede theorem. V

298. Embeddings from Mm (C) into Mn (C). V

299. The operator norm of a nilpotent matrix of order two. VII

300. Contractions and ToeplizToepliz matrices. VI, VII

301. Positive operators that sum up to the identity. VI

16
302. A characterization of unitary matrices. V

303. Anticommutative systems. VII

304. Singular values vs restriction over subspaces. XI

305. Unit HessenbergHessenberg matrix with prescribed RitzRitz values. III

306. Equivalence classes under unitary conjugation in M2 (C). V

307. From a projection to an orthogonal projection. V, VI

308. Conjugation of a convex function over Hn . VI

309. The resolvant and the numerical range. V

310. A characterization of nilpotent matrices. III

311. The real part of a nilpotent matrix. V

312. The orthogonal of a positive definite Hermitian matrix. VI

313. The method of alternate projections. V, VI

314. HadamardHadamard matrices of size 2m . III

315. Zero trace matrix is a commutator. III

316. The unipotent group. III

317. Non-isomorphic groups whose elements have the same orders. III

318. Elements in GLn (Z) ∩ (In + pMn (Z)). III

319. Matrices AB + BA with A, B ∈ HPDn . VI

320. Subsets with odd cardinals and even intersections. III

321. Formula for eA+B when [A, B] = A. X

322. Generators of MarkovianMarkov semi-groups. VIII

323. ‘Log-convexity’ of the Perron–FrobeniusPerronFrobenius eigenvalue. VIII

324. A simple algebra. III, IV, VI

325. The Cecioni–FrobeniusCecioniFrobenius theorem. IX

326. Symmetric matrices that are compatible, in the sense of linear elasticity. VI

17
327–330. HornHorn!Roger & Johnson’sJohnson!Charles R. Theorem 4.1.7: matrices similar
to real ones, matrices similar to their Hermitian adjoint, products of Hermitian matrices.
V, VI

331. Sums of squares in Mn (R). V

332. Sums of permutation matrices. VIII

333. The maximum of diagonal sums. VIII

334. Operator monotone functions; LoewnerLoewner’s Theorem. VI

335. The relative gain array (after C. R. JohnsonJohnson!Charles R. & H. ShapiroShapiro).

III

336. The relative gain array. Positive definite case. III, VI

337. A bound for the permanent. III

338. Similarity in GL2 (Z). III, IX

339. Diagonalizable “companion” matrix. IX

340. Test for diagonalizability. III

341. The inverse of a strictly diagonally dominant matrix. V, XII

342. The relative gain array. Strictly diagonally dominant case. V

343. The relative gain array. The case of permutation matrices. VIII

344. Shifted QR method (Francis’Francis algorithm.) XIII

345. Matrix commuting with a non-derogatory matrix. IX

346. Matrix commuting with a non-derogatory matrix and its transpose. V, IX

347. Semipositive definite symmetric matrices that have non-negative entries. VI

348. Parametization of SO3 (R) by unitary quaternions. X

349. Explicit LU factorization. XI

350. Tridiagonal bistochastic matrices. VIII

351. The instability of factorization M = Hn · Hn for M ∈ Mn (R). V, VI, VII

352. Maximal positive / negative subspaces of a Hermitian matrix and its inverse. VI

353. Irrelevance of the choice of angle in the method of JacobiJacobi. XIII

18
354. A square root cannot be defined over GL2 (C). V, X

355. The Gauss–SeidelGaussSeidel algorithm for positive semi-definite matrices. XII

356. The numerical range of A and A−1 . V

357. In orientable Riemannian geometry, the HodgeHodge star operator is an isometry. X, XI

358. Five-fold symmetry of lattices (after M. Křı́žek, J. Šolc and A. ŠolcováKriz@Křı́žeksolc@Šolcsolco@Šolcov

III

359. The similarity of a matrix and its transpose (after O. Taussky and H. ZassenhausTausskyZassenhaus).
III, IX

360. When a real matrix is conjugate to its transpose through a symmetric positive definite
matrix. V, VI, IX

361. The convex cone spanned by the squares of skew-symmetric matrices. V

362. Computing the multiplicity of an eigenvalue from a polynomial equation p(M ) = 0. III

363. Singular values vs eigenvalues. V, XI

364. Tr eH eK vs Tr eH+K when H, K are Hermitian. VI

365. Linear differential equation whose matrix has non-negative entries. X

366. The supporting cone of SOn (R) at In . V

367. The spectrum of (B, C) 7→ ((BX − XC)X T , (X T (XC − BX)). V

368. A proof of the Böttcher–Wenzelbottcher@BöttcherWenzel Inequality. VII

369. The numerical radius of a commutator. V

370. Cholesky vs polar factorization. X, XI

371. LoewnerLoewner’s Theory of operator monotone functions. VI

372. Subspaces formed of singular matrices. III

373. A problem about coins and weights.

374. The rank of some integer matrices. III

375. Mixing two idempotent matrices. III

376. Spectrum and the equation M = A + A∗ . V, VI

377. Finding 0n by multiplying factors sA + tB, (s, t) 6= (0, 0). V

19
378. Polynomials P such that P (A) is diagonalizable for every A. IX

379. A universal annihilator on Mn (Fp ). III

380. Non-negative matrices and the Hilbert distance. VIII

381. Left- and right- annhilators of A ∈ Mn (R) where R is finite.

382. Linear forms over quadratic forms.

383. The cone of non-negative symmetric matrices with positive entries. VIII

384. The determinant and the permanent as eigenvalues. III

385. FrobeniusFrobenius determinant formula. III, V

386. An other proof of HadamardHadamard Inequality. XI

387. An integral fraction. III

388. Special composition of symmetries. III

389. A nonlinear eigenvalue problem. VI, VIII

390. Idempotent matrices that are sums of idempotent matrices. III

391. A third proof of Fuglede’sFuglede Theorem. VI, X

392. Iterative method for a linear system: the case of positive matrices. VIII, XII

393. A matrix related to Boolean algebra. III

394. Singular values of bistochastic matrices. VIII, XI

395. T ∗ T and T T ∗ are unitarily similar. V, XI

396. A decomposition lemma in H+

n . VI

397. Numerical range vs diagonal. V

398. Achieving a prescribed characteristic polynomial. III

399. The WaringWaring problem for matrices.

400. A proof that S 7→ log det S and S 7→ (det S)1/n are concave over SDPn . VI

401. Numerical analysis of the polar decomposition. X

402. An identity involving the cofactor matrix. III

403. Geometric multiplicity vs GershgorinGershgorin disks. V

20
404. HadamardHadamard product vs A 7→ AA∗ . V

405. HadamardHadamard factorization of positive semidefinite matrices. VI

406. Yosida approximation of S 7→ − log det S. VI

407. M -matrices. V, VIII

408. Convex conjugate of H 7→ (det H)1/n . VI

409. Inconditional convergence of the diagonal in the method of JacobiJacobi. XIII

410. Lightning. III

411. The p-norm of circulant matrices. VII

412. A FarkasFarkas Lemma for symmetric matrices. VI

413. Unit distance representation of a graph in Euclidian space. VI

414. A maximization problem in the unit disk. V

415. Polynomial identities over alternate matrices of size n ≤ 4; after B. KostantKostant and
L. H. RowenRowen. III

416. A variant of Cayley–HamiltonCayleyHamilton Theorem for alternate matrices. III

417. Singular values of blocks of LorentzLorentz transformations. X, XI

418. The product of two monotone matrices. V

419. The spread of the diagonal of a normal matrix. V

420. Lewis’ TheoremLewis. VII

421. Wielandt’sWielandt Theorem for positive primitive matrices. VIII

422. Sharpness in Wielandt’sWielandt Theorem. VIII

423. Monotonicity of S 7→ Ŝ over SPDn . VI

424. A determinantal inequality for three positive symmetric matrices. VI

425. FibonacciFibonacci numbers in the powers of a 0, 1-matrix. III

426. A convex body in Sym3 contained in | det | ≤ 1. VI

427. Can a vector space be the finite union of proper subspaces ?

428. A family of positive semi-definite matrices. VI

21
429. A parametrization of 2 × 2 matrices with real entries. V, VII

430. Large random matrix with prescribed determinant in Fq . III

431. The Jordan form of the square of a Jordan block. IX

432. Similar permutation matrices (Brauer). VIII, IX

433. The positive rank of a non-negative matrix. VIII

434. A CapelliCapelli identity (2 × 2) case.

435. Matrices commuting with A + B and AB. III

436. Isomorphic orthogonal groups. X

437. Unitary matrices vs numerical radius. V

438. Normality of A, B, AB and BA (bis). V

439. Eigenvalue on the boundary of the numerical range. V

440. A polynomial formula for the adjugate matrix. III

441. The characteristic polynomial of S 7→ X T SX. III

442. An equation in the unitary group. V

443. Computing the characteristic polynomial of real or complex matrices. III, V

444. Are A and AT orthogonaly similar ? III, V

445. The orthogonal of bistochastic matrices.

446. Positive definite completion of a symmetric tridiagonal matrix.

447. Strang transform.

448. Comparing P 2 and (P + Q)2 when P and Q are orthogonal projections.

449. The maximal order in GLn (Fp ).

450. Real square roots of symmetric matrices.

451. The determinant of some complex and real matrices.

452. Ch. Davis’Davis proof of Toeplitz–HausdorffToeplitzHausdorff Theorem.

453. A Pfaffian calculation.

454. A maximal cone over which det ≥ 0.

22
455. A proof of Banach’sBanach Formula using the Cayley–HamiltonCayleyHamilton Theo-
rem.

456. The persistance of eigenvalues after rank-one perturbations (after HörmanderHörmander

& MelinMelin).

457. Elementary divisors of some integerwise matrix.

458. An estimate of the spread of a complex matrix.

459. Normal completion.

460. The cut-norm.

461. The Cayley–Menger determinant.

462. Constant-coefficient ODEs vs HankelHankel matrices.

463. A determinantal calculation.

464. A convex function

465. The case H, K ∈ Sym+

n in the product HK.

466. Arithmetic-geometric inequality for the spectrum of a matrix product.

467. Two determinants of physical relevance.

468. The Inverse Eigenvalue Problem.

469. Real matrices whose square is symmetric.

470. Commuting matrices.

471. Another proof of Hadamard’s Inequality.

472. The 3-dimensional area of the unit sphere of M2 (R).

473. Joint spectral radius, again.

474. Eigenvalues and eigenvectors of normal matrices.

475. The action of the group of invertible quaternions over SO3 (k).

476. The spectral radius of QA when Q is unitary.

477. Mean-by-cofactors for positive definite matrices.

23
Exercises
1. Let K be a field and M, N ∈ Mn (K). Let k be the subfield spanned by the entries of
M and N . Assume that M is similar to N in Mn (K). Show that M is similar to N in
Mn (k). Compare with Exercise 2, page 55.
2. (a) When M, N ∈ GLn (k), show that rk(M − N ) = rk(M −1 − N −1 ).
(b) If A ∈ GLn (k) and x, y ∈ k n are given, let us define B = A + xy T . If B ∈ GLn (k),
show that B −1 = A−1 − B −1 xy T A−1 . Compute B −1 x and deduce an explicit formula
for B −1 , the ShermanSherman–MorrisonMorrison formula.
(c) We now compute explicitly det B.
i. We begin with the case where A = In . Show that det(In + xy T ) = 1 + y T x
(several solutions).
ii. We continue with the case where A is non-singular. Deduce that det B =
(det A)(1 + y T A−1 x).
iii. What is the general algebraic expression for det B ? Hint: Principle of algebraic
identities. Such an identity is unique and can be deduced form the complex case
k = C.
iv. Application: If A is alternate and x ∈ k n , prove that det(A + txxT ) ≡ det A.
(d) Let t vary in k. Check that A + txy T is invertible either for all but one values of t, or
for all values, or for no value at all. In particular, the set GL+ n (R) of real matrices
with positive determinant is rank-one convex (see the next exercise), in the sense
that if A, B ∈ GL+ n (R) and rk(B − A) = 1, then the interval (A, B) is contained in
GL+ n (R).
(e) We now specialize to k = R. Check that det(A + xy T ) det(A − xy T ) ≤ det A2 . Show
that when the rank of P is larger than one, det(A + P ) det(A − P ) can be larger
than det A2 .
3. Given a map f : GLn (k) → k, we define f ∗ : GLn (k) → k by f ∗ (A) := f (A−1 ) det A.

(a) Check that f ∗∗ = f .

(b) If f is a linear combination of minors, prove that f ∗ is another such combination.
Mind that the void minor, which is the constant function equal to 1, is allowed in
these combinations. More precisely, prove that for every sets I and J of row and
column indices, with |I| = |J|, the (I, J)-minor of A−1 is given by the formula
   c 
−1 I J
(det A)A = (I, J) A
J Ic
for an appropriate sign (I, J).
(c) Let GL+ +
n (R) be the set of real matrices with det A > 0. We say that f : GLn (R) → R
is rank-one convex if its restrictions to segments (A, B) is convex whenever B − A
has rank one. Show that, if f is rank-one convex, then f ∗ is rank-one convex (use
the previous exercise).

24
 (d) According to J. BallBall, we say that f : GL+n (R) → R is polyconvex if there exists
a convex function g : RN → R such that f (A) = g(m(A)), where m(A) is the list of
minors of A. Show that, if f is polyconvex, then f ∗ is polyconvex. Hint: A convex
function is the supremum of some family of affine functions.

4. Assume that the characteristic of the field k is not equal to 2. Given M ∈ GLn (k), show
that the matrix
0n M −1
 
A :=
M 0n
is diagonalisable. Compute its eigenvalues and eigenvectors. More generally, show that
every involution (A2 = I) is diagonalisable.

5. Let P ∈ k[X] have simple roots in k. If A ∈ Mn (k) is such that P (A) = 0n , show that A
is diagonalisable. Exercise 4 is a special case of this property.

6. If n = 2, find a matrix P ∈ GL2 (k) such that, for every matrix A ∈ M2 (k), there holds
P −1 AP = Â (Â is the matrix of cofactors.)
Nota: If A ∈ SL2 (k), we thus have P −1 AP = A−T , meaning that the natural represen-
tation of SL2 (k) into k 2 is self-dual.

7. Prove the norm inequality (1 < p < ∞)

1/p 0
kAkp ≤ kAk1 kAk1/p
∞ , A ∈ Mn (C),

by a direct comptutation using only the Hölderhold@Hölder inequality and the explicit
formulae for the norms kAk1 and kAk∞ . Remark: Exercise 20 of Chapter 4 corresponds
to the case p = 2, where Hölder is nothing but Cauchy–SchwarzCauchySchwarz.

8. (See also Exercise 119). Let µ be a probability measure on R, with a compact support.
We assume that this support is not finite. Define its moments
Z
mk := xk dµ(x), k ∈ N,
R

and the determinants

m0 m1 · · · mn
m0 m1 ··· mn
m1 m2 · · · mn+1
m1 m2 ··· mn+1 .. ..
Dn := .. .. , Dn (x) := . . .
. .
mn−1 mn · · · m2n−1
mn mn+1 · · · m2n
1 x ··· xn

Define at last
1
pn (x) = √ Dn (x).
Dn−1 Dn

25
 (a) Prove that the leading order term of the polynomial pn is cn xn for some cn > 0.
Then prove that Z
pn (x)pm (x)dµ(x) = δnm .
R

In other words, the family (pn )n∈N consists in the orthonormal polynomials relatively
to µ.
(b) What happens if the support of µ is finite ?
(c) Prove the formula
Z
1 Y
Dn = (xi − xj )2 dµ(x0 ) · · · dµ(xn ).
(n + 1)! Rn+1 0≤i<j≤n

Hint: Compute Dn , a determinant of size n + 1, in (n + 1)! different ways. Then

use the expression of a VandermondeVandermonde determinant.

9. Find a short proof of Birkhoff’s TheoremBirkhoff (Theorem 5.5.1), using the wedding
Lemma, also known as Hall’s TheoremHall!Ph.: let n be a positive integer and let G
(girls) and B (boys) two sets of cardinality n. Let R be a binary relation on G × B
(gRb means that the girl g and the boy b appreciate each other). Assume that, for every
k = 1, . . . , n, and for every subset B 0 of B, of cardinality k, the image G0 of B 0 (that is,
the set of those g such that gRb for at least one b ∈ B 0 ) has cardinality larger than or
equal to k. Then there exists a bijection f : B → G such that f (b)Rb for every b ∈ B.
Nota: the assumption is politically correct, in the sense that it is symmetric in B and G
(though it is not clear at a first sight). The proof of the wedding Lemma is by induction.
It can be stated as a result on matrices: given an n × n matrix M of zeroes and ones,
assume that, for every k and every set of k lines, the ones in these lines correspond to at
least k columns. Then M is larger than or equal to a permutation matrix. Hint: Here
are the steps of the proof. Let M be bistochastic. Prove that it cannot contain a null
block of size k × l with k + l > n. Deduce, with the wedding lemma, that there exists
a permutation σ such that miσ(i) > 0 for every i. Show that we may restrict to the case
mii > 0 for every i. Show also that we may restrict to the case where 0 < mii < 1. In
this case, show that (1 − )M + In is bi-stochastic for || small enough. Conclude.

10. Let k be a field. Define the symplectic group Spm (k) as the set of matrices M in M2m (k)
that satisfy M T Jm M = Jm , where
 
0m Im
Jm := .
−Im 0m

Check that the word “group” is relevant. Using the Pfaffian, prove that every symplectic
matrix (that is, M ∈ Spm (k)) has determinant +1. Compare with Corollary 7.6.1.

11. Set n = 2m.

26
 (a) Show the following formula for the Pfaffian, as an element of Z[xij ; 1 ≤ i < j ≤ n] :
X
Pf(X) = (−1)σ xi1 i2 · · · xi2m−1 i2m .

Hereabove, the sum runs over all the possible ways the set {1, . . . , n} can be parti-
tionned in pairs :
{1, . . . , n} = {i1 , i2 } ∪ · · · ∪ {i2m−1 i2m }.
To avoid redundancy in the list of partitions, one normalized by

i2k−1 < i2k , 1 ≤ k ≤ m,

and i1 < i3 < · · · < i2m−1 (in particular, i1 = 1 and i2m = 2m). At last, σ is the
signature of the permutation (i1 i2 · · · i2m−1 i2m ).
Compute the number of monomials in the Pfaffian.
(b) Deduce an “expansion formula with respect to the i-th row” for the Pfaffian: if i is
given, then X
Pf(X) = α(i, j)(−1)i+j+1 xij Pf(X ij ),
j(6=i)

where X ij ∈ Mn−2 (k) denotes the alternate matrix obtained from X by removing
the i-th and the j-th rows and columns, and α(i, j) is +1 if i < j and is −1 if j < i.
(c) In particular, we have
n
X
Pf(X) = (−1)j x1j Pf(X 1j ).
j=2

Comment. This formula provides an alternate adjoint X

b with the following prop-
erties:
• the formula X X
b = Pf(X)In ,
• the entries of X
b are homogeneous polynomials of degree m − 1 in those of X.
(d) Deduce Pf(An ) = 1, where An denotes the alternate matrix whose upper-diagonal
entries are 1s. Hint: Induction.

12. (a) Let A, B be n × n Hermitian positive definite matrices. Denote by λ1 (A) ≤ λ2 (A) ≤
·√· · and
√ λ1 (B) ≤ λ2 (B) ≤ · · · their eigenvalues. Remarking that AB is similar to
BA B, show that the spectrum of AB is real, positive, and satisfies

λi (A)λ1 (B) ≤ λi (AB) ≤ λi (A)λn (B).

Hint: Use Theorem 3.3.2.

(b) Compare with point a) of Exercise 6, Chapter 7. Show also that the conclusion still
holds if A, B are only positive semi-definite.

27
 (c) More generally, prove the inequalities

λj (A)λk (B) ≤ λi (AB) ≤ λj (A)λl (B),

whenever j + k ≤ i + 1 and j + l ≥ i + n.
(d) Set n = 2. Let a1 ≤ a2 , b1 ≤ b2 , µ1 ≤ µ2 be non-negative numbers, satisfying
µ1 µ2 = a1 a2 b1 b2 and the inequalities

a1 b1 ≤ µ1 ≤ min{a1 b2 , a2 b1 }, max{a1 b2 , a2 b1 } ≤ µ2 ≤ a2 b2 .

Prove that there exist 2 × 2 real symmetric matrices A and B, with spectra {a1 , a2 }
and {b1 , b2 }, such that {µ1 , µ2 } is the spectrum of AB. Hint: Begin with the
case where some of the inequalities are equalities. Then use the intermediate value
Theorem.
(e) Set n ≥ 2, and assume that the Hermitian matrix B has eigenvalues b1 = 0, b2 =
· · · = bn = 1. Show that

λ1 (AB) = 0, λi−1 (A) ≤ λi (AB) ≤ λi (A).

Conversely, if
µ1 = 0 ≤ a1 ≤ µ2 ≤ a2 ≤ · · · ≤ µn−1 ≤ an ,
show that there exists a Hermitian matrix A with eigenvalues a1 , . . . , an , such that
AB has eigenvalues µ1 , . . . , µn .
Generalize to the case b0 = · · · = bk = 0, bk+1 = · · · = bn = 1.

13. (From J. GroßGroß and G. TrenklerTrenkler.) Given A, B two Hermitian positive semi-
definite matrices, show that Cn = R(AB) ⊕ ker(AB).

14. (a) Assume that A ∈ Mn×p (C) is injective. Prove that H := A∗ A is positive definite.
Show that AH −1 A∗ is the orthogonal projector onto R(A).
(b) Given two injective matrices Aj ∈ Mn×pj (C), define Hj as above. Define also F :=
A∗1 A2 and then M := H2−1 F ∗ H1−1 F . Using the previous exercise, show that the
eigenvalues of M belong to [0, 1].

15. Let A ∈ Mn (C) and H ∈ HPDn be given. If θ ∈ R, define φ(θ) := H + eiθ A + e−iθ A∗ .
Define a matrix Mk ∈ Hkn by
 
H A 0n
 ∗ .. ... ... 
 A . 
Mk = 
 0 ... .. .
 n . A 
.. ∗
. A H

(a) Decomposing vectors x ∈ Ckn as k blocks xj ∈ Cn , write x∗ Mk x as a sum of terms

of the form y ∗ φ(2`π/k)y (` = 1, . . . , k), and x∗k Ax1 .

28
 (b) We assume that φ(2`π/k) > 0n (l = 1, . . . , k). Show that there exist two positive
constants ck , dk such that

x∗ Mk x ≥ ck kxk2 − dk kxk k kx1 k.

Deduce that there exists a tk > 0, such that adding tk In in the bottom-right block,
moves Mk to a positive definite matrix.
(c) Under the same assumption as above, prove that M1 , . . . , Mk−1 are positive definite.

16. Let A ∈ Mn (C) and H ∈ HPDn be given, with A 6= 0n . We are interested in the equation

X + A∗ X −1 A = H,

where X ∈ HPDn is the unknown.

(a) Show that a necessary condition for X to exist is (Property (P))

• the Hermitian matrix φ(θ) := H + eiθ A + e−iθ A∗ is positive semi-definite for
every θ,
• the map θ 7→ det(φ(θ)) does not identically vanish.
Hint: Factorize φ(θ), and more generally H + zA + z −1 A∗ .
We shall prove later on that Property (P) is also a sufficient condition. The proof
relies more or less upon infinite dimensional matrices called ToeplizToepliz matri-
ces. For a general account of the theory, see M. RosenblumRosenblum & J. Rov-
niakRovniak, HardyHardy classes and operator theory, Oxford U. Press, 1985).
(b) Check that (P) is fulfilled in the case where H = In and A is real skew-symmetric.
Then show that a solution does exist, which is not unique. Hint: First, solve the
case n = 2. The nature of the solution depends on the sign of t2 − 1/4, where
 
0 t
A= .
−t 0

(c) Let A be invertible. Find a transformation that reduces the equation to the case
H = In (though with a different A). Verify that this transformation preserves the
validity or the failure of Property (P).
(d) From now on, we assume that H = In . Show that X is a solution of X + A∗ X −1 A =
In , if and only if In − X is a solution of Y + AY −1 A∗ = In .
(e) We temporarily assume that the equation admits at least one solution X ∈ HPDn .
Here is an algorithm for the approximation of a solution (the largest one):

X0 = In , Xn+1 := In − A∗ Xn−1 A.

i. First, show that X ≤ In in Hn .

ii. Prove inductively that Xk ≥ X for every k.

29
 iii. Prove inductively that Xk is non-increasing in HPDn . Deduce that it converges
to some limit, and that this limit is a solution.
iv. Deduce that the equation admits a largest solution.
v. Show that the equation also admits a smaller solution in HPDn .
(f) We now turn to the existence of a solution.
i. Define a block-tridiagonal matrix Mk ∈ Hkn by
 
In A 0n
 ∗ ... ... ... 
 A 
Mk =  .. .. .
 0
 n . . A 
.. ∗
. A In

If φ(0) > 0n and (P) holds, show that Mk ∈ HPDkn for every k ≥ 1. Hint: use
Exercise 15.
ii. Then show that there exists a unique blockwise lower-bidiagonal matrix Lk , with
diagonal blocks in HPDn , such that Lk L∗k = Mk .
iii. Then prove that there exist matrices Bj ∈ HPDn and Cj ∈ Mn (C), such that,
for every k ≥ 1, there holds
 
B1 0n
. ..
 C1 . . .
 

Lk =  ... .. .
 n0 . 0 
n 
...
Ck−1 Bk

iv. Write the recursion satisfied by (Bj , Cj ), and check that Xk := Bk2 satisfies the
algorithm above. Then, show that Xk converges as k → +∞, and that its limit
is a solution of our equation (therefore the greatest one).
v. Assuming (P) only, show that we may assume φ(0) > 0n (consider the matrix
eiα A instead of A, with α suitably chosen). Conclude.

17. (a) Let a ∈ Rn have positive entries. Recall (Exercise 20.a of Chapter 5) that the
extremal points of the convex set defined by the “inequality” b  a are obtained
from a by permutation of its entries.
Show
Y Y
(1) (b  a) =⇒ ( bj ≥ ai ).
j i

(b) Deduce an other proof of Proposition 3.4.2, with the help of Theorem 3.4.1. (One
may either deduce (1) from Proposition 3.4.2 and Theorem 3.4.2. These are rather
long proofs for easy results !)

30
18. Let A ∈ Mn (k) be invertible and define M ∈ M2n (k) by

0n A−1
 
M := .
−AT A−1

Show that M is symplectic: M T JM = J, with

 
0n In
J := .
−In 0n

19. (The Banach–MazurBanachMazur distance.)

(a) Let N and N 0 be two norms on k n (k = R or C).
If A ∈ Mn (k), we may define norms

N 0 (Ax) N (A−1 x)
kAk→ := sup , kA−1 k← := sup .
x6=0 N (x) x6=0 N 0 (x)

Show that kAk→ kA−1 k← achieves its upper bound.

We shall denote by δ(N, N 0 ) the minimum value.
Verify

0 ≤ log δ(N, N 00 ) ≤ log δ(N, N 0 ) + log δ(N 0 , N 00 ).

Stefan Banach.
Banach, Stefan (Poland) When N = k · kp , we shall write `p instead. If in
addition N 0 = k · kq , we write k · kp,q for k · k→ .
(b) In the set N of norms on k n , let us consider the following equivalence relation:
N ∼ N 0 if and only if there exists an A ∈ GLn (k) such that N 0 = N ◦ A. Show that
log δ induces a metric d on the quotient set Norm := N / ∼. This metric is called
the Banach–MazurBanachMazur distance.How many classes of Hermitian norms are
there ?
(c) Compute kIn kp,q for 1 ≤ p, q ≤ n (there are two cases, depending on the sign of
q − p). Deduce that
δ(`p , `q ) ≤ n| p − q | .
1 1

0 0
(d) Show that δ(`p , `q ) = δ(`p , `q ), where p0 , q 0 are the conjugate exponents.
∗
(e) i. Given H ∈ H+ n , find that the average of x Hx, as x runs over the set defined
by |xj | = 1 for all j’s, is Tr H (the measure is the product of n copies of the
normalized Lebesgue measure on the unit disk). Deduce that
√
Tr M ∗ M ≤ kM k∞,2

for every M ∈ Mn (k).

31
 ii. On the other hand, prove that

kAkp,∞ = max kA(i) kp0 ,

1≤i≤n

where A(i) denotes the i-th row vector of A.

√
iii. Deduce that δ(`2 , `∞ ) = n.
iv. Using the triangle inequality for log δ, deduce that

δ(`p , `q ) = n| p − q |
1 1

whenever p, q ≥ 2, and then for every p, q such that (p − 2)(q − 2) ≥ 0. Nota:

The exact value of δ(`p , `q ) is not known when (p − 2)(q − 2) < 0.
v. Remark that the “curves” {`p ; 2 ≤ p ≤ ∞} and {`p ; 1 ≤ p ≤ 2} are geodesic,
in the sense that the restrictions of the Banach–Mazur distance to these curves
satisfy the triangular equality.
(f) When n = 2, prove that δ(`1 , `∞ ) = 1. On the other hand, if n ≥ 3, then δ(`1 , `∞ ) >
1.
(g) A Theorem proved by F. JohnJohn states that the diameter of (Norm, d) is precisely
1
2
log n. Show that this metric space is compact. Nota: One may consider the norm
whose unit ball is an m-agon in R2 , with m even. Denote its class by Nm . It seems
that d(`1 , Nm ) = 12 log 2 when 8|m.

20. (Continuation of Exercise 19.) We study here classes of norms (in Norm) that contain a
pair {k · k, k · k∗ }. We recall that the dual norm of k · k is defined by

<(y ∗ x)
kxk∗ := inf .
x6=0 kyk

As shown in the previous exercise, k · k and k · k∗ are in the same class if and only if there
exists an A ∈ GLn (C) such that

(2) kAxk∗ = kxk, ∀x ∈ Cn .

(a) Let A, k · k and k · k∗ satisfy (2). Show that A−∗ A is an isometry of (Cn , k · k), where
A−∗ := (A∗ )−1 . Deduce that A−∗ A is diagonalizable, with eigenvalues of modulus
one. Nota: The whole exercise is valid with the field R instead of C, but the latter
result is a bit more difficult to establish.
(b) Let P ∈ GLn (C) be such that D := P −1 A−∗ AP is diagonal. Define a norm N ∼ k · k
by N (x) := kP xk. Show that D is an isometry of (Cn , N ), and that

N∗ (Bx) = N (x), ∀x ∈ Cn ,

where B := P ∗ AP .

32
 (c) Using Exercise 7 of Chapter 3 (page 56), prove that the class of k · k contains a norm
N such that
N∗ (∆x) = N (x), ∀x ∈ Cn
for some diagonal matrix ∆. Show also that N (∆−∗ ∆x) ≡ N (x). Show that one
may choose ∆ unitary.
(d) Find more than one example of such classes of norms on C2 .
21. (Numerical range.)
Given A ∈ Mn (C), define rA (x) = (Ax, x) = x∗ Ax. The numerical range of A is
H(A) = {rA (x) ; kxk2 = 1}.

(a) We show that if n = 2, then H(A) is an ellipse whose foci are the eigenvalues of A.
i. First check that it suffices to consider the cases of matrices
   
0 2a 1 2a
, , a ∈ R+ .
0 0 0 −1
ii. Treat the first case above.
iii. From now on, we treat the second case. First prove p that H(A) is the union of
circles with center p ∈ [−1, 1] and radius r(p) = a 1 − p2 .
iv. We define the (full) ellipse E ∈ C ∼ R2 by the inequality
x2 y2
+ ≤ 1.
1 + a2 a2
Show that H(A) ⊂ E.
v. Define p 7→ g(p) := y 2 + (x − p)2 − r(p)2 over [−1, 1]. If (x, y) ∈ E, prove that
min g ≤ 0 ; deduce that g vanishes, and thus that (x, y) ∈ H(A). Conclude.
vi. Show that for a general 2 × 2 matrix A, the area of A equals
π
|det[A∗ , A]|1/2 .
4
(b) Deduce that for every n, H(A) is convex (Toeplitz–HausdorffToeplizHausdorff The-
orem). See an application in Exercise 131.
(c) When A is normal, show that H(A) is the convex hull of the spectrum of A.
(d) Let us define the numerical radius w(A) by
w(A) := sup{|z| ; z ∈ H(A)}.
Prove that
w(A) ≤ kAk2 ≤ 2w(A).
Hint: Use the polarization principle to prove the second inequality.
Deduce that A 7→ w(A) is a norm (there are counter-examples showing that it is not
submultiplicative.)

33
 (e) Let us assume that there exists a complex number λ ∈ H(A) such that |λ| = kAk2 ,
say that λ = rA (x) for some unit vector x. Prove that λ is an eigenvalue of A, with
x an eigenvector. Thus ρ(A) = kAk2 = w(A). Prove also that x is an eigenvector of
A∗ , associated with λ̄. Hint: Show at first that λA∗ x + λ̄Ax = 2|λ|2 x.
(f) If A2 = 0n , show that w(A) = 12 kAk2 (see also Exercise 100.)

22. (JacobiJacobi matrices.) Let A ∈ Mn (R) be tridiagonal, with ai,i+1 ai+1,i > 0 when
i = 1, . . . , n − 1.

(a) Show that A is similar to a tridiagonal symmetric matrix S with si,i+1 < 0.
(b) Deduce that A has n real and simple eigenvalues. We denote them by λ1 < · · · < λn .
(c) For j = 1, . . . , n, let A(j) be the principal submatrix, obtained from A by keeping
the first j rows and columns. Without loss of generality, we may assume that the
off-diagonal entries of A are non-positive, and denote bj := −aj,,j+1 > 0. If λ is an
eigenvalue, show that
−1
x := (b−1
1 · · · bj−1 Dj−1 (λ))1≤j≤n , Dj (X) := det(A(j) − XIj )

is an eigenvector associated with λ.

(d) Deduce that the sequence of coordinates of the eigenvector associated with λj has
exactly j − 1 sign changes (one says that this eigenvector has j − 1 nodes). Hint:
Use the fact that (D1 , . . . , Dn ) is a SturmSturm sequence.
How could this be proven rapidly when j = 1 ?

23. Let k be a field with zero characteristic. Let A ∈ Mn (k) and l ≥ 1 be given. We form
the block-triangular matrix
 
A In 0n . . . 0n
. . . .. 
 0n . . . . . . . 

 . .
.. ... ...

M :=  .. 0n  ∈ Mnl (k).
 . ... ...
 ..

In 
0n . . . . . . 0n A

(a) Let P ∈ k[X] be a polynomial. Show that P (M ) is block-triangular, with the

(i, i + r)-block equal to r!1 P (r) (A).
(b) If an irreducible polynomial q divides a polynomial Q, together with its derivatives
Q(r) up to Q(s) , prove that q s+1 divides Q.
(c) Assume that the minimal polynomial of A is irreducible. Compute the minimal
polynomial of M . Compute the lists of its invariant polynomials.
(d) Deduce an alternate and somehow simpler second canonical form of square matrices.

34
 24. Given M ∈ Mn (k), denote |Mji | the minor obtained by removing the i-th row and the
i,l
j-column. We define similarly |Mj,k |.

(a) Prove Desnanot–JacobiDesnanotJacobi formula, also called DodgsonDodgson (see

Lewis C.)Lewis Caroll1 condensation formula,
1,n
|M1,n | det M = |M11 | |Mnn | − |Mn1 | |M1n |.

See Exercise 144 for a generalization.

Left: Charles L. Dodgson.

(Strangely enough, this is the only

stamp with his portrait.)

(b) Deduce a recursive algorithm for the

computation of det M , in about 4n3 /3
operations.
(c) Does this algorithm always produce a
result?
(d) (GantmacherGantmacher &
KreinKrein, Kotelyanskiı̆Kotelyanskiı̆.)
If S ∈ Symn , one defines |S(I)| as the
minor of S corresponding to lines and
columns of indices i, j ∈ I.
If S is positive semi-definite, show that
|S(I ∩ J)| · |S(I ∪ J)| ≤ |S(I)| · |S(J)|.

Hint: One may always assume that

I ∪ J = [[1, n]]. By a density argument,
one may assume that S is positive defi-
nite. Argue by induction upon the car-
dinal of the symmetric difference I∆J.
If I \ J and J \ I are singletons, apply
Desnanot–Jacobi.
25. Consider the oriented graph whose vertices are the points of Z2 and edges are the hor-
izontal (left to right) and vertical (downwards) segments. Given two vertices A and B,
denote by n(A, B) the number of paths from A to B. Thus n(A, B) ≥ 1 iff A1 ≤ B1 and
A2 ≥ B2 .
Given 2m points A1 , . . . , Am and B 1 , . . . , B m , we assume that Ai+1 (respectively B i+1 ) is
strictly upper right with respect to Ai (resp. B i ) and that B m is lower right with respect
to A1 .
1
Charles Lutdwidge Dodgson, English mathematician. Educated people, as well as children, prefer to call
him Lewis Caroll.

35
 (a) Consider m-tuples of paths γj , each one joining Aj to B j . Prove that the number of
such m-tuples, for which the γj ’s are pairwise disjoint, equals the determinant of

N := n(Ai , B j ) 1≤i,j≤m .

(b) Prove that the matrix N is totally positive.

26. Let U ∈ Un be upper HessenbergHessenberg. Up to a multiplication by a unitary diagonal

matrix, we may assume that the entries βk := uk+1,k are real non-negative. Prove that
there exist numbers αk ∈ C such that |αk |2 + βk2 = 1 for k = 1, . . . , n − 1, |αn | = 1 and
U = G1 (α1 ) · · · Gn (αn ), where
   
−αk βk
Gk (αk ) := diag Ik−1 , , In−k−1 , k = 1, . . . , n − 1,
βk ᾱk

and
Gn (αn ) := diag(1, . . . , 1, −αn ).
This is the SchurSchur parametrization. Notice that the matrices Gk (αk ) are unitary.

27. (P. D. LaxLax, H. F. WeinbergerWeinberger (1958).) Let V be a linear subspace of

Mn (R), with the property that the spectrum of every matrix M ∈ V is real. We shall
denote λ1 (M ) ≤ · · · ≤ λn (M ) the eigenvalues of M ∈ V , repeated with multiplicities.

(a) Prove that the functions M 7→ λk (M ) are continous over V .

(b) Let A, B ∈ V , with λ1 (B) > 0 (one says that B is positive.) Given λ ∈ R, show that
the polynomial x 7→ det(λIn −A−xB) has n real roots (counting with multiplicities.)
Nota: This is really a difficult question, but just assume that the functions λk are
infinitely differentiable away from the origin, a fact that is true when the eigenvalues
are simple for every non-zero M ∈ V .
(c) With A, B as above, prove that µ 7→ λk (A + µB) is strictly increasing. Deduce that
µ 7→ λk (A + µB) − µλ1 (B) is non-decreasing, and that µ 7→ λk (A + µB) − µλn (B)
is non-increasing.
(d) Given X, Y ∈ V , show that

λk (X) + λ1 (Y ) ≤ λk (X + Y ) ≤ λk (X) + λn (Y ).

Deduce that λ1 and λn are respectively a concave and a convex functions.

(e) Prove that the subset of positive matrices is a convex cone in V . Deduce that the
relation A ≺ B, defined on V by

λk (A) ≤ λk (B), ∀k = 1, . . . , n,

is an order relation.

36
 (f) Give an example of such a subspace V , of dimension n(n + 1)/2. Did we already
know all the results above in this case ?
28. Denote by SL+2 (Z) the set of non-negative matrices with entries in Z and determinant
+1. Denote by E, F the “elementary” matrices:
   
1 1 1 0
E= , F = .
0 1 1 1
Show that SL+
2 (Z) is the disjoint union of

{I2 }, E · SL+ +
2 (Z) and F · SL2 (Z).

Deduce that SL+ 2 (Z) is the free monoid spanned by E and F , meaning that for every
+
A ∈ SL2 (Z), there exists a unique word m in two letters, such that A = m(E, F ). Notice
that I2 corresponds to the void word.
Show that  
1 m m
Am :=  m 1 + m2 0 , m∈N
2 4
m 0 1+m +m
is an element of SL+ 3 (Z), which is irreducible, in the sense that Am = M N and M, N ∈
+
SL3 (Z) imply that M or N is a permutation matrix (the only invertible elements in
SL+ +
3 (Z) are the matrices of even permutations.) Deduce that SL3 (Z) cannot be generated
by a finite number of elements.
29. (R. M. MayMay, C. JeffriesJeffries, D. LogofetLogofet, UlianovUlianov.) We distinguish
three signs −, 0, + for real numbers, which exclude each other. In mathematical terms,
− = (−∞, 0), 0 = {0} and + = (0, +∞). The product of signs is well-defined.
Two given matrices A, B ∈ Mn (R) are said sign-equivalent if the entries aij and bij have
same sign, for every pair (i, j). Sign-equivalence is obviously an equivalence relation.
An equivalence class is written as a matrix S, whose entries are signs. Here are three
examples of sign-classes:
 
  0 − 0
0 −
S1 = diag(−, · · · , −), S2 = , S3 =  + 0 0  .
+ −
+ + −
In some applications of dynamical system theory, we are concerned with the asymptotic
stability of the origin in the system ẋ = Ax. For some reason, we do know the signs
of the entries of A, but the magnitude of non-zero entries is unknown. This arises for
instance in ecology or in the study of chemical reactions. Hence we ask whether the
sign structure of A (its sign-class) ensures that the whole spectrum lies in the half-space
Σ := {z ∈ C ; <z < 0}, or not. If it does, then we say that this class (or this matrix) is
sign-stable.
Given a sign-class S, we denote G(S) the oriented graph whose vertices are the indices
j = 1, . . . , n, and arrows correspond to the non-zero entries sij with i 6= j.

37
(a) Show that the classes S1 and S2 above are sign-stable, but that S3 is not.
(b) Actually (Hint), S3 is reducible. Show that the determination of the stable sign-
classes reduces to that of the stable irreducible sign-classes.
From now on, we restrict to irreducible sign-classes.
(c) Given a class S, we denote by S̄ its closure, where − = (−∞, 0) is replaced by
(−∞, 0], and similarly for +.
If a class S is sign-stable, show that S̄ is weakly sign-stable, meaning that its elements
A have their spectra in the closed half-space Σ̄. Deduce the following facts:
i. sii ≤ 0 for every i = 1, . . . , n,
ii. G(S) does not contain cycles of length p ≥ 3.
(d) We restrict to sign-classes that satisfy the two necessary conditions found above.
i. Considering the trace of a matrix in S, show that sign-stability requires that
there exists a k such that skk = −.
ii. Deduce that, for a class to be stable, we must have, for every pair i 6= j, either
sij = sji = 0 or sij sji < 0. In ecology, one says that the matrix is a predation
matrix. Hint: Use the irreducibility and the absence of cycle of length p > 2.
iii. Under the two additional restrictions just found, show that every monomial (σ
a permutation)
(σ)s1σ(1) · · · snσ(n)
is either 0 or (−)n . Deduce that the sign of the determinant is not ambiguous:
Either every element of S satisfies (−)n det A > 0, or every element of S satisfies
det A = 0. In the latter case, every monomial in det S must vanish.
iv. Check that sign-stability requires that the sign of the determinant be (−)n .
(e) Check that the following class satisfies all the necessary conditions found above, but
that it is not sign-stable because it contains an element A with eigenvalues ±i:
 
0 − 0 0 0
 + 0 − 0 0 
 
 0 + − − 0 .
S5 =  
 0 0 + 0 − 
0 0 0 + 0

(f) Show that the following class satisfies all the necessary conditions found above, but
one (det S7 = 0):  
0 − 0 0 0 0 0
 + − − 0 0 0 0 
 
 0 +
 − − − 0 0 

S7 =  0 0 + 0 0 0 0 .

 0 0
 + 0 0 − 0 

 0 0 0 0 + − − 
0 0 0 0 0 + 0

38
 Actually, show that every element A of S7 satisfies Ax = 0 for some non-zero vector
x with x1 = x2 = x3 = x6 = 0.
30. Let G be a classical group of real or complex n×n matrices. We only need that it satisfies
Proposition 7.3.1. Let G1 be a compact subgroup of G, containing G ∩ Un .
(a) Let M be an element of G1 , with polar decomposition QH. Verify that H belongs
to G1 ∩ HPDn .
(b) Using the fact that H m ∈ G1 for every m ∈ Z, prove that H = In .
(c) Deduce that G1 = G ∩ Un . Hence, G ∩ Un is a maximal compact subgroup of G.
31. Following C. R. JohnsonJohnson!Charles R. and C. J. HillarHillar (SIAM J. Matrix Anal.
Appl., 23, pp 916-928), we say that a word with an alphabet of two letters is nearly
symmetric if it is the product of two palindromes (a palindrome can be read in both senses;
for instance the French city LAVAL is a palindrome). Thus ABABAB = (ABABA)B
is nearly symmetric. Check that every word in two letters of length ` ≤ 5 is nearly
symmetric. Show that if a word m(A, B) is nearly symmetric, then the matrix m(S1 , S2 )
is diagonalizable with positive real eigenvalues, for every symmetric, positive definite
matrices S1 , S2 (see Exercise 258).
32. In R1+m we denote the generic point by (t, x)T , with t ∈ R and x ∈ Rm . Let C + be the
cone defined by t > kxk. Recall that those matrices of O(1, m) that preserve C + form the
subgroup G+± . The quadratic form (t, x) 7→ kxk2 − t2 is denoted by q.
Let M belong to G+± .

(a) Given a point x in the unit closed ball B of Rm , let (t, y)T be the image of (1, x)T
under M . Define f (x) := y/t. Prove that f is a continous map from B into itself.
Deduce that it has a fixed point. Deduce that M has at least one real positive
eigenvalue, associated with an eigenvector in the closure of C + . Nota: If m is odd,
one can prove that this eigenvector can be taken in the light cone t = kxk.
(b) If M v = λv and q(v) 6= 0, show that |λ| = 1.
(c) Let v = (t, x) and w = (s, y) be light vectors (that is q(v) = q(w) = 0), linearly
independent. Show that v ∗ Jw 6= 0.
(d) Assume that M admits an eigenvalue λ of modulus different from 1, v being an
eigenvector. Show that 1/λ is also an eigenvector. Denote by w a corresponding
eigenvector. Let < v, w >◦ be the orthogonal of v and w with respect to q. Using
the previous question, show that the restriction q1 of q to < v, w >◦ is positive
definite. Show that < v, w >◦ is invariant under M and deduce that the remaining
eigenvalues have unit modulus.
(e) Show that, for every M ∈ G+± , ρ(M ) is an eigenvalue of M .

33. Assume that A ∈ Mn (C) is tridiagonal, with an invertible diagonal part D. Assume that
the relaxation method converges for every parameter ω in the disc |ω − 1| < 1.

39
 (a) Show that, for every ω in the circle |ω − 1| = 1, the spectrum of Lω is included in
the unit circle.
(b) Deduce that the spectrum of the iteration matrix J of the JacobiJacobi method is
included in the interval (−1, 1). Compare with Theorem 9.4.1 and Exercise 7 of the
book.

34. Let A ∈ Mn (C) be given, and U (t) := exp(tA).

(a) Show that kU (t)k ≤ exp(tkAk) for t ≥ 0 and any matrix norm. Deduce that the
integral Z +∞
e−2γt U (t)∗ U (t) dt
0
converges for every γ > kAk.
(b) Denote Hγ the value of this integral, when it is defined. Computing the derivative at
h = 0 of h 7→ U (h)∗ Hγ U (h), by two different methods, deduce that Hγ is a solution
of

(3) A∗ X + XA = 2γX − In , X ∈ HPDn .

(c) Let γ be larger than the supremum of the real parts of eigenvalues of A. Show
that Equation (3) admits a unique solution in HPDn , and that the above integral
converges.
(d)
In particular, if the spectrum of M has positive
real part, and if K ∈ HPDn is given, then the
LyapunovLyapunov equation

M ∗ H + HM = K, H ∈ HPDn

admits a unique solution.

Let x(t) be a solution of the differential equation
ẋ + M x = 0, show that t 7→ x∗ Hx decays, and
strictly if x 6= 0.

Alexandr M. Lyapunov.
Lyapunov, Alexandr M. (USSR)

35. Show that if M ∈ Mn (C) and if Tr M ∗ M ≤ n, then | det M | ≤ 1, with equality if, and
only if, M is unitary.

40
36. Let k be a field of characteristic zero, meaning that 1 spans an additive subgroup, iso-
morphic to Z. By a slight abuse of notation, this subgroup is denoted by Z. We call Λ a
lattice of rank n if there exists a basis {x1 , . . . , xn } of k n such that

Λ = Zx1 ⊕ · · · ⊕ Zxn .

Such a basis of k n is called a basis of Λ.

Let Λ and Λ0 be two lattices of rank n, with Λ0 ⊂ Λ. Prove that there exist a ba-
sis {y 1 , . . . , y n } of Λ, together with integers d1 , . . . , dn , such that d1 |d2 , d2 |d3 , . . . , and
{d1 y 1 , . . . , dn y n } is a basis of Λ0 . Show that d1 , . . . , dn are uniquely defined by this prop-
erty, up to their signs. Finally, prove that the product d1 · · · dn equals the order [Λ : Λ0 ]
of the quotient group Λ/Λ0 .

37. (From E. S. KeyKey.) Given the companion matrix of a polynomial X n −a1 X n−1 −· · ·−an ,
in the form  
0 1 0 ···
 .. . .
. ... 0  ,

 .
 
 0 ··· 0 1 
an · · · · · · a1
and given a root x of P , compute an eigenvector associated with x. Deduce that, if
P1 , . . . , Pk have a common root z, then z k is an eigenvalue of the product of their com-
panion matrices.

38. Define the wedge product in k 3 in the same way as in R3 . Given a non-zero vector a in
k 3 , find all matrices A ∈ M3 (k) with the property that (Aa) ∧ x = A(a ∧ x) = a ∧ (Ax)
for every x ∈ k 3 . Hint: The result depends on whether a · a = a21 + a22 + a23 vanishes or
not.

39. (From Y. TianTian, Yongge.) Let P be a projector (that is P 2 = P ) on a real or complex

finite dimensional space.

(a) Prove that In − P ∗ P is positive semi-definite if and only if P is an orthogonal

projector, that is R(P )⊥ ker(P ).
(b) In general, prove the equalities

ker(In − P ∗ P ) = R(P ) ∩ R(P ∗ ) = ker(2In − P − P ∗ ).

and deduce that

R(In − P ∗ P ) = ker P + ker P ∗ = R(2In − P − P ∗ ).

40. Let k · k be a unitary invariant norm on Mn (C), and let A ∈ HPDn and B ∈ Mn (C)
be given. Recall that A admits a unique logarithm log A in Hn , a matrix such that
exp(log A) = A. For complex numbers z, we thus define Az := exp(z log A).

41
 (a) Define F (z) := Az BA1−z . Show that kF (z)k = kF (<z)k, then that F is bounded
on the strip 0 ≤ <z ≤ 1.
(b) If kABk ≤ 1 and kBAk ≤ 1, deduce that kA1/2 BA1/2 k ≤ 1. More generally,

kA1/2 BA1/2 k2 ≤ kABk · kBAk.

Compare with Exercise 21, page 79.

(c) Replacing B by B 0 := (1 + ckBk)−1 B and A by A0 := A + cIn with c > 0, show that
the same result holds true if we suppose only that A ∈ Hn is positive semi-definite.
(d) More generally, if H, K are Hermitian positive semi-definite, prove that

kHBKk2 ≤ kH 2 Bk · kBK 2 k.

41. Let H ∈ Hn be positive semidefinite and have bandwidth 2b − 1, meaning that |j − i| ≥ b

implies mij = 0.

(a) Prove that the CholeskyCholesky factorization H = LL∗ inherits the band-property
of M .
(b) Deduce that the largest eigenvalue of H is lower than or equal to the maximum
among the sums of b consecutive diagonal entries. Compare to Exercise 20, page 59
of the book.

42. (Preconditionned Conjugate Gradient Method.)

Let Ax = b be a linear system whose matrix A is symmetric positive definite (all entries
are real.) Recall that the convergence ratio of the Conjugate gradient method is the
number p
K(A) − 1
τGC = − log p ,
K(A) + 1
p
that behaves like 2/ K(A) when K(A) is large, as it uses to be in the real life. The
number K(A) := λmax (A)/λmin (A) is the condition number of A.
Preconditioning is a technique that reduces the condition number, hence increases the
convergence ratio, through a change of variables. Say that a new unknown is y := B T x,
so that the system is equivalent to Ãy = b̃, where

à := B −1 AB −T , b = B b̃.

For a given preconditioning, we associate the matrices C := BB T and T := In − C −1 A.

Notice that preconditioning with C or α−1 C is essentially the same trick if α > 0, although
T = T (α) differs significantly. Thus we mere associate to C the whole family

{T (α) = In − αC −1 A ; α > 0}.

(a) Show that à is similar to C −1 A.

42
 (b) Consider the decomposition A = M − N with M = α−1 C and N = α−1 C − A. This
yields an iterative method

C(xk+1 − xk ) = b − αAxk ,

whose iteration matrix is T (α). Show that there exist values of α for which the
method is convergent. Show that the optimal parameter (the one that maximizes
the convergence ratio) is
2
αopt = ,
λmin (Ã) + λmax (Ã)
with the convergence ratio

K(Ã) − 1
τopt = − log .
K(Ã) + 1

(c) When K(Ã) is large, show that

τGCP
q
∼ K(Ã),
τopt
where τGCP stands for the preconditioned conjugate gradient, that is the conjugate
gradient applied to Ã.
Conclusion ?
43. (Continuation.) We now start from a decomposition A = M − N and wish to construct
a preconditioning.
Assume that M + N T , obviously a symmetric matrix, is positive definite. We already
know that kM −1 N kA < 1, where k · kA is the Euclidean norm associated with A (Lemma
9.3.1.)
(a) Define T := (In − M −T A)(In − M −1 A). Prove that kT kA < 1. Deduce that the
“symmetric” method

M xk+1/2 = N xk + b, M T xk+1 = N T xk+1/2 + b

is convergent (remark that A = M T − N T .)

This method is called symmetric S.O.R., or S.S.O.R. when M is as in the relaxation
method.
(b) From the identity T = In − M −T (M + N T )M −1 A, we define C = M (M + N T )−1 M T .
Express the corresponding preconditioning C(ω) when M and N come from the
S.O.R. method:
1
M = D − E, ω ∈ (0, 2).
ω
This is the S.S.O.R. preconditioning.

43
 (c) Show that λmax (C(ω)−1 A) ≤ 1, with equality when ω = 1.
(d) Compute ρ(T ) and K(Ã) when A is tridiagonal with aii = 2, ai,i±1 = −1 and aij = 0
otherwise. Compare the S.S.O.R. method and the S.S.O.R. preconditioned conjugate
gradient method.

44. In control theory of linear systems, we face differential equations

ẋ = Ax + Bu,

where A ∈ Mn (R) and B ∈ Mn×m (R). We call x(t) the state and u(t) the control.
Controllability is the property that, given a time T > 0, an initial state x0 and a final
state xf , it is possible to find a control t 7→ u(t) such that x(0) = x0 and x(T ) = xf .

(a) Assume first that x(0) = 0. Express x(T ) in terms of Bu and etA (0 ≤ t ≤ T ) in a
closed form. Deduce that controllability is equivalent to xf ∈ H, where
n−1
H := + R(Ak B).
k=0

This is Kalman’sKalman criterion.

(b) Reversing the time, show that controllability from a general x(0) to xf = 0 is equiv-
alent to Kalman’sKalman criterion. Conclude that controllability for general initial
and final states is equivalent to Kalman’s criterion.
(c) Prove the following forms of Kalman’s criterion:
i. ker B T does not contain any eigenvector of AT .
ii. For every complex number z, the matrix
 T 
A − zIn
BT

has rank n.
(d) Assume m = 1: The control is scalar. We shall denote b instead of B, since it is a
vector. Furthermore, assume controllability. Show that there exists a vector c ∈ Rn
such that cT (A+In )−1 b = −δ1k for k = 1, . . . , n. Deduce that the spectrum of A+bcT
reduces to {1}. Hence the feedback u(t) = c(t) · x(t) yields stabilization, since then
x(t) decays exponentially for every initial data.

45. Consider a zero-sum repeated game between two players A and B. Each player chooses
one object among a list O1 , . . . , On . When A chooses Oi and B chooses Oj , the payoff
is mij ∈ R, which is positive if A wins, negative if B wins. Obviously, the matrix M is
skew-symmetric.
Players play a large number of games. One may represent their strategies by vectors
A B A
x
P , xA , where xi is the probability that A chooses the i-th object. Hence xA ≥ 0 and
B
i xi = 1, and the same for x . Given a pair of strategies, the expectation has the form

44
 φ(xA , xB ) where φ(x, y) := xT M y. Player A tries to maximize, while B tries to minimize
the expectation.
A NashNash equilibrium (x̄, ȳ) is a pair of strategies that is optimal for both players, in
the sense that, for all strategies x and y:

φ(x, ȳ) ≤ φ(x̄, ȳ) ≤ φ(x̄, y).

(a) Prove that a NashNash equilibrium always exists, and that the set of NashNash
equilibria is a product C × D of convex subsets.
(b) Deduce that, for every skew-symmetric matrix with real entries, there exists a non-
negative vector x 6= 0, such that M x is non-negative, and xj (M x)j = 0 for each
j = 1, . . . , n.
(c) Example: The list of objects consists in scissors, a stone, a hole and a sheet of
paper. The payoff is ±1, according to the following natural rules. Scissors win
against paper, but looses against the hole and the stone. Paper wins against the
hole and the stone. The hole wins again the stone.
Find the (unique and rather counter-intuitive) NashNash equilibrium.
Remark: In many countries, this game is more symmetric, with only scissors, stone
and the sheet of paper. It was illustrated during WWII, when the leaders of Great
Britain, USSR and Germany each had their own choice. Ultimately, the cissors and
the stone defeated the sheet of paper. During the Cold War, there remained the
cissors and the stone, untill Staline’sStaline death in 1953 and Churchill’sChurchill
loss of 1955 elections. One had to wait untill 1991 to see the cissors defeating the
stone, unlike in the usual game.

46. (P. Van den Driesschevanden@van den Driessche, H. K. WimmerWimmer.)

(a) Characterize the complex numbers a, b, c and the vectors X ∈ Cn−1 , such that the
following matrix is unitary

aX ∗
 
c
U := .
In−1 − bXX ∗ X

(b) Let C be a companion matrix, given in the form

0∗
 
m
C= , m ∈ C, V ∈ Cn−1 .
−In−1 V

Find the polar decomposition C = QH. Hint: Q equals −U , where U is as in the

previous question.

47. Let E be an invariant subspace of a matrix M ∈ Mn (R).

(a) Show that E ⊥ is invariant under M T .

45
 (b) Prove the following identity between characteristic polynomials:

(4) PM (X) = PM |E (X)PM T |E ⊥ (X).

48. (See also YakubovichYakubovich & StarzhinskiiStarzhinskii, Linear differential equations

with periodic coefficients. Wiley & Sons, 1975.)
Let M belong to Spn (R). We recall the notations of Chapter 7: M T JM = J and
J 2 = −I2n as well as J T = −J.

(a) Show that the characteristic polynomial is reciprocal:

 
2n 1
PM (X) = X PM .
X

Deduce a classification of the eigenvalues of M .

(b) Define the quadratic form
q(x) := 2xT JM x.
Verify that M is a q-isometry.
(c) Let (e−iθ , eiθ ) be a pair of simple eigenvalues of M on the unit circle. Let Π be the
corresponding invariant subspace:

Π := ker(M 2 − 2(cos θ)M + I2n ).

i. Show that JΠ⊥ is invariant under M .

ii. Using the formula (4) above, show that e±iθ are not eigenvalues of M |JΠ⊥ .
iii. Deduce that R2n = Π ⊕ JΠ⊥ .
(d) (Continued.)
i. Show that q does not vanish on Π \ {0}. Hence q defines a Euclidian structure
on Π.
ii. Check that M |Π is direct (its determinant is positive.)
iii. Show that M |Π is a rotation with respect to the Euclidian structure defined by
q, whose angle is either θ or −θ.
(e) More generally, assume that a plane Π is invariant under a symplectic matrix M ,
with corresponding eigenvalues e±iθ , and that Π is not Lagrangian: (x, y) 7→ y T Jx is
not identically zero on Π. Show that M |Π acts as rotation of angle ±θ. In particular,
if M = J, show that θ = +π/2.
(f) Let H be an invariant subspace of M , on which the form q is either positive or
negative definite. Prove that the spectrum of M |H lies in the unit circle and that
M |H is semisimple (the JordanJordan!Camille form is diagonal).

46
 (g) Equivalently, let λ be an eigenvalue of M (say a simple one) with λ 6∈ R and |λ| =6 1.
Let H be the invariant subspace associated with the eigenvalues (λ, λ̄, 1/λ, 1/λ̄).
Show that the restriction of the form q to H is neither positive nor negative definite.
Show that the invariant subspace K associated with the eigenvalues λ and λ̄ is
q-isotropic. Thus, if q|H is non-degenerate, its signature is (2, 2).

49. (From L. TartarTartar.) In Mn (R), prove the inequality

(Tr M )2 ≤ (rk M ) Tr(M T M ).

Hint: Apply Schur’sSchur trigonalization Theorem.

Use the latter to built as many as possible non-trivial quadratic forms on Mn (R), non-
negative on the cone of singular matrices.

50. The minors of general matrices are not independent on each other. For instance, each
entry is a minor (of order one) and the determinant (an other minor) is defined in terms of
entries. An other instance is given by the row- or column-expansion of the determinant.
See also Exercise 24 above. Here is another relation.
Denote P2m the set of partitions I ∪J of {1, . . . , 2m} into two sets I and J of equal lengths
m. If (I, J) ∈ Pm , let σ(I, J) be the signature of the permutation (i1 , . . . , im , j1 , . . . , jm ),
where
I = {i1 ≤ · · · ≤ im }, J = {j1 ≤ · · · ≤ jm }.
Prove that, for every matrix A ∈ M2m×m (k), there holds
X
σ(I, J)A(I)A(J) = 0,
(I,J)∈Pm

with  
1 ··· m
A(I) := A .
i1 · · · im
Find other algebraic relations (syzygies) between minors.

51. (a) Let Σ belong to SPDn . Prove that the linear map σ 7→ σΣ+Σσ is an automorphism
of Symn (R). Hint: Consider the spectrum of σ 2 Σ. Show that it is real non-negative
on one hand, non-positive on the other hand. Then conclude.
(b) Let Σ belong to SPDn . Compute the differential of the map Σ 7→ Σ2 .
√
(c) Deduce that the square root map S 7→ S is analytic on SPDn . Remark: The
same result holds true on HPDn , same proof.

52. We consider real symmetric n × n matrices. We use the Schur–FrobeniusSchurFrobenius

norm k · kF . The result would be the same for complex Hermitian matrices.

47
 (a) Given two matrices A = diag(a1 , . . . , an ) and B = P diag(b1 , . . . , bn )P T , with P an
orthogonal matrix, verify that
X
kB − Ak2F = p2il (ai − bl )2 .
i,l

(b) Assume that both A and B are positive definite and as above. Prove that
√ √
k B − Ak4F ≤ nkB − Ak2F .
√ √
Hint: Use | b − a|2 ≤ |b − a| for positive real numbers, together with Cauchy–
Schwarz inequality.
(c) Deduce that the square root map, defined on SPDn , is Hölderian with exponent
1/2. Verify that the supremum of
√ √
k B − Ak2F kB − Ak−1
F ,

taken either on SPDn or on the subset of diagonal matrices, takes the same value.
Remark: See the improvement of this result in Exercise 110.

53. The electromagnetic field (E, B) must be understood as an alternate 2-form:

ω = dt ∧ (E · dx) + B1 dx2 ∧ dx3 + B2 dx3 ∧ dx1 + B3 dx1 ∧ dx2 .

In coordinates (t, x1 , x2 , x3 ), it is thus represented by the alternate matrix

 
0 E1 E2 E3
 −E1 0 B3 −B2 
A=  ∈ M4 (R).
 −E2 −B3 0 B1 
−E3 B2 −B1 0

In another choice of LorentzianLorentz coordinates, ω is thus represented by the new

matrix A0 := M T AM . Matrix M ∈ O(1, 3) is that of change of variables. With A0
are associated (E 0 , B 0 ). Namely, the decomposition of the electro-magnetic field into an
electric part and a magnetic one depends on the choice of coordinates. The purpose of
this exercise is to find Lorentzian invariants, that is quantities associated with ω that do
not depend on the choice of coordinates.

(a) If M = diag(±1, Q) with Q ∈ O3 (R), express E 0 , B 0 in terms of E, B and Q.

(b) Verify that M belongs to O(1, 3) ∩ SPD4 if and only if there exists a unit vector q
and numbers c, s with c2 − s2 = 1, such that
 
c sq T
M= .
sq I3 + (c − 1)q q T

Find how M transforms E and B.

48
 (c) Verify that |B|2 − |E|2 and |E · B| are Lorentzian invariants.
(d) We now show that |B|2 − |E|2 and |E · B| are the only Lorentzian invariants. Thus
let (Ei , Bi ) and (Ef , Bf ) be two pairs of vectors in R3 , defining two 2-forms ωi and
ωf . We assume that

|Bf |2 − |Ef |2 = |Bi |2 − |Ei |2 =: , |Ef · Bf | = |Ei · Bi | =: δ.

i. Choose q ∈ S2 that is orthogonal to both Ei and Bi . If θ ∈ R is given, define

f (θ) := |(cosh θ)E + (sinh θ)B ∧ q|2 .

√
Show that the range of f covers exactly the interval [ 21 ( + 2 + 4δ), +∞).
ii. Deduce that there exist a pair (Em , Bm ) of vectors, both belonging to the plane
spanned by Ei and Bi and with

|Bm |2 = |Bf |2 , |Em |2 = |Ef |2 , |Em · Bm | = δ,

and a symmetric positive definite Lorentzian matrix Mi such that

ωi (Mi x, Mi y) ≡ ωm (x, y).

iii. Show also that there exists an orthogonal Lorentzian matrix Rf such that

ωm (Rf x, Rf y) ≡ ωf (x, y).

Conclude.

54. Let A ∈ Hn be given blockwise in the form

 
B C
A= , B ∈ Hm , D ∈ Hn−m .
C∗ D

Assume that the least eigenvalue λ of B is greater than the highest eigenvalue µ of D.
Prove that the spectrum of A is included into (−∞, µ]∪[λ, +∞). Prove also that [λ, +∞)
contains exactly m eigenvalues, counting with multiplicities.

55. Let k be a field and A, B, C be non-colinear matrices in Mn (k), such that A − B, B − C

and C − A have rank one. Show that every distinct matrices P, Q in the affine plane
spanned by A, B, C differ from a rank-one matrix.
On the contrary, find in M2 (k) four matrices A, B, C, D such that A − B, B − C, C − D,
D − A have rank one, but for instance A − C has rank two.

56. Given an abelian ring A, recall that for a square matrix M ∈ Mn (A), adjM denotes the
transpose of the cofactors of M , so that the following identity holds

M (adjM) = (adjM )M = (det M )In .

49
 (a) Prove that the rank of adjM equals either n, 1 or 0. Characterize the three cases by
the rank of M .
(b) Describe in more details adjM when rkM = n − 1 and A is a field. Show that if 0
is a simple eigenvalue, then the right and left eigenvectors (` and r), normalized by
`r = 1, are related by
det M (i)
`i ri = P (j)
,
j det M

where M (j) is the matrix obtained from M by removing its j-th row and column.
Justify that the denominator is non-zero.
(c) Given n ≥ 2, show that det(adjM ) = (det M )n−1 and adj(adjM ) = (det M )n−2 M .
In the special case n = 2, the latter means adj(adjM ) = M . In particular, if n ≥ 3
(but not when n = 2), we have

(det M = 0) =⇒ (adj(adjM ) = 0n ).

Hint: First prove the formulæ when A is a field and det M 6= 0. Then, considering
that the required identities are polynomial ones with integral coefficients, deduce
that they hold true in Z[X11 , . . . , Xnn ] by choosing A = Q, and conclude.

57. (A. MajdaMajda, ThomannThomann) Let A := diag{1, −1, . . . , −1} be the matrix of the
standard scalar product h·, ·i in Lorentzian geometry. The vectors belong to R1+n , and
read x = (x0 , . . . , xn )T . The forward cone K + is defined by the inequalities

hx, xi > 0, x0 > 0.

The words non-degenerate and orthogonal are used with respect to the Lorentzian scalar
product.

(a) Let q, r ∈ K + be given. Show the “reverse Cauchy–SchwarzCauchySchwarz inequal-

ity”

(5) hq, qi hr, ri ≤ hq, ri2 ,

with equality if, and only if, q and r are colinear. This is a special case of a deep
result by L. GårdingGaa@Gårding.
(b) More generally, if q ∈ K + and r ∈ R1+n , prove (5).
(c) Given µ ∈ R and q, r ∈ K + , define the quadratic form

Hµ (x) := µhq, xi hr, xi − hq, ri hx, xi.

50
 i.
In the case q = r = e0 := (1, 0, . . . , 0)T ,
check that Hµ is positive definite when-
ever µ > 1.
Deduce that if µ > 1 and if q, r are
colinear, then Hµ is positive definite.
Hint: Use a LorentzLorentz transfor-
mation to drive q to e0 .
Lorentz, Hendrik (Holland)

Hendrik Lorentz.
ii. We now assume that q and r are not colinear. Check that the plane P spanned
by q and r is non-degenerate and that the LorentzLorentz “norm” is definite on
P ⊥ . Deduce that Hµ is positive definite if, and only if, its restriction to P has
this property.
iii. Show that Hµ is not positive definite for µ ≤ 1, as well as for large positive µ’s.
On the other hand, show that it is positive definite for every µ in a neighbour-
hood of the interval
2hq, ri2
 
2, .
hq, ri2 − hq, qi hr, ri
58. (a) Given a matrix M ∈ Mn (C) that have a positive real spectrum, show that
√
det(In + M ) ≥ 2n det M .

(b) Deduce that, for every positive definite Hermitian matrices H, K, there holds

(det(H + K))2 ≥ 4n (det H)(det K).

Nota: This inequality can be improved into

(det(H + K))1/n ≥ (det H)1/n + (det K)1/n ,

the proof of which being slightly more involved. This can be viewed as a consequence
of an inequality of L. GårdingGaa@Gårding about hyperbolic polynomials. See
Exercise 218 for a proof without polynomials. See Exercise 219 for an improvement
of the inequality above.
(c) Show that the map
H 7→ − log det H
is strictly convex upon HPDn . Notice that Gårding’sGaa@Gårding inequality tells
us the better result that H 7→ (det H)1/n is concave. This is optimal since this map
is homogeneous of degree one, and therefore is linear on rays R+ H. However, it
depends on n, while H 7→ log det H does not.

51
 (d) Deduce that the set of positive definite Hermitian matrices such that det H ≥ 1 (or
greater or equal to some other constant) is convex.
59. Let N ∈ Mn (C) be given, such that every matrix of the form N + H, H Hermitian, has
a real spectrum.
(a) Prove that there exists a matrix M in N + Hn that has only simple eigenvalues.
(b) Because of simplicity, the eigenvalues M + H 7→ λj are C ∞ -functions for H small.
Compute their differentials:
Yi∗ KXi
dλj (M ) · K = ,
Yi∗ Xi
where Xi , Yi are eigenvectors of M and M ∗ , respectively.
(c) Show that Yi and Xi are colinear. Hint: Take K of the form xx∗ with x ∈ Cn .
(d) Deduce that N is Hermitian.
60. One wishes to prove the following statement: A matrix M ∈ Mn (C), such that M + H is
diagonalizable for every H ∈ Hn , has the form iaIn + K where a ∈ R and K is Hermitian.
(a) Prove the statement for n = 2.
(b) Let M satisfy the assumption. Show that there exists a Hermitian matrix K such
that M +K is upper triangular with pure imaginary diagonal entries. Then conclude
by an induction.
61. (P. D. LaxLax) Let A ∈ Mn (k) be given, with k = R or C. Assume that for every x ∈ k n ,
the following bound holds true:
|hAx, xi| ≤ kxk2 .
Deduce that the sequence of powers (Am )m∈N is bounded. Hint: Prove that the unitary
eigenvalues are semi-simple. Then use Exercise 10 of Chapter 4.
62. Let e ∈ Rn denote the vector (1, . . . , 1)T . A square matrix M ∈ Mn (R) is called a
“Euclidean distance matrix” (EDM) if there exist vectors p1 , . . . , pn in a Euclidean vector
space E, such that mij = kpi − pj k2 for every pair (i, j).

(a) Show that every Euclidean distance matrix,

besides being non-negative and symmetric
with a diagonal of zeroes (obvious condi-
tions), also defines a non-positive quadratic
form on the hyperplane e⊥ .

Euclid (Vatican state)

52
 Left: Euclid

(b) Given a symmetric matrix M with a diagonal of zeroes, assume that it defines a
non-positive quadratic form on the hyperplane e⊥ . Check that M ≥ 0 (in the sense
that the entries are nonnegative.) Prove that there exists a vector v ∈ Rn such that
the quadratic form
1
q(x) := (v · x)(e · x) − xT M x
2
is non-negative. Let S be the symmetric matrix associated with q, and let P be a
symmetric square root of S. Prove that M is an EDM, associated with the column
vectors p1 , . . . , pn of P .
(c) Prove that the minimal dimension r of the Euclidean space E, equals the rank of
JM J, where J is the orthogonal projection onto e⊥ :
1 T
J := In − ee .
n
P
63. (F. HansenHansen, G. PedersenPedersen) Hereafter, we denote by (x, y) = j xj ȳj the
usual Hermitian product in Cn . Given a numerical function f : I → R defined on an
interval, and given a Hermitian n × n matrix H, with Sp(H) ⊂ I, we define f (H) in
the following natural way: Let H = U ∗ DU be a diagonalization of H in a unitary basis,
D = diag{d1 , . . . , dn }, then f (H) := U ∗ f (D)U , where
f (D) = diag{f (d1 ), . . . , f (dn )}.
(a) Find a polynomial P ∈ R[X], that depends only on f and on the spectrum of H,
so that f (H) = P (H). Deduce that the definition above is not ambiguous, namely
that it does not depend on the choice of the unitary eigenbasis.
(b) Let m be any positive integer and H1 , . . . , Hm be Hermitian. We also give m matrices
A1 , . . . , Am in Mn (C), with the property that
A∗1 A1 + · · · + A∗m Am = In .
Finally, we define
H := A∗1 H1 A1 + · · · + A∗m Hm Am .
i. Let I be an interval of R that contains all the spectra of H1 , . . . , Hm . Show that
H is Hermitian and that I contains Sp(H).
ii. For each λ ∈ I, we denote by Ek (λ) the orthogonal projector on ker(Hk − λ). If
ξ is a unit vector, we define the (atomic) measure µξ by
m X
X
µξ (S) = (Ek (λ)Ak ξ, Ak ξ).
k=1 λ∈S

Show that µξ is a probability. Also, show that if ξ is an eigenvector of H, then

Z
(Hξ, ξ) = λ dµξ (λ).

53
 iii. Under the same assumptions, show that
Z 
(f (H)ξ, ξ) = f λ dµξ (λ) .

iv. If f is convex on I, deduce that

m
!
X
Trf (H) ≤ Tr A∗k f (Hk )Ak .
k=1

Hint: Use Jensen’sJensen inequality, plus the fact that

n
X
TrM = (M ξl , ξl ),
l=1

for every unitary basis {ξ1 , . . . , ξn }, for instance an eigenbasis if M is Hermitian.

64. We deal with complex n × n matrices. We denote by σ(A) the spectrum of A and by
ρ(A) its complement, the resolvant set of A. We use only the canonical Hermitian norm
on Cn and write kAk for the induced norm on Mn (C) (we wrote kAk2 in the book). We
denote dist(z; F ) the distance from a complex number z to a closed subset F in C.

(a) Prove that for every matrix A ∈ Mn (C) and complex number z ∈ ρ(A), there holds
1
(6) k(z − A)−1 k ≤ .
dist(z; σ(A))

(b) When A is normal, prove that the equality holds in (6).

(c) Conversely, we consider a matrix A such that the equality holds in (6).
i. Show that we may assume, up to a unitary conjugation, that A be block-
triangular
λ X∗
 
A= ,
0 B
with X ∈ Cn−1 . Hint: Apply Theorem 3.1.3 (SchurSchur).
ii. When A is block-triangular as above, compute the inverse of z − A blockwise,
when z is close to (but distinct from) λ. Establish the following inequality

2< ᾱX ∗ (z − B)−1 v + |z − λ|2 k(z − B)−1 vk2 ≤ kvk2 ,

for every complex number α and v ∈ Cn−m . Deduce that X = 0.

iii. Conclude, by an induction, that A is diagonal. Finally, show that a matrix
satisfies for every z the equality in (6), if and only if it is normal.

54
65. We use the notations of the previous exercise. In addition, if  > 0 we define the -
pseudospectrum as
 
−1 1
σ (A) := σ(A) ∪ z ∈ ρ(A) ; k(z − A) k ≥ .

We recall (Exercise 21 in this list) that the numerical range

H(A) := {rA (x) ; kxk2 = 1}

is a convex compact subset.

(a) Prove that [
σ (A) = σ(A + B).
kBk≤

(b) Prove also that

σ (A) ⊂ {z ∈ C ; dist(z; H(A)) ≤ }.
66. (From notes by M. CosteCoste.) This exercise shows that a matrix M ∈ GLn (R) is the
exponential of a real matrix if, and only if, it is the square of another real matrix.
(a) Show that, in Mn (R), every exponential is a square.
(b) Given a matrix A ∈ Mn (C), we denote A the C-algebra spanned by A, that is the
set of matrices P (A) as P runs over C[X].
i. Check that A is commutative, and that the exponential map is a homomorphism
from (A , +) to (A∗ , ×), where A∗ denotes the subset of invertible matrices (a
multiplicative group.)
ii. Show that A∗ is an open and connected subset of A.
iii. Let E denote exp(A), so that E is a subgroup of A∗ . Show that E is a neigh-
bourhood of the identity. Hint: Use the Implicit Function Theorem.
iv. Deduce that E is closed in A∗ ; Hint: See Exercise 21, page 135. Conclude that
E = A∗ .
v. Finally, show that every matrix B ∈ GLn (C) reads B = exp(P (B)) for some
polynomial P .
(c) Let B ∈ GLn (R) and P ∈ C[X] be as above. Show that

B 2 = exp(P (B) + P̄ (B)).

Conclusion ?
67. (The Le Verrier–Faddeevlever@Le VerrierFaddeev method.) Given A ∈ Mn (k), we define
inductively a sequence (Aj , aj , Bj )1≤j≤n by
1
Aj = ABj−1 (or A1 = A), aj = − Tr Aj , Bj = Aj + aj In .
j

55
 Show that the characteristic polynomial of A is

X n + a1 X n−1 + · · · + an .

Apply Cayley–Hamilton’sCayleyHamilton theorem and

compare with Exercise 25, page 37.

Hamilton,
Rowan (Eire)

Rowan Hamilton.

68. Given A ∈ Mn (k), with its characteristic polynomial

PA (X) = X n + a1 X n−1 + · · · + an ,

we form a sequence of polynomials by the Horner’sHorner rule:

p0 (X) := 1, p1 (X) := X + a1 , pj (X) = Xpj−1 (X) + aj , . . .

Prove that
n−1
1 X
(XIn − A)−1 = pj (A)X n−j−1 .
PA (X) j=0

69. Let A ∈ Mn (C) be given, with eigenvalues λj and singular values σj , 1 ≤ j ≤ n. We

choose the decreasing orders:

|λ1 | ≥ |λ2 | ≥ · · · ≥ |λn |, σ1 ≥ σ2 ≥ · · · ≥ σn .

Recall that the σj ’s are the square roots of the eigenvalues of A∗ A.

We wish to prove the inequality
k
Y k
Y
|λj | ≤ σj , 1 ≤ k ≤ n.
j=1 j=1

(a) Prove directly the case k = 1. Show the equality in the case k = n.
(b) Working within the exterior algebra, we define A∧p ∈ End(Λp (Cn )) by

A∧p (x1 ∧ · · · ∧ xp ) := (Ax1 ) ∧ · · · ∧ (Axp ), ∀x1 , . . . , xp ∈ Cn .

Prove that the eigenvalues of A∧p are the products of p terms λj with pairwise
distinct indices. Deduce the value of the spectral radius.

56
 (c) We endow Λp (Cn ) with the natural Hermitian norm in which the canonical basis
made of ei1 ∧ · · · ∧ eip with i1 < · · · < ip , is orthonormal. We denote by h·, ·i the
scalar product in Λp (Cn ).
i. If x1 , . . . , xp , y1 , . . . , yp ∈ Cn , prove that

hx1 ∧ · · · ∧ xp , y1 ∧ · · · ∧ yp i = det (x∗i yj )1≤i,j≤p .

ii. For M ∈ Mn (C), show that the Hermitian adjoint of M ∧p is (M ∗ )∧p .

iii. If U ∈ Un , show that U ∧p is unitary.
iv. Deduce that the norm of A∧p equals σ1 · · · σp .
(d) Conclude.

70. Use Exercise 20.a of Chapter 5 to prove the theorem of R. HornHorn!Roger & I. SchurSchur
: The set of diagonals (h11 , . . . , hnn ) of Hermitian matrices with given spectrum (λ1 , . . . , λn )
is the convex hull of the points (λσ(1) , . . . , λσ(n) ) as σ runs over the permutations of
{1, . . . , n}.

71. (A theorem by P. D. LaxLax.)

Assume that a subspace V of Mn (R) has dimension 3, and that its non-zero elements
have a real spectrum, with pairwise distinct eigenvalues. When M ∈ V and M 6= 0n ,
denote
λ1 (M ) < · · · < λn (M )
its eigenvalues. We equip Rn with the standard Euclidean norm.

(a) Verify that

λj (−M ) = −λn−j+1 (M ).
(b) Prove that there exists a continuous map

M 7→ B(M ) = {r1 (M ), . . . , rn (M )},

defined on V \ {0n }, such that B(M ) is a unitary basis of M . Hint: The domain
V \ {0n } is simply connected.
(c) Let choose M0 a non-zero element of V . We orient Rn in such a way that B(M0 ) be
a direct basis. Show that B(M ) is always direct.
(d) Show also that for every j, there exists a constant ρj = ±1 such that, for every
non-zero M , there holds
rj (−M ) = ρj rn−j+1 (M ).
(e) From the former questions, show that if n ≡ 2, 3 (mod 4), then
n
Y
ρj = −1.
j=1

57
 (f) On another hand, show that there always holds

ρj ρn−j+1 = 1.

Deduce that n 6≡ 2 (mod 4).

72. (The exchange, gyration or sweep operator.) Given a matrix M ∈ Mn (k) in block form
 
A B
M= , A ∈ GLp (k),
C D

we define a matrix (q := n − p)
−1 
−A−1 B
   
Ip 0p×q A B A
exc(M ) := × = .
C D 0q×p Iq CA−1 D − CA−1 B

(a) Show that the exchange map is an involution:

exc(exc(M )) = M.

(b) If D is non-singular, prove that exc(M ) is non-singular, with

exc(M −1 ) = exc(M )−1 .

(c) Let J := diag{Ip , −Iq }. Show that

exc(JM J)T = exc(M T ).

(d) We restrict to k = R. Recall that O(p, q) is the orthogonal group associated with J.
Show that the exchange map is well defined on O(p, q). With the previous formulæ,
prove that it maps O(p, q) on a subset of On (R).
(e) Show that the image of the exchange map is a dense open subset of On (R).

73. Using the quadratic forms of Rn that are preserved by elements of the groups O(p, q) and
On (R), find a simpler proof of the fact that the exchange map maps the former into the
latter.

74. Given a function f : (0, +∞) → R, we may define a map

M →7 f (M ),
SPDn → Symn (R)

in the same way as we defined the square root. The uniqueness is proved with the same
argument (see for instance Exercise 63.a). We say that f is a monotone matrix function
if, whenever 0n < M < N in the sense of quadratic forms, there holds f (M ) < f (N ).

(a) Prove that f (s) := −1/s is a monotone matrix function.

58
(b) Verify that the set of monotone matrix functions is a convex cone. Deduce that, for
every nonnegative, non-zero measure m,
Z +∞
dm(t)
f (s) := −
0 s+t
is a monotone matrix function.
(c)
Prove that, given numbers a ≥ 0, b ∈ R and a non-
negative bounded measure m, such that (a, m) 6=
(0, 0),
Z +∞
1 − st
(7) f (s) := as + b − dm(t)
0 s+t

is a monotone matrix function (Loewner’sLoewner

theorem asserts that every monotone matrix func-
tion is of this form ; such functions have a holo-
morphic extension to the domain C \ R− , and send
the PoincaréPoincaré half-space =z > 0 into itself.
The formula above is the NevanlinnaNevanlinna
representation of such functions.)

Poincaré,
Henri (France)

Henri Poincaré.

(d) Prove that, given a non-negative measure ν 6= 0,

Z +∞ −1
dν(t)
s 7→
0 s+t

is a monotone matrix function (this function is the inverse of the CauchyCauchy

transform of the measure ν.)
(e) Compute Z ∞
dt
.
0 tα (s + t)
Deduce that fα (s) = sα is a monotone matrix function for every α ∈ (0, 1).
(f) Prove that 0n < M ≤ N implies log M ≤ log N . Hint: Consider the map α−1 (fα −
1).
(g) Find two matrices M, P ∈ SPD2 (R) such that M P + P M 6∈ SPD2 (R). Deduce
that f2 (s) = s2 is not a monotone matrix function.

59
75. Given S ∈ SPDn (R), prove the formula

π n/2
Z
T Sx
√ = e−x dx.
det S Rn

Deduce that
S 7→ log det S
is concave on SPDn (R). Nota: an other proof was given in Exercise 58.

76. Let M ∈ Mn (R) be non-negative and let us choose numbers s1 , . . . , sn in (1, +∞). For
λ ∈ R, define
S λ := diag{sλ1 , . . . , sλn }.
Assume that M − S λ is non-singular for every λ ≥ 0. prove that ρ(M ) < 1.

77. (From T. BarbotBarbot.) Let S ∈ Mm (R) and R ∈ Mp×m (R) be given, with S symmetric.
Our goal is to prove that there exists a symmetric Σ ∈ Mp (R) such that
   
S RT S
(8) ≤ =: ρ.
R Σ 2
R 2

Indeed, since the reverse inequality is always true, we shall have an equality. Nota: This
is a particular case of Parrott’sParrott lemma. See Exercise number 87.
This property may be stated as a vector-valued version of the Hahn–BanachHahnBanach
theorem for symmetric operators: Let u : F → E, defined on a subspace of E, a Euclidian
space, with the symmetric property that hu(x), yi = hx, u(y)i for every x, y ∈ F , then
there exists a symmetric extension U ∈ L(E), such that kU k ≤ kuk (and actually kU k =
kuk) in operator norm.
By homogeneity, we are free to assume ρ = 1 from now on. In the sequel, matrix
inequalities hold in the sense of quadratic forms.

(a) Show that S 2 + RT R ≤ Im .

(b) For |µ| > 1, show that

Hµ := µIp − R(µIm − S)−1 RT

is well-defined, and that the map µ 7→ Hµ is monotone increasing on each of the

intervals (−∞, −1) and (1, +∞).
(c) We begin with the case where ρ(S) < 1. Then (Im − S 2 )−1 is well-defined and
symmetric, positive definite. Prove that kR(Im − S 2 )−1/2 k2 ≤ 1. Deduce that
k(Im − S 2 )−1/2 RT k2 ≤ 1. Conclude that H−1 ≤ H1 .
(d) Let Σ be symmetric with H−1 ≤ Σ ≤ H1 . For instance, Σ = H±1 is convenient.
Prove the inequality (8). Hint: Consider an eigenvector (x, y)T for an eigenvalue
µ > 1. Compute y T Σy and reach a contradiction. Proceed similarly if µ < −1.

60
 (e) In the general case, we have by assumption ρ(S) = kSk2 ≤ 1. Replace S by tS
with 0 < t < 1 and apply the previous result. Then use a compactness argument as
t → 1− .

78. Let K be a compact subgroup of GLn (R). We admit the existence of a HaarHaar measure,
that is a probability µ on K, with the left-invariance property:
Z Z
φ(gh) dµ(h) = φ(h) dµ(h), ∀φ ∈ C(K), ∀g ∈ K.
K K

(a) Let | · | denote the canonical Euclidian norm, and (·, ·) its scalar product. For
x, y ∈ Rn , define
Z
hx, yi := (hx, hy) dµ(h), kxk := hx, xi1/2 .
K

Show that h·, ·i is a scalar product, for which every element of K is an isometry.
(b) Deduce that K is conjugated to a subgroup of On (R). Similarly, prove that every
compact subgroup of GLn (C) is conjugated to a subgroup of Un .

79. (Thanks to P. de la Harpedela@de la Harpe and E. GhysGhys.) Let p, q ≥ 1 be integers.

We endow Cp and Cq with the canonical Hermitian scalar products hy, zi := y1 z̄1 + · · · .
On Cp+q , we consider their difference. The corresponding Hermitian form is

Q(z) = |z1 |2 + · · · − |zp+1 |2 − · · · .

Denote by X the set of linear subspaces of Cp+q on which the restriction of Q is positive
definite, and which are maximal for this property.

(a) Show that E ∈ X if, and only if, it is the graph

{(x, M x) | x ∈ Cp }

of a matrix M ∈ Mq×p (C) with kM k < 1 ; this norm is taken with respect to the
Hermitian norms of Cp and Cq , in particular, kM k2 = ρ(M ∗ M ) = ρ(M M ∗ ).
(b) Let Z ∈ U(p, q), written blockwise as
 
A B
Z= .
C D

Verify that, if E ∈ X , then ZE ∈ X . Deduce that if M lies in the unit ball B of

Mq×p (C), then so does the matrix

σZ (M ) := (AM + B)(CM + D)−1 .

(c) Show that σ : U(p, q) → Bij(B) is a homomorphism. In other words, it is a group

action of U(p, q) over B.

61
 (d) Let M ∈ B be given. Prove that there exists a unique element Z of U(p, q) ∩ HPDn
such that M = σZ (0q×p ). This shows that the group action is transitive.
(e) Find the stabilizer of 0q×p , that is the set of Z’s such that σZ (0q×p ) = 0q×p .
(f) Deduce that B is diffeormorphic to the homogeneous space

U(p, q)/(Up × Uq ).

80. Given a function φ : R → R, we admit that there is a unique way to define a Φ :

Symn (R) → Symn (R) such that Φ(OT SO) = OT SO if O is orthogonal, and Φ(D) =
diag{φ(d1 ), . . . , φ(dn )} in the diagonal case (see for instance Exercise 63.a). It is clear
that the spectrum of Φ(S) is the image of that of S under φ.
Let B ∈ Mn (R) be symmetric and A ∈ Mn (R) be diagonal, with diagonal entries
a1 , . . . , an . If φ is of class C 2 , prove Balian’sBalian formula:
X X φ0 (aj ) − φ0 (ai )
lim t−2 Tr(Φ(A + tB) + Φ(A − tB) − 2Φ(A)) = b2ii φ00 (ai ) + b2ij .
t→0
i
aj − ai
i,j (j6=i)

Hint: Prove the formula first in the case where φ(t) = tm for some integer m. Then pass
to general polynomials, then to C 2 functions.

81. We denote by k · kF the FrobeniusFrobenius norm: kAk2F = Tr(A∗ A).

(a) Show that the set Nn of n × n normal matrices is closed. Deduce that if A ∈ Mn (C),
there exists an N in Nn for which kA − N kF is minimum.
(b) Given h ∈ Hn and t ∈ R, exp(ith) is unitary. Therefore we have

kA − N kF ≤ kA − eith N e−ith kF .

By letting t → 0, deduce that

(9) (A − N )N ∗ − N ∗ (A − N ) ∈ Hn .

(c) Using a unitary conjugation, show that we may assume that N is diagonal.
In that case, write N = diag{d1 , . . . , dn }. Then:
i. Show that dj = ajj . Hint: Compare with other diagonal matrices.
ii. Suppose that dj = dk (=: d) for some pair (j, k) (j 6= k). Verify that

kB − dI2 kF ≤ kB − nkF , ∀n ∈ N2 ,

where  
ajj ajk
B := .
akj akk
Deduce that ajk = akj = 0 (in other words, B = dI2 ).

62
 iii. From (9) and the previous question, deduce that one can define a Hermitian
matrix H, such that

hjk (dk − dj ) = ajk , ∀j, k (j 6= k).

(d) In conclusion, prove that for every A in Mn (C), there exist a normal matrix N (the
one defined above) and a Hermitian one H, such that

A = N + [H, N ].

82. (von Neumannvonneu@von Neumann inequality.)

Let M ∈ Mn (C) be a contraction, meaning that

kM k2 ≤ 1. In other words, there holds M ∗ M ≤
In in the sense of Hermitian matrices. We re-
call that kM ∗ k2 = kM k2 , so that
√ we also have
∗ ∗
M M √≤ In . We denote S = In − M M and
T := In − M M ∗ . Such positive square roots do
vonneu@von
exist, from unitary diagonalisation ; they turn out
Neumann, John!(USA)
to be unique, but we do not use this fact.
John von Neumann.

Given an integer k ≥ 1 and K = 2k + 1, we define a matrix Vk ∈ MKn (C) blockwise:

 .. 
.
In
 
 
In
 
 
∗
S −M
 
 
Vk =  M T ,
 
In
 
 
In
 
 

 . .. 


In

where the dots represent blocks In , while missing entries are blocks 0n . The column and
row indices range from −k to k. In particular, the central block indexed by (0, 0) is M .
All the other diagonal blocks are null.

(a) We begin with the easy case, where M is normal. Prove that

kp(M )k2 = max{|p(λ)| ; λ ∈ Sp(M )}.

(b) We turn to the general case. Check that M S = T M and SM ∗ = M ∗ T . Deduce that
Vk is unitary.

63
 (c) Show that, whenever q ≤ 2k, the central block of the q-th power Vkq equals M q .
Deduce that if p ∈ C[X] has degree at most 2k, then the central block of p(Vk )
equals p(M ).
(d) Then, still assuming do p ≤ 2k, show that kp(M )k2 ≤ kp(Vk )k2 .
(e) Deduce von Neumannvonneu@von Neumann inequality:

kp(M )k2 ≤ max{|p(λ)| ; λ ∈ S 1 },

where S 1 is the unit circle.

83. Let k be field and A ∈ Mn (k) be given. We denote B := adjA = ÂT the transpose
of the cofactors matrix. We recall BA = AB = (det A)In . Denote also the respective
characteristic polynomials

pA (X) = X n − a1 X n−1 + · · · + (−1)n an , pB (X) = X n − b1 X n−1 + · · · + (−1)n bn .

(a) Prove the identity a 

n n n
bn pA (X) = (−1) X pB .
X
(b) Deduce that, if det A 6= 0, there holds bj = an−j aj−1
n for j = 1, . . . , n.
(c) Extend these formulas to the general case. Hint: Apply the previous question when
k is replaced by the ring A[a11 , . . . , ann ], where the indeterminates aij are the entries
of a general matrix A. See for instance the proof of Theorem 2.1.1.
(d) Conclude that the spectrum of B is given, counting with multiplicities, by

λ1 · · · λj−1 λj+1 · · · λn , j = 1, . . . , n,

where λ1 , . . . λn are the eigenvalues of A.

(e) Compare with the additional exercise 56.

84. We consider an n×n matrix X whose entries xij are independent indeterminates, meaning
that the set of scalars is the ring A := Z[x11 , x12 , . . . , xnn ]. We embed A into its field of
fractions k = Z(x11 , x12 , . . . , xnn ).

(a) Prove that X is non-singular.

(b) Let 1 ≤ p ≤ n − 1 be an integer. Consider the block forms
   
Xp · −1 · ·
X= , X = ,
· · · Yn−p

where Xp ∈ Mp (A) and Yn−p ∈ Mn−p (k). Prove the identity

det Xp
det Yn−p = .
det X

64
85. Let k be a field and V a finite dimensional k-vector space. A flag in V is a sequence
V = (V1 , . . . , Vn = V ) of subspaces with the properties dim Vm = m and Vm ⊂ Vm+1 .
In particular, n = dim V . A basis {X1 , . . . , Xn } is adapted to the flag if for every m,
{X1 , . . . , Xm } is a basis of Vm . Obviously, every flag admits an adapted basis, and con-
versely, an adapted basis determines uniquely the flag. Two adapted bases differ only by
a “triangular” change of basis:

Ym = amm Xm + am,m−1 Xm−1 + · · · + am1 X1 , amm 6= 0.

Identifying V to k n , we deduce that the set of flags is in one-to-one correspondence

with the set of right cosets GLn (k)/Tsup (k), where Tsup (k) denotes the subgroup of
upper triangular matrices whose diagonal is non-singular. Therefore, questions about
flags reduce to questions about GLn (k)/Tsup (k).

(a) Consider the statement

(B): given two flags V and V 0 in V , there exists a basis adapted to V, of
which a permutation is adapted to V 0 .
Prove that (B) is equivalent to
(M): given A ∈ GLn (k), there exists T, T 0 ∈ Tsup (k) and a permutation
matrix P such that A = T P T 0 .
(b) In the statement (M), prove that the permutation is necessarily unique.
(c) We turn to the existence part. Thus we give ourselves A in GLn (k) and we look for
T, T 0 and a permutation σ, such that
n
X
aij = tiσ(r) t0rj .
r=1

Show the necessary condition

σ(1) = max{i ; ai1 6= 0}.

(d) Show that there exists an index i 6= σ(1) such that (we use the notation of Section
2.1 for minors)  
σ(1) i
A 6 0.
=
1 2
Hint: This amounts to finding syzygies between minors taken from two given
columns, here the first and the second.
Then prove that σ(2) must be the maximum of such indices i.
(e) By induction, prove the necessary condition that σ(j) is the largest index i with the
properties that i 6= σ(1), . . . , σ(j − 1) and
 
σ(1) · · · σ(j − 1) i
A 6= 0.
1 ··· j−1 j

65
 (f) Deduce the theorem that for every two flags V and V 0 , there exists a basis adapted
to V, of which a permutation is adapted to V 0 .
(g) In the particular case k = R or C, prove that, for every A in a dense subset of GLn (k),
(M) holds true, with P the matrix associated with the permutation σ(j) = n + 1 − j.

86. Show that a complex matrix A ∈ Mn (C) is Hermitian if and only if hAx, xi is real for
every x ∈ Cn , where h·, ·i is the standard scalar product.

87. Let k = R or C. The matrix norms that we consider here are subordinated to the `2 -norms
of k d .
Given three matrices A ∈ Mp×q (k), B ∈ Mp×s (k) and C ∈ Mr×q (k), we consider the
affine set W of matrices W ∈ Mn×m (k) of the form
 
A B
W = ,
C D

where D runs over Mr×s (k). Thus n = p + r and m = q + s.

Denoting  
I
P = , Q=( I 0 )
0
the projection matrices, we are going to prove (Parrott’sParrott Lemma) that

(10) min{kW k ; W ∈ W} = max{kQW k, kW P k},

where the right hand side does not depend on D:

 
A
WP = , QW = ( A B )
C

(a) Check the inequality

inf{kW k ; W ∈ W} ≥ max{kQW k, kW P k}.

(b) Denote µ(D) := kW k. Show that the infimum of µ on W is attained.

(c) Show that it is sufficient to prove (10) when s = 1.
(d) From now on, we assume that s = 1, and we consider a matrix D0 ∈ Mr×1 (k) such
that µ is minimal at D0 . We denote by W0 the associated matrix. Let us introduce
a function D 7→ η(D) = µ(D)2 . Recall that η is the largest eigenvalue of W ∗ W . We
denote f0 its multiplicity when D = D0 .
i. If f0 ≥ 2, show that W0∗ W0 has an eigenvector v with vm = 0. Deduce that
µ(D0 ) ≤ kW P k. Conclude in this case.

66
 ii. From now on, we suppose f0 = 1. Show that η(D) is a simple eigenvalue for
every D in a small neighbourhood of D0 . Show that D 7→ η(D) is differentiable
at D0 , and that its differential is given by
2
∆ 7→ 2
< [(QW0 y)∗ ∆Qy] ,
kyk

where y is an associated eigenvector:

W0∗ W0 y = η(D0 )y.

iii. Deduce that either Qy = 0 or QW0 y = 0.

iv. In the case where Qy = 0, show that µ(D0 ) ≤ kW P k and conclude.
v. In the case where QW0 y = 0, prove that µ(D0 ) ≤ kQW k and conclude.

Nota: ParrottParrott is the name of a mathematician. Therefore, Parrott’sParrott

Lemma has nothing to do with the best seller Le Théorème du Perroquet, written by
the mathematician Denis GuedjGuedj.

88. For a Hermitian matrix A, denote by Pk the leading principal minors:

a11 · · · a1k
Pk := det ... . . . .. .
.
ak1 · · · akk

When k = 0, we also set P0 = 1. Finally, we set

Pk
k := sign ∈ {−1, 0, 1}, k = 1, . . . , n.
Pk−1

(a) We assume that Pk 6= 0 for all k. Prove that the number of positive eigenvalues of
A is precisely the number of 1’s in the sequence 1 , . . . , n .
Hint: Argue by induction, with the help of the interlacing property (Theorem 3.3.3).
(b) We assume only that Pn = det A 6= 0. Prove that the number of negative eigenvalues
of A is precisely the number of sign changes in the sequence 1 , . . . , n (the zeros are
not taken in account).

89. We define a HilbertHilbert space H2 of holomorphic functions on the unit disc D, endowed
with the scalar product Z 2π
1
hf, gi := f (eiθ )g(eiθ ) dθ.
2π 0
Don’t worry about this loosy definition. You may view H2 as the completion of the space
of polynomials under the norm p
kf k = hf, f i,

67
 or as the set of L2 -functions on the unit circle T that have a holomorphic extension to D.
In other words, L2 -functions f : T → C such that
Z 2π
f (eiθ )eimθ dθ = 0, m = 0, 1, 2, . . .
0

We define the SzegöSzegö kernel

1
k(λ, µ) = .
1 − λµ̄
When λ ∈ D, we define a holomorphic function

kλ (z) := k(z, λ).

(a) If u ∈ H2 , prove
hkλ , ui = u(λ).
(b) For φ holomorphic and bounded on D, the operator Mφ : u 7→ φu is bounded on H2 .
Prove
Mφ∗ kλ = φ(λ) kλ .
(c) Prove also that
kMφ∗ k = kMφ k = sup{|φ(z)| ; z ∈ T}.
(d) We assume moreover that φ : D → D. Given N distinct numbers λ1 , . . . , λN in D,
we form the Hermitian matrix A with entries

aij := (1 − wi w̄j )k(λi , λj ),

where wj = φ(λj ). Prove that A is semi-definite positive.

Hint: Write that Mφ Mφ∗ ≤ I on the space spanned by the kλj ’s.
Nota: Pick’sPick Theorem tells us that, given the λj ’s and the wj ’s, A ≥ 0N
is equivalent to the existence of a holomorphic function φ : D → D such that
φ(λj ) = wj . This is an interpolation problem. The space H2 is a HardyHardy
space.

90. Let A ∈ Hn be a semi-definite positive Hermitian matrix, with ajj > 0 and ajk 6= 0 for
every (j, k). Let us form the Hermitian matrix B such that bjk := 1/ajk . Assume at last
that B is semi-definite positive too.
√
(a) Prove that |ajk | = ajj akk , using principal minors of rank 2.
(b) Using principal minors of rank 3, show that ajk akl ājl is real positive.
(c) Deduce that A is rank-one: There exists a v ∈ Cn such that A = vv ∗ .
(d) What does it tell in the context of the previous exercise ?

68
91. Let A ∈ Mn (C) be a normal matrix. We decompose A = L + D + U in strictly lower,
diagonal and strictly upper triangular parts. Let us denote by `j the Euclidean length of
the j-th column of L, and by uj that of the j-th row of U .
(a) Show that
k k j−1
k X
X X X
u2j ≤ lj2 + u2mj , k = 1, . . . , n − 1.
j=1 j=1 j=1 m=1

(b) Deduce the inequality √

kU kS ≤ n − 1kLkS ,
where k · kS is the Schur–FrobeniusSchurFrobenius norm.
(c) Prove also that
1
kU kS ≥ √ kLkS .
n−1
(d) Verify that each of these inequalities are optimal. Hint: Consider a circulant matrix.
92. The ground field is R.
(a) Let P and Q be two monic polynomials of respective degrees n and n − 1 (n ≥ 2).
We assume that P has n real and distinct roots, strictly separated by the n − 1 real
and distinct roots of Q. Show that there exists two real numbers d and c, and a
monic polynomial R of degree n − 2, such that
P (X) = (X − d)Q(X) − c2 R(X).

(b) Let P be a monic polynomial of degree n (n ≥ 2). We assume that P has n real and
distinct roots. Build sequences (dj , Pj )1≤j≤n and (cj )1≤j≤n−1 , where dj , cj are real
numbers and Pj is a monic polynomial of degree j, with
Pn = P, Pj (X) = (X − dj )Pj−1 (X) − c2j−1 Pj−2 (X), (2 ≤ j ≤ n).
Deduce that there exists a tridiagonal matrix A, which we can obtain by algebraic
calculations (involving square roots), whose characteristic polynomial is P .
(c) Let P be a monic polynomial. We assume that P has n real roots. Prove that
one can factorize P = Q1 · · · Qr , where each Qj has simple roots, and the factoriza-
tion requires only finitely many operations. Deduce that there is a finite algorithm,
involving no more than square roots calculations, which provides a tridiagonal sym-
metric matrix A, whose characteristic polynomial is P (a tridiagonal symmetric
companion matrix).
93. (D. KnuthKnuth.) Let A be an alternate matrix in Mn+1 (k), k a field. We index its
rows and columns from 0 to n (instead of 1, . . . , n + 1), and form the matrices Aji by
removing from A the i-th row and the j-th column. We denote also by PF[l1 , . . . , l2m ]
the Pfaffian of the alternate matrix obtained by retaining only the rows and columns of
indices l1 , . . . , l2m .

69
 (a) Prove the following formulas, either (if n is even):

det A01 = PF[0, 2, . . . , n] PF[1, 2, . . . , n],

or (if n is odd):
det A01 = PF[0, 1, 2, . . . , n] PF[2, . . . , n].
Hint: If n is odd, expand the identity det(A + XB) = (Pf(A + XB))2 where
b12 = −b21 = 1 and bij = 0 otherwise. Then use Exercise 11.b) in the present list. If
n is even, expand the identity det M (X, Y ) = (PfM (X, Y ))2 , where
 
0 −X −Y 0 · · · 0
 X 
 
 Y 
M (X, Y ) :=  0 .
 
 A 
 .. 
 . 
0

(b) With the first formula, show that if A is an m × m alternate matrix, then the
transpose matrix of cofactors adjA is symmetric for m odd! Prove that in fact adjA
has the form ZZ T where Z ∈ k m . Compare with the additional exercise 56.
(c) On the contrary, show that adjA is alternate when m is even. Prove also that the
entries of the matrix
1
à := adjA
PfA
are polynomials in the entries of A (every entry of adjA is a multiple of the Pfaffian).
This matrix à plays an intermediate role between adjA and A−1 in that the inverse
of A is given by
1
A−1 = Ã.
PfA
94. Let k be a field and n an even integer. If x, y ∈ k n , denote by x ∧ y the alternate matrix
xy T − yxT . Show the formula

Pf(A + x ∧ y) = (1 + y T A−1 x) PfA

for every non-singular alternate n × n matrix A.

Hint: We recall the formulæ
1
det(M +xy T ) = (1+y T M −1 x) det M, (M +xy T )−1 = M −1 − M −1 xy T M −1 .
1 + y T M −1 x

95. Let k be a field, n be an even integer and A be an n × n non-singular alternate matrix.

Using the odd case of Exercise 93 above, prove the formula
1
A−1 = α(i, j)(−1)i+j+1 PfAij 1,≤i,j≤n ,


PfA
70
 where Aij is obtained from A be removing the i-th and j-th rows and columns, and α(i, j)
is ± according to the sign of j − i. Compare this formula with Exercise 11.b) above.
In particular, show that PfA divides, as a polynomial, every entry of adjA.
96. (BanachBanach.) Let p ∈ [1, +∞] be such that p 6= 2. We consider matrices M ∈ Mn (R)
which are isometries, namely
kM xkp = kxkp , ∀x ∈ Rn .
(a) Let us begin with the case 2 < p < +∞. We give x, y ∈ Rn such that xi yi = 0 for
every i ≤ n. Define u = M x and v = M y, and let H be the set of indices j such
that uj vj 6= 0.
i. Show that the function
X
θ(s) := (|suj + vj |p − |suj |p − |vj |p )
j∈H

vanishes identically.
ii. Computing the second derivative of θ, show that H is void.
(b) (Continued, 2 < p < +∞.) Let mk be the number of non-zero entries in the k-th
column of M , and Ek be the vector space spanned by the other columns. Using
the previous question, show that dim Ek ≤ n − mk . Then deduce that mk = 1. At
last, show that M is the product of a diagonal matrix diag(±1, . . . , ±1) and of a
permutation matrix.
(c) If 1 < p < 2, prove the same conclusion. Hint: Apply the previous result to M T .
(d) If p = 1, prove directly that if xi yi = 0 for every index i, then (M x)j (M y)j = 0 for
every index j. Hint: Use θ(1) = θ(−1) = 0. Conclude.
(e) If p = +∞, conclude by applying the case p = 1 to M T .
97. (LemmensLemmens & van Gaansvangaa@van Gaans.) We endow Rn with some norm
k · k. Let M ∈ Mn (R) be non-expansive: kM xk ≤ kxk for every x ∈ Rn .
(a) Let B be the unit ball. Show that
\
D := M kB
k≥1

is a compact symmetric convex set. We denote by E the vector space spanned by

D.
(b) Show that M E = E and that the restriction of M to E is an isometry.
(c) Let (kj )j∈N be increasing sequence such that M kj → A. Prove that AB = D.
(d) Show that there exists an increasing sequence (kj )j∈N such that M kj converges.
Prove that M kj+1 −kj converges towards a projector P whose range is E, and which
is non-expansive.

71
 98. Given an alternate 4 × 4 matrix A, verify that its characteristic polynomial equals
X
X4 + X2 a2ij + Pf(A)2 .
i<j

We define

R+ (A) = (a12 + a34 )2 + (a23 + a14 )2 + (a31 + a24 )2 ,

R− (A) = (a12 − a34 )2 + (a23 − a14 )2 + (a31 − a24 )2 .

Factorize PA in two different ways and deduce the following formula for the eigenvalues
of A, in characteristic different from 2:
i p p 
± R+ (A) ± R− (A) ,
2
where the signs are independent of each other.

99. (a) Verify that the characteristic polynomial PV of a real orthogonal matrix V can be
factorized as  
r s m 1
PV (X) = (X − 1) (X + 1) X Q X +
X
where Q ∈ R[X] a monic polynomial whose roots lie in (−2, 2).
(b) Conversely, we give a monic polynomial Q ∈ R[Y ] of degree m, whose roots lie in
(−2, 2), and we consider
 
m 1
P (X) = X Q X + .
X

Let A be a tridiagonal symmetric matrix whose characteristic polynomial is Q (see

Exercise 92.)
i. Prove that Im − 41 A2 is positive definite.
ii. Let us define r
1
B := In − A2 .
4
Prove that  
A B
V :=
−B A
is orthogonal.
iii. Prove PV = P (V is a companion matrix of P .)

100. (Inspired by M. T. KaraevKaraev.) Let E be a finite dimensional HilbertHilbert space.

We are interested in nilpotent endomorphisms. Recall that u ∈ L(E) is nilpotent of order
m if um = 0E but uk 6= 0E if k < m.

72
 (a) Let F be the orthogonal of ker u and let G be u(F ). Prove that there exists an
orthonormal basis of F , whose image is an orthogonal basis of G. Hint: This is
essentially the Singular Value Decomposition.
(b) Deduce that, if m = 2, there exists an orthonormal basis of E, in which the matrix
of u has the “JordanJordan!Camille” form
 
0 a2 0 · · · 0
 .. . . . . . . .
 . . . . ..


(11)
 ... 
.

 0 
 . ...
 ..

an 
0 ··· ··· 0

(c) Arguing by induction, prove the same result for every m ≥ 2.

(d) Deduce that if M ∈ Mn (C) is nilpotent, then M is unitarily similar to a matrix of
the form (11).
(e) If M and N are unitarily similar, check that their numerical ranges (see Exercise
21) are equal, and that kM k2 = kN k2 .
(f) Let M ∈ Mn (C) be nilpotent of order m. Prove that H(M ) is a disk centered at
the origin. Show that its radius is less than or equal to (HaagerupHaagerup–de la
Harpedela@de la Harpe inequality)
π
kM k2 cos .
m+1
Hint: It is enough to work in the case M is of the form (11) and the aj ’s are non-
zero. Then it is easy to show that H(M ) is rotationaly invariant. Since it is convex
(Exercise 21), it is a disk. The triangular inequality leads to the computation of the
spectral radius of  
0 1/2 0 · · · 0
. . . .. 
 1/2 . . . . . . . 

 . .. . ..

.
 0 0 

 . .. ..
 ..

. . 1/2 
0 ··· 1/2 0
101. (From de Oliveiradeol@de Oliveira.) Let (α1 , . . . , αn ) ∈ Cn and (β1 , . . . , βn ) ∈ Cn be
given. Let us form the diagonal matrices ∆ := diag{α1 , . . . , αn } and D := diag{β1 , . . . , βn }.
(a) If V is unitary, show that
 
α1
Tr(∆V DV ∗ ) = βn S  ...  ,


β1 · · ·

αn
for some orthostochastic matrix S.

73
 (b) Using Birkhoff’sBirkhoff Theorem, deduce that Tr(∆V DV ∗ ) belongs to the convex
hull of the set of numbers
n
X
αj βσ(j) , σ ∈ Sn .
j=1

(c) More generally, given two normal matrices A, B ∈ Mn (C) whose respective spectra
are (α1 , . . . , αn ) and (β1 , . . . , βn ), prove that Tr(AB) belongs to this convex hull.
Nota: See Exercise 139 for a related result.
(d) With the same notations as above, prove that
n
X n
X
min |αj − βσ(j) |2 ≤ kA − Bk2F ≤ max |αj − βσ(j) |2 ,
σ∈Sn σ∈Sn
j=1 j=1

where k · kF is the Frobenius norm. The first inequality constitutes Hoffman–

WielandtHoffmanWielandt theorem. The case for Hermitian matrices was found
earlier by Loewner (Löwner)Loewner.
102. (Y. Shizuta, S. KawashimaShizutaKawashima.) Let A, B be n × n Hermitian matrices,
with B ≥ 0n . We denote by a1 , . . . , ar the distinct
P eigenvalues of A and by P1 , . . . , Pr the
corresponding eigenprojectors, so that A = j aj Pj .
(a) Prove that Pj is Hermitian and that Pj Pk = 0n when k 6= j.
(b) Let us assume that there exists a skew-Hermitian matrix K such that B + [K, A] is
positive definite. Show that ker B does not contain any eigenvector of A.
(c) Let us assume that the intersection of the kernels of B, [B, A], [[B, A], A],... (take
successive commutators with A) equals {0}. Prove that ker B does not contain any
eigenvector of A.
(d) Conversely, we assume that ker B does not contain any eigenvector of A.
i. Define X 1
K := Pi BPj .
i6=j
ai − aj

Show that K is skew-Hermitian, and that B + [K, A] is positive definite.

ii. Show that the intersection of the kernels of B, [B, A], [[B, A], A],... equals {0}.
103. Let A ∈ Mn×m (k) be given, with k = R or k = C. We put the singular values σ1 ≥ σ2 ≥
· · · in decreasing order. We endow Mn×m (k) with the norm k·k2 . Prove that the distance
of A to the set Rl of matrices of rank less than or equal to l, equals σl+1 (if l = min{n, m},
put σl+1 = 0). Hint: Use Theorem 7.7.1 of Singular Value Decomposition.
104. (a) Let T0 ∈ Mn (C) be such that kT0 k2 ≤ 1. Prove
1
(12) k(T0 − wIn )−1 k2 ≤ , ∀w ∈ C ; |w| > 1
|w| − 1

74
 Hint: Use von Neumannvonneu@von Neumann Inequality (Exercise 82), while ap-
proximating the function f (z) := (z − w)−1 by polynomials, uniformly on a neigh-
bourhood of Sp(T0 ).
(b) Let T ∈ Mn (C) be of the form T = aIn + N where a ∈ C and N 2 = 0n .
i. Prove that
(kT k2 ≤ 1) ⇐⇒ kN k2 + |a|2 ≤ 1 .

Deduce that kT k2 is the largest root of the quadratic equation

r2 − kN k2 r − |a|2 = 0.

ii. Prove that

   
1 1
k(T − wIn )−1 k2 ≤ , ∀w ∈ C ; |w| > 1 ⇐⇒ kN k2 + |a| ≤ 1 .
|w| − 1 2

iii. Deduce that the converse of (12) does not hold if n ≥ 2. In particular, one
cannot replace the assumption kT k2 ≤ 1 by (12) in the inequality of von Neu-
mannvonneu@von Neumann.
iv. However, prove that for such a T = aIn + N , (12) implies kT k2 ≤ 2, and the
equality is achieved for some (a, N ).
(c) Let T ∈ Mn (C) satisfy Property (12). Prove that

kT k2 ≤ e.

Hint: Use the formula

Z
1
T = z k (zIn − T )−k dz.
2ikπ Γr

(Interestingly enough, this part of the exercise is true not only for the norm k · k2 ,
but also in every BanachBanach algebra.)

105. Let A, B be two positive semi-definite Hermitian matrices.

(a) Prove that for every x ∈ Cn , kBxk22 ≤ kBk2 hBx, xi.

(b) Deduce that BAB ≤ kAk2 kBk2 B in the sense of Hermitian matrices.

106. (From L. TartarTartar.)

n
P
(a) P
Let f (z) = n≥0 an z be a series, converging for |z| < R. We denote F (z) :=
n
n≥0 |an |z . Check that whenever A is a BanachBanach algebra and a, b ∈ A, there
holds
F (kbk) − F (kak)
kf (b) − f (a)k ≤ kb − ak.
kbk − kak

75
 (b) Deduce that if A, B ∈ Hn (C), then

keiB − eiA k2 ≤ kB − Ak2 .

Hint: Choose an integer m ≥ 1. Apply the previous result to f (z) = eiz/m , and
decompose eiB − eiA as a sum of m products having eiB/m − eiA/m as one factor and
unitary matrices otherwise. Then let m → +∞.

107. Let k be R or C. We consider a norm on k n such that the induced norm has the property
kA2 k = kAk2 for every Hermitian, positive semi-definite matrix A.

(a) Show that kAk = ρ(A) for every A ∈ H+

n.

(b) Deduce that, for every x ∈ k n , there holds kxk22 = kxk kxk∗ (recall that k · k∗ is the
dual norm of k · k).
(c) Let Σ be the unit sphere of norm k · k. Given x ∈ Σ, show that Σ is on one side of
the plane defined by <(y ∗ x) = kxk2 .
(d) Deduce that k · k is proportional to k · k2 . Hint: Given a point x0 ∈ Σ, show
that the smallest convex cone with vertex x0 , which contains Σ, is the half-plane
<(y ∗ x) ≤ kxk2 (this uses the previous question). Deduce that Σ is a differentiable
manifold of codimension one. Then conclude with the previous question.
√
(e) Let k · k be and induced norm on Mn (k) such that the square root A 7→ A is
1/2-Holderian on HPDn (or SPDn if k = R), with constant one (compare with
Exercise 52):
√ √
(13) k B − Ak ≤ kB − Ak1/2 .

Prove that the norm on k n is proportional to k · k2 . Comment: For the validity of

(13) for k · k2 , see Exercise 110.

108. (Ky FanKy Fan norms.) Let sj (M ) (1 ≤ j ≤ n) denote the singular values of a matrix
M ∈ Mn (R), labelled in increasing order: s1 (M ) ≤ · · · ≤ sn (M ).

(a) We define σj := sj + · · · + sn . Prove the formula

σn−j+1 (M ) = sup{Tr(P M Q) ; P ∈ Mj×n (R), Q ∈ Mn×j (R)

s.t. P P T = Ij , QT Q = Ij }.

Deduce that σj is a convex function, and thus is a norm. Do you recognize the norm
σn = sn ?
(b) (Thanks to M. de la Sallede la Salle) Deduce that there exists two norms N± over
Mn (R), with the property that

(det M = 0) ⇐⇒ (kM k+ = kM k− ).

76
 (c) Let a := (a1 , · · · , an ) be a given n-uplet of non-negative reals numbers with a1 ≤
· · · ≤ an . We define

E(a) := {M ∈ Mn (R) ; sj (M ) = aj , for all 1 ≤ j ≤ n}.

Verify that the set E 0 (a) defined below is convex, and that it contains the convex
hull of E:

E 0 (a) := {M ∈ Mn (R) ; σj (M ) ≤ aj + · · · + an , for all 1 ≤ j ≤ n}.

(d) Show that the extremal points of E 0 (a) belong to E(a), and deduce that E 0 (a) is
the convex hull of E(a). Hint: The set ext(E 0 (a)) is left- and right-invariant under
multiplication by orthogonal matrices. Thus one may consider diagonal extremal
points.
(e) Deduce that the convex hull of On (R) is the unit ball of k · k2 . Remark: Here,
the convex hull of a set of small dimension (n(n − 1)/2) has a large dimension n2 .
Thus it must have faces of rather large dimension ; this is precisely the contents of
Corollary 5.5.1, when applied to the induced norm k · k2 . See Exercise 137 below for
a more accurate description.

109. (Continuation.) We say that a subset X of Mn (R) is rank-one convex if whenever A, B ∈

X and B − A is of rank one, then the segment (A, B) is included in X. Rank-one
convexity is preserved under intersection. If Y is a subset of Mn (R), its rank-one-convex
hull is the smallest rank-one convex subset that contains Y . We recall that a function
f : X → R ∪ {+∞} is rank-one convex if it is convex on every segment of the type above.

(a) Let f : Mn (R) → R ∪ {+∞} be rank-one convex. Show that the sets defined by
f (M ) ≤ α are rank-one convex.
(b) Verify that M 7→ | det M | is rank-one convex.
(c) Prove the formula for products of singular values:

sn (M ) · · · sn−j+1 (M ) = sup{det(P M Q) ; P ∈ Mj×n (R), Q ∈ Mn×j (R)

s.t. P P T = Ij , QT Q = Ij }.

(d) Deduce that πj := sn · · · sj is rank-one convex.

(e) Deduce that the rank-one-convex hull of the set E(a) (see the previous exercise) is
included in

E 00 (a) := {M ∈ Mn (R) ; πj (M ) ≤ aj · · · an , for all 1 ≤ j ≤ n}.

Comment: A result of B. DacorognaDacorogna tells us that E 00 (a) is actually equal

to the rank-one-convex hull of E(a).

110. Let A, B be n × n positive semi-definite Hermitian matrices.

77
 (a) Using the formula (TaylorTaylor expansion)
B 2 − A2 = (B − A)2 + A(B − A) + (B − A)A,
prove that ρ(B − A)2 ≤ kB 2 − A2 k2 . Hint: Use an eigenvector of B − A. The
formula above might not be the useful one in some case.
(b) Deduce that√(13) holds
√ true for the 1/2operator norm k · k2 . Comment: The resulting
inequality k B − Ak2 ≤ kB − Ak2 is much more powerful than that of Exercise
52, since it does not depend on the dimension n. In particular, it holds true for
bounded self-adjoint operators in Hilbert spaces.
(c) Let x 7→ S(x) be a map of class C 2 from the unit ball Bd of Rd to
p SPDn . Assume
that the second derivatives are bounded over Bd . Prove that x 7→ S(x) is Lipschitz
continuous.
111. (Continuation.) Likewise, write the TaylorTaylor expansion
B 3 = A3 + · · · + H 3 , H := B − A.
Then, using an eigenvector e of H, associated with ρ(H), show that
e∗ (B 3 − A3 )e − ρ(H)3 kek22 = 2ρ(H)kAek22 + e∗ AHAe + 3ρ(H)2 e∗ Ae.
Prove the bound |e∗ AHAe| ≤ ρ(H)kAek22 and deduce that
kB − Ak32 ≤ kB 3 − A3 k2 .
What about the map A 7→ A1/3 over HPDn ? Any idea about A 7→ A1/4 (simpler than
what you think in a first instance) ?
112.
Given a real polynomial

P = X n − a1 X n−1 + · · · + (−1)n an ∈ R[X],

whose roots are x1 , . . . , xn (repeted with multiplic-

ities), we define the NewtonNewton sums
X
sk := xkα (s0 = n).
α

We recall that sk is a polynomial in a1 , . . . , ak , with

integer coefficients.
Let us form the HankelHankel matrix
 
s0 s1 · · · sn−1
 s1 s2 · · · sn 
Newton, H :=  .. ..  .
 
Isaac!(France)  . . 
sn−1 sn · · · s2n−2
Isaac Newton.

78
 Let Q be the quadratic form associated with H. Write Q as a sum of squares. Deduce
that the rank of H equals the number of distinct complex roots of P , while the index
of H (the number of positive squares minus the number of negative squares) equals the
number of distinct real roots of P . Conclude that the roots of P are all real if, and only
if, H is positive semi-definite.

113. (Continuation. NewellNewell (1972/73), IlyushechkinIlyushechkin (1985), LaxLax (1998),

DomokosDomokos (2011) ; special thanks to L. TartarTartar.) In the previous exercise,
set P := det(XIn − A) where A is a general matrix in Symn (R). Check that the sk ’s
are polynomials in the entries of A. Show that H is positive semi-definite. Deduce that
every principal minor of H,  
i1 . . . ir
H ,
i1 . . . ir
is a polynomial in the entries of A, and that it takes only non-negative values.
According to Artin’sArtin Theorem (see J. BochnakBochnak M. CosteCoste & M.-F.
RoyRoy: Real algebraic geometry, Springer-Verlag (1998), Theorem 6.1.1), this property
ensures that these minors are sums of squares of rational functions. HilbertHilbert pointed
out that not every non-negative polynomial is a sum of squares of polynomials. However,
it turns out that these principal minors are sums of squares of polynomials : Let us endow
Symn (R) with the scalar product hA, Bi := Tr(AB).

(a) Check that H is a GramGram matrix: hij = hWi , Wj i for some Wi ∈ Symn (R).
(b) Show that the exterior algebra of Symn (R) is naturally endowed with a scalar prod-
uct.
(c) Show that  
i1 . . . ir
H = kWi1 ∧ · · · ∧ Wir k2
i1 . . . ir
and conclude. Remark: The scalar product of a Euclidean space E extends in a
natural way to the exterior algebra Λr E.

Nota: This formulation, due to IlyushechkinIlyushechkin, gives the discriminant over

Symn (R) as the sum of n! squares of polynomials. This upper bound has been improved
by DomokosDomokos into    
2n − 1 2n − 3
− .
n−1 n−1
The minimal number of squares is not yet know, except for n = 2 and 3, where 2 and 5
squares suffice.

114. Let A be a principal ideal domain. If M ∈ Mn (A) and M = P DQ with P, Q ∈ GLn (A)
and D diagonal, prove the following equality about cofactors matrices:

co(M ) = (det P )(det Q)P −T co(D)Q−T .

79
 Prove
D` (co(M )) = (det M )`−1 Dn−` (M )
and deduce the value of d` (co(M )), the `-th invariant factor of co(M ). Compare with the
result of Exercise 56.

115. (Potter.) Let k be a field and ω an element of k.

(a) Prove that there exists polynomials Pr,j ∈ Z[X] such that, for every integer n ≥ 1,
every element ω in k and every pair of matrices A, B ∈ Mn (k) such that

(14) AB = ωBA,

there holds
r
X
r
(15) (A + B) = Pr,j (ω)B j Ar−j .
j=0

Matrices satisfying (14) are said to ω-commute. Remark that they satisfy Ap B q =
ω pq B q Ap , a formula that is a discrete analogue of the Stone–von Neumann for-
mulaStonevonneu@von Neumann.
(b) Define polynomials
l
Y
φl (X) = (1 + X + · · · + X s−1 ).
s=1

Show the formula

φj φr−j Pr,j = φr .
Hint: Proceed by induction over r.
(c) Assume that ω is a primitive root of unity, of order r (then r is not the characteristic
of k). Deduce that (14) implies (Potter’s TheoremPotter)

(A + B)r = Ar + B r .

Remark. It is amazing that when r is prime, the identity (15) can occur in two
cases : either A and B ω-commute with respect to a primitive root of unity of order
r, or r is the characteristic of k and [A, B] = 0n ; and both cases exclude each other !
(d) Let B be a cyclic matrix in the sense of Section 5.4:
 
0 M1 0 ··· 0
 .. ... ... ... ..
 . .


 . .. .. 
B :=  .. . . 0 .

 .. 
 0 . Mr−1 
Mr 0 ··· ··· 0

80
 We recall that the diagonal blocks are square (null) matrices, of respective sizes
n1 , . . . , nr . Let us define A := diag(In1 , ωIn2 , . . . , ω r−1 Inr ). We assume again that ω
is a primitive root of unity, of order r. Prove that (B, A) ω-commute and deduce
that
(A + B)r = In + diag(N1 , . . . , Nr ),
where Nj := Mj Mj+1 · · · Mj−1 . For instance,
N1 = M1 M2 · · · Mr , Nr = Mr M1 M2 · · · Mr−1 .

116. Let H ∈ HPDn and h ∈ Hn be given. We recall that M := hH −1 is diagonalizable with

real eigenvalues (see Exercise 258). We recall that a matrix is normal if and only if its
eigenvectors form a unitary basis. Thus the angles between eigenvectors of M measure
the deviation of M from normality. We compute here the minimum angle θ∗ as h runs
over Hn .
The space Cn is endowed with its Hermitian norm k · k2 .
(a) Let v and w be two eigenvectors of M associated with distinct eigenvalues. Prove
that
v⊥H w (i.e. w∗ Hv = 0).
(b) Let v ∈ Cn be given. We define a linear form L(w) := v ∗ w with domain the H-
orthogonal to v. Check that
kLk ≤ inf kv + rHvk2 .
r∈R

Deduce that
|v ∗ w|
 
K(H) − 1
Λ := sup ; w⊥H v ≤ inf ρ(In + rH) = ,
kvk2 kwk2 r∈R K(H) + 1
where K(H) is the condition number of H.
(c) Looking for a specific pair (v, w), prove that there holds actually
K(H) − 1
Λ= .
K(H) + 1

(d) Deduce that

2
sin θ∗ = .
K(H) + 1
117. Let us define Kn := ∆n ∩ Symn (R), the set of symmetric, bistochastic matrices.
(a) Show that Kn is the convex hull of the set of matrices of the form
1
Qσ := (Pσ + Pσ−1 ), (σ ∈ Sn )
2
where Pσ denotes the permutation matrix associated with σ.

81
 (b) If σ = (1, 2, . . . , n) is a cycle, prove that Qσ is extremal in Kn if and only if either
n is odd or n = 2. Hint: If n is even, show that Qσ = 21 (Q+ + Q− ) where Q± are
permutation matrices associated with involutions, and Q+ 6= Q− if n ≥ 4. If n is
odd, consider the graph Γ of pairs (i, j) for which qij 6= 0 in Qσ ; an edge between
(i, j) and (i0 , j 0 ) means that either i = i0 or j = j 0 . The graph Γ is a cycle of length
2n, and d((i, j), (j, i)) = n is odd. If Qσ = 12 (R + S) with R, S ∈ Kn , show that
rji + rij = 1 along Γ and conclude.
(c) Deduce that ext(Kn ), the set of extremal points of Kn , consists in the matrices Qσ
for the permutations σ that are products of disjoint cycles of lengths as above (either
odd or equal to two).

118. Let L : Mn (C) → C be a linear form with the properties that if H is Hermitian, then
L(H) is real, and if moreover H ≥ 0n , then L(H) ≥ 0. Prove that L(A∗ ) = L(A) for
every A ∈ Mn (C). Then prove, for every pair of matrices,

|L(A∗ B)|2 ≤ L(A∗ A)L(B ∗ B).

119. (See also Exercise 8)

(a) Prove the determinantal identity (Cauchy’sCauchy double alternant)

Q Q
1 i<j (aj − ai ) k<l (bk − bl )
= Q .
ai + bj 1≤i,j≤n i,k (ai + bk )

Hint: One may assume that

a1 , . . . , an , b1 , . . . , bn are indetermi-
nate, and then work in the field
Q(a1 , . . . , an , b1 , . . . , bn ). This determi-
nant is a homogeneous rational function
whose denominator is quite trivial. Some
specialisations make it vanishing ; this
gives an accurate information about the Cauchy,
numerator. There remains to find a scalar Augustin (France)
factor. That can be done by induction on n,
with an expansion with respect to the last Augustin Cauchy.
row and column.

82
 (b)
Define the HilbertHilbert matrix of order n as the
GramGram matrix

hij := hxi−1 , xj−1 iL2 (0,1) , (1 ≤ i, j ≤ n)

where
Z 1
hf, giL2 (0,1) := f (x)g(x) dx.
0

Use the formula above to compute (H −1 )nn .

Hilbert,
David (D. R. of Congo)

David Hilbert.

(c) In particular, find that kH −1 k2 ≥ 32n

1
16n (asymptotically, there holds a better bound
1
8π
16n .) Hint: The binomial (2m)!/(m!)2 is larger than 22m /(2m + 1).
(d) Deduce that the HilbertHilbert matrix is pretty much ill-conditionned:
1
κ(H) := kHk2 kH −1 k2 ≥ 16n .
32n
√
Remark: Hilbert’s matrix satisfies kHk2 ≤ π, and this estimate is the best one to
be uniform, in√the sense that, denoting Hn the Hilbert matrix of size n × n, one has
supn kHk2 = π. See for instance G. PólyaPol@Pólya & G. SzegöSzegö, Problems
and theorems in analysis, vol. I. Fourth edition (1970), Springer-Verlag. Part III,
chapter 4, exercise 169.

120. (I. KovacsKovacs, D. SilverSilver & S. WilliamsWilliams) Let A ∈ Mpq (k) be in block
form
A11 A12 · · ·
 
.
A =  A21 . .
 

..
.
where the blocks are p × p. Let us assume that these blocks commute pairwise. Prove
that q
X Y
det A = det ∆, ∆ := (σ) Amσ(m) .
σ∈Sq m=1

83
 Notice that ∆ is nothing but the determinant of A, considered as a q × q matrix with
entries in the abelian ring R generated by the Aij ’s. One may therefore write

detpq A = detp detR A,

where the subscripts indicate the meaning of each determinant. Hint: Use Schur’sSchur
formula for the determinant of a matrix with four blocks. Argue by induction over q.
121. Let us define the tridiagonal and block-tridiagonal matrices
   
0 1 Jp Ip
. ... . .
 1 .. 0  Ip . . . . 0
   
 

Jp :=  .. .. .. 
 ∈ Mp (k), Apq := 
 .. .. .. 
 ∈ Mpq (k).
 . . .   . . . 
 .. ..   .. .. 
 0 . . 1   0 . . Ip 
1 0 Ip Jp
Denote by Tp the polynomial

Tp (X) := det(XIp + Jp ).

Using the previous exercise, prove that the characteristic polynomial Ppq (Y ) of Apq is the
resultant
Res(Tq (· − Y ), T̂p ), T̂ (X) := T (−X).
Nota: Since these matrices have integer coefficients, their characteristic polynomials do
not depend of the scalar field k. It is therefore enough to consider the real case.
122. Let A ∈ Mn (C) have no purely imaginary eigenvalue (one says that A is hyperbolic). The
aim of this exercise is to prove the existence and uniqueness of a GreenGreen matrix. This
is a matrix-valued function G : R → Mn (C) that is bounded, differentiable for t 6= 0,
which has left and right limits G(0±), and satisfies
dG
(t) = AG(t), (t 6= 0), G(0+) − G(0−) = In .
dt
(a) We begin with the case where the eigenvalues of A have negative real part. We recall
that there exists a positive ω and a finite C such that k exp(tA)k ≤ Ce−tω for t > 0.
Prove that 
0n , t < 0,
G(t) :=
exp(tA), t > 0
defines a Green matrix.
Prove that there are constants as above such that k exp(tA)k ≥ C 0 e−tω for t < 0
(Hint: Use the inequality 1 ≤ kM k kM −1 k). Deduce that the Green matrix is
unique.
(b) Treat the case where A = diag(B, C), where the eigenvalues of B (resp. of C) have
negative (resp. positive) real parts.

84
 (c) Treat the general case.
(d) Show that actually the Green matrix decays exponentially fast at infinity.
(e) Let f : R → Cn be bounded continuous and define
Z
y(t) := G(t − s)f (s).
R

Show that y is the unique bounded solution of

y 0 (t) = Ay(t) + f (t).

123. Let A ∈ Mn (C) be given. The spectrum of A is split into three parts σ− , σ+ , σ0 according
to the sign of the real part of the eigenvalues. For instance, σ0 is the intersection of the
spectrum with the imaginary axis.

(a) Show that the following definitions of a subspace are equivalent:

• The sum of the generalized eigenspaces associated with the eigenvalues of neg-
ative real part,
• The largest invariant subspace on which the spectrum of the restriction of A lies
in the open left half-space of C.
• The set of data a ∈ Cn such that the solution of the Cauchy problem

x0 (t) = Ax(t), x(0) = a

tends to zero as t → +∞.

This subspace is called the stable invariant subspace (more simply the stable sub-
space) of A and denoted by S(A). Prove that if a ∈ S(A), the solution of the Cauchy
problem above actually decays exponentially fast.
(b) Let us define the unstable subspace by U (A) := S(−A). Give characterizations of
U (A) similar to above.
(c) Show that the following definitions of a subspace are equivalent:
• The sum of the generalized eigenspaces associated with the purely imaginary
eigenvalues,
• A subspace C(A), invariant under A and such that

Cn = S(A) ⊕ C(A) ⊕ U (A),

• The set of data a ∈ Cn such that the solution of the Cauchy problem above
is polynomially bounded for t ∈ R: there exists m ≤ n − 1 and c0 such that
kx(t)k ≤ c0 (1 + |t|m ).
The subspace C(A) is called the central subspace of A.

85
 (d) Prove that the spectra of the restrictions of A to its stable, unstable and central
subspaces are respectively σ− , σ+ and σ0 .
(e) If A ∈ Mn (R), prove that the stable, unstable and central subspaces are real, in the
sense that they are the complexifications of subspaces of Rn .
(f) Express S(A∗ ), . . . in terms of S(A), . . ..
Nota: The case where σ0 is void corresponds to a hyperbolic matrix, in the sense
of the previous exercise.

124. In control theory, one meets a matrix H ∈ M2n (C) given by the formula (notations are
equivalent but not identical to the standard ones)

A BB ∗
 
H := .
C ∗ C −A∗

Hereabove, A, B, C have respective sizes n × n, n × m and m × n. Without loss of

generality, one may assume that B is one-to-one and C is onto ; hence m ≤ n.
One says that the pair (A, B) is stabilizable if the smallest invariant subspace of A,
containing the range of B, contains C(A) ⊕ U (A). One also says that the pair (A, C) is
detectable if the largest invariant subspace of A, contained in ker C, is contained in S(A).

(a) Prove that (A, C) is detectable if and only if (A∗ , C ∗ ) is stabilizable.

(b) From now on, we assume that (A, B) is stabilizable and (A, C) is detectable. If
ρ ∈ R, show that (A − iρIn , B) is stabilizable and (A − iρIn , C) is detectable.
(c) Prove that H is non-singular. Hint: Let (x, y)T belong to ker H. Find an a priori
estimate and deduce that

Cx = 0, By = 0, Ax = 0, A∗ y = 0.

(d) Deduce that H is hyperbolic (see the previous exercise).

(e) Let (0, y0 )T belong to S(H), the stable subspace. Define (x(t), y(t))T the solution of
the Cauchy problem z 0 (t) = H(t), z(0) = (0, y0 )T . Establish an integral estimate ;
prove that B ∗ y ≡ 0, Cx ≡ 0 and

x0 (t) = Ax(t), y 0 (t) = −A∗ y(t).

Deduce that x ≡ 0. Using the fact that (A, B) is stabilizable, prove that y ≡ 0.
Hint: The space spanned by the values y(t) is invariant under A∗ and annihilated
by B ∗ .
(f) Likewise, if (x0 , 0) belongs to U (H), use the detectability of (A, C) to prove that
x0 = 0.
(g) Deduce the Lopatinskiı̆Lopatinskiı̆ condition

C2n = ({0} × Cn ) ⊕ S(H).

86
125. Let A, B, C be matrices with complex entries, such that the product ABC makes sense
and is a square matrix. Prove

|Tr(ABC)| ≤ kAkS kBkS kCkS ,

√
with kAkS = TrA∗ A the Schur–FrobeniusSchurFrobenius norm. Hint: Apply repeat-
edly the Cauchy–SchwarzCauchySchwarz inequality.
Nota: This is the discrete analogue of the following inequality for functions defined on
the plane:
Z
f (x, y)g(y, z)h(z, x)dx dy dz ≤ kf kL2 (R2 ) kgkL2 (R2 ) khkL2 (R2 ) .
R3

126. Let H be a tridiagonal Hermitian 3 × 3 matrix:

 
a1 b1 0
H=  b̄1 a2 b 2  (aj ∈ R, bj ∈ C).
0 b̄2 a3

We denote the characteristic polynomial of H by P .

(a) Prove the formula (a3 − a1 )|b1 |2 = P (a1 ). Therefore, the signs of P (a1 ) and P (a3 )
are determined by that of a3 − a1 .
(b) Construct a pair (a, λ) ∈ R3 × R3 such that a  λ, with a1 < a3 and
Y
(λj − a1 ) < 0.
j

(c) Deduce that there exists a pair (a, λ) ∈ R3 × R3 such that a  λ (and therefore there
is a Hermitian matrix with diagonal a and spectrum λ, from Theorem 3.4.2), but no
such matrix being tridiagonal.
Nota: A theorem of M. AtiyahAtiyah (Sir M.) asserts that the Hermitian matrices
with given diagonal and given spectrum form a connected set (not true for real
symmetric matrices). A strategy could have been to show that in such a set, there
exists a tridiagonal element. False alas !
(d) Prove Atiyah’s result for 2 × 2 Hermitian matrices. Prove also that it becomes false
in 2 × 2 real symmetric matrices.

127. If N is a square matrix, we denote N̂ the transpose of the matrix of its cofactors. Recall
that N N̂ = N̂ N = (det N )In .
Let A, B, C, D be given 2 × 2 matrices. We form a 4 × 4 matrix
 
A B
M := .
C D

87
 (a) Prove that
det M = det(AD) + det(BC) − 2Tr(B D̂C Â).
(b) Deduce that  
(det D)Â − ĈDB̂ · · ·
M̂ = .
··· ···
Compare with formula of Corollary 8.1.1.
(c) Show that rkM ≤ 2 holds if and only if ĈDB̂ = (det D)Â, ... or equivalently
B D̂C = (det D)A, ... Deduce that

(rkM ≤ 2) =⇒ (det(AD) = det(BC)).

(d) More generally, let A, B, C, D be given in Mn (k), such that the matrix M defined
blockwise as above is of rank at most n (this is a (2n) × (2n) matrix). Prove that
det(AD) = det(BC). Hint: Use the rank decomposition.
Nota: This ensures that, for every matrix M 0 equivalent to M (M 0 = P M Q with
P, Q non-singular), there holds det(A0 D0 ) = det(B 0 C 0 )
128. Let f : (0, +∞) → R be a continous monotone matrix function (see Exercise 74). We
recall that f has a representation of the form (7) with a ≥ 0, b ∈ R and m a non-negative
bounded measure, with (a, m) 6= (0, 0).
(a) Show that f is concave increasing.
(b) From now on, we assume that f is continous at s = 0, meaning that f (0+) > −∞.
Prove that for every s, t ≥ 0, there holds f (s + t) ≤ f (s) + f (t) − f (0). Deduce that
if A ∈ Hn is non-negative, and if s ≥ 0, then f (A + sIn ) ≤ f (A) + (f (s) − f (0))In .
(c) Deduce that is A, B ∈ Hn are non-negative and if s ≥ 0, then A ≤ B + sIn implies
f (A) ≤ f (B) + (f (s) − f (0))In .
(d) Prove at last that for every non-negative Hermitian A, B, there holds

kf (B) − f (A)k2 ≤ f (kB − Ak2 ) − f (0).

Hint: Choose s cleverly in the above inequality.

Compare with the results of Exercises 110 and 111. We recall (Exercise 74) that
s 7→ sα is a monotone matrix function for 0 < α ≤ 1.
129. (KrattenthalerKrattenthaler) Let X1 , ..., Xn , A2 , ..., An and B1 , ..., Bn−1 be indetermi-
nates. Prove the identity
Y Y
det ((Xi + An ) · · · (Xi + Aj+1 )(Xi + Bj−1 ) · · · (Xi + B1 )) = (Xi − Xj ) (Bi − Aj ).
1≤i,j≤n
i<j i<j

Hint: Prove that the right-hand side divides the left-hand side. Then compare the
degrees of these homogeneous polynomials. At last, compute the ratio by specializing the
indeterminates.

88
130. (R. PichéPiché.) We recall (see Exercise 49 for the case with real entries) that the Hermi-
tian form h(M ) := (n − 1) Tr(M ∗ M ) − | Tr M |2 takes non-negative values on the cone of
singular matrices in Mn (C). Use this result to prove that the spectrump of a general matrix
A ∈ Mn (C) is contained in the disk of center n−1 Tr A and radius n−1 | Tr A|2 + nh(A).
Check (on n = 2 for instance) that this result is not implied by Gershgorin’sGershgorin
Theorem.
131. Given a complex n × n matrix A, show that there exists a unitary matrix U such that
M := U ∗ AU has a constant diagonal:
1
mii = Tr A, ∀i = 1, ..., n.
n
Hint: Use the convexity of the numerical range (see Exercise 21).
In the Hermitian case, compare with Schur’sSchur Theorem 3.4.2.
132. We aim at computing the number of connected components in the subsets of Mn (R)
or of Symn (R) made of matrices with simple eigenvalues. The corresponding sets are
denoted hereafter by sMn (R) and sSymn (R). We recall that GLn (R) has two connected
components, each one characterized by the sign of the determinant. We denote by GL+n
the connected set defined by det M > 0.
(a) Let A ∈ sMn (R) be given. Show that there exists a matrix P ∈ GL+ n such that
−1
P AP is block-diagonal, with the diagonal blocks being either scalars (distinct real
numbers), or 2 × 2 matrices of rotations of distinct nonzero angles.
(b) Let A, B be given in sMn (R), with the same number of pairs of complex conjugate
eigenvalues.
i. Find a path from the spectrum of A to that of B, such that each intermediate
set is made of distinct real numbers and distinct pairs of complex conjugate
numbers.
ii. By using the connectedness of GL+ n , prove that A and B belong to the same
connected component of sMn (R).
(c) Deduce that sMn (R) has exactly n+1
 
2
connected components.
(d) Following the same procedure, prove that sSymn is connected. Comment: Here
is a qualitative explanation of the discrepancy of both results. The complement
of the set of matrices with simple eigenvalues is the zero set of a homogeneous
polynomial A 7→ ∆(A), the discriminant of the characteristic polynomial. In the
general case, this polynomial takes any sign, and the complement of sMn (R) is
an algebraic hypersurface. It has codimension one and splits Mn (R) into several
connected components. In the symmetric case, ∆ takes only non-negative values,
because the spectra remain real. It therefore degenerates along its zero set, which
turns out to be of algebraic codimension two (see V. I. ArnoldArnold, Chapitres
supplémentaires de la théorie des équations différentielles ordinaires, Mir (1980)).
Consequently sSymn is connected.

89
133. In paragraph 6.3.1, one shows that the minimal polynomial of a companion matrix equals
its characteristic polynomial. On another hand, Proposition 10.1.1 tells us that eigen-
values of an irreducible Hessenberg matrix are geometrically simple. Show that these
results are variants of the more general one: The minimal polynomial of an irreducible
Hessenberg matrix equals its characteristic polynomial.
134. Given A ∈ Mn×m (k) and B ∈ Mm×n (k), we form the matrix M ∈ M2n+m (k):
 
In A 0
M=  0 Im B  .
0 0 In

Compute the inverse M −1 . Deduce that if we need dN α (d, α independent of N ) operations

to invert an N × N matrix, we can multiply two matrices such as A and B in d(2n + m)α
operations. In particular, the converse of Proposition 8.1.3 holds true.
135. Let n be an even integer (n = 2m). You check easily that given a row x and a column y
of n scalars, their product satisfies the relation
m
X X X
xy = (x2k−1 + y2k )(x2k + y2k−1 ) − x2k x2k−1 − y2k y2k−1 .
k=1 k k
3
Deduce a way to compute a matrix product in Mn (k) in n2 + O(n2 ) multiplications and
n3 +O(n2 ) additions, instead of n3 +O(n2 ) of each by the naive method. Comment: This
is Winograd’sWinograd calculation. In the 60’s, the computational cost of a multiplication
was two or three times that of an addition. Thus it was valuable to divide the number
of multiplications by some factor (here by two), keeping the number of additions roughly
the same.
Can this idea be used recursively, as for Strassen’sStrassen multiplication ?
136. Let k be R or C and k · k be a unitary invariant norm on Mn (k). Prove that for every
matrix A ∈ Mn (k), there holds kA∗ k = kAk.
137. Given a norm k · k on Rn , we denote by S the unit sphere of the corresponding induced
norm on Mn (R).
(a) Given e, f ∈ Rn such that kf k = kek =
6 0, we define

K(e, f ) := {M ∈ Mn (R) ; kM k ≤ 1 and M e = f }.

Prove that K(e, f ) is a convex subset of S.

(b) Conversely, let K be a convex subset of S. We assume that the norm k · k in Rn is
strictly convex, meaning that if kx + yk = kxk + kyk, then x and y are proportional
(with the same sense of course). Prove that there exists a pair (e, f ) as above, such
that K = K(e, f ). Hint: Consider an internal point N of K and a unit vector e for
which kN ek = 1.

90
 (c) In this question, the norm of Rn is k · k2 and the sphere S is denoted by S2 .
i. Prove that every K(f 0 , f ) is linearly isometric to K(e, e) with e = (1, . . . , 1)T .
ii. Show that K(e, e) is the set of matrices such that M e = e, M T e = e and the
restriction of M to e⊥ (an endomorphism of course) has operator norm at most
one. Hence K(e, e) is linearly isometric to the unit ball of Mn−1 (R) equipped
with the induced norm k · k2 . Compare this result with Corollary 5.5.1.
iii. In particular, show that K(e, e), and therefore each K(e, f ), is maximal among
the convex subsets of S2 .

138. (Continuation.) We keep e = (1, . . . , 1)T but consider the norm k · kp , where 1 ≤ p ≤ ∞.
The corresponding set K(e, e) is denoted by Kp .

(a) For p = 1, show that K1 reduces to the set ∆n of bistochastic matrices.

(b) For p = ∞, show that K∞ is the set of stochastic matrices, defined by M e = e and
M ≥ 0.
(c) If r ∈ (p, q), show that Kp ∩ Kq ⊂ Kr .
(d) Assume that 1 < p < ∞. Making a Taylor expansion of ke + ykp and of ke + M ykp ,
as  → 0, prove that Kp ⊂ K2 .
(e) Prove that K∞ is not a subset of K2 . Hint: Compare the dimensions of these convex
sets.
(f) Show that p 7→ Kp is non-decreasing on [1, 2] and non-increasing on [2, ∞) (we have
seen above that the monotonicity fails on [2, ∞]).
(g) If 1 ≤ p < 2, prove the “right-continuity”
\
Kp := Kq .
p<q≤2

(h) If 1 < p < ∞, show that M 7→ M T is an isometry from Kp onto Kp0 , where p0 is
the conjugate exponent (the calculations above show that this is false for p = 1 or
p = ∞). Deduce that q 7→ Kq is “left-continuous” on (2, ∞) (of course it is not at
q = ∞ since K∞ is much too big).
(i) Deduce that \
Kq = ∆n .
2≤q<∞

139. Let d, δ ∈ Rn be given, together with unitary matrices Q, R. We form the diagonal
matrices D = diag(d) and ∆ = diag(δ).

(a) Show that Tr(DQ∆R) equal dT Sδ, where the matrix of moduli |S| is majorized by
a bi-stochastic matrix M (see also Exercise 101).

91
 (b)
Deduce von Neumann’s inequalityvonneu@von
Neumann in Mn (C):
X
(16) |Tr(AB)| ≤ si (A)si (B),
i

where s1 (C) ≤ · · · ≤ sn (C) denote the singular

vonneu@von
Neumann, John!(Hungary) values of C.

John von Neumann.

Nota: B. DacorognaDacorogna and P. MaréchalMaréchal have proven the more gen-
eral inequality
X
(17) Tr(AB) ≤ (sign det(AB))s1 (A)s1 (B) + si (A)si (B).
i≥2

140. Let k be R or C. Given a bounded subset F of Mn (k), let us denote by Fk the set of
all possible products of k elements in F . Given a matrix norm k · k, we denote kFk k the
supremum of the norms of elements of Fk .
(a) Show that kFk+l k ≤ kFk k · kFl k.
(b) Deduce that the sequence kFk k1/k converges, and that its limit is the infimum of the
sequence.
(c) Prove that this limit does not depend on the choice of the matrix norm.
This limit is called the joint spectral radius of the family F , and denoted ρ(F ). This
notion is due to G.-C. RotaRota and G. StrangStrang.
(d) Let ρ̂(F ) denote the infimum of kF k when k · k runs over all matrix norms. Show
that ρ(F ) ≤ ρ̂(F ).
(e) Given a norm N on k n and a number  > 0, we define for every x ∈ k n
∞
X
kxk := (ρ(F ) + )−l max{N (Bx) ; B ∈ Fl ).
l=0

i. Show that the series converges, and that it defines a norm on k n .

ii. For the matrix norm associated with k · k, show that kAk ≤ ρ(F ) +  for every
A ∈ F.
iii. Deduce that actually ρ(F ) = ρ̂(F ). Compare with Householder’sHouseholder
Theorem.
141. (G.-C. RotaRota & G. StrangStrang.) Let k be R or C. Given a subset F of Mn (k),
we consider the semi-group F generated by F . It is the union of sets Fk defined in the
previous exercise, as k runs over N. We have F0 = {In }, F1 = F , F2 = F · F ,. . .

92
 If F is bounded, prove that there exists a matrix norm k · k such that kAk ≤ 1 for every
A ∈ F . Hint: In the previous exercise, take a sup instead of a series.

142. Let define the two matrices

   
1 1 1 0
A0 = , A1 = .
0 1 1 1

Given a map s : {1, ..., r} → {0, 1} (i.e. a word in two letters), we define

A(s) := As(1) As(2) · · · As(r) , Â(s) := As(r) As(r−1) · · · As(1)

(Â(s) is the palindrome of A(s)). Show that

 
m(s) 0
Â(s) − A(s) = ,
0 −m(s)

where m(s) is an integer.

143. Let T ∈ Symn (R) be ToeplizToepliz, meaning that

 
a1 a2 a3 an
.
 a2 a1 a2 . .
 


T =  a3 a2 a1 . .
 . 
.


 . . .
.. .. ..


an a1

We denote by ∆j the principal minors, in particular ∆1 = a and ∆n = det T .

(a) Prove the inequality

∆n ∆n−2 ≤ ∆2n−1 .
Hint: Use Desnanot–JacobiDesnanotJacobi formula of Exercise 24.
(b) If ∆n−2 6= 0 and a1 , . . . , an−1 are given, show that there exists a unique value an such
that the above inequality becomes an equality.
(c) When T is positive definite, deduce the inequalities
1/(n−1)
∆1/n
n ≤ ∆n−1 ≤ · · · ≤ ∆1 .

(d) Let us assume that T is positive definite. Prove that for every N > n, T may be
completed into an SDPN Toepliz matrix.

144. Let A ∈ Mn (k) be given, with n = p + q. For 1 ≤ i, j ≤ p, let us define the minor
 
i p + 1 ··· n
dij := A .
j p + 1 ··· n

93
 With the entries dij , we form a matrix D ∈ Mp (k). Prove the Desnanot–JacobiDesnanotJacobi
formula (see Exercise 24 for the case p=2)
 
p−1 p + 1 ··· n
det D = δ det A, δ := A .
p + 1 ··· n

Hint: Develop dij with the help of Schur’sSchur determinant formula. Then apply once
more Schur’s formula to det A.

145. (B. Perthame and S. Gaubert.)PerthameGaubert Given a non-negative matrix N ∈

Mn (R) that is irreducible, let denote M := N − ρ(N )In . From Perron–Frobenius The-
orem, λ = 0 is a simple eigenvalue of M , associated with a positive eigenvector X. Let
also Y denote a positive eigenvector of M T , again for the zero eigenvalue.
Given an initial data x0 ∈ Rn , let t 7→ x(t) be the solution of the ODE ẋ = M x, such
that x(0) = x0 .

(a) Show that

(18) Y · x(t) ≡ Y · x0 .

(b) Let H be a C 1 -function over R. Show that

      
d X xj X xj xk
(19) Yj Xj H = X k Yj H −H
dt j Xj j,k
Xj Xk
   
xk xj xj
+ − H0 .
Xk Xj Xj

(c) Show that the kernel of the quadratic form

X Yj
(x, z) 7→ mjk (xj zk − xk zj )2
j,k
Xj

is exactly the line spanned by z.

(d) Let s := x · Y /X · Y . Deduce that the expression
X Yj
D(t) := (xj − sXj )2
j
X j

satisfies a differential inequality of the form

Ḋ + D ≤ 0,

where  > 0 depends only on M . Hint: Use both (18) and (19).
(e) Verify that x(t) converges towards sX, exponentially fast.

94
 (f) State a result for the solutions of ẋ = N x.

146. To do this exercise, you need to know about the exterior algebra ΛE. Recall that if E is a
K-vector space of dimension n, the exterior algebra ΛE is the direct sum of the subspaces
Λk E of the tensor algebra, spanned by the vectors x1 ∧· · ·∧xk , where x∧y := x⊗y −y ⊗x
whenever x, y ∈ E, and ∧ is associative. We have
 
k n
dim Λ E = .
k

If {e1 , . . . , en } is a basis of E, then a basis of Λk E is given by the vectors ej1 ∧ · · · ∧ ejk

as 1 ≤ j1 < · · · < jk ≤ n.

(a) Let u be an endomorphism in E. Prove that there exists a unique endomorphism in

Λk E, denoted by u(k) , such that

u(k) (x1 ∧ · · · ∧ xk ) = u(x1 ) ∧ · · · ∧ u(xk ), ∀x1 , . . . , xk ∈ E.

(b) Let {e1 , . . . , en } be a basis of E and A be the matrix of u in this base. Show that
the entries of the matrix associated with u(k) are the minors
 
i1 · · · ik
A
j1 · · · jk

with i1 < · · · < in and j1 < · · · < jn .

(c) Let x1 , . . . , xk be linearly independent vectors of u, associated with the eigenvalues
λ1 , . . . , λk . Prove that x1 ∧· · ·∧xk is an eigenvector of u(k) . What is the corresponding
eigenvalue ?
(d) If K = C, show that the spectrum of u(k) is made of the products
Y
λi , |I| = k,
i∈I

where {λ1 , . . . , λn } is the spectrum of u. Hint: Consider first the case where
λ1 , . . . , λn are distinct. Then proceed by density.
(e) Prove that the above property is true for every scalar field K. Hint: The case
K = C provides an algebraic identity with integral coefficients.

147. (Continuation.) We now take K = R. Recall that a matrix M ∈ Mn (R) is totally positive
if all the minors  
i1 · · · ik
M
j1 · · · jk
with k ≤ n, i1 < · · · < in and j1 < · · · < jn are positive. Total positiveness implies
positiveness.

95
 (a) If M is positive, prove that ρ(M ) is a positive simple eigenvalue with the property
that ρ(M ) > |λ| for every other eigenvalue of M .
(b) Let M ∈ Mn (R) be totally positive. Prove that its eigenvalues are real, positive and
pairwise distinct (thus simple). Hint: Proceed by induction on n. Use the fact that
λ1 · · · λn is the unique eigenvalue of M (n) .
(c) Likewise, show that the singular values of M are pairwise distinct.

148. (LoewnerLoewner.)

(a) Let A be a tridiagonal matrix with non-negative off-diagonal entries (i 6= j).

i. Let us consider a minor  
i1 · · · ik
A
j1 · · · jk
with k ≥ 2, i1 < · · · < ik and j1 < · · · < jk . Show that it is either the
product of minors of smaller sizes (reducible case), or it is a principal minor
(j1 = i1 , . . . , jk = ik ).
ii. Deduce that there exists a real number x such that xIn + A is totally non-
negative. Hint: Choose x large enough. Argue by induction on n.
(b) Let B ∈ Mn (R) be tridiagonal with non-negative off-diagonal entries. Deduce from
above and from Trotter’sTrotter formula
 k
1
exp M = lim In + M ,
k→+∞ k
that the semi-group (exp(tB))t≥0 is made of totally non-negative matrices.
(c) Conversely, let A ∈ Mn (R) be given, such that the semi-group (exp(tA))t≥0 is made
of totally non-negative matrices.
i. Show that the off-diagonal entries of A are non-negative.
ii. If j > i + 1, show that aij ≤ 0. Hint: consider the minor
 
i i+1
Mt
i+1 j

of Mt := exp(tA).
iii. Deduce that A is tridiagonal, with non-negative off-diagonal entries.
149. We denote by e the vector of Rn whose every component equals one.

(a) Let A be a Hermitian matrix, semi-positive definite, and denote a ∈ Rn the vector
whose components are the diagonal entries of A. Show that B := aeT + eaT − 2A is a
Euclidean distance matrix (see Exercise 62). Hint: Use a factorization A = M T M .
(b) Conversely, show that every Euclidean distance matrix is of this form for some semi-
positive definite Hermitian matrix.

96
 (c) Deduce that the set of Euclidean distance matrices is a convex cone. Find also a
direct proof of this fact.
(d) Let s ∈ Rn be such that eT s = 1. Let F (s) be the cone of semi-positive definite
Hermitian matrices T such that T s = 0. Show that the restriction to F (s) of the
map A 7→ aeT + eaT − 2A is injective, and that its inverse is given by
1
M 7→ − (In − esT )M (In − seT ).
2
150. Let A ∈ Mn×m (R) and b ∈ Rn be given. Define two sets

X := {x ∈ Rm ; Ax ≤ b and x ≥ 0}, Y := {y ∈ Rn ; AT y ≥ 0, y ≥ 0 and b · y < 0},

where the inequalities stand for vectors, as in Chapter 5.

Prove that exactly one of both sets is void (Farkas’Farkas Lemma).

151. Consider the homogeneous polynomial pr (X) := X02 − X12 − · · · − Xr2 for some integer
r ≥ 1.

(a) For r = 1 and r = 2, show that there exist real symmetric matrices S0 , . . . , Sr such
that
pr (X) = det(X0 S0 + · · · + Xr Sr ).
(b) When r ≥ 3 show that there does not exist matrices A0 , . . . , Ar ∈ M2 (R) such that

pr (X) = det(X0 A0 + · · · + Xr Ar ).

Hint: Consider a vector v ∈ Rr+1 such that v0 = 0 and the first row of v1 A1 + · · · +
vr Ar vanishes.

152. Here is another proof of Birkhoff’sBirkhoff Theorem. Let A be a bistochastic matrix of

size n.

(a) Prove that there does not exist a submatrix of size k×l with null entries and k+l > n.
Hint: Count the sum of all entries of A.
Then Exercise 10 of Chapter 2, page 32, tells you that there exists a permuta-
tion σ such that aiσ(i) 6= 0 for every i = 1, ..., n (this result bears the name of
Frobenius–König TheoremFrobeniusKonig@König). In the sequel, we denote by P
σ(i)
the permutation matrix with entries pij = δj .
(b) Let a be the minimum of the numbers aiσ(i) 6= 0, so that a ∈ (0, 1]. If a = 1, prove
that A = P . Hint: Again, consider the sum of all entries.
(c) If a < 1, let us define
1
B= (A − aP ).
1−a
Show that B is bistochastic. Deduce that if A is extremal in ∆n , then A = P .

97
153. This is a sequel of Exercise 26, Chapter 4 (#23, Chap. 7 in the second edition ; this
exercise does not exist in the French edition). We recall that Σ denotes the unit sphere of
M2 (R) for the induced norm k · k2 . Also recall that Σ is the union of the segments [r, s]
where r ∈ R := SO2 (R) and s ∈ S, the set of orthogonal symmetries. Both R and S are
circles. At last, two distinct segments may intersect only at an extremity.

(a) Show that there is a unique map ρ : Σ \ S → R, such that M belongs to some
segment [ρ(M ), s) with s ∈ S. For which M is the other extremity s unique ?
(b) Show that the map ρ above is continuous, and that ρ coincides with the identity
over R. We say that ρ is a retraction from Σ \ S onto R.
(c)
Let f : D → Σ be a continuous function, where
D is the unit disk of the complex plane, such that
f (exp(iθ)) is the rotation of angle θ. Show that
f (D) contains an element of S.
Hint: Otherwise, there would be a retraction of D
onto the unit circle, which is impossible (an equiv-
alent statement to BrouwerBrouwer Fixed Point
Theorem).
Meaning. Likewise, one finds that if a disk D0 is
immersed in Σ, with boundary S, then it contains
an element of R. We say that the circles R and S
of Σ are linked.
Luitzen Brouwer.

154. Recall that ∆3 denotes the set of 3×3 bistochastic matrices. As the convex hull of a finite
set (the permutation matrices), it is a polytope. It thus has k-faces for k = 0, 1, 2, 3. Of
course, 0-faces are vertices. Justify the following classification:

• There are 6 vertices,

• 15 1-faces, namely all the segments [P, Q] with P, Q permutation matrices,
• 18 2-faces, all of them being triangles. Each one is characterized by an inequality
mij + mi0 j 0 ≤ 0, where i 6= i0 and j 6= j 0 ,
• 9 3-faces, all of them being 3-simplex. Each one is characterized by an inequality
mij ≤ 0 for some pair (i, j).

98
 Hint: To prove that a convex subset of dimension k ≤ 3 is
a face, it is enough to characterize it by a linear inequality
within ∆3 . Notice that the alternate sum 6 − 15 + 18 − 9
vanishes, as the EulerEuler–PoincaréPoincaré characteristics
of the sphere S3 is zero. Be cautious enough to prove that
there is not any other face.

Euler,
Leonhard (Switzerland)

Leonhard Euler.

155. (From V. Blondel & Y. Nesterov.) Let F = {A1 , . . . , Am } be a finite subset of Mn (k)
(k = R or k = C). We denote by ρ(A1 , . . . , Am ) the joint spectral radius of F (see Exercise
140 for this notion). Prove that
1
ρ(A1 + · · · + Am ) ≤ ρ(A1 , . . . , Am ).
m
Suppose that k = R and that A1 , . . . , Am are non-negative. Prove that

ρ(A1 , . . . , Am ) ≤ ρ(A1 + · · · + Am ).

156. Let n = lm and A1 , . . . , Am ∈ Ml (k) be given matrices. Let us form the matrix

0l · · · · · · 0l A1
 
 2 ..
 A . 0l 


A :=  .. .. .. 
 l 0 . . .
. 
 . . . . .
 .. .. .. . . .. 


0l · · · 0l Am 0l

Prove that
det(In − A) = det(Il − Am · · · A1 ).
Deduce the formula involving characteristic polynomials:

PA (X) = PAm ···A1 (X m ).

157. Let H be a Hermitian matrix, given in block form as

 
A B
H= .
B∗ C

99
 Assume that rk(H) = rk(A) (we say that H is a flat extension of A). Prove that the
number of positive (resp. negative) eigenvalues of A and H are equal. In particular:

(H ≥ 0) ⇐⇒ (A ≥ 0).

158. Let A ∈ Mn (C) be a normal matrix. We define B ∈ Mn−1 (C) by deleting the last row
and the last column from A. Let {λ1 , . . . , λn } be the spectrum of A, and {µ1 , . . . , µn−1 }
be that of B. Finally, denote e := (0, . . . , 0, 1)T .

(a) Show the identity

det(λIn−1 − B)
h(λ − A)−1 e, ei = .
det(λIn − A)
(b) Deduce that the rational function

det(λIn−1 − B)
R(λ) :=
det(λIn − A)

has simple poles with non-negative residues.

(c) Conversely, let {λ1 , . . . , λn } and {µ1 , . . . , µn−1 } be given tuples of complex numbers,
such that the rational function
Qn−1
j=1 (λ − µj )
R(λ) := Qn
k=1 (λ − λk )

has simple poles with non-negative residues. Prove that there exists a normal matrix
A such that {λ1 , . . . , λn } is the spectrum of A, and {µ1 , . . . , µn−1 } is that of B.

159. Let n = lm and A1 , . . . , Am , B 1 , . . . , B m ∈ Ml (k) be given matrices. Let us form the

matrix  
0l · · · · · · 0l B 1 A1
 2 ..
. 0l B 2 

 A
. ..
 
 3 .. .
 B 0l 

A :=  ..  .
 0 ... ..
.
..
. . 
 l 
 .. . . . . . . . . . . . .. 
 . . . 
0l · · · 0l B m Am 0l
We define also the product
   
0l XIl 0l XIl
M (X) := ··· .
B m Am B 1 A1

100
  
s
(a) Let λ ∈ k and r := ∈ k 2l be such that M (λ)r = λm r. Construct an
t
eigenvector T of A, for the eigenvalue λ, such that we have in block form
 
t1
 .. 
 . 
T =  tm−2  .
 
 m−1 
 t =s 
tm = t

(b) Conversely, let λ be an eigenvalue of A. Assume that λ is not zero. Prove that λm
is an eigenvalue of M (λ).
(c) Let us define the multivariate polynomial
    
0l Xm Il 0l X1 Il
Q(X1 , . . . , Xm ) := det X1 · · · Xm I2l − ··· .
B m Am B 1 A1
Show that if at least one of the Xj ’s vanishes, then Q vanishes. Therefore X1 · · · Xm
factorizes in Q(X1 , . . . , Xm ).
(d) In the case where A has n distinct non-vanishing eigenvalues, deduce that
(20) X n PA (X) = det(X m I2l − M (X)),
where PA is the characteristic polynomial of A.
(e) Using the principle of algebraic identities, show that (20) holds for every scalar field
k and every matrices A1 , . . . , B m .
160. Let S be a set and m, n be positive integers. Let (As )s∈S and (Bs )s∈S be two families
indexed by S, with As ∈ Mm (k), Bs ∈ Mn (k). We assume that the only subspaces of k m
(respectively k n ) invariant by every As (respectively Bs ), that is As E ⊂ E, are {0} and
k m (respectively k n ) itself.
Let M ∈ Mn×m (k) be such that Bs M = M As for every s. Prove (Schur’s LemmaSchur)
that either M = 0n×m , or m = n and M is non-singular. Hint: Consider the range and
the kernel of M .
161. Let n = p + q with 0 < p < q be given. We denote by A the subset of Mn (k) made of
the matrices with block form  
0p 0p×q
.
A 0q
Likewise, B is made of the matrices
 
0q 0q×p
.
B 0p
Both A and B are subalgebras of Mn (k), with dimension pq and the property that M N =
0n for every two elements (of the same algebra). Prove that A and B are not conjugated
in Mn (k). Show however that B is conjugated to AT in Mn (k).

101
162. (von Neumannvonneu@von Neumann)

Recall that a norm k · k on Mn (C) is unitarily

invariant if kU AV k = kAk for every U, V ∈ Un
and A ∈ Mn (C). In this exercise, we define
σ : Mn (C) → Rn by σ(A) = (s1 (A), . . . , sn (A)),
where s1 (A) ≤ · · · ≤ sn (A) are the singular values
of A.
vonneu@von
Neumann, John!(Tatarstan)

John von Neumann.

(a) Let k · k be a unitarily invariant norm. Prove that there exists a unique function
g : Rn → R, even with respect to each coordinates and invariant under permutations,
such that k · k = g ◦ σ.
In the sequel, such an invariant norm on Rn is called a gauge.
(b) What are the gauges associated with k · k2 and to the Frobenius norm (Tr(A∗ A))1/2 ?
(c) Conversely, let g be a gauge on Rn . We denote by g∗ its dual norm on Rn . Verify
that g∗ is a gauge. Then prove that
< Tr(A∗ B)
(21) g∗ ◦ σ(A) ≤ sup .
B6=0n g(σ(B))

Hint: First consider the case where A is diagonal and non-negative, and use only
the B’s that are diagonal an non-negative.
(d) Prove that (21) is actually an equality. Hint: Use von Neumann’svonneu@von
Neumann inequality of Exercise 139.
(e) Deduce that if G is a gauge, then G ◦ σ is a unitarily invariant norm on Mn (C).
Hint: Apply the results above to g = G∗ .

163. Given A ∈ Mn (k), we define a linear map TA by X 7→ TA X := AT XA for X ∈ Symn (k).

The goal of this exercise is to compute the determinant of TA .

(a) If A = QT DQ with Q orthogonal, prove that TA and TD are conjugate to each other.
(b) Compute det TD when D is diagonal.
(c) Verify TQS = TS ◦ TQ , whence det TQS = det TQ det TS .
(d) Consider the case k = R. Deduce from above that if A is itself symmetric, then
det TA = (det A)2 .
(e) (Case k = R, continuing). Let Q be real orthogonal. Show that In is a cluster point
of the sequence (Qm )m∈N . Deduce that the identity of Symn (R) is a cluster point of

102
 the sequence (TQm )m∈N . Thus det TQ = ±1. Using the connectedness of SOn , show
that actually det TQ = 1.
(f) (Case k = R, continuing). Using the polar decomposition of A ∈ GLn (R), prove
that det TA = (det A)2 . Show that this formula extends to every A ∈ Mn (R).
(g) Check that the formula det TA = (det A)2 is a polynomial identity with integer
coefficients, thus extends to every scalar field.

164. (BoydBoyd, DiaconisDiaconis, SunSun, Jun & XiaoXiao, Lin.) Let P be a symmetric
stochastic n × n matrix:
X
pij = pji ≥ 0, pij = 1 (i = 1, . . . , n).
j

We recall that λ1 = 1 is an eigenvalue of P , which is the largest in modulus (Perron–

Frobenius). We are interested in the second largest modulus µ(P ) = max{λ2 , −λn } where
λ1 ≥ · · · ≥ λn is the spectrum of P ; µ(P ) is the second singular value of P .

(a) Let y ∈ Rn be such that kyk2 = 1 and j yj = 0. Let w, z ∈ Rn be such that

P

 
1 1
(pij 6= 0) =⇒ (zi + zj ) ≤ yi yj ≤ (wi + wj ) .
2 2
P P
Show that λ2 ≥ j zj and λn ≤ j wj . Hint: Use Rayleigh ratio.
(b) Taking !
(2j−1)π
r
2 (2j − 1)π 1 π cos n
yj = cos , zj = cos + ,
n 2n n n cos πn
deduce that µ(P ) ≥ cos πn for every tridiagonal symmetric stochastic n × n matrix.
(c) Find a tridiagonal symmetric stochastic n × n matrix P ∗ such that
π
µ(P ∗ ) = cos .
n
Hint: Exploit the equality case in the analysis, with the y and z given above.
(d) Prove that P 7→ µ(P ) is a convex function over symmetric stochastic n × n matri-
ces. Comment: S.-G. HwangHwang, Suk-Geun and S.-S. PyoPyo, Sung-Soo prove
conversely that, given real numbers λ1 = 1 ≥ λ2 ≥ · · · ≥ λn satisfying
1 λ2 λ3 λn
+ + + ··· + ≥ 0,
n n(n − 1) (n − 1)(n − 2) 2·1

there exists a symmetric bistochastic matrix whose spectrum is {λ1 , . . . , λn }. This

matrix is not necessarily tridiagonal, since µ may be smaller than the bound cos πn .
It may even vanish, for the matrix n1 11T .

103
165. We deal with non-negative matrices in the sense of Perron–Frobenius theory. We say
that a non-negative matrix A ∈ Mn (R) is primitive if it is irreducible – thus the spectral
radius is a simple eigenvalue – and this eigenvalue is the only one of maximal modulus.
Let us denote by x and y positive eigenvectors of A and AT , according to Perron–Frobenius
TheoremPerronFrobenius. We normalize them by y T x = 1.
(a) Assume first that ρ(A) = 1. Remarking that ρ(A − xy T ) is less than one, prove that
Am − xy T tends to zero as m → +∞. Deduce that Am > 0 for m large enough.
(b) Deduce the same result without any restriction over ρ(A).
Wielandt’sWielandt Theorem asserts that Am > 0 for m ≥ n2 − 2n + 2.
(c) Conversely, prove that if A is non-negative and irreducible, and if Am > 0 for some
integer m, then A is primitive.
166. Let A, B, C be complex matrices of respective sizes n × r, s × m and n × m. Prove that
the equation
AXB = C
is solvable if, and only if,
AA† CB † B = C.
In this case, verify that every solution is of the form

A† CB † + Y − A† AY BB † ,

where Y is an arbitrary r×s matrix. We recall that M † is the Moore–PenroseMoorePenrose

inverse of M .
167. Let k be a field and A ∈ Mn (k), B ∈ Mm (k) be given matrices.
(a) Suppose that the spectra of A and B are disjoint : σ(A) ∩ σ(B) = ∅. Prove that,
given C ∈ Mn×m (k), the equation AX − XB = C is uniquely solvable in Mn×m (k)
(SylvesterSylvester–RosenblumRosenblum Theorem).
(b) We go back to a general pair (A, B) and we assume that the equation AX −XB = C
is solvable. Prove that the following (n + m) × (n + m) matrices are similar :
   
A 0 A C
D := , T := .
0 B 0 B

(c) We now prove the converse, following FlandersFlanders & WimmerWimmer. This
constitutes Roth’sRoth Theorem.
i. We define two homomorphisms φj of Mn+m (k):

φ0 (K) := DK − KD, φ1 (K) := T K − KD.

Prove that the kernels of φ1 and φ0 are isomorphic, hence of equal dimensions.
Hint: This is where we use the assumption.

104
 ii. Let E be the subspace of Mm×(n+m) (k), made of matrices (R, S) such that

BR = RA, BS = SB.

Verify that if  
P Q
K := ∈ ker φj (j = 0 or 1),
R S
then (R, S) ∈ E. This allows us to define the projections µj (K) := (R, S), from
ker φj to E.
iii. Verify that ker µ0 = ker µ1 , and therefore R(µ0 ) and R(µ1 ) have equal dimen-
sions.
iv. Deduce that µ1 is onto.
v. Show that there exists a matrix in ker φ1 , of the form
 
P X
.
0 −Im

Conclude.
Remark. With the theory of elementary divisors, there is a finite algorithm which
computes a matrix conjugating given similar matrices. However, the knowledge of
such a conjugator between D and K does not give an explicit construction of a
solution.

168. (WielandtWielandt.) As in Exercise 146, we use the exterior algebra ΛE = ⊕nk=0 Λk E,

where now E = Cn . For a given matrix M ∈ Mn (C), we define M(k) ∈ End(Λk E) by

M(k) x1 ∧ · · · ∧ xk = (M x1 ) ∧ x2 ∧ · · · ∧ xk + · · · + x1 ∧ · · · ∧ xk−1 ∧ (M xk ).

(a) Let λ1 , . . . , λn be the eigenvalues of M . Verify that the spectrum of M(k) consists of
the numbers λI , |I| = k, where
X
λI := λj .
j∈I

(b) The canonical scalar product in E extends to Λk E ; its definition over decomposable
vectors is
hx1 ∧ · · · ∧ xk , y 1 ∧ · · · ∧ y k i = det(hxi , yj i)1≤i,j≤k .
If M is Hermitian, show that M(k) is Hermitian.
(c) From now on, we assume that A and B are Hermitian, with respective eigenvalues
(they must be real) µ1 ≥ · · · ≥ µn and ν1 ≥ · · · ≥ νn . We from C := A + B, whose
eigenvalues are λ1 ≥ · · · ≥ λn .
i. What is the largest eigenvalue of B(k) ?

105
 ii. Show that there is a permutation I 7→ I 0 of subsets of {1, . . . , n} of cardinal k,
such that
Xk
λI ≤ µI 0 + νi .
i=1

iii. We now assume that A has simple eigenvalues: µ1 > · · · > µn , and that B is
small, in the sense that

(22) max |νi | < min(µj − µj+1 ).

i j

Show that the permutation mentionned above is the identity. Deduce the set of
inequalities
k
X
(23) λI ≤ µI + νi , ∀I ; |I| = k.
i=1

(d) Conversely, we give ourselves two lists of real numbers µ1 > · · · > µn and ν1 ≥ · · · ≥
νn , satisfying the smallness assumption (22). Let λ1 ≥ · · · ≥ λn be given, satisfying
the list of inequalities (23), together with
X X X
(24) λi = µi + νi .
i i i

We define κ0i := λI − µi , and κ is a re-ordering of κ0 , so that κ1 ≥ · · · ≥ κn .

i. Show that κ ≺ ν. Deduce that there exists a orthostochastic matrix M , such
that κ = M ν. We recall that an orthostochastic matrix is of the form mij =
|uij |2 , where U is unitary.
ii. Show that λ is the spectrum of C = A + B, where A := diag(µ) and B :=
U ∗ diag(ν)U , where U is as above. Therefore the inequalities (23), together with
(24), solve the A. Horn’sHorn!Alfred problem when B is “small”.
169. Let k be a field, A = k[Y, Z] be the ring of polynomials in two indeterminates and
K = k(Y, Z) be the corresponding field of rational fractions.

(a) We consider the matrix

 
X +Y Z
M (X, Y, Z) := ∈ M2 (A[X]).
Z X −Y

From Theorem 6.2.1, we know that there exist P, Q ∈ GL2 (K[X]) such that
 
1 0
PM = Q.
0 X2 − Y 2 − Z2

Show that one cannot find such a pair with P, Q ∈ GL2 (A[X]), namely with poly-
nomial entries in (X, Y, Z). Hint: The top row entries would vanish at the origin.

106
 (b) Let us now consider the matrix
 
X +Y 1
N (X, Y ) := ∈ M2 (A0 [X]), A0 := k[Y ].
1 X −Y
Show that there exist S, T ∈ GL2 (A0 [X]) such that
 
1 0
SN = T.
0 X2 − Y 2 − 1
Hint: multiply N left and right by appropriate elementary matrices.
(c) One has S, T ∈ GL2 (k[X, Y ]). Explain why this does not contradict the previous
result.
170. (HeisenbergHeisenberg Inequality.) Let A and B be two Hermitian matrices of size n.
We employ the canonical Hermitian product h·, ·i on Cn and the associated norm.
(a) For every x ∈ Cn , prove
|h[A, B]x, xi|2 ≤ 4kAxk2 kBxk2 .

(b) If kxk = 1 and C is Hermitian (quantum physi-

cists say that C is an observable), we define the
expectation and the variance of C by

E(C; x) := hCx, xi, V (C; x) := kCx−E(x)xk2 .

Prove the Heisenberg Inequality: for x, A and B

as above,

|E(i[A, B]; x)|2 ≤ 4V (A; x) V (B; x).

Heisenberg,
Werner (Germany) The Heisenberg inequality is therefore a manifesta-
tion of the non-commutativity between operators.
Werner Heisenberg.

171. A lattice in Rm is a discrete subgroup of maximal rank. Equivalently, it is the set of vectors
of which the coordinates over a suitable basis of Rm are integers. It is thus natural to
study the bases of Rm .
We consider the case where m = 2n and Rm is nothing but Cn . Given a family B :=
{v 1 , . . . , v 2n } of vectors in Cn , we form the n × (2n) matrix Π := (v 1 , . . . , v 2n ). Prove that
B is an R-basis if, and only if, the (2n) × (2n) matrix
 
Π
Π
is non-singular, where Π denotes the matrix with complex conjugate entries.

107
172. Prove Schur’sSchur Pfaffian identity
 
aj − ai Y aj − ai
Pf = .
ai + aj 1≤i,j≤2n i<j aj + ai

See Exercise 119 for a hint.

173.
Check the easy formula valid whenever the inverses
concern regular n × n matrices:

(In + A−1 )−1 + (In + A)−1 = In .

Deduce Hua IdentityHua, Loo-Keng

(B + BA−1 B)−1 + (A + B)−1 = B −1 .

Hua,
Loo-Keng (China)

Loo-Keng Hua.

Hint: transport the algebra structure of Mn (k) by the linear map M 7→ BM . This
procedure is called isotopy ; remark that the multiplicative identity in the new structure
is B. Then apply the easy formula.

174. In Mn (k), we define the JordanJordan!Pascual product by

1
A • B := (AB + BA).
2
Of course, we assume that the characteristic of the field k is not equal to 2. We warn
the reader that the bullet is not an associative product. We notice that the square A • A
coincides with A2 .

(a) Prove Jordan Identity

A2 • (A • B) = A • (A2 • B).
(b) Deduce that there is no ambiguity in the definition of A•m when m ∈ N. In other
words, A generates an associative as well as commutative bullet-algebra.
(c) For every matrices A, B, we define two linear maps UA and VA,B by

UA (B) := 2A • (A • B) − A2 • B

and
VA,B := 4UA•B − 2(UA ◦ UB + UB ◦ UA ).

108
 Prove Thedy’sThedy Identity:

UVA,B (C) = VA,B ◦ UC ◦ VA,B .

Note: One should not confuse Pascual Jordan, a German physicist, with the French
mathematician Camille JordanJordan!Camille, whose name has been given to a
canonical form of matrices over an algebraically closed field. Amazingly, simple
Euclidean (P.) Jordan algebras obey a spectral theorem, where every element is
“diagonalizable with real eigenvalues” ; hence their (C.) Jordan’s form is trivial.
Somehow, we can say that there is an exclusion principle about Jordan concepts.
Besides Jordan and Thedy identities, there are two more complicated identities,
due to GlennieGlennie, valid in Mn (k). It has been an important discovery, in
the theory of Jordan algebras, that Glennie’s and Thedy’s identities do not follow
from Jordan’s. As a matter of fact, there exists a Jordan algebra, namely that of
3 × 3 Hermitian matrices over Cayley’sCayley octonions (a non-associative division
algebra), in which all these three identities are violated. This exceptional object is
called Albert algebraAlbert.

175. Let P, Q ∈ Mn (R) be rank-one projectors. Prove that the matrices xP + yQ are diago-
nalizable with real eigenvalues for every x, y ∈ R if, and only if, either 0 < Tr(P Q) < 1
or P Q = QP = 0n or Q = P .

176. In Mn (C), we define endomorphisms LA and PA (the linear and quadratic representations
of the underlying JordanJordan!Pascual algebra) when A ∈ Mn (C) by
1
LA (M ) := (AM + M A), PA (M ) := AM A.
2
Show that
Pexp(tA) = exp(2tLA ).

177. Let A ∈ Mn (k) and p ∈ k[X] be given. Show that the minimal polynomial divides
p (that is p(A) = 0n ) if, and only if, there exists a matrix C ∈ Mn (k[X]) such that
p(X)In = (XIn − A)C(X).

178. (DelvauxDelvaux, Van Barelvanbar@Van Barel) Let m, n ≥ 1 and A ∈ GLn (k), B ∈

GLm (k), G, H ∈ Mm×n (k) be given. Let us define R := A − GT BH and S := B −1 −
HA−1 GT .

(a) Show that the following matrices are equivalent within Mm+n (k):
 −1   −1   
B H B 0 S 0
, , .
GT A 0 R 0 A

(b) Deduce the equality

rkR − rkS = n − m.

109
179. (Higham & al.)Higham Let M ∈ GLn (k) be given. We define the classical group G ⊂
GLn (k) by the equation
AT M A = M.
Let p ∈ k[X] be given, with p 6≡ 0. Let us form the rational function f (X) :=
p(X −1 )p(X)−1 .

(a) Prove (again) that for every A ∈ G, the following identities hold true:

M p(A) = p(A−1 )T M, p(AT )M = M p(A−1 ).

(b) Deduce that

(A ∈ G) =⇒ (f (A) ∈ G),
whenever p(A) is non-singular.
(c) Exemple: every MœbiusMœbius (Möbius) function preserves a complex group de-
fined by an equation
A∗ M A = M.

180. A matrix pencil is a polynomial

L(X) := X m A0 + X m−1 A1 + · · · + XAm−1 + Am ,

whose coefficients are n × n matrices. It is a monic pencil if A0 = In . If in addition the

scalar field is C and the matrices Ak are Hermitian, we say that the pencil is Hermitian.
Finally, a Hermitian pencil is hyperbolic if the roots of the polynomial

Pu (X) := hL(X)u, ui

are real and simple (notice that Pu ∈ R[X]) for every non-zero vector u ∈ Cn .

(a) Let A ∈ Hn and B ∈ Mn (C) be given. We assume that there exist x, y ∈ Cn such
that
hAx, xi < 0 < hAy, yi and hBx, xi = hBy, yi = 0.
Show that there exists a non-zero vector z ∈ Cn such that hAz, zi = hBz, zi = 0.
Hint: Apply the Toeplitz–HausdorffToeplizHausdorff Theorem from Exercise 21 to
A + iB.
(b) Let L(X) be a hyperbolic Hermitian pencil. We denote λ1 (u) < · · · < λn (u) the
roots of Pu when u 6= 0. Recall that the λj ’s are smooth functions. We denote also

∆j := λj (S n−1 )

the image of the unit sphere under λj . This is obviously a compact interval [λ− +
j , λj ]
of R.
i. Check that λ− − + +
j ≤ λj+1 and λj ≤ λj+1 .

110
 ii. Show that Pu0 (λj (u)) is not equal to zero, that its sign j does not depend on u,
and that j j+1 = −1.
iii. Assume that λ− + − +
j+1 ≤ λj . Let us choose t ∈ [λj+1 , λj ]. Prove that there exists a
unit vector z such that hL(t)z, zi = hL0 (t)z, zi = 0. Reach a contradiction.
iv. Deduce that the root intervals ∆j are pairwise disjoint.
Nota: A Hermitian pencil is weakly hyperbolic if the roots of every Pu are real,
not necessarily disjoint. For such pencils, two consecutive root intervals intersect at
most at one point, that is
−
λ+
j ≤ λj+1 .

181. We prove here the converse of Exercise 127, when the scalar field k is infinite. We recall
that in an infinite field, a polynomial function vanishes identically on k n if, and only if,
the associated polynomial is zero. Actually, if a product of polynomial functions vanishes
identically, one of the polynomials at least is zero.
We thus assume that, for every matrix M 0 equivalent to M (M 0 = P M Q with P, Q
non-singular), there holds det(A0 D0 ) = det(B 0 C 0 ), where
 0
A B0

0
M = .
C 0 D0
Using the rank decomposition, we may assume that M = diag{In , J}, where J is a
quasi-diagonal n × n matrix.
(a) Choosing P an Q appropriately, show that for every X, Y ∈ Mn (k), there holds
det(XY +J) = (det X)(det Y ). Deduce that det(In +JR) = 1 for every non-singular
R.
(b) Deduce that the polynomial X 7→ det(In + JX) − 1 (a polynomial in n2 indetermi-
nates) is zero, and thus that the polynomial X 7→ Tr(JX) is zero.
(c) Show at last that the rank of M is at most n.
Nota: This, together with Hilbert’sHilbert Nullstellensatz, implies that for every minor
µ of M of size r ∈ (n, 2n], viewed as a polynomial of the entries, there exist an integer m
such that µm belongs to the ideal spanned by the polynomials ∆P,Q := (det A0 )(det D0 ) −
(det B 0 )(det C 0 ) with A0 , ... the blocks of M 0 := P M Q. Clearly, m ≥ 2, but the least
integer m(r) is not known, to our knowledge.
182. Recall that if V is a vector space of dimension n, a complete flag in V is a set of subspaces
F0 = {0} ⊂ F1 · · · ⊂ Fn = V , with dim Fk = k for each index k. Let P be a subset of
{1, . . . , n}, say that P = {p1 < · · · < pr }. Given a complete flag F and an index subset
P of cardinality r, we define
ΩP (F ) = {L ∈ Xr | dim Fk ∩ L ≥ k},
with Xr the set of subspaces of V of dimension r. Let A be an n × n Hermitian matrix,
and take V = Cn .

111
 (a) Let L ∈ Xr be given. Show that the quantity
r
X
hAxk , xk i
k=1

does not depend on the choice of a unitary basis {x1 , . . . , xm } of L. We call it the
trace of A over L, and denote TrL A.
(b) Let a1 ≤ · · · ≤ an be the eigenvalues of A, counting with multiplicities. Let B =
(u1 , . . . , un ) the corresponding unitary basis and F∗ (A) be the complete flag spanned
by B: Fk = uk ⊕ Fk−1 . Let P := (p1 , . . . , pr ) be as above. Show that
X
ap = sup TrL A.
p∈P L∈ΩP (F∗ (A))

Hint: To show that every such TrL A is less than or equal the left-hand side, find an
adapted basis of L. To show the inequality left≤right, choose an appropriate L.

183. (From G. PisierPisier).

Let Z ∈ Mn (C) be given, whose entries are unit numbers: |zjk | = 1. We define the linear
map over Mn (C)

F : M 7→ F (M ) := Z ◦ M, (F (M ))jk := zjk mjk .

(a) Show that

√
kF (M )k1 = kM k1 , kF (M )k∞ = kM k∞ , kF (M )k2 ≤ nkM k2 .

(b) Deduce that

kF (M )kp ≤ nα kM kp , p ∈ [1, ∞], α := min{1/p, 1/p0 }.

Hint: Use Riesz–ThorinRieszThorin Interpolation Theorem.

(c) Given N ∈ Mn (C), let Abs(N ) be its absolute value: aij := |mij |. Prove that

kAbs(N )kp ≤ nα kN kp , p ∈ [1, ∞].

Hint: Find a Z adapted to this N .

(d) Define ξ := exp(2iπ/n) and Ω ∈ Mn (C) by ωjk := ξ jk . Show that Ω∗ Ω = nIn .
Deduce that √
kAbs(Ω)k2 = nkΩk2 ,
√
and therefore the constant n is optimal in the inequality

kAbs(M )k2 ≤ C2 kM k2 .

112
 √
Nota: The analogous problem in Mn (R) is less trivial. One can show that n
is optimal if and only if there exists a unitary matrix U with entries of constant
modulus, namely n−1/2 . In the real case, this means that the entries are ±n−1/2 .
This amounts to the existence of a HadamardHadamard matrix. Such matrices
exist only for a few values of n. For instance, they do not for n = 3, 5, 6, 7. The
determination of sizes n for which a Hadamard matrix exists is still an open question.
(e) More generally, we show now that nα is the optimal constant in the inequality

kAbs(M )kp ≤ Cp kM kp , ∀M ∈ Mn (C).

i. Check that if p ≤ 2, then

1 1
kxk2 ≤ kxkp ≤ n p − 2 kxk2 ,

while if p ≥ 2, then
1 1
kxk2 ≥ kxkp ≥ n p − 2 kxk2 .
ii. Deduce that kΩkp ≤ n1−α , and therefore

kAbs(Ω)kp ≥ nα kΩkp .

Conclude.
iii. Deduce in particular that kΩkp = n1−α .

184. Let A, B ∈ Hn be given, such that A + B is positive definite.

(a) Show that there exists a Hermitian matrix, denoted by A2B, such that

inf {hA(x − y), x − yi + hBy, yi}

y∈Cn

equals hA2Bx, xi for every x ∈ Cn .

We call A2B the inf-convolution of A and B.
(b) Check that 2 is symmetric and associative.
(c) Compute A2B in closed form. Check that

A2B = B(A + B)−1 A

(surprizingly enough, this formula is symmetric in A and B, and it defines a Hermi-

tian matrix).
(d) If A and B are non-degenerate, show that

A2B = (A−1 + B −1 )−1 .

113
185. (T. YamamotoYamamoto.)
Given A ∈ Mn (C), we denote by s1 (A) ≥ · · · ≥ sn (A) its singular values, and by λ1 (A), . . .
its eigenvalues ordered by non-increasing modulus:

|λ1 (A)| ≥ · · · ≥ |λn (A)|.

In this list, it is useful that an eigenvalue with multiplicity ` corresponds to ` consecutive

eigenvalues of the list. In this exercise, we prove

lim sk (Am )1/m = |λk (A)|, ∀k = 1, . . . , m.

m→+∞

Remark that for k = 1, this is nothing but the fundamental Lemma of Banach algebras,
namely
lim kAm k1/m = ρ(A),
m→+∞

where the operator norm is k · k2 .

(a) Prove that there exist subspaces G (of dimension n − k + 1) and H (of dimension
k) that are invariant under A, such that the spectrum of the restriction of A to G
(respectively to H) is λk (A), . . . , λn (A) (resp. λ1 (A), . . . , λk (A)). Hint: Use Jordan
reduction.
(b) Verify the formula

kAxk2
(25) sk (A) = sup inf ,
∗
dim F =k x∈ F kxk2

where x ∈∗ F means that x runs over the non-zero vectors of F . Hint: Both sides
are unitarily invariant. Use the decomposition in singular values.
(c) Deduce the bounds
−1
(A|H )−1 ≤ sk (A) ≤ kA|G k .
(d) Apply these bounds to Am and conclude.

186. (R. C. ThompsonThompson.) Let R be a principal ideal domain. Let M ∈ Mn×m (R) be
given, with r := min{n, m}. Let M 0 ∈ Mp×q (R) be a submatrix of M , with p + q > r.
Show that Dp+q−r (M 0 ) divides Dr (M ) (recall that Dk (M ) is the g.c.d. of all minors of
size k of M ). Hint: If p + q = r + 1, have a look to Exercise 10 of Chapter 2.
Let k be a field and A ∈ Mn (k) be given. Let M 0 ∈ Mp×q (k[X]) be a submatrix of
XIn − A, with p + q > n. We denote the invariant factors of M 0 by α1 |α2 | · · · . Deduce
that the product α1 · · · αp+q−n divides the characteristic polynomial of A.

187. The first part is an exercise about polynomials. The second part is an application to
matrices, yielding DunfordDunford decomposition. The field k has characteristic zero.

114
 (a) Let P ∈ k[X] be monic of degree n. Suppose first that P splits, say
s
Y
P (X) = (X − aj )mj , .
j=1

i. Show that there exists a unique polynomial Q of degree n, such that for every
j = 1, . . . , s, one has
Q(aj ) = aj and Q(`) (aj ) = 0, ∀1 ≤ ` ≤ mj − 1.
ii. Define the polynomial
s
Y
π(X) := (X − aj ).
j=1

Show that aj is a root of π ◦ Q of order mj at least. Deduce that P divides π ◦ Q.

iii. If P does not split, let K be an extension of k in which P splits. Let Q and π
be defined as above. Show that π ∈ k[X] and Q ∈ k[X].
(b) Let M ∈ Mn (k) be given. We apply the construction above to P = PM , the
characteristic polynomial. We let R[X) := X − Q(X), and we define D := Q(M )
and N := R(M ).
i. What is the spectrum of N ? Deduce that N is nilpotent.
ii. Show that π(D) = 0n . Deduce that D is diagonalisable in a suitable extension
of k.
iii. Deduce the Dunford decomposition: M writes as D + N for some diagonalisable
D and nilpotent N , with [D, N ] = 0n . Both D and N have entries in k, though
D could be diagonalisable only in a suitable extension, that one containing all
the eigenvalues of M .
188.
Let φ denote the EulerEuler indicator, φ(m) is the
number of integers less than m that are prime to
m. We recall the formula
X
φ(d) = n.
d|n

In the sequel, we define the n × n matrices G (for

Euler, gcd), Φ and D (for divisibility) by
Leonhard (Switzerland)

Leonhard Euler.

1 if i|j,
gij := gcd(i, j), Φ := diag{φ(1), . . . , φ(n)}, dij :=
0 else.

115
 (a) Prove that DT ΦD = G.
(b) Deduce the SmithSmith determinant formula:

det(( gcd(i, j) ))1≤i,j≤n = φ(1)φ(2) · · · φ(n).

(c) Compute the invariant factors of G as a matrix of Mn (Z), for small values of n. Say
up to n = 10.

189. (from D. FerrandFerrand.) Let σ be a permutation of {1, . . . , n} and P σ be the associated

permutation matrix.

(a) We denote by cm the number of cycles of length m in σ. Show that the characteristic
polynomial of P σ is Y
(X m − 1)cm .
m≥1

(b) Let σ and τ be two permutations. We assume that P σ and P τ are similar in Mn (k),
the field k having characteristic zero. Show that for every m, there holds ck (σ) =
ck (τ ). Deduce that σ and τ are conjugated as permutations.

190. Let M ∈ Mn×m (A) be given. Verify that det M T M equals the sum of the squares of the
minors of M of size m.

191. One begins with the following observation: if A, B ∈ GLn (k) are such that A2 = B 2 ,
then M := A−1 B is conjugated to its inverse M −1 . Verify!
We prove now the converse, and even slightly better; namely, if M and M −1 are similar,
then there exist A, B ∈ GLn (k) such that A2 = B 2 = In and M := A−1 B.

(a) Show that both the assumption and the conclusion are invariant under conjugation.
(b) Assume n = 2m. We define S, T ∈ GLn (k) by
   
Im 0m Im Im
S= , T = ,
J(4; m) −Im 0m −Im

where J(a; m) stands for the Jordan block with eigenvalue a. Check that S and T
are involutions, and show that S −1 T has only one eigenvalue and one eigendirection.
Conclude that the result holds true for J(1; 2m) (notice that J(1; n) is always similar
to its inverse).
(c) We keep the notations of the previous question. Show that the intersection of
ker(S T + In ) and ker(T T + In ) is a line. Deduce that there is a hyperplane H
that is stable under both S and T . Show that the restrictions of S and T to H
are involutions, and that of S −1 T is similar to J(1; 2m − 1). Hint: the latter is
the restriction of J(1; 2m) to a stable hyperplane; it cannot be something else than
J(1; 2m − 1).

116
 (d) We have thus proven that for every n, J(1; n) satisfies the claim. Use this to prove
that J(−1; n) satisfies it too.
(e) Let a ∈ k be given, with a 6= 0, ±1. Check that
 
0m J(a; m)
J(a; m)−1 0m

is an involution. Deduce that diag{J(a; m), J(a; m)−1 } satisfies the claim. Conclude
that diag{J(a; m), J(a−1 ; m)} satisfies it too.
(f) If the characteristic polynomial of M splits on k (for instance if k is algebraically
closed), prove the claim. Hint: Apply JordanJordan!Camille decomposition.
(g) In order to solve the general case, we use the second canonical form in Frobe-
niusFrobenius reduction. Let P be a power of an irreducible monic polynomial
over k.
i. Show that the inverse of the companion matrix BP is similar to the companion
matrix of P̂ , the polynomial defined by
1
P̂ (X) = X degP P (1/X).
P (0)

ii. Show that diag{BP , BP−1 } is the product of two involutions. Hint: Mimic the
case of J(a; m).
iii. Conclude. Hint: Verify that if the list of elementary divisors of an invertible
matrix M is p1 , . . . , pr , then the elementary divisors of M −1 are p̂1 , . . . , p̂r . Mind
that one must treat the cases (X ± 1)s apart.

192. Let M ∈ Mn (R) have the following properties:

• M is irreducible,
• For every pair i 6= j, one has mij ≥ 0,
• eT M = 0, where eT = (1, . . . , 1).

(a) Show that λ = 0 is a simple eigenvalue of M , associated with a positive eigenvector

V > 0, and that the other eigenvalues have a negative real part.
(b) Let us denote D := diag{1/v1 , . . . , 1/vn }. Prove that the symmetric matrix DM +
M T D is negative semi-definite, its kernel being spanned by V .

193. (Tan LeiLei, Tan). The following is an easy but not so well-known consequence of Perron–
FrobeniusPerronFrobenius Theorem.
Let A be a non-negative n × n matrix. Then ρ(A) < 1 if, and only if, there exists a
positive vector x > 0, such that Ax < x.

117
194. In Le Verrier’slever@Le Verrier method, and even in Fadeev’sFadeev variant (see Exercise
67), the complexity of the computation of the characteristic polynomial of an n×n matrix
M in characteristic 0 is n4 if we use the naive way to multiply matrices. It is nα+1 if
we now multipling matrices in O(nα ) operations. Here is an improvement, found by F.
PreparataPreparata and D. SarwateSarwate in 1978.
We still compute the NewtonNewton sums of the eigenvalues, which are equal to the
traces of the powers M k , k = 0, ..., m − 1. However we do not√compute all
√ the matrix
powers. Let r be the smallest integer larger than or equal to m (r = m if m is a
square).

(a) What is the complexity of the calculation of the

powers M , . . . , M r ?
(b) What is the complexity of the calculation of the
powers M 2r , . . . , M r(r−1) ?
(c) How many operations do we need to compute the
Newton sums S0 , . . . , Sm−1 once we now the powers
computed above ? Hint: To compute Tr(BC), one
does not need to compute BC.
(d) Prove that the complexity of the computation of
the characteristic polynomial is at most O(nα+1/2 ).

Urbain Le Verrier.
Le Verrier, Urbain (France)

118
195.
In the theory of differential equations, one
says that a matrix A ∈ Mn (R) is stable if
its eigenvalues lie in the open left half-plane
{z ; <z < 0}. It is strongly stable if A −
D is stable for every diagonal non-negative
matrix D. The lack of strong stability yields
what is called a Turing instabilityTuring. We
propose here necessary conditions for A to be
strongly stable.
Let us define M = −A, so that the spectrum
of M + D has a positive real part for all D
as above.
Alan Turing (with von Neumann).
Turing, Alan (Portugal)vonneu@von Neu-
mann, John!(Portugal)

(a) Let i1 < · · · < ir be indices between 1 and n. Prove that the principal submatrix
 
i1 · · · ir
M
i1 · · · ir

has its spectrum in {<z ≥ 0}. Hint: take dj = 0 if j is an index is , dj = y otherwise.

Take the limit as y → +∞ and use Schur’sSchur complement formula.
(b) Deduce that every principal minor of M must be non-negative.
(c) Show that the polynomial P (X) := det(XIn + M ) has positive coefficients. Hint:
This real polynomial has roots of positive real parts. Deduce that for every size
1 ≤ r ≤ n, there exists a principal minor of order r in M which is positive.
(d) What does all that mean for the principal minors of A ?

119
 R. A. SatnoianuSatnoianu and P. van den
Driesschevanden@van den Driessche (2005)
provided an example which shows that these
necessary conditions are not sufficient when
n ≥ 4. They turn out to be so for n = 1, 2, 3.

Schur’s complement formula is associated

with the Banachiewicz’s inversion formula
(Corollary 8.1.1).Banachiewicz

Left: Tadeusz Banachiewicz.

Banachiewicz, Tadeusz (Poland)

196. Let A ∈ Mn (C) be given, with characteristic polynomial X n − a0 X n−1 − · · · − an−1 . Let
φ be the solution of the differential equation

y (n) = a0 y (n−1) + · · · + an−1 y,

with the initial conditions φ(0) = φ0 (0) = · · · = φ(n−2) (0) = 0 and φ(n−1) (0) = 1 (this is the
fundamental solution of the ODE). Finally, define matrices A0 , . . . , An−1 by Aj = hj (A)

120
 with

h0 (z) = 1, h1 (z) = z − a0 , . . .
hj (z) = zhj−1 (z) − aj−1 , . . . hr (z) = PA (z).

Prove that

exp(tA) = φ(n) (t)A0 +φ(n−1) (t)A1 +· · ·+φ(t)An−1 .

Nota. The sequence A0 , . . . , An−1 is the

Fibonacci–Horner basisFibonacciHorner of the al-
gebra spanned by A. It is actually a basis only if
PA is the minimal polynomial of A. The interest
of the formula is that it is valid even if the minimal
polynomial has a lower degree.

Leonardo Fibonacci.
Fibonacci, Leonardo (Dominica)

197. We apply the JacobiJacobi method to a real 3 × 3 matrix A. Our strategy is that called
“optimal choice”.

(a) Let (p1 , q1 ), (p2 , q2 ), . . . , (pk , qk ), . . . be the sequence of index pairs that are chosen at
consecutive steps (recall that one vanishes the off-diagonal entry of largest modulus).
Prove that this sequence is cyclic of order three: It is either one of the sequences
. . . , (1, 2), (2, 3), (3, 1), (1, 2), . . ., or . . . , (1, 3), (3, 2), (2, 1), (1, 3), . . .
(b) Assume now that A has simple eigenvalues. At each step, one of the three off-diagonal
entries is null, while the two other ones are small, since the method converges. Say
that they are 0, xk , yk with 0 < |xk | ≤ |yk | (if xk vanishes then we are gone because
one diagonal entry is an eigenvalue). Show that yk+1 ∼ xk and xk+1 ∼ 2xk yk /δ,
where δ is √ a gap between two eigenvalues. Deduce that the method is of order
ω = (1 + 5)/2, the golden ratio (associated notably with the Fibonacci sequence),
meaning that the error k at step k satisfies

k+1 = O(k k−1 ).

This behaviour ressembles that of the secant method for the resolution of algebraic
equations.

121
 Illustration for the golden ratio (left) and the Fibonacci sequence (right).
golden ratio (Macao)

198. (PuszPusz and WoronowiczWoronowicz). Let A, B ∈ H+ n two given positive semi-definite

matrices. We show here that among the positive semi-definite matrices X ∈ H+ n such
that  
A X
H(X) := ≥ 02n ,
X B
there exists a maximal one. The latter is called the geometric mean of A and B, and is
denoted by A#B. Then we extend properties that were well-known for scalars.

(a) We begin with the case where A is positive definite.

i. Prove that H(X) ≥ 02n is equivalent to XA−1 X ≤ B (see Exercise 6, Chapter
8).
ii. Deduce that A−1/2 XA−1/2 ≤ (A−1/2 BA−1/2 )1/2 . Hint: The square root is oper-
ator monotone over R+ . See the Additional Exercise 74.
iii. Deduce that among the matrices X ∈ H+ n such that H(X) ≥ 02n , there exists a
maximal one, denoted by A#B. Write the explicit formula for A#B.
iv. If both A, B are positive definite, prove that (A#B)−1 = A−1 #B −1 .
(b) We now consider arbitrary elements A, in H+
n.

i. Let  > 0 be given. Show that H(X) ≥ 02n implies X ≤ (A + In )#B.
ii. Prove that  7→ (A + In )#B is non-decreasing.
iii. Deduce that A#B := lim→0+ (A + In )#B exists, and that it is the largest
matrix in H+n among those satisfying H(X) ≥ 02n . In particular,

lim (A + In )#B = lim+ A#(B + In ).

→0+ →0

The matrix A#B is called the geometric mean of A and B.

(c) Prove the following identities. Hint: Don’t use the explicit formula. Use instead
the definition of A#B by means of H(X).

122
 • A#B = B#A;
• If M ∈ GLn (C), then M (A#B)M ∗ = (M AM ∗ )#(M BM ∗ ).
(d) Prove the following inequality between harmonic, geometric and arithmetic mean:
1
2(A−1 + B −1 )−1 ≤ A#B ≤ (A + B).
2
Hint: Just check that
 
−1 −1 −1

1
H 2(A +B ) ≤ 02n and H (A + B) ≥ 02n .
2
√
In the latter case, use again the fact that s 7→ s is operator monotone.
(e) Prove that the geometric mean is “operator monotone”:

(A1 ≤ A2 and B1 ≤ B2 ) → (A1 #B1 ≤ A2 #B2 ),

and that it is a “operator concave”, in the sense that for every θ ∈ (0, 1), there holds

(θA1 + (1 − θ)A2 )#(θB1 + (1 − θ)B2 ) ≥ θ(A1 #B1 ) + (1 − θ)(A2 #B2 ).

Note that the latter property is accurate, since the geometric mean is positively
homogeneous of order one. Note also that the concavity gives another proof of
the arithmetico-geometric inequality, by taking A1 = B2 = A, A2 = B1 = B and
θ = 1/2.
(f) Prove the identity between arithmetic, harmonic and geometric mean:

A+B
2(A−1 + B −1 )−1 # = A#B.
2
Hint: Use the fact that M #N is the unique solution in H+
n of the RicattiRicatti
equation XM −1 X = N . Use it thrice.

(Continuation.) Each positive definite Hermitian matrix A defines a norm kxkA :=

199. √
x∗ Ax. If M ∈ Mp×q (C) and A1 ∈ HDPq , A2 ∈ HDPq , we denote by kM kA1 ←A2
the norm
sup{kM xkA1 ; kxkA2 = 1}.

(a) Show that the dual norm of the Hermitian norm k · kA is k · kA−1 .
(b) Prove the following interpolation result (compare with the Riesz–ThorinRieszThorin
Theorem): For every matrix M ∈ Mp×q (C) and every A1 , B1 ∈ HDPq , A2 , B2 ∈
HDPq , we have
1/2 1/2
kM kA1 #B1 ←A2 #B2 ≤ kM kA1 ←A2 kM kB1 ←B2 .

Hint: Again, use the definition of the geometric mean, not the formula.

123
 Comment: In the terminology of interpolation theory, one writes for A, B ∈ HDPn

(Cn ; A#B) = [(Cn ; A), (Cn ; B)]1/2 ,

where 1/2 is the interpolation parameter. Recall that

[(Cn ; A), (Cn ; B)]0 = (Cn ; B), [(Cn ; A), (Cn ; B)]1 = (Cn ; A).

More generally, [(Cn ; A), (Cn ; B)]θ can be computed for every diadic θ = m2−k by
means of iterated geometric mean. For instance

[(Cn ; A), (Cn ; B)]3/4 = [(Cn ; A), [(Cn ; A), (Cn ; B)]1/2 ]1/2 = [(Cn ; A), (Cn ; A#B)]1/2 .

(c) Following the idea presented above, show that there exists a unique continuous curve
s 7→ H(s) for s ∈ [0, 1], with the property that H(0) = B, H(1) = A and
 
s+t
H = H(s)#H(t), ∀0 ≤ s, t ≤ 1.
2
This curve is defined by the formula

H(s) = A1/2 (A−1/2 BA−1/2 )1−s A1/2 .

We denote [A, B]s = H(s). Verify that [A, B]1−s = [B, A]s .
200. In periodic homogenization (a chapter of Applied Partial Differential Equations) of elliptic
PDEs, one has a continuous map x 7→ A(x) from Rn into SDPn , which is Λ-periodic, Λ
being a lattice. The homogenized matrix Ā is defined as follow. For every vector e ∈ Rn ,
we admit that there exists a unique (up to an additive constant) solution w : Rn → R of
the problem

div(A(x)∇w) = 0 in Rn , w(x + ω) = w(x) + e · ω, ∀ω ∈ Λ.

(Notice that the last property implies that ∇w is periodic, and its average equals e.)
Then we define
Āe :=< A∇w >,
where < · > denotes the average over Y .
In this exercise we consider the simple case where n = 2 and A depends only on the first
coordinate x1 . In particular, one can take Λ spanned by P~e1 (P the period of x1 7→ A(x1 ))
and by ~e2 .
(a) Let w be as above. Prove that ∂w/∂x2 is a constant. More precisely, show that
∂w/∂x2 ≡ e2 .
(b) Likewise, prove that
∂w
+ a12 e2
a11
∂x1
is a constant and compute that constant.

124
 (c) Finally, prove the following formula
!
[a11 ] [a12 ]
Ā = 2
[a12 ] < (det A)/a11 > + <a<1/a
12 /a11 >
11 >

where [f ] :=< f /a11 > / < 1/a11 >. In particular, verify that det Ā = [det A].

201. We recall that the symplectic group Spn (k) is defined as the set of matrices M in M2n (k)
which satisfy M T Jn M = Jn , where
 
0n In
Jm := .
−In 0n

(a) Let  
A B
M=
C D
be a symplectic matrix. Check that AB T and AT C are symmetric, and that AT D −
C T B = In .
(b) Let X ∈ Symn (k) be given and M be as above. Show that (CX + D)T (AX + B) is
symmetric. Deduce that if CX + D is non-singular, then M · X := (AX + B)(CX +
D)−1 is symmetric.
(c) Show that the matrices R(X) := C T (AX + B)AT C and S(X) := AT (CX + D)AT C
are symmetric for every symmetric X.
(d)
From now on, we choose k = R. We
say that M is a Hamiltonian matrix if
AB T and AT C are positive definite.

Rowan Hamilton.
Hamilton, Rowan (Eire)

i. Let X ∈ SPDn be given. Show that R(X) and S(X) are positive definite.
ii. Show also that A−1 C ∈ SPDn .
iii. Deduce that Y is similar to the product of three positive definite symmetric
matrices.
iv. Conclude that Y is positive definite (see Exercise 6 of Chapter 7).
To summarize, a Hamiltonian matrix M acts over SPDn by X 7→ M ·X (SiegelSiegel,
Bougerol)Bougerol.

125
 (e) Prove that the set of Hamiltonian matrices is a semi-group: The product of two
Hamiltonian matrices is Hamiltonian (WojtkowskiWojtkowski).
202. For every A ∈ Mn×m (C) and every t ∈ R, show that
n − m = Tr exp(−tAA∗ ) − Tr exp(−tA∗ A).
Comment: This is a special case of a formula giving the index of a FredholmFredholm
operator T , thus in infinite dimension:
indT = Tr exp(−tT T ∗ ) − Tr exp(−tT ∗ T ), ∀t > 0.
Notice that in general the difference exp(−tT T ∗ ) − exp(−tT ∗ T ) does not make sense.
203. Let n, r ≥ 2 be two integers. If A1 , . . . , Ar ∈ Mn (k) are given, one defines
X
Tr (A1 , . . . , Ar ) := (σ)Aσ(1) · · · Aσ(r) ,
σ∈Sr

where Sr is the group of permutations of {1, . . . , r} and  : Sr → {−1, +1} is the signature.
(a) Verify that Tr : Mn (k)r → Mn (k) is an alternate r-linear map.
(b) We consider the case k = R and we endow Mn (R) with the Frobenius norm. We
thus have a Euclidean structure, with scalar product hA, Bi = Tr(B T A).
i. Show that the supremum τ (r, n) of
kTr (A1 , . . . , Ar )k
kA1 k · · · kAr k
over A1 , . . . , Ar 6= 0n is reached. We choose an r-uplet (M1 , . . . , Mr ) at which
this maximum is obtained. Check that one is free to set kMj k = 1 for all j.
 
a b
ii. Let ∈ SL2 (R) be given. Show that
c d
kTr (M1 , M2 , M3 , . . . , Mr )k = kTr (aM1 + bM2 , cM1 + dM2 , M3 , . . . , Mr )k.
Deduce that if τ (r, n) 6= 0, then kaM1 + bM2 k kcM1 + dM2 k ≥ 1.
iii. Derive from above that τ (r, n) 6= 0 implies that hMi , Mj i = 0 for every pair
i 6= j.
iv. Conclude that Tr ≡ 0n for every r ≥ n2 + 1
(c) We go back to a general field of scalars k. Prove that for every r ≥ n2 + 1 and every
A1 , . . . , Ar ∈ Mn (k), one has
Tr (A1 , . . . , Ar ) = 0n .
Hint: Apply the principle of algebraic identities. Use the fact that R is infinite.
Comment. The vanihing of Tr is called a polynomial identity over Mn (k). The
above result is far from optimal. The Theorem of AmitsurAmitsur and LevitzkiLevitzki
tells us that Tr vanishes identically over Mn (k) if, and only if, r ≥ 2n. See Exercise
289.

126
204. This exercise yields a lemma of MinkowskiMinkowski.
(a) Show that, if either p = 2 and α ≥ 2, or p ≥ 3 and α ≥ 1, then pα ≥ α + 2.
(b) Define q := pβ , with p a prime number and β ≥ 1 if p is odd, β ≥ 2 if p = 2. Deduce
from above that for every 2 ≤ j ≤ k, p2β+vp (k) (vp (k) is the power to which p divides
k, its p-valuation) divides  
k
pjβ .
j
(c) Let q be as above and B ∈ Mn (Z) be such that p does not divide B in Mn (Z). Let
k ≥ 2 be an integer and form A := In + qB. Show that

Ak ≡ In + kqB mod p2β+vp (k) .

(d) Deduce that if A is as above and if Ak = In , then A = In .

(e) More generally, let A ∈ Mn (Z) be given. Prove that if m divides A − In and if
Ak = In for some integers m ≥ 3 and k ≥ 1, then A = In . In other words, the kernel
of the homomorphism GLn (Z) → GLn (Z/mZ) is torsion free.
(f) Show that the statement is false when m = 2. Find a matrix A ∈ I2 + 2M2 (Z) such
that A2 = I2 and A 6= I2 .
205. Let A ∈ Mn (k), B ∈ Mm (k) and M ∈ Mn×m (k) be such that AM = M B. It is well-
known that if n = m and M is non-singular, then the characteristic polynomials of A and
B are equal: PA = PB . Prove that gcd{PA , PB } has a factor of degree r = rk M . Hint:
Reduce to the case where M is quasidiagonal.
206. Let k be a field. Given two vectors X, Y in k 3 , we define the vector product as usual:
 
x2 y3 − x3 y2
X × Y :=  x3 y1 − x1 y3  .
x1 y2 − x2 y1

Prove the following identity in M3 (k):

X(Y × Z)T + Y (Z × X)T + Z(X × Y )T = det(X, Y, Z) I3 , ∀X, Y, Z ∈ k 3 .

207. Let k be a field and 1 ≤ p ≤ m, n be integers.

(a) Let M ∈ Mn×m (k) be given, with rkM = p. Show that there exist two matrices
X ∈ Mn×p (k) and Y ∈ Mp×m (k) such that M = XY .
(b) We write such a rank−p matrix in block form:
 
A B
M= ,
C D

with A ∈ Mp×p (k). If A is non-singular, show that D = CA−1 B.

127
 (c) We assume that M as rank p and that det(J + M ) = 0, where
 
0p 0
J= .
0 In−p

Show that the block A is singular. Deduce that there exists a non-zero vector
z ∈ k n−p such that either (J + M )Z = 0, or (J + M T )Z = 0, where
 
0
Z := .
z

208. (Continuation.) We assume that n = kp and M has rank p at most. We can therefore
factorize M in the form  
A1
 .. 

M =  .  B1 · · · Bk .
Ak
Let X1 , . . . , Xk be indeterminates, and define X := diag{X1 Ip , . . . , Xk Ip }. Show that
!
X
det(X + M ) = det X1 · · · Xk Ip + X̂j Bj Aj ,
j

where X̂j denotes X1 · · · Xj−1 Xj+1 · · · Xk . Hint: As usual, Schur formula is useful.

209. We show here that H 7→ (det H)1/n is concave over HPDn .

(a) Recall that, given H and K in HPDn , the product HK is diagonalizable with real,
positive, eigenvalues (see Exercise 258).
(b) Deduce that
1
(det H)1/n (det K)1/n ≤ Tr HK.
n
Hint: Use the arithmetic-geometric inequality.
(c) Show that
 
1/n 1
(det H) = min Tr HK ; K ∈ HPDn and det K = 1 .
n

(d) Deduce concavity.

128
210. The notation comes from a nonlinear electrodynamics called the Born–Infeld modelBornInfeld.

In the canonical Euclidian space R3 (but Rn works as

well), we give ourselves two vectors E and B, satisfying

kEk2 + (E · B)2 ≤ 1 + kBk2 .

Prove the following inequality between symmetric ma-

trices
EE T + BB T ≤ (1 + kBk2 )I3 .

Max Born.
Born, Max (Rép. de Guinée)

211. Fix two integers 0 ≤ m ≤ n − 1. We give ourselves complex numbers ajk for every
1 ≤ j, k ≤ n such that |k − j| ≤ m (2m + 1 diagonals). We assume that akj = ajk .
Prove that we can complete this list of entries so as to make a matrix A ∈ HDPn if, and
only if, every principal submatrix of size m + 1,
 
ajj ··· aj,j+m
 .. .. .. 
 . . . 
aj+m,j · · · aj+m,j+m

is positive definite.

212. We denote    
1 1 1 0
X := , Y := .
0 1 1 1

(a) Let M ∈ SL2 (Z) be a non-negative matrix (that is an element of SL2 (N)). If M 6= I2 ,
show that the columns of N are ordered: (m11 − m12 )(m21 − m22 ) ≥ 0.
(b) Under the same assumption, deduce that there exists a matrix M 0 ∈ SL2 (N) such
that either M = M 0 X or M = M 0 Y . Check that Tr M 0 ≤ Tr M . Under which
circumstances do we have Tr M 0 < Tr M ?
(c) Let M ∈ SL2 (N) be given. Arguing by induction, show that there exists a word w0
in two letters, and a triangular matrix T ∈ SL2 (N), such that M = T w0 (X, Y ) ∈
SL2 (N).

129
 (d) Conclude that for every M ∈ SL2 (N), there exists a word w in two letters, such that
M = w(X, Y ).

Comment. One can show that every element of SL2 (Z), whose trace is larger than 2, is
conjugated in SL2 (Z) to a word in X and Y . This word is not unique in general, since if
M ∼ w0 (X, Y )w1 (X, Y ), then M ∼ w1 (X, Y )w0 (X, Y ) too.

213. Let M, N ∈ Mn (k) be given.

(a) Show that there exists a non-zero pair (a, b) ∈ k̄ 2 (k̄ the algebraic closure of k) such
that det(aM + bN ) = 0.
(b) Let (a, b) be as above and x be an element of ker(aM + bN ). Show that (M ⊗ N −
N ⊗ M )x ⊗ x = 0. Deduce that det(M ⊗ N − N ⊗ M ) = 0 for every M, N ∈ Mn (k).
(c) We assume that M and N commute to each other. Show that det(M ⊗ N − N ⊗ M )
can be computed as a nested determinant (see exercise 120).
(d) When N = In , show that det(M ⊗ In − In ⊗ M ) = 0 also follows from the Cayley–
HamiltonCayleyHamilton theorem. Hint: Use the previous question.

214. Let M ∈ Mm (k) and N ∈ Mn (k) be given. Prove that

det M ⊗ N = (det M )n (det N )m .

Hint: Use again Exercise 120.

Likewise, let M ∈ Mp×q (k) and N ∈ Mr×s (k) be given, with pr = qs. Check that M ⊗ N
is a square matrix. Show that rk(M ⊗ N ) ≤ (min{p, q})(min{r, s}), and deduce that
det(M ⊗ N ) = 0 when p 6= q. The case p = q is covered by the identity above.

215. (P. FinslerFinsler, J. MilnorMilnor.) Let F : Symn (R) → R be a C 1 -function. Let us

define the symmetric matrix M by
∂F
mij := (0n ).
∂aij
Let µ− be the smallest eigenvalue of M . Prove that
 
F (A) − F (0)
µ− = liminf | A > 0, TrA → 0 .
TrA

216. Let H be a Hermitian matrix, such that for every K ∈ HPDn , there holds

det(H + K) ≥ det H.

Prove that H is positive semi-definite.

217. If A ∈ Symn (R), we denote qA the associated quadratic form. In the sequel, A and B
are too real symmetric matrices.

130
(a) Assume that there exists a positive definite matrix among the linear combinations
of A and B (the pencil spanned by A and B). Prove that there exists a P ∈ GLn (R)
such that both P T AP and P T BP are diagonal. Hint: A classical result if B itself
is positive definite.
(b) We assume instead that qA (x) = qB (x) = 0 implies x = 0, and that n ≥ 3.
i. Show that R(B) ∩ A(ker B) = {0}.
ii. Let ∆(λ) be the determinant of A + λB. Using a basis of ker B, completed as a
basis of Rn , show that the degree of ∆ equals the rank of B. Deduce that there
exists a non-degenerate matrix in this pencil.
(c) We keep the assumption that qA = qB = 0 implies x = 0. From the previous
question, we may assume that B is non-degenerate.
i. Let us define
qA (x) + iqB (x)
Z(x) := p ∈ C, ∀x 6= 0.
qA (x)2 + qB (x)2

Show that there exists a differentiable real-valued map x 7→ θ(x) over Rn \ {0},
such that Z(x) = exp(iθ(x)) for every x 6= 0 in Rn .
ii. Let x be a critical point of θ. Show that qA (x)Bx = qB (x)Ax. Show that there
exists such a critical point x1 . Show that qB (x1 ) 6= 0.
iii. We define E1 := (Bx1 )⊥ . Show that the restriction of B to E1 is non-degenerate.
Prove that a critical point x2 of the restriction of θ to E1 \ {0} again satisfies
qA (x2 )Bx2 = qB (x2 )Ax2 , as well as (x1 )T Bx2 = 0 and qB (x2 ) 6= 0.
iv. Arguing by induction, construct a sequence x1 , . . . , xn of vectors of Rn with the
properties that Axj k Bxj , and (xj )T Bxk vanishes if and only if j 6= k.
v. Conclusion: There exists a P ∈ GLn (R) such that both P T AP and P T BP are
diagonal.
(d) We are now in the position that both A and B are diagonal, and still qA = qB = 0
implies x = 0. We wish to show that there exists a linear combination of A and B
that is positive definite
i. We argue by contradiction. Suppose that none of the vectors λdiagA + µdiagB
is positive (in the sense of Chapter 5) when (λ, µ) run over R2 . Show that there
exists a hyperplane in Rn , containing diagA and diagB, but no positive vector.
Hint: Apply Hahn–BanachHahnBanach.
Deduce thatPthere exists a non-negative, non-zero vector y ∈ Rn such that
ii. P
j yj ajj = j yj bjj = 0.
iii. Show that this implies y = 0.
In conclusion, we have found the equivalence (as long as n ≥ 3) of the conditions:
• If qA (x) = qB (x) = 0, then x = 0,
• There exists a positive definite linear combination of qA and qB ,

131
 and they imply a third one
• There exists a basis of Rn , orthogonal for both qA and qB .
(e) Provide a counter-example when n = 2: Find A and B such that qA = qB = 0
implies x = 0, but there does not exist a basis simultaneously orthogonal for A and
B. In particular, combinations of A and B cannot be positive definite.

218. This may be a new proof of Gårding’sGaa@Gårding Theorem.

(a) Let F : Ω → R+ be a positive, homogeneous function of degree α, over a convex

cone Ω of RN . We assume that F is quasi-concave over Ω; by this we mean that at
every point x ∈ Ω, the restriction of the Hessian D2 F (x) to ker dF (x) is non-positive.
Prove that F 1/α is concave over Ω. Hint: Use repeatedly EulerEuler Identity for
homogenous functions.
(b) We now focus to RN ∼ Hn (and thus N = n2 ), with F (M ) = det M and Ω = HPDn .
i. If M ∈ HPDn is diagonal, compute explicitly dF (M ) and D2 F (M ). Then check
that F is quasi-concave at M .
ii. Extend this property to all M ∈ HPDn , using the diagonalisability through a
unitary conjugation.
iii. Deduce a particular case of Gårding’s Theorem: M 7→ (det M )1/n is concave
over HPDn .
Gårding’s Theorem is that if F is a hyperbolic polynomial, homogeneous of degree
n, and Ω is its forward cone, then F 1/n is concave over Ω. This may be proved in
full generality with the argument above, together with the fact that the quadratic
form Z 7→ D2 F (X)Z ⊗2 is of signature (1, N − 1). See an other proof below.
Notice that the first concavity result given above can be written even for non-smooth
functions F , thus without invoquing first- an second-order differential. Prove that
if G : Ω → R+ is homogeneous of degree one, and if K := {x ∈ Ω | G(x) ≥ 1} is
convex (by homogeneity, this amounts to quasi-concavity), then G is concave.

219. Here is an other proof of the concavity of det1/n over HPDn .

(a) Given non-negative real numbers a1 , . . . , 

an , we denote σk (a) the k-th elementary
n
symmetric polynomial, which is a sum of monomials. Prove that
k
 
n
σk (a) ≥ σn (a)k/n .
k

Hint: Use the arithmetico-geometric inequality.

(b) Prove that for every K ∈ HPDn , one has

n
det(In + K) ≥ 1 + (det K)1/n .

132
 (c) Deduce the inequality
(det(H + K))1/n ≥ (det H)1/n + (det K)1/n
for every H, K ∈ HPDn . Conclude.
220. We consider the Hermitian norm k · k2 over Cp and Cq . We denote by B the unit ball
(the set of linear contractions) in Mp×q (C). Recall that a contraction is a map satisfying
kf (x) − f (y)k2 < kx − yk2 whenever y 6= x.
(a) Show that M ∈ Mp×q (C) is a contraction if, and only if, kM k2 < 1. Deduce that
M ∗ is also a contraction.
(b) Let H ∈ Hq and P ∈ GLq (C) be given. Show that P −∗ HP −1 < Iq is equivalent to
H < P ∗P .
(c) Given a matrix U ∈ U(p, q), written in block form
 
A B
U= ,
C D
we define a map F over Mp×q (C) by
F (Z) := (AZ + B)(CZ + D)−1 .
Show that F maps B into itself.
(d) Show that the set of maps F form a group (denoted by Γ) as U runs over U(p, q),
and that the map U 7→ F is a group homomorphism.
(e) Show that for every Z given in B, there exists an F as above, such that F (Z) = 0p×q .
Deduce that the group Γ acts transitively over B.
221. (Loo-Keng HuaHua, Loo-Keng.) This exercise uses the previous one, and in particular
has the same notations. We define the following function over B × B:
| det(Iq − W ∗ Z)|2
φ(W, Z) := .
det(Iq − W ∗ W ) det(Iq − Z ∗ Z)
(a) Of course, φ(0p×q , Z) ≥ 1 for every contraction Z, with equality only if Z = 0p×q .
(b) Show that if U ∈ U(p, q) and F is defined as above, then
φ(F (W ), F (Z)) = φ(W, Z).
We say that φ is invariant under the action of Γ.
(c) Deduce Hua’s Inequality: For any two contractions W and Z, one has
det(Iq − W ∗ W ) det(Iq − Z ∗ Z) ≤ | det(Iq − W ∗ Z)|2 ,
with equality if and only if W = Z. Hint: Use transitivity ; then it is enough to
treat the case W = 0p×q .

133
 In other words, φ(W, Z) ≥ 1, with equality only if W = Z. The quantity

det(Iq − W ∗ Z)
K(W, Z) := ,
det(Iq − W ∗ W )1/2 det(Iq − Z ∗ Z)1/2

whose square modulus is φ(W, Z), is the Bergman kernelBergman of the symmetric do-
main B.

222. We use the notations of Exercise 146. We assume K = C. The exterior algebra Λk E is
naturally endowed with a Hermitian structure, in which a unitary basis is given by the
vectors ei1 ∧ · · · ∧ eik with 1 ≤ i1 < · · · < ik ≤ n, {e1 , . . . , en } being the canonical basis of
E = Cn .

(a) Prove that A(k) B (k) = (AB)(k) .

∗
(b) Prove that A(k) = (A∗ )(k) .
(c) Deduce that if U ∈ Un , then U (k) is unitary too.
(d) Let s1 , . . . , sn denote the singular values of a matrix A. Prove that |TrA| ≤ s1 +
· · · + sn .
(e) Let λ1 , . . . , λn denote the eigenvalues of A. Deduce for every 1 ≤ k ≤ n the inequality

|σk (λ)| ≤ σk (s),

where σk is the elementary symmetric polynomial if degree k in n arguments. The

case k = 1 has been established in the previous question. The case k = n is trivial.
Hint: Apply the case k = 1 to A(k) , and use Exercise 146.
(f) Use this result to prove that

| det(In + A)| ≤ det(In + |A|),

√
where |A| := A∗ A is the non-negative symmetric part in the polar decomposition.

223. (L. DinesDines.) Let A, B ∈ Symn (R) be given. Show that the range of the map
x 7→ (xT Ax, xT Bx) is a convex subset of R2 .
Compare with the Toeplitz–Hausdorff Lemma (Exercise 21).

224. Let A ∈ Mn (k) be a block triangular matrix. Show that A is nilpotent if, and only if, its
diagonal blocks are nilpotent.

225. A subspace V of Mn (k) is nilpotent if every element of V is nilpotent. Let V be a nilpotent

subspace.

(a) If A ∈ V , prove that Tr A2 = 0.

i. If chark 6= 2, deduce that Tr(AB) = 0 for every A, B ∈ V .

134
 ii. More generally, check that

Tr(AB) = (Tr A)(Tr B) + σ2 (A) + σ2 (B) − σ2 (A + B),

and deduce that Tr(AB) = 0 for every A, B ∈ V , even in characteristic 2.

Hereabove, σ2 (M ) denotes the second elementary symmetric polynomial in the
eigenvalues of M .
(b) Let U denote the subspace of upper triangular matrices, and U + that of strictly
upper triangular. Likewise, we denote L and L− for lower triangular matrices. One
has
Mn (k) = L− ⊕ U = L ⊕ U + .
Let φ be the projection over L− , parallel to U . We denote K := V ∩ ker φ and
R := φ(V ).
i. Show that K ⊂ U + .
ii. If M ∈ R and B ∈ K, prove that Tr(M B) = 0. In other words, RT and K are
orthogonal subspaces of U + , relatively to the form hM, N i := Tr(M T N ).
iii. Deduce that
dim R + dim K ≤ dim U + .
(c) In conclusion, show that the dimension of a nilpotent subpace of Mn (k) is not larger
than
n(n − 1)
.
2
Nota. Every nilpotent subspace of dimension n(n − 1)/2 is conjugated to U + .
However, it is not true that every nilpotent subspace is conjugated to a subspace of
U + . For instance, let n = 3: Prove that the space of matrices
 
0 0 x
0 0 y 
y −x 0

is nilpotent, but is not conjugated to a space of triangular matrices. Hint: the

kernels of these matrices intersect trivially.

226. We use the scalar product over Mn (C), given by hM, N i = Tr(M ∗ N ). We recall that
the corresponding norm is the Schur–Frobenius√norm k · kF . If T ∈ GLn (C), we denote
T = U |T | the polar decomposition, with |T | := T ∗ T and U ∈ Un . The AluthgeAluthge
transform ∆(T ) is defined by

∆(T ) := |T |1/2 U |T |1/2 .

(a) Check that ∆(T ) is similar to T .

(b) If T is normal, show that ∆(T ) = T .

135
 (c) Show that k∆(T )kF ≤ kT kF , with equality if, and only if, T is normal.
(d) We define ∆n by induction, with ∆n (T ) := ∆(∆n−1 (T )).


i. Given T ∈ GLn (C), show that the sequence ∆k (T ) k∈N is bounded.
ii. Show that its limit points are normal matrices with the same characteristic poly-
nomial as T (Jung, Jung, Il Bong KoKo, Eungil & PearcyPearcy, or AndoAndo).
iii. Deduce that when T has only one eigenvalue µ, then the sequence converges
towards µIn .
Comment: The sequence does converge for every initial T ∈ Mn (C), according to
J. AntezanaAntezana, E. R. PujalsPujals and D. Strojanoff.Strojanoff
(e) If T is not diagonalizable, show that these limit points are not similar to T .

227. (After C. de Lellisdelel@de Lellis & L. SzékelyhidiSzékelyhidi Jr.)

If x ∈ Rn , we denote x ⊗ x := xxT . We also use the standard Euclidian norm. The
purpose of this exercise is to prove that the convex hull of
 
n 1
K := (v, S) | v ∈ R , |v| = 1 and S = v ⊗ v − In
n
equals  
1
C= (v, S) | S ∈ Symn (R), Tr S = 0 and v ⊗ v − In ≤ S .
n

Left: Mmmh ! This Camillo de Lellis is not

the mathematician.

dele@de
Lellis, Camillo (Vatican
state)

(a) Show that every (v, S) in C satisfies

 
1
|v| ≤ 1 and S ≤ 1 − In .
n

(b) Check that K ⊂ C.

136
 (c) Prove that C is a convex compact subset of Rn × Symn (R).
(d) Let (v, S) ∈ C be given, such that
1
v⊗v− In 6= S.
n
i. Show that |v| < 1.
ii. Let µ be the largest eigenvalue of S. Show that µ ≤ 1 − 1/n. In case of equality,
show that there exists a unit vector w such that
1
S =w⊗w− In .
n
iii. In the latter case (µ = 1 − 1/n), show that v = ρw for some ρ ∈ (−1, 1). Deduce
that (v, S) is not an extremal point of C.
iv. We now assume on the contrary that µ < 1 − 1/n. Let N denote the kernel of
1
S−v⊗v+ In .
n
If N ⊂ ker S 0 for a symmetric, trace-less S 0 , show that (v, S + S 0 ) ∈ C for ||
small enough.
If dim N ≤ n − 2, deduce again that (v, S) is not an extremal point of C.
v. We still assume µ < 1−1/n, and we now treat the case where N is a hyperplane.
Show that there exists a vector z 6= 0 such that
1
S =v⊗v+z⊗z− In .
n
Show that there exists a non-zero pair (α, β) ∈ R2 such that, defining w = αz
and s = z ⊗ w + w ⊗ z + βz ⊗ z, one has (v, S) ± (w, s) ∈ C. Deduce that (v, S)
is not an extremal point in C.
(e) Deduce that every extremal point of C belongs to K.
(f) Conclude, with the help of KreinKrein-MilmanMilman’s Theorem.
(g) Show that in particular, 0n is a relative interior point of the convex set C (that is,
an interior point of C as a subset of the affine space spanned by C).

228. We recall that the Pfaffian of a 4 × 4 alternate matrix A is a12 a34 + a13 a42 + a14 a23 .
Show that Alt4 (R) ∩ GL4 (R) has two connected components, each one homeomorphic to
S 2 × S 2 × R2 , with S 2 the two-dimensional sphere.

229. Let M ∈ Mn (k) be given in block form

 
A B
M= ,
C D

137
 where the diagonal blocks have square shape. Recall that if A is non-singular, then its
SchurSchur complement is defined as

Ac := D − CA−1 B.

Let us restrict ourselves to the case k = C and to Hermitian matrices. If M is positive

definite, then so is A, and the Schur complement is well-defined.
Is instead M is positive semi-definite, show that R(B) ⊂ R(A), and thus there exists
a (not necessarily unique) rectangular matrix X such that B = AX. Show that the
product X ∗ AX does not depend on the particular choice of X above. Deduce that the
map A 7→ Ac extends continuously to the closure of HPDn , that is to the cone of positive
semi-definite matrices. Finally, show that det M = det A · det Ac .

230. After R. B. BapatBapat, we define a doubly stochastic n-tuple as an n-uplet A = (A1 , . . . , An )

of Hermitian semi-positive definite matrices Aj ∈ H+ n (yes, the same n) with the proper-
ties that

• for every j = 1, . . . , n, Tr Aj = 1,
• and moreover, nj=1 Aj = In .
P

The set of doubly stochastic n-tuples is denoted by Dn .

(a) Given A as above and V ∈ Un , we define A(V ) ∈ Mn (R) by

∗
A(V )jk := v j Ak v j ,

where v 1 , . . . , v n are the columns of V .

Show that A(V ) is a doubly-stochastic matrix.
(b) Conversely, let A = (A1 , . . . , An ) be an n-uplet of n × n complex matrices. Define
A(W ) as above. Show that, if A(W ) is doubly stochastic for every unitary V , then
A is a doubly stochastic n-uplet.
(c) Check that the subset of Dn made of those A’s such that every Aj is diagonal, is
isomorphic to the set of doubly stochastic matrices.
(d) Remark that Dn is a convex compact subset of (Hn )n . If each Aj has rank one,
prove that A is an extremal point of Dn .
Nota: there exist other extremal points if n ≥ 4. However, the determination of all
the extremal points of Dn is still an open question.

231. (Continuation.) We examine some of the extremal points A = (A1 , . . . , An ) of Dn .

(a) If A = 21 (B + C) with B, C ∈ Dn , prove that R(B j ) and R(C j ) are contained in

R(Aj ) for every j.
(b) If A is not extremal, show that one may choose B, C as above, such that R(B i ) 6=
R(Ai ) for some index i.

138
 (c) In the situation described above, assume that Ai has rank two. Prove that B i = xi x∗i
and C i = yi yi∗ , where {xi , yi } is a unitary basis of R(Ai ).
(d) We now assume that every Aj has rank two, and that for every pair j 6= k, the planes
R(Aj ) and R(Ak ) are not orthogonal. Prove that all the matrices B j and C j have
the forms xj x∗j and yj yj∗ respectively, with {x1 , . . . , xn } and {y1 , . . . , yn } two unitary
bases of Cn .
(e) We set n = 4 and choose two unitary bases {v1 , . . . , v4 } and {w1 , . . . , w4 } of C4 . We
define
1 1
A1 = (v1 v1∗ + w3 w3∗ ), A2 = (v2 v2∗ + w4 w4∗ ),
2 2
1 1
A = (w1 w1∗ + w2 w2∗ ),
3
A = (v3 v3∗ + v4 v4∗ ).
4
2 2
Check that A ∈ D4 . Find a choice of the v’s and w’s such that there does not exist
a unitary basis {x1 , . . . , x4 } with xj ∈ R(Aj ). Deduce that A is an extremal point.

232. (From M. H. MehrabiMehrabi.) Let A, B ∈ Mn (R) be such that [A, B] is non-singular

and verify the identity
A2 + B 2 = ρ[A, B],
for some ρ ∈ R.
Show that
ρ−i
ω := p
1 + ρ2
is a (2n)-th root of unity.

233. We are given three planes E1 , E2 and E3 in the Euclidian space R3 , of respective equations
zj · x = 0. We are searching an orthogonal basis {v1 , v2 , v3 } such that vj ∈ Ej for each j.
Prove that such a basis exists if, and only if,

∆ := (det(z1 , z2 , z3 ))2 − 4(z1 · z2 )(z2 · z3 )(z3 · z1 )

is non-negative. When ∆ is positive, there exist two such bases, up to scaling.

234. Let M ∈ Mn (k) be given. Use Theorem 2.3.1 of the book to calculate the Pfaffian of the
alternate matrix  
0n M
.
−M T 0n
Warning. Mind that the Pfaffian of the same matrix when M = In is (−1)n(n−1)/2 .

235. In a vector space V of dimension n, we are given n subspaces Ej , j = 1, . . . , n. We

examine the question whether there exists a basis {v1 , . . . , vn } of V with vj ∈ Ej for every
j.

139
 (a) Check the obvious necessary condition: for every index subset J,
 
(26) dim + Ej ≥ |J|.
j∈J

In the sequel, we want to prove that it is also a sufficient condition. We shall argue
by induction over n.
(b) Let us assume that the result is true up to the dimension n − 1.
From now on we assume that dim V = n and that the property (26) is fulfilled.
i. Prove the claim for the given set E1 , . . . , En , in the case where there exists a
index subset J such that
 
dim + Ej = |J|, 1 ≤ |J| ≤ n − 1.
j∈J

Hint: Apply the induction hypothesis to both

W := + Ej and Z := V /W.
j∈J

ii. There remains the case where, for every J with 1 ≤ |J| ≤ n − 1, one has
 
dim + Ej ≥ |J| + 1.
j∈J

Select a non-zero vector v1 ∈ E1 . Define Fj = (kv1 + Ej )/kv1 , a subspace of

W := V /kv1 . Check that F2 , . . . , Fn and W satisfy the assumption (26). Apply
the induction hypothesis.
(c) Conclude that a necessary and sufficient condition for the existence of a basis
{v1 , . . . , vn } of V with vj ∈ Ej for every j = 1, . . . , n, is that (26) is fulfilled for
every index subset J.

236. We consider differential equations over Mn (C), of the form

dM
(27) = [A(t), M ].
dt
Hereabove, A(t) can be given a priori, or be determined by M (t). As usual, the bracket
denotes the commutator.

(a) Show that  

d T dM
det M = Tr M̂ ,
dt dt
where M̂ is the adjugate of M . Deduce that t 7→ det M (t) is constant.

140
 (b) Define P (t), the solution of the Cauchy problem

dP
= −P A, P (0) = In .
dt
Show that M (t) = P (t)−1 M (0)P (t). In particular, the spectrum of M (t) remains
equal to that of M (0) ; one speaks of an isospectral flow.
(c) First example: take A := M ∗ . Show that t 7→ kM (t)kF (Frobenius norm) is constant.
(d) Second example: take A := [M ∗ , M ]. Show that t 7→ kM (t)kF is monotonous non
increasing. Deduce that the only rest points are the normal matrices.

237. This problem and the following one examine the dimension of the commutant of a given
matrix A ∈ Mn (k) when k is algebraically closed. We begin with the equation

(28) BY = Y C,

where B = J(0; p) and C = J(0 : q) are nilpotent matrices in Jordan form. The unknown
is Y ∈ Mp×q (k). We denote by {e1 , . . . , eq } and {f1 , . . . , fp } the canonical bases of k q and
k p , respectively. We thus have Cej = ej−1 , and likewise with B and fj .

(a) Show that for every solution Y of (28), there exist scalars y1 , . . . , yq such

Y ej = y1 fj + y2 fj−1 + · · · + yj f1 .

(b) If p < q, explain why y1 , . . . , yq−p must vanish.

(c) Deduce that the solution space of (28) has dimension min(p, q).

238. (Continuation.) Let A ∈ Mn (k) be given. We denote com(A) the space of matrices
X ∈ Mn (k) such that AX = XA.

(a) If A0 is similar to A, show that com(A) and com(A0 ) are conjugated, and thus have
the same dimension.
(b) Assuming that k is algebraically closed, we thus restrict to the case where A =
diag{A1 , . . . , Ar }, where Aj has only one eigenvalue λj and j 6= k implies λj 6= λk .
We decompose X blockwise accordingly:

X = (Xjk )1≤j,k≤r .

Show that X commutes with A if, and only if, Aj Xjk = Xjk Ak for every j, k. In
particular, j 6= k implies that Xjk = 0: X is block-diagonal too. Therefore,
r
X
dim com(A) = dim com(Aj ).
j=1

141
 (c) We are thus left with the special case where A has only one eigenvalue λ, and has
invariant polynomials (X − λ)m1 , . . . , (X − λ)mn with

m1 ≤ · · · ≤ mn and m1 + · · · + mn = n.

i. Show that the commutant of A does not depend on λ, but only on m1 , . . . , mn .

ii. Use the previous exercise to find
n−1
X
dim com(A) = (2j + 1)mn−j .
j=0

(d) Deduce that for every A ∈ Mn (k), one has

dim com(A) ≥ n.

239. (After C. HillarHillar & Jiawang NieNie, Jiawang.) Let S ∈ Symn (Q) be given. We
assume that S ≥ 0n in the sense of quadratic forms. The purpose of this exercise is to
prove that S is a sum of squares (Aj )2 with Aj ∈ Symn (Q).

Recall that if n = 1, this follows from the La-

grange’sLagrange Theorem: every positive integer
is the sum of four squares of integers.

We recall the key fact that in an Abelian ring, the

product of two sums of squares is a sum of squares.

We denote by p the minimal polynomial of S,

which we write

p(X) = X s − as−1 X s−1 + · · · + (−1)s as .

Joseph-Louis Lagrange.
Lagrange, Joseph-Louis (France)
(a) Check that p has simple roots. In particular, a0 and a1 cannot vanish both.
(b) Verify that the coefficients ak are non-negative. Deduce that a1 6= 0.
(c) We decompose (−1)s−1 p(X) = q(X 2 )X − r(X 2 ). Check that T := q(S 2 ) is a sum
of squares, and that it is invertible. Show then that T −2 r(S 2 ) is a sum of squares.
Conclude.
240. Let A, B ∈ Mn (C) be given. We assume that there exists a non-singular matrix P such
that
AP = P B, A∗ P = P B ∗ .
Prove that A and B are unitarily similar: there exists a U ∈ Un such that AU = U B.
Hint: Using the polar decomposition to P , show that we may assume P ∈ HPDn , and
then P diagonal.

142
 This statement is a part of the proof of Specht TheoremSpecht: A and B are unitarily
similar if, and only if, the equality Tr w(A, A∗ ) = Tr w(B, B ∗ ) holds true for every word
w in two letters. The rest of the proof involves representation theory and is beyond the
scope of our book.
241. (From C. VillaniVillani.) Let t 7→ R(t) ∈ Symn (R) be a continuous function over [0, T ].
We denote J0 (t) and J1 (t) the matrix-valued solutions of the differential equation
d2 J
+ R(t)J = 0,
dt2
uniquely determined by the Cauchy data
J0 (0) = In , J00 (0) = 0n , J1 (0) = 0n , J10 (0) = In .
We wish to prove that S(t) := J1 (t)−1 J0 (t) is symmetric, whenever J1 (t) is non-singular.
(a) Let uj (t) (j = 1, 2) be two vector-valued solutions of the ODE
(29) u00 + R(t)u = 0.
Verify that t 7→ hu01 (t), u2 (t)i − hu02 (t), u1 (t)i is constant.
(b) For a solution u of (29), show that
J0 (t)u(0) + J1 (t)u0 (0) = u(t).

(c) For t ∈ [0, T ), let us define the space Vt of the solutions of (29) such that u(t) = 0.
Show that it is an n-dimensional vector space. If J1 (t) is non-singular, verify that
an alternate definition of Vt is the equation
u0 (0) = S(t)u(0).

(d) Deduce the symmetry of S(t) whenever it is well-defined.

242. This is a sequel of Exercise 26, Chapter 4 (see also Exercise 153 in this list). We recall
that Σ denotes the unit sphere of M2 (R) for the induced norm k · k2 . Also recall that Σ is
the union of the segments [r, s] where r ∈ R := SO2 (R) and s ∈ S, the set of orthogonal
symmetries. Two distinct segments may intersect only at an extremity.
We construct the join J(R, S) as follows: in the construction above, we replace the
segments [r, s] by the lines that they span. In other words, J(R, S) is the union of the
(affine) lines passing through one rotation and one orthogonal symmetry. Of course,
Σ ⊂ J(R, S).
(a) Prove that J(R, S) is the algebraic set defined by the equation
kAk2F = 1 + (det A)2 ,
where kAk2F := Tr(AT A) (Frobenius norm).

143
 (b) If A ∈ J(R, S) \ O2 (R), show that A belongs to a unique line passing through a
rotation and an orthogonal symmetry (these rotation and symmetry are unique).
(c) Show that for every matrix A ∈ J(R, S), one has kAk2 ≥ 1.
(d) Find a diffeomorphism from a neighborhood of I2 to a neighborhood of 02 , which
maps the quadrics of equation (Y + Z)2 = 4ZT onto J(R, S).

243. (Continuation.)

(a) Interpret the equation of J(R, S) in terms of the singular values of A.

(b) More generally, show that the set USVn of matrices M ∈ Mn (R), having s = 1 as
a singular value, is an irreducible algebraic hypersurface.
(c) Show that the unit sphere of (Mn (R), k · k2 ) is contained in USVn . Therefore USVn
is the ZariskiZariski closure of the unit sphere.

244. Let A, B ∈ M2 (k) be given, where the characteristic of k is not 2. Show that [A, B]2 is a
scalar matrix λI2 . Deduce the following polynomial identity in M2 (k):

[A, B]2 , C = 02 .
 
(30)

Remark. Compared with T4 (A, B, C, D) = 02 (AmitsurAmitsur & LevitzkiLevitzkii

Theorem, see Exercise 289), the identity (30) has one less argument, but its degree is one
more.

245. In Mn (C), prove the equivalence between

det M = 0,

and

There exists a matrix A ∈ Mn (C) such that A−zM ∈ GLn (C) for every z ∈ C.

Hint: Use the rank decomposition (Theorem 6.2.2); show that M is equivalent to a
nilpotent matrix in Mn (C).

246. Let k be a field and  

A B
P =
C D
be an orthogonal matrix, with A and D square.
Prove that
det D = ± det A.
Hint: multiply P by  T 
A CT
.
0 I

Extend this result to elements P of a group O(p, q).

144
247. We endow Mn (C) with the induced norm k · k2 . Let G a subgroup of GLn (C) that is
contained in the open ball B(In ; r) for some r < 2.

(a) Show that for every M ∈ G, there exists an integer p ≥ 1 such that M p = In .
Hint: The eigenvalues of elements of G must be of unit modulus and semi-simple
(otherwise G is unbounded); they may not approach −1.
(b) Let A, B ∈ G be s.t. Tr(AM ) = Tr(BM ) for all M ∈ G. Prove that A = B. Hint:
Choose M in the subgroup spanned by B.
(c) Deduce that G is a finite group.
(d) On the contrary, find an infinite subgroup of GLn (C), contained in B(In ; 2).

248. Let z1 , . . . , zn ∈ C have positive real parts. Prove that the Hermitian matrix A with
entries
1
ajk :=
z̄j + zk
is positive definite.
Hint: Look for a Hilbert space H and elements f1 , . . . , fn ∈ H such that

ajk = hfj , fk i.

249. Let m ∈ N∗ be given. We denote Pm : A 7→ Am the m-th power in Mn (C). Show that
the differential of Pm at A is given by
m−1
X
DPm (A) · B = Aj BAm−1−j .
j=0

Deduce the formula Z 1

D exp(A) · B = e(1−t)A BetA dt.
0

250. Let f , an entire analytic function, be given.

If D = diag(d1 , . . . , dn } ∈ Mn (C), we define a matrix f [1] (D) ∈ Mn (C) by
 0
 f (dj ), k = j,
[1]
f (D)jk =
f (dj )−f (dk )
, k 6= j,

dj −dk

where we identify
f (b) − f (a)
:= f 0 (a),
b−a
if b = a.

145
 • If f = Pm (notations of the previous exercise), check that

(31) Df (D)B = f [1] (D) ◦ B,

where A ◦ B denotes the HadamardHadamard product.

• Prove that (31) holds true for every analytic function f (Daletskiı̆–KreinDaletskiı̆Krein
Formula).
Hint: Use polynomial approximation.

251. (B. G. ZaslavskyZaslavsky & B.-S. TamTam, Bit-Shun.) Prove the equivalence of the
following properties for real n × n matrices A:

Strong PerronPerron–FrobeniusFrobenius. The spectral radius is a simple eigen-

value of A, the only one of this modulus ; it is associated with positive left and right
eigenvectors.
Eventually positive matrix. There exists an integer N ≥ 1 such that k ≥ N implies
Ak > 0n .

252. (M. GoldbergGoldberg.) In a finite dimensional associative algebra A with a unit, every
element has a unique minimal polynomial (prove it). Actually, associative may be weak-
ened into power-associative: the powers ak are defined in a unique way. You certainly
think that if B is a sub-algebra and a ∈ B, then the minimal polynomial of a is the same
in A and B. So try this ....
Here A = Mn (k). Select a matrix M 6= In , 0n such that M 2 = M . What is its minimal
polynomial (it is the one in the usual, matricial, sense) ?
Then consider
B := M AM = {M AM ; A ∈ Mn (k)}.
Check that B is a subalgebra of A, and that M is the unit element of B. What is its
minimal polynomial in B ?
The explanation of this paradox lies in the notion of subalgebra. The equality of minimal
polynomials is guarranted if the subalgebra and the algebra have the same unit, which is
not the case hereabove.

253. (C. A. BergerBerger!C. A.’s theorem, proof by C. PearcyPearcy.) Recall (see Exercise
21) that the numerical radius of A ∈ Mn (C) is the non-negative real number

w(A) := max{|x∗ Ax| ; x ∈ Cn }.

The numerical radius is a norm, which is not submultiplicative. We show that it satisfies
however the power inequality.
In what follows, we use the real part of a square matrix
1
ReM := (M + M ∗ ),
2
146
 which is Hermitian and satisfies

x∗ (ReM )x = <(x∗ M x), ∀ x ∈ Cn .

(a) Show that w(A) ≤ 1 is equivalent to the fact that Re(In − zA) is semi-definite
positive for every complex number z in the open unit disc.
(b) From now on, we assume that w(A) ≤ 1. If |z| < 1, verify that In − zA is non-
singular. Hint: The numerical radius dominates the spectral one.
(c) If M ∈ GLn (C) has a non-negative real part, prove that Re (M −1 ) ≥ 0n . Deduce
that Re(In − zA)−1 ≥ 0n whenever |z| < 1.
(d) Let m ≥ 1 be an integer and ω be a primitive m-th root of unity in C. Check that
the formula
m−1
1 1 X 1
m
=
1−X m k=0 1 − ω k X
can be recast as a polynomial identity.
Deduce that
m−1
m −1 1 X
m
(In − z A ) = (In − ω k zA)−1 ,
m k=0
whenever |z| < 1.
(e) Deduce from above that
Re(In − z m Am )−1 ≥ 0n ,
whenever |z| < 1. Going backward, conclude that for every complex number y in
the open unit disc, Re(In − yAm ) ≥ 0n and thus w(Am ) ≤ 1.
(f) Finally, prove the power inequality

w(M m ) ≤ w(M )m , ∀M ∈ Mn (C), ∀ m ∈ N.

Nota: A norm which satisfies the power inequality is called a superstable norm.
It is stable if there exists a finite constant C such that kAm k ≤ C kAkm for every
A ∈ Mn (k) and every m ≥ 1. Induced norms are obviously superstable.
(g) Let ν ≥ 4 be a given constant. Prove that N (A) := νw(A) is a submultiplicative
norm over Mn (C) (GoldbergGoldberg & TadmorTadmor). Use the matrices
   
0 1 0 0
A= , B=
0 0 1 0

to show that this becomes false for ν < 4.

254. (J. DuncanDuncan.) We denote hx, yi the usual sesquilinear product in Cn . To begin
with, let M ∈ GLn (C) be given. Let
√ us write M = U√H = KU the left and right polar
decomposition. We thus have H = M ∗ M and K = M M ∗ .

147
 √ √
(a) Prove that U H = K U .
(b) Check that √ √
hM x, yi = h H x, U ∗ K yi, ∀ x, y ∈ Cn .
Deduce that
|hM x, yi|2 ≤ hHx, xi hKy, yi.
(c) More generally, let a rectangular matrix A ∈ Mn×m (C) be given. Prove the gener-
alized CauchyCauchy–SchwarzSchwarz inequality
√ √
|hAx, yi|2 ≤ h A∗ A x, xi h AA∗ y, yi, ∀ x, y ∈ Cn .

Hint: Use the decompositions

Cm = ker A ⊕⊥ R(A∗ ), Cn = ker A∗ ⊕⊥ R(A),

then apply the case above to the restriction of A from R(A∗ ) to R(A).

255. (a) Let A, B ∈ Mn (C), with A normal. If B commutes with A, prove that B commutes
with A∗ . This is B. FugledeFuglede’s theorem. Hint: Use the spectral theorem for
normal operators. See also Exercise 297
(b) More generally, let A1 , A2 be normal and B rectangular. Assume that A1 B = BA2 .
Prove that A∗1 B = BA∗2 (Putnam’sPutnam theorem). Hint: Use the matrix
 
A1 0
0 A2
and apply Fuglede’s theorem.
(c) Let A, B ∈ Mn (C) be given. Assume that the span of A and B is made of normal
matrices. Prove that [A, B] = 0n (H. RadjaviRadjavi & P. RosenthalRosenthal).
Hint: Use the matrices C = A + B and D = A + iB to prove [A∗ , B] = 0n , then
apply Fuglede’s theorem.

256. Let k be a field of characteristic zero. We consider two matrices A, B ∈ Mn (k) satisfying

[[A, B], A] = 0n .

In other words, ∆2 B = 0n , where

∆ : M 7→ [A, M ]

is a linear operator over Mn (k).

(a) Check that ∆ is a derivation:

∆(M N ) = (∆M )N + M (∆N ).

Therefore ∆ obeys a LeibnizLeibniz formula.

148
 (b) Deduce that for every m ≥ 1, one has

(32) ∆m (B m ) = m! (∆B)m .

(c) Deduce from (32) that ∆m (B j ) = 0n whenever m > j.

(d) Using the CayleyCayley–HamiltonHamilton Theorem, infer that ∆n (B n ) = 0n .
(e) Back to (32), establish that [A, B] is nilpotent (N. JacobsonJacobson).
(f) Application: let A ∈ Mn (C) satisfy [[A, A∗ ], A] = 0n . Prove that A is normal.

Nota: If k = C or R, one can

deduce from (32) and Proposition
4.4.1 that ρ(∆B) = 0, which is
the required result.

Left: Gottfried Leibniz.

Leibniz, Gottfried (Romania)

257. (Thanks to A. GuionnetGuionnet) Let n > m (≥ 1) be two integers. If V ∈ Un , we

decompose blockwise  
P Q
V = ,
R T
with P ∈ Mn (C). Notice that T is the matrix of a contraction.
If z ∈ C, of unit modulus, is not an eigenvalue of T ∗ , we define

W (z) := P + zQ(In−m − zT )−1 R ∈ Mm (C).

Show that W (z) is unitary.

If n = 2m, build other such rational maps from dense open subsets of Un to Um .

258. It seems that I have taken for granted the following fact:

149
 If H ∈ HPDn and h ∈ Hn are given, then the product Hh is diagonalizable
with real eigenvalues. The list of signs (0, ±) of the eigenvalues of Hh is the
same as for those of h.

Here is a proof:
√ √
(a) Show that Hh is similar to H h H.
(b) Use Sylvester’s inertia for the Hermitian form associated with h.

259. (L. MirskyMirsky.)

For a Hermitian matrix H with smallest and largest eigenvalues λ± , we define the spread

s(H) := λ+ − λ− .

Show that
s(H) = 2 max |x∗ Hy|,
where the supremum is taken over the pairs of unit vectors x, y ∈ Cn that are orthogonal:
x∗ y = 0.

260. For a given A ∈ GLn (C), we form M := A−1 A∗ . Let (λ, x) be an eigen-pair: M x = λx.

(a) Show that either |λ| = 1, or x∗ Ax = 0.

(b) Let us assume that x∗ Ax 6= 0. Prove that λ is a semi-simple eigenvalue.
(c) Find an A ∈ GL2 (C) such that the eigenvalues of M are not on the unit circle.
(d) Show that there exists matrices A ∈ GLn (C) without a bilateral polar decomposition
A = HQH, where as usual Q ∈ Un and H ∈ HPDn .

261. If A ∈ Mn (C) is given, we denote s(A) ∈ Rn+ the vector whose components s1 ≤ s2 ≤
· · · ≤ sn are the singular values of A.
Warning. This exercise involves two norms on Mn (C), namely the operator norm k · k2
and the Schur–Frobenius norm k · kF .

(a) Using von Neumann’svonneu@von Neumann inequality (16), prove that for every
matrices A, B ∈ Mn (C), we have

ks(A) − s(B)k2 ≤ kA − BkF .

(b) Deduce the following property. For every semi-definite positive Hermitian matrix H,
the projection (with respect to the distance d(A, B) := kA − BkF ) of In + H over
the unit ball of k · k2 is In .

150
262. Let A ∈ GLn (C) be given, and U DV ∗ be an SVD of A. Identify the factors Q and H of
the polar decomposition of A, in terms of U, V and D.
Let us form the sequence of matrices Xk with the rule
1
X0 = A, Xk+1 := (Xk + Xk−∗ ).
2
Show that Xk has the form U Dk V ∗ with Dk diagonal, real and positive. Deduce that

lim Xk = Q
k→+∞

and that the convergence is quadratic.

263. Let H ∈ Hn be positive semi-definite.

(a) Prove that H exp(−H) is Hermitian and satisfies the inequality

1
H exp(−H) ≤ In .
e

(b) Deduce that the solutions of the ODE

dx
+ Hx = 0
dt
satisfy
1
kHx(t)k2 ≤ kx(0)k2 .
et
Nota. This result extends to evolution equations in Hilbert spaces. For instance,
the solutions of the heat equation (a partial differential equation)

∂u
= ∆x u, x ∈ Rd , t ≥ 0
∂t
satisfy the inequality
Z Z
2 1
|∆x u(x, t)| dx ≤ |u(x, 0)|2 dx.
Rd et Rd

264. Let A, B ∈ HPDn be given. Prove that for every vector h ∈ Cn , we have
√ √
h∗ (A]B)h ≤ h∗ Ah h∗ Bh,

where A]B is the geometric mean of A and B (see Exercise 198).

Hint: Use the explicit formula

1/2
A]B = A1/2 A−1/2 BA−1/2 A1/2 .

151
265. (BatailleBataille.) Let A, B ∈ Mn (k) be such that A2 B = A. We assume moreover that
A and B have the same rank.

(a) Show that ker A = ker B.

(b) Prove BAB = B.
(c) Show that AB and A have the same rank and deduce k n = R(B) ⊕ ker A.
(d) Finally, prove that B 2 A = B.

266. Let us denote by N the set of nilpotent matrices in Mn (k). We also denote Gn the set
of polynomials p ∈ k[X] of degree less than n, such that p(0) = 0 and p0 (0) 6= 0. In other
words, p ∈ Gn if and only if

p(X) = a1 X + · · · + an−1 X n−1 , a1 6= 0.

For p, q ∈ Gn , we define p ◦ q as the unique element r ∈ Gn such that

r(X) ≡ p(q(X)), mod X n .

(a) Verify that (Gn , ◦) is a group, and that (p, N ) 7→ p(N ) is a group action over N .
(b) Apply this to prove that if k has characteristic 0, then for every j = 1, 2, . . . and
every N ∈ N , the matrix In + N admits a j-th-root in In + N , and only one in this
class.
(c) Denote this j-th-root by (In + N )1/j . If k = C, prove that

lim (In + N )1/j = In .

j→+∞

267. (M. CavachiCavachi, Amer. Math. Monthly 191 (2009)) We consider a matrix A ∈
GLn (Z) with the property that for every k = 1, 2, . . ., there exists a matrix Ak ∈ Mn (Z)
such that A = (Ak )k . Our target is to prove that A = In .

(a) Show that the distance of the spectrum of Ak to the unit circle tends to zero as
k → +∞. Deduce that the sequence of characteristics polynomials of the Ak ’s takes
finitely many values.
(b) Prove that there exists two integers (1 ≤)j|k such that j 6= k, while Aj and Ak have
the same characteristic polynomials. Show that their roots actually belong to the
unit circle, and that they are roots of unity, of degree less than or equal to n.
(c) Show that in the previous question, one may choose j and k such that k is divisible
by n!j. Deduce that the spectrum of Aj reduces to {1}.
(d) Verify that, with the terminology of Exercise 266, A belongs to In +N and Aj = A1/j .
(e) Verify that in the previous question, one may choose j arbitrarily large.
(f) For j as in the previous question and large enough, show that Aj = In . Conclude.

152
268. For a subgroup G of Un , prove the equivalence of the three properties

(P1) G is finite,
(P2) there exists a k ≥ 1 such that M k = In for every M ∈ G,
(P3) the set {Tr M | M ∈ G} is finite.

More precisely,

(a) Prove that (P 1) =⇒ (P 2) =⇒ (P 3).

(b) Let us assume (P 3). Choose a basis {M1 , . . . , Mr } of the subspace spanned by G,
made of elements of G. Every element M of G writes
r
X
M= αj Mj .
j=1

Express the vector α~ in terms of the vector of components Tr(M Mj∗ ). Deduce that
α
~ can take only finitely many values. Whence (P 1).
(c) The assumption that G ⊂ Un is crucial. Find an infinite subgroup T of GLn (C), in
which the trace of every element equals n. Thus T satisfies (P 3) though not (P 1).

269. This is about the numerical range, defined in Exercise 21. We recall that the numerical
range H(A) of a matrix A ∈ Mn (C) is a convex compact set, which contains the spectrum
of A.

(a) Let λ ∈ C be given, such that ρ(A) < |λ|. Show that there exists a conjugate P −1 AP
such that λ 6∈ H(P −1 AP ). Hint: Use the HouseholderHouseholder Theorem.
(b) Use the case above to show Hildebrant’sHildebrant Theorem: the intersection of
H(P −1 AP ), as P runs over GLn (C), is precisely the convex hull of the spectrum of
A. Hint: separate this convex hull from an exterior point by a circle.

270. (Suggested by L. BergerBerger!L..) Here is a purely algebraic way to solve the problem
raised in Exercise 267.

(a) Show that det A = 1.

(b) Let p be a prime number. Show that (Ak )o(p,n) ≡ In , mod p, where o(p, n) is the
order of GLn (Z/pZ).
(c) Deduce that A ≡ In , mod p. Conclude.

271. Given H ∈ HPDn (R), prove the formula

πn
Z
∗ Hz
= e−z dz,
det H Cn

153
 where we identify Cn to R2n . Hint: One may split z = x + iy, H = S + iA where
S ∈ SPDn and A is skew-symmetric, and apply Exercise 75 to the integral with respect
to x first, then to the integral with respect to y.
Extend the formula above to non-Hermitian matrices such that H +H ∗ is positive definite.
Hint: Use holomorphy.

272. (R. BellmanBellman.) Let A1 , . . . , Ar ∈ Mn×m (C) be strict contractions, meaning that
A∗j Aj < Im . According to Exercise 221, this implies that for every pair 1 ≤ i, j ≤ r, the
matrix Im − A∗i Aj is non-singular.
The purpose of this exercise is to prove that the Hermitian matrix B whose (i, j) entry is
1
det(Im − A∗i Aj )

is positive definite. This generalizes Loo-Keng Hua’sHua, Loo-Keng inequality

(a) Let z ∈ Cn be given. Show that the matrix of entries z ∗ A∗i Aj z is positive semi-
definite.
(b) With the help of Exercise 21, Chapter 3, prove that the matrix of entries (z ∗ A∗i Aj z)`
is positive semi-definite for every ` ∈ N. Deduce the same property for the matrix
of entries exp(z ∗ A∗i Aj z).
(c) Express the matrix B as an integral, with the help of Exercise 271. Conclude.

273. (See also Exercise 201.) Let A, B, Γ, ∆ ∈ Mn (R) be given. We consider the transformation
over Mn (C) (we warn the reader that we manipulate both fields R and C)

T 7→ T 0 := (A + T Γ)−1 (B + T ∆).

In the sequel, we shall use the 2n × 2n matrix

 
A B
F := .
Γ ∆

In order to ensure that T 0 is defined for every T but the elements of a dense open subset
of Mn (C), we assume a priori that F is non-singular.

(a) We are interested only in those transformations that map symmetric matrices T onto
symmetric matrices T 0 . Show that this is equivalent to the identity
 
T 0n In
F JF = λJ, J := ,
−In 0n

for some λ ∈ R∗ .

154
 (b) Prove the identity

T 0 − T 0 = λ(A + T Γ)−1 (T − T )(A + T Γ)−∗ .

Deduce that if λ > 0, and if the imaginary part of T is positive definite, then the
imaginary part of T 0 is positive definite. Nota: Complex symmetric matrices with
positive definite imaginary form the SiegelSiegel domain Hn .
(c) Let T ∈ Mn (C) be given. Check that
 
In T
X :=
In T
is non-singular if, and only if, the imaginary part of T is non-singular. Write in
closed form its inverse.
(d) From now on, we assume that λ > 0. We are looking for fixed points (T 0 = T ) of
the transformation above, especially whether there exists such a fixed point in the
Siegel domain. Remark that we may assume λ = 1, and therefore F is a symplectic
matrix.
i. Let T ∈ Symn (C) be a fixed point. We define X as above. Show that
 
−1 N 0n
XF X =
0n N

for some N ∈ Mn (C).

ii. We assume now that this fixed point is in the Siegel domain. Find a matrix
K ∈ Mn (C) such that  
K 0n
Y := X
0n K
is symplectic. Show that Y F Y −1 is symplectic and has the form
 
M 0n
.
0n M
iii. Show that M is unitary. Deduce a necessary condition for the existence of a
fixed point in the Siegel domain: the eigenvalues of F have unit modulii.
Nota: Frobenius showed that this existence is equivalent to i) the eigenvalues of
F have unit modulii, ii) F is diagonalizable. The uniqueness of the fixed point
in Hn is much more involved and was solved completely by FrobeniusFrobenius.
Let us mention at least that if F has simple eigenvalues, then uniqueness holds
true. This existence and uniqueness problem was posed by KroneckerKronecker.
See a detailed, albeit non-technical account of this question and its solution, in
T. Hawkins’Hawkins article in Arch. Hist. Exact Sci. (2008) 62:23–57.
274. We are interested in matrices M ∈ Sym3 (R) with mjj = 1 and |mij | ≤ 1 otherwise. In
particular, there exist angles θk such that mij = cos θk whenever {i, j, k} = {1, 2, 3}.

155
 (a) Prove that det M = 0 if, and only if, there exists signs such that

±θ1 ± θ2 ± θ3 ∈ 2πZ.

(b) We give ourselves a non-zero vector x ∈ R3 . We ask whether there exists a matrix
M as above (obviously with det M = 0) such that M x = 0.
i. Prove the necessary condition that |xi | ≤ |xk | + |xj | for every pairwise distinct
i, j, k.
ii. Prove that this condition is also sufficient. Hint: Reduce the problem to the
case where x is non-negative. Then there exists a triangle whose edges have
lengths x1 , x2 , x3 .
275. (F. HollandHolland) Let A1 , . . . , Ar ∈ HPD2 be given. The eigenvalues of Aj are denoted
λ1 (Aj ) ≤ λ2 (Aj ).
Let Q1 , . . . , Qr ∈ U2 be given. Prove the inequality
r
! r
! r
!
X X X
det Q∗j Aj Qj ≥ λ1 (Aj ) λ2 (Aj ) .
j=1 j=1 j=1
Pr
Hint: Use the WeylWeyl Inequalities for the eigenvalues of j=1 Q∗j Aj Qj .

276. (WiegmannWiegmann) Let M be a complex, normal matrix.

(a) If the diagonal entries of M are its eigenvalues (with equal multiplicities), show that
M is diagonal. Hint: compute the FrobeniusFrobenius norm of M .
(b) More generally, consider a block form of M , with diagonal blocks M`` , ` = 1, . . . , r.
Let us assume that the union of the spectra of the diagonal blocks equal the spectrum
of M , with the multiplicities of equal eigenvalues summing up to the multiplicity as
an eigenvalue of M . Prove that M is block-diagonal.

277. Let two matrices A, B ∈ Mn (k) be given. We say that (A, B) enjoys the property L if
the eigenvalues of λA + µB have the form λαj + µβj (j = 1, . . . , n) for some fixed scalars
αj , βj . Necessarily, these scalars are the respective eigenvalues of A and B.

(a) If A and B commute, show that (A, B) enjoys property L.

(b) Let us assume that k has characteristic zero, that A is diagonalizable and that (A, B)
enjoys property L. Up to a conjugation (applied simultaneously to A and B), we
may assume that A is diagonal, of the form diag{a1 Im1 , . . . , ar Imr } with a1 , . . . , ar
pairwise distinct. Let us write B blockwise, with the diagonal blocks B`` of size
m` × m` .
Prove that (MotzkinMotzkin & TausskyTaussky)
r
Y
det(XIn − λA − µB) = det((X − λa` )Im` − µB`` ).
`=1

156
 Hint: Isolate one diagonal block of XIn − λA − µB. Then compute the determinant
with the help of Schur’sSchur complement formula. Then look at its expansion about
the point (a` , 1, 0). One may simplify the analysis by translating a` to zero.
(c) We now assume that k = C, and A and B are normal matrices. If (A, B) enjoys prop-
erty L, prove that A and B commute (WiegmannWiegmann.) Hint: Use Motzkin
& Taussky’ result, plus Exercise 276.

278. (a) Let u1 , . . . , un−1 ∈ k n and x0 , . . . , xn ∈ k n be given. Prove that

n
X
det(u1 , . . . , un−1 , x` ) det X
c` = 0,
`=0

c` denotes the matrix whose columns are x0 , . . . , xn ∈ k n , x` being omitted.

where X
(b) Deduce the following formula for matrices M, N ∈ Mn (k):
n
X
det(M N ) = det MkN det NM
k
,
k=1

where MkN denotes the matrix obtained from M by replacing its last column by the
k-th column of N , and NkM denotes the matrix obtained from N by replacing its
k-th column by the last column of M .
(c) More generally, prove Sylvester’sSylvester Lemma: given 1 ≤ j1 < · · · < jr ≤ n,
then det(M N ) equals the sum of those products det M 0 det N 0 where M 0 is obtained
by exchanging
  r columns of N by the columns of M of indices j1 , . . . , jr . There are
n
choices of the columns of N , and the exchange is made keeping the order
r
between the columns of M , respectively of N .

279. Recall that the HadamardHadamard product of two matrices A, B ∈ Mp×q (k) is the
matrix A ◦ B ∈ Mp×q (k) of entries aij bij with 1 ≤ i ≤ p and 1 ≤ j ≤ q. If A ∈ Mn (k) is
given blockwise  
a11 A12
A= ,
A21 A22
and if a11 is invertible, then the SchurSchur complement A22 −A21 a−1
11 A12 is denoted A|a11
and we have the formula det A = a11 det(A|a11 ).

(a) Let A, B ∈ Mn (k) be given blockwise as above, with a11 , b11 ∈ k ∗ (and therefore
A22 , B22 ∈ Mn−1 (k).) Prove that
1
(A ◦ B)|a11 b11 = A22 ◦ (B|b11 ) + (A|a11 ) ◦ E, E := B21 B12 .
b11

157
 (b) From now on, A and B are positive definite Hermitian matrices. Show that

det(A ◦ B) ≥ a11 b11 det (A22 ◦ (B|b11 )) .

Deduce Oppenheim’sOppenheim (Sir A.) Inequality:

n
!
Y
det(A ◦ B) ≥ aii det B.
i=1

Hint: Argue by induction over n.

(c) In case of equality, prove that B is diagonal.
(d) Verify that Oppenheim’s Inequality is valid when A and B are only positive semi-
definite.
(e) Deduce that
det(A ◦ B) ≥ det A det B.
See Exercise 285 for an improvement of Oppenheim’s Inequality.

280. Let φ : Symm (R) → Symn (R) (or as well φ : Hm → Hn ) be linear. We say that φ is
positive if A ≥ 0m implies φ(A) ≥ 0n , and that it is unital if φ(Im ) = In .

(a) Let φ be positive and unital. If A is positive semi-definite (resp. definite), prove
that  
In φ(A)
≥ 0n
φ(A) φ(A2 )
or, respectively,
φ(A−1 ) In
 
≥ 0n
In φ(A)
Hint: Use a spectral decomposition of A.
(b) Deduce that φ(A)2 ≤ φ(A2 ) or, respectively, φ(A)−1 ≤ φ(A−1 ).

281. (Exercises 281 to 284 are taken from P. HalmosHalmos, Linear algebra. Problem book,
MAA 1995.) Let A, B ∈ Mn (C) be given. We prove here that if A, B and AB are normal,
then BA is normal too.

(a) Let us define C := [B, A∗ A]. Expand C ∗ C and verify that Tr C ∗ C = 0. Deduce that
B commutes with A∗ A.
(b) Let QH be the polar decomposition of A. Recall that, since A is normal, Q and H
commute. Prove that B commutes with H. Hint: H is a polynomial in H 2 .
(c) Deduce the formula Q∗ (AB)Q = BA. Conclude.

282. Let A, B ∈ Mn (R) be unitary similar (in Mn (C)) to each other:

∃ U ∈ Un s.t. AU = U B.

158
 (a) Show that there exists an invertible linear combination S of the real and imaginary
parts of U , such that AS = SB and A∗ S = SB ∗ , simultaneously.
(b) Let QH be the polar decomposition of S. Prove that A and B are actually orthog-
onally similar:
AQ = QB.

283. Let A1 , . . . , Ar ∈ Mn (k) and p1 , . . . , pr ∈ k[X] be given. Prove that there exists a
polynomial p ∈ k[X] such that

p(Aj ) = pj (Aj ), ∀j = 1, . . . , r.

Hint: This is a congruence problem in k[X], similar to the Chinese remainder lemma.

284. Prove that for every matrix A ∈ Mn (R) with n ≥ 2, there exists an invariant plane
Π ⊂ Rn (dim Π = 2 and AΠ ⊂ Π).

285. (S. FallatFallat & C. JohnsonJohnson!Charles R..) Let A and B be n × n, Hermitian

positive semi-definite matrices. According to Exercise 279, both
n
! n
!
Y Y
aii det B and bii det A
i=1 i=1

admit the upper bound

det A ◦ B.
Thanks to the HadamardHadamard inequality, they also have the lower bound det A det B.
In order to find a more accurate inequality than Oppenheim’sOppenheim (Sir A.), as well
as a symmetric one, it is thus interesting to compare
n
! n
!
Y Y
det A ◦ B + det A det B vs aii det B + bii det A.
i=1 i=1

Using induction over the size n, we shall indeed prove that the latter is less than or equal
to the former.

(a) If either n = 2, or B is diagonal, or

 
1 ··· 1
 .. ..  ,
B = . .
1 ··· 1

show that actually

n
! n
!
Y Y
det A ◦ B + det A det B = aii det B + bii det A.
i=1 i=1

159
 (b) We turn to the general case. To begin with, it is enough to prove the inequality for
positive definite matrices A, B.
In the sequel, A and B will thus be positive definite.
(c) We decompose blockwise
a11 x∗ b11 y ∗
   
A= , B= .
x A0 y B0

Let F := e1 eT1 be the matrix whose only non-zero entry is f11 = 1. Prove that
det A
à := A − F
det A0
is positive semi-definite (actually, it is singular). Hint: Use Schur’sSchur comple-
ment formula.
(d) Apply the Oppenheim inequality to estimate det à ◦ B.
(e) Using Exercise 2, deduce the inequality
n
! n
! !
Y det A Y
det A ◦ B ≥ aii det B + b11 det A0 ◦ B 0 − aii det B .
i=1
det A0 i=2

(f) Apply the induction hypothesis. Deduce that

n
! n
!
Y Y
det A ◦ B ≥ aii det B + bii det A − det A det B
i=1 i=1
n
!
det A 0 0
Y
+ (det B − b 11 det B ) det A − aii .
det A0 i=2

Conclude.
286. (S. SedovSedov.) Let x be an indeterminate. For n ≥ 1, let us define the snail matrix
Sn (x) by  
1 x · · · xn−1
x4n−5 · · · xn 
Sn (x) :=  .. ..  ,
 
 . . 
3n−3 2n−2
x ··· ··· x
2 −1
where the powers 1, x, . . . , xn are arranged following an inward spiral, clockwise.
(a) Prove that for n ≥ 3,
2 −9n+12
det Sn (x) = x4n 1 − x4n−6 1 − x4n−10 det Sn−2 (x).

Hint: Make a combination of the first and second rows. Develop. Then make a
combination of the last two rows.

160
 (b) Deduce the formula
n−2
Y
(n−1)(n−2)/2 f (n)
1 − x4k+2 ,


det Sn (x) = (−1) q
k=0

where the exponent is given by

1
f (n) = (2n3 − 6n2 + 13n − 6).
3

287. Let p ≥ 2 be a prime number. We recall that Fp denotes the field Z/pZ. Let A ∈ Mn (Fp )
be given. Prove that A is diagonalizable within Mn (Fp ) if, and only if Ap = A.
Hint: The polynomial X p − X vanishes identically over Fp and its roots are simple.

288. Let R be a commutative ring containing the rationals (was it a field, it would be of
characteristic zero), and let A ∈ Mn (R) be given. Let us assume that Tr(A) = Tr(A2 ) =
· · · = Tr(An ) = 0. Prove that An = 0n .

Hint: Begin with the

case where R is a
field and use the New-
ton’sNewton sums.

Left: Isaac Newton.

Newton,
Isaac!(Paraguay)

289. We present the proof by S. RossetRosset!Shmuel of the Amitsur–LevitzkiAmitsurLevitzki

Theorem (for users of the 2nd edition, it is the object of Section 4.4).
We need the concept of exterior algebra (see Exercise 146). Let {e1 , . . . , e2n } be a basis
of k 2n . Then the monomials ej1 ∧ · · · ∧ ejr with j1 < · · · < jr (the sequence may be
empty) form a basis of the exterior algebra Λ(k 2n ). This is an associative algebra, with
the property that for two vectors e, f ∈ k 2n , one has e ∧ f = −f ∧ e.
We denote by R the subalgebra spanned by the 2-forms ei ∧ ej .

(a) Check that R is a commutative sub-algebra.

(b) If A1 , . . . , A2n ∈ Mn (k), let us define

A := A1 e1 + · · · + A2n e2n ∈ Mn (Λ(k 2n )).

161
 i. Show that for every ` ≥ 1,
X
A` = S` (Ai1 , . . . , Ai` )ei1 ∧ · · · ∧ ei` ,
i1 <···<i`

where the standard polynomial S` in non-commutative indeterminates X1 , . . . , X`

is defined by X
S` (X1 , . . . , X` ) := (σ)Xσ(1) · · · Xσ(`) .
σ

Hereabove, the sum runs over the permutations of {1, . . . , `}, and (σ) denotes
the signature of σ.
ii. When ` is even, show that Tr S` (B1 , . . . , B` ) = 0, for every B1 , . . . , B` ∈ Mn (k).
iii. If k has characteristic zero, deduce that A2n = 0n . Hint: Use the previous
exercise.
(c) Whence the Theorem of Amitsur & Levitzki: for every A1 , . . . , A2n ∈ Mn (k), one
has
S2n (A1 , . . . , A2n ) = 0n .
Hint: First assume that k has characteristic zero. Then use the fact that S2n is a
polynomial with integer coefficients.
(d) Prove that S2n−1 (A1 , . . . , A2n−1 ) does not vanish identically over Mn (k). Hint:
Specialize with the matrices E 11 , E 12 , . . . , E 1n , E 21 , . . . , E n1 , where E m` is the matrix
whose (i, j)-entry is one if i = m and j = `, and zero otherwise.

290. Prove the following formula for complex matrices:

∞
X (−1)k+1
log det(In + zA) = Tr(Ak )z k .
k=0
k

Hint: Use an analogous formula for log(1 + az).

291. Let A, B ∈ Hn be such that

det(In + xA + yB) ≡ det(In + xA) det(In + yB), ∀x, y ∈ R.

(a) Show that for every k ∈ N,

Tr((xA + yB)k ) ≡ xk Tr(Ak ) + y k Tr(B k ).

Hint: Use Exercise 290.

(b) Infer
2 Tr(ABAB) + 4 Tr(A2 B 2 ) = 0.
(c) Deduce that AB = 0. This is the theorem of CraigCraig & SakamotoSakamoto.
Hint: Set X := AB. Use the fact that (X + X ∗ )2 + 2X ∗ X is semi-positive definite.

162
292. We recall that the function H 7→ θ(H) := − log det H is convex over HPDn . We extend
θ to the whole of Hn by posing θ(H) = +∞ otherwise. This extension preserves the
convexity of θ. We wish to compute the LegendreLegendre transform

θ∗ (K) := sup {Tr(HK) − θ(H)}.

H∈Hn

(a) Check that θ∗ (U ∗ KU ) = θ∗ (K) for every unitary U .

(b) Show that
θ∗ (K) ≥ −n − log det(−K).
In particular, θ∗ (K) is infinite, unless K is negative definite. Hint: Use diagonal
matrices only.
(c) Show that, for every positive definite Hermitian matrices H and H 0 , one has

log det H + log det H 0 + n ≤ Tr(HH 0 ).

Hint: HH 0 is diagonalizable with positive real eigenvalues.

(d) Conclude that θ∗ (K) ≡ −n + θ(−K).
(e) Let χ(H) := θ(H) − n/2. Verify that χ∗ (H) = χ(−H). Do you know any other
convex function on a real space having this property ?

293. Let A ∈ Hn be given. Show, by an explicit construction, that the set of matrices H ∈ Hn
satisfying H ≥ 0n and H ≥ A admits a least element, denoted by A+ :

∀H ∈ Hn , (H ≥ 0n and H ≥ A) ⇐⇒ (H ≥ A+ ).

Let B be an other Hermitian matrix. Deduce that the set of matrices H ∈ Hn satisfying
H ≥ B and H ≥ A admits a least element. We denote it by A ∨ B and call it the
supremum of A and B. We define the infimum by A ∧ B := −((−A) ∨ (−B)).
Prove that
A ∨ B + A ∧ B = A + B.

294. Let J ∈ M2n (R) be the standard skew-symmetric matrix:

 
0n −In
J= .
In 0n

Let S ∈ Sym2n (R) be given. We assume that dim ker S = 1.

(a) Show that 0 is an eigenvalue of JS, geometrically simple, but not algebraically.
(b) We assume moreover that there exists a vector x 6= 0 in R2n such that the quadratic
form y 7→ y T Sy, restricted to {x, Jx}⊥ is positive definite. Prove that the eigenvalues
of JS are purely imaginary. Hint: Use Exercise 258 in an appropriate way.

163
 (c) In the previous question, S has a zero eigenvalue and may have a negative one. On
the contrary, assume that S has one negative eigenvalue and is invertible. Show that
JS has a pair of real, opposite eigenvalues. Hint: What is the sign of det(JS) ?
295. Let A ∈ Mn (C) be given. Denoting its minimal polynomial by πA , let us define a
differential operator  
d
LA := πA .
dt
The degree of πA is denoted by r.
(a) Prove that there exists functions fj (t) for j = 0, . . . , r − 1, such that
exp(tA) = f0 (t)In + f1 (t)A + · · · + fr−1 (t)Ar−1 , ∀t ∈ R.

(b) Prove that t 7→ fj (t) is C ∞ . Hint: This requires proving a uniqueness property.
(c) Applying LA to the above identity, show that these functions satisfy the differential
equation
LA fj = 0.
Deduce that X
fj = etλα pjα (t),
where the λα ’s are the distinct eigenvalues and pjα are polynomials.
(d) Determine the initial conditions for each of the fj ’s. Hint: Use the series defining
the exponential.
296. (After KrishnapurKrishnapur.) Let A ∈ Mn×m (C) be given. We assume that m ≤ n and
denote the columns of A by C1 , . . . , Cm , that is Cj := Aej with {e1 , . . . , em } the canonical
basis of Cm . Let also σ1 , . . . , σm be the singular values of A.
Besides, we define Hi the subspace of Cn spanned the Cj ’s for j 6= i. At last, we denote
by di the distance of Ci to Hi .
(a) To begin with, we restrict to the case where A has full rank: rk A = m. Check that
A∗ A is non-singular.
(b) Let us define the vector Vj := (A∗ A)−1 ej . Show that AVj · Ci = δij for all i, j =
1, . . . , m.
(c) Let us decompose Vj = vj1 e1 + · · · + vjn en . Show that AVj is orthogonal to Hj and
that its norm equals vjj dj .
(d) Deduce the identity vjj d2j = 1.
(e) Prove the Negative second moment identity
X X
σj−2 = d−2
j .
j j

Hint: Compute the trace of (A∗ A)−1 .

164
 (f) What do you think of the case where rk A < m ?
297. Here is another proof of Fuglede’sFuglede Theorem (see Exercise 255), due to von Neu-
mannvonneu@von Neumann.

(a) Given A, B, C, D ∈ Mn (C), check

tity

Tr([A, B][C, D]+[A, C][D, B]+[A,

(b) Deduce that if [A, B] = 0n and if e

B is normal, then [A, B ∗ ] = 0n .
vonneu@von
Neumann, John!(Hungary)

John von Neumann.

298. (After J. von Neumannvonneu@von Neumann.) An embedding from Mm (C) to Mn (C) is

an algebra homomorphism f with the additional properties that f (Im ) = In and f (A∗ ) =
f (A)∗ for every A ∈ Mm (C).
(a) Prove that f sends GLm (C) into GLn (C).
(b) Deduce that the spectrum of f (A) equals that of A.
(c) Likewise, prove that the minimal polynomial of f (A) divides that of A. In particular,
if A is semi-simple, then so is f (A).
(d) If f (A) = 0n and A is Hermitian, deduce that A = 0m . Use this property to prove
that if f (M ) = 0n , then M = 0m (injectivity). Deduce that m ≤ n.
(e) Let P ∈ Mm (C) be a unitary projector: P 2 = P = P ∗ . Prove that f (P ) is a unitary
projector.
(f) Let x ∈ Cm be a unit vector. Thus f (xx∗ ) is a unitary projector, whose rank is
denoted by k(x). Prove that k is independent of x. Hint: An embedding preserves
conjugacy.
(g) Show that m divides n. Hint: Decompose Im as the sum of unitary projectors xx∗ .
(h) Conversely, let us assume that n = mp. Prove that there exists an embedding from
Mm (C) to Mn (C). Hint: Use the tensor product of matrices.
299. Let A ∈ Mn (C) be a nilpotent matrix of order two: A2 = 0n . This exercise uses the
standard operator norm k · k2 .
(a) Using standard properties of this norm, verify that kM k22 ≤ kM M ∗ + M ∗ M k2 for
every M ∈ Mn (C).
(b) When k is a positive integer, compute (AA∗ + A∗ A)k in close form. Deduce that
kAA∗ + A∗ Ak2 ≤ 21/k kAk22 .

165
 (c) Passing to the limit as k → +∞, prove that
1/2
(33) kAk2 = kAA∗ + A∗ Ak2 .

300. Let T ∈ Mn (C) be a contraction in the operator norm: kT k2 ≤ 1. Let us form the matrix
S ∈ HnN :
In T ∗ · · · T ∗N
 
... ... .. 
T . 

S :=  . ... ... .
 .. T∗ 
TN ··· T In
We also define the matrix R ∈ MnN (C) by
 
In 0n · · · · · · 0n
... .. 
T .


R := 0n . . . . .
.

. . . .

. . .
 .. . . . . . . . . .. 


0n · · · 0n T In

Express S in terms of R. If (In − R)y = x, show that

x∗ Sx = kyk22 − kRyk22 .

Deduce that S is positive semi-definite.

301. Let H1 , . . . , Hr ∈ Hn be positive semi-definite matrices such that

r
X
Hj = In .
j=1

Let also z1 , . . . , zr ∈ C be given in the unit disc: |zj | ≤ 1.

Prove that
r
X
zj Hj ≤ 1.
j=1 2
Hint: Factorize
Pr 
j=1 zj Hj 0n · · ·

 0n 0n · · · ∗
 = M diag{z1 In , . . . , zr In }M
.. ..
. .

with M appropriate.

166
302. (Sz. NagySz. Nagy.) If M ∈ Mn (C) is similar to a unitary matrix, show that
(34) sup kM k k2 < ∞.
k∈Z

Conversely, let M ∈ GLn (C) satisfy (34). Show that the eigenvalues of M belong to the
unit circle and are semi-simple. Deduce that M is similar to a unitary matrix.
303. Let M1 , . . . , Mr ∈ Mn (C) satisfy the anticommutation relations
Mi Mj + Mj Mi = 0n , Mi∗ Mj + Mj∗ Mi = δij In , ∀1 ≤ i, j ≤ r.
Let us define r
X
M (z) = zj Mj , ∀z ∈ Rr .
j=1

Verify that M (z) is nilpotent of order 2. Prove the identity

kM (z)k2 = kzk2 .
Hint: Use Exercise 299.
304. (S. Joshi & S. BoydJoshiBoyd.) Given A ∈ Mn (C) and a (non trivial) linear subspace2
V of Cn , we define
G(A|V ) := sup {kAxk | x ∈ V, kxk = 1} , H(A|V ) := inf {kAxk | x ∈ V, kxk = 1} ,
where we use the canonical Hermitian norm. Obviously, we have H(A|V ) ≤ G(A|V ),
thus
G(A|V )
κV (A) := ≥ 1.
H(A|V )
Notice that when A is singular, κV (A) may be infinite.
Given an integer k = 1, . . . , n, we wish to compute the number
θk (A) := inf {κV (A) | dim V = k} .
(a) What is κV (A) when V = Cn ?
(b) Show that θk is unitary invariant: if B = U AV with U and V unitary, then θk (B) =
θk (A). Deduce that θk (A) depends only upon the singular values of A.
(c) From now on, we assume that A is non-singular. From the previous question, we
may assume that A is diagonal and positive. We denote (0 <)σn ≤ · · · ≤ σ1 its
singular values, here its diagonal entries. Show first that
G(A|V ) ≥ σn−k+1 , H(A|V ) ≤ σk .
Deduce that
 
σn−k+1
(35) θk (A) ≥ max ,1 .
σk
2
This exercise is unchanged when replacing the scalar field by R.

167
 (d) We wish to show that (35) is actually an equality. Let {~e1 , . . . , ~en } be the canonical
basis of Cn . Show that given i ≤ j and σ ∈ [σi , σj ], there exists a unit vector
y ∈ Span(~ei , ~ej ) such that kAyk = σ.
(e) Let N be least integer larger than or equal to n+1 2
, that is the integral part of 1+n/2.
Show that each plane Span(~eN −` , ~eN +` ) with ` = 1, . . . , n − N contains a unit vector
y` such that kAy` k = σN . Deduce that θn−N +1 (A) = 1. In other words, θk (A) = 1
for every k ≤ n − N + 1.
Nota: This can be recast as follows. Every ellipsoid of dimension n − 1 contains a
sphere of dimension d, the largest integer strictly less than n/2.
(f) Assume on the contrary that k > n−N +1. Using the same trick as above, construct
a linear space W of dimension k such that κW (A) = σn−k+1 /σk . Conclude.
305. Among the class of HessenbergHessenberg matrices, we distinguish the unit ones, which
have 1’s below the diagonal:
 
∗ ··· ··· ∗
 .. .. 
1 . .
 . 
M = 0 . .
 .

. . . . .
. . . . .

. . . . .
0 ··· 0 1 ∗

(a) Let M ∈ Mn (k) be a unit Hessenberg matrix. We denote by Mk the submatrix

obtained by retaining the first k rows and columns. For instance, Mn = M and
M1 = (m11 ). We set Pk the characteristic polynomial of Mk .
Show that (B. KostantKostant & N. WallachWallach, B. ParlettParlett & G. StrangStrang)

Pn (X) = (X − mnn )Pn−1 (X) − mn−1,n Pn−2 (X) − · · · − m2n P1 (X) − m1n .

(b) Let Q1 , . . . , Qn ∈ k[X] be monic polynomials, with deg Qk = k. Show that there
exists one and only one unit Hessenberg matrix M such that, for every k = 1, . . . , n,
the characteristic polynomial of Mk equals Qk . Hint: Argue by induction over n.
Nota: The list of roots of the polynomials P1 , . . . , Pn are called the Ritz valuesRitz of
M.
306. (Partial converse of Exercise 21.) Let A and B be 2 × 2 complex matrices, which have
the same spectrum. We assume in addition that

det[A∗ , A] = det[B ∗ , B].

Prove that A and B are unitarily similar. Hint : Prove that they both are unitarily
similar to the same triangular matrix.
Deduce that two matrices in M2 (C) are unitary similar if, and only if they have the same
numerical range.

168
307. Let P ∈ Mn (C) be a projection. Let us define

Π := P (P ∗ P + (In − P )∗ (In − P ))−1 P ∗ .

Prove that Π vanishes over (R(P ))⊥ . Then prove that Πx = x for every x ∈ R(P ). In
other words, Π is the orthogonal projection onto R(P ). Hint: If x ∈ R(P ), then x = P x,
thus you can find the solution of

(P ∗ P + (In − P )∗ (In − P ))y = P ∗ x.

308. Compute the conjugate of the convex function over Hn


−(det H)1/n , if H ≥ 0n ,
H 7→
+∞, otherwise.

Hint: Use Exercise 209. Remark that the conjugate of a positively homogeneous function
of degree one is the characteristic function of some convex set.

309. (a) Let A ∈ Mn (C) and λ ∈ C be given. Show that if λ is not in the numerical range
W (A), then λIn − A is invertible, and its inverse verifies
1
k(λIn − A)−1 k2 ≤ .
dist(λ, W (A))

(b) Conversely, let us consider compact subsets X of the complex plane, such that
1
(36) k(λIn − A)−1 k2 ≤ , ∀ λ 6∈ X.
dist(λ, X)

Obviously, such an X is contained in the resolvant set C \ Sp(A). If  > 0 tends to

zero, show that

kIn + Ak2 = 1 +  sup{<z | z ∈ W (A)} + O(2 ).

Deduce that
sup{<z | z ∈ W (A)} ≤ sup{<z | z ∈ X}.
Finally, prove that W (A) is the convex hull of X. In particular, if A = J(2; 0) is
a 2 × 2 Jordan block, then X contains the circle C1/2 . When A is normal, Sp(A)
satisfies (36).

310. Let k be a field of characteristic zero. We consider a matrix A ∈ Mn (k). If X ∈ Mn (k),

we define the linear form and the linear map

τX (M ) := Tr(XM ), adX (M ) = XM − M X.

(a) We assume that A is nilpotent.

169
 i. If AM = M A, show that AM is nilpotent.
ii. Verify that ker adA ⊂ ker τA .
iii. Deduce that there exists a matrix B such that τA = τB ◦ adA .
iv. Show that A = BA − AB.
(b) Conversely, we assume instead that there exists a B ∈ Mn (k) such that A = BA −
AB.
i. Verify that for k ∈ N, BAk − Ak B = kAk .
ii. Deduce that A is nilpotent. Hint: adB has finitely many eigenvalues.
(c) If A = J(0; n) is the nilpotent JordanJordan!Camille block of order n, find a diagonal
B such that [B, J] = J.
311. (From E. A. HermanHerman.) Let A ∈ Mn (C) be such that A2 = 0n . We denote its
rank, kernel and range by r, N and R.
(a) By computing k(A + A∗ )xk22 , show that ker(A + A∗ ) = N ∩ R⊥ .
(b) Verify that R ⊂ N and N ⊥ ⊂ R⊥ . Deduce that N ∩ R⊥ is of dimension n − 2r and
that the rank of 21 (A + A∗ ) (the real part of A) is 2r.
(c) Show that N is an isotropic subspace for the Hermitian form x 7→ x∗ (A + A∗ )x,
contained in R(A + A∗ ). Deduce that the number of positive / negative eigenvalues
of A + A∗ are both equal to r.
(d) Example: Take n = 2r and
 
0r B
A := , B ∈ GLr (C).
0r 0r
∗
Find the equations
√ of the stable and the unstable subspaces of A + A . Hint: the
formula involves BB ∗ .
312. Let E be a hyperplane in Hn . We use the scalar product hA, Bi := Tr(AB). Prove the
equivalence of the following properties.
• Every nonzero matrix K ∈ E has at least one positive and one negative eigenvalues.
• E ⊥ is spanned by a positive definite matrix.
313. (After von Neumannvonneu@von Neumann, HalperinHalperin, AronszajnAronszajn, Kay-
alarKayalar & WeinertWeinert.) We equip Cn with the standard scalar product, and
Mn (C) with the induced norm. Let M1 and M2 be two linear subspaces, and P1 , P2
the corresponding orthogonal projections. We recall that Pj∗ = Pj = Pj2 . We denote
M := M1 ∩ M2 and P the orthogonal projection onto M . Finally, we set Q := In − P
and Ωj := Pj Q. Our goal is two prove
lim (P2 P1 )m = P
m→+∞

(von Neumann–Halperin Theorem), and to give a precise error bound.

170
 (a) Show that Pj P = P and P Pj = P .
(b) Deduce that Ωj = Pj − P . Verify that Ωj is an orthogonal projection.
(c) Show that (Ω2 Ω1 )m = (P2 P1 )m − P .
(d) Deduce that

k(P2 P1 )m − P k2 = k(Ω1 Ω2 Ω1 )2n−1 k = kΩ1 Ω2 Ω1 k2n−1 .

(e) Let x ∈ Cn be given. If Ω1 x = Ω2 x = x, show that x = 0. Deduce that kΩ1 Ω2 Ω1 k < 1

and establish the von Neumann–Halperin Theorem.

314. A HadamardHadamard matrix is a matrix M ∈

Mn (Z) whose entries are ±1s and such that
M T M = nIn . The latter means that √1n M is or-
thogonal.
We construct inductively a sequence of matrices
Hm ∈ M2m (Z) by
 
Hm −Hm
H0 = (1), Hm+1 = .
Hm Hm
Hadamard,
Jacques (Rép. de Guinée) Verify that each Hm is a Hadamard matrix.
Nota: It is unknown whether there exists or not
Jacques Hadamard. a Hadamard whose size is not a power of 2.

315. (M. RossetRosset!Myriam and S. RossetRosset!Shmuel.) Let A be a Principal Ideal Do-

main. We define the vector product X × Y in A3 with the same formulæ as in R3 . For
instance, the first coordinate is x2 y3 − x3 y2 .

(a) We admit for a minute that the map φ : (X, Y ) 7→ X × Y is surjective from A3 × A3
into A3 . Prove that every matrix M ∈ M2 (A) with zero trace is a commutator
BC − CB with B, C ∈ M2 (A).
 
a
(b) We now prove the surjectivity of φ. Let Z = b  ∈ A3 be given. Show that there

  c
x
exists X = y  such that ax + by + cz = 0 and gcd{x, y, z} = 1. Then Bézout gives

c 

u
a vector U = v  such that ux + vy + wz = 1. Set Y = Z × U . Then Z = X × Y .

w

316. Let k be a field, U Tn (k) be the set of upper triangular matrices with 1s along the diagonal.
It is called the unipotent group.

171
 (a) Prove that U Tn (k) is a subgroup of GLn (k).
(b) If G is a group, D(G) is the group generated by the commutators xyx−1 y −1 . It is
a normal subgroup of G. Show that D(U Tn (k)) consists of the matrices such that
mi,i+1 = 0 for every i = 1, . . . , n − 1.
(c) Let G0 = Un (k) and G1 = D(U Tn (k)). We define Gr by induction; Gr+1 is the group
generated by the commutators where x ∈ G0 and y ∈ Gr . Describe Gr and verify
that Gn = {In }. One says that the group U Tn (k) is a nilpotent.

317. (Follow-up of the previous exercise. Thanks to W. ThurstonThurston.) We take k = Z/3Z

and n = 3.

(a) Show that every element M 6= I3 of U T3 (Z/3Z) is of order 3.

(b) What is the order of U T3 (Z/3Z) ?
(c) Find an abelian group of order 27, with 26 elements of order 3.
(d) Deduce that the number of elements of each order does not characterizes a group in
a unique manner.

318. Let p ≥ 3 be a prime number. M ∈ GLn (Z) be of finite order (M r = In for some r ≥ 1),
and such that M ≡ In mod p.

(a) Prove that M = In . Hint: If M 6= In , write M = In + pα A with A not divisible b

p. Likewise, r = pβ ` with β ≥ 0 and ` not divisble by p. Verify that for each k ≥ 2,
 
kα r
p
k

is divisible by pα+β+1 . Deduce that M r ≡ In + `pα+β A mod pα+β+1 . Conclude.

(b) Let G be a finite subgroup of GLn (Z). Deduce that the reduction mod p is injective
over G. Yet, this reduction is not injective over GLn (Z).
(c) The result is false if p = 2: find a matrix In + 2A of finite order, with A 6= 0n .

319. (After C. S. BallantineBallantine.) We prove that every Hermitian matrix H with strictly
positive trace can be written as H = AB + BA with A, B ∈ HPDn .

(a) We first treat the case where the diagonal entries hjj are strictly positive. Prove that
such a pair (A, B) exists with A diagonal. Hint: Induction over n. Choose an > 0
small enough.
(b) Conclude, with the help of Exercise 131.
(c) Conversely, if A, B ∈ HPDn are given, prove that Tr(AB + BA) > 0.

320. Let S be a finite set of cardinal n. Let E1 , . . . , Em be subsets of S with the properties
that

172
 • every Ej has an odd cardinal,
• for every i 6= j, Ei ∩ Ej has an even cardinal.

(a) Let us form the matrix A with n columns (indexed by the elements of S) and m
rows, whose entry ajx equals 1 if x ∈ Ej , and 0 otherwise. Prove that AAT = Im .
Hint: Yes, Im ! Leave some room to your imagination.
(b) Deduce m ≤ n.

321. Let A, B ∈ Mn (C) satisfy the relation [A, B] = A. Let us define the following matrix-
valued functions of t ∈ R:

X(t) := Be−tA , Y (t) = Ae−tB , Z(t) = et(A+B) e−tA e−tB .

(a) Find a differential equation for X. Deduce that

[B, e−tA ] = tAe−tA .

(b) Find a differential equation for Y . Deduce that

Ae−tB = e−t e−tB A.

(c) Find a differential equation for Z. Deduce that

et(A+B) = eτ (t)A etB etA , τ (t) := 1 − (t + 1)e−t .

Remarks.
• From Exercise 256, we know that A is nilpotent.
• This result can be used in order to establish an explicit formula for the semi-
group generated by the Fokker–PlanckFokkerPlanck equation ∂t f = ∆v f + v ·
∇v f .
322. Let A ∈ Mn (R). We denote 1 the vector whose coordinates are ones. Prove the equiva-
lence of the following properties.
• The semi-group (Mt := etA )t≥0 is MarkovianMarkov, meaning that Mt is stochastic
(Mt ≥ 0n and Mt 1 = 1).
• The off-diagonal entries of A are non-negative, and A1 = 0.
• A1 = 0, and for every X ∈ Rn , we have

A(X ◦ X) ≥ 2X ◦ (AX),

where we use the HadamardHadamard product, (X ◦ Y )j = xj yj .

323. (S. BoydBoyd & L. VandenbergheVandenberghe). The spectral radius of a non-negative
matrix A ∈ Mn (R) is an eigenvalue (Perron–FrobeniusPerronFrobenius Theorem), which
we denote λpf (A).

173
 (a) When A > 0n , prove that
1
log λpf (A) = lim log 1T Ak 1,
k→+∞ k
where 1 is the vector whose entries are all ones.
p
(b) Let A, B be positive matrices and let us define C by cij := aij bij (the square root,
in the sense of Hadamard product).
i. Show that for every m ≥ 1,
q
(C m )ij ≤ (Am )ij (B m )ij .
p
ii. Deduce that λpf (C) ≤ λpf (A)λpf (B) .
(c) More generally, show that log λpf (A) is a convex function of the variables log aij .
Nota: This is not what is usually called log-convexity.

324. Given B ∈ Mn (C), assume that the only subspaces F invariant under both B and B ∗
(that is BF ⊂ F and B ∗ F ⊂ F ) are Cn and {0}. Let us denote by L the sub-algebra of
Mn (C) spanned by B and B ∗ , i.e. the smallest algebra containing B and B ∗ .

(a) Verify that ker B ∩ ker B ∗ = {0}.

(b) Construct a matrix H ∈ L that is Hermitian positive definite.
(c) Deduce that In ∈ L : L is unital.
(d) We show now that L does not admit a proper two-sided ideal (we say that the unital
algebra L is simple). So let J be a two-sided ideal of L.
i. We define \
K= (ker M ∩ ker M ∗ ).
M ∈J

Show that there exist finitely many elements Mj ∈ J such that

\ X
K = (ker Mj ∩ ker Mj∗ ) = ker (Mj Mj∗ + Mj∗ Mj ).
j j

ii. Verify that BK ⊂ K and B ∗ K ⊂ K.

iii. Deduce that either J = (0n ), or J contains an invertible element.
iv. Conclude.

Comment. This can be used to prove the following statement, which interpolates the
unitary diagonalization of normal matrices and the AmitsurAmitsur–LevitskiLevitski the-
orem that the standard noncommutative polynomial S2n vanishes over Mn (k).

174
 Let us say that a matrix A ∈ Mn (C) is r-normal if the standard polynomial in 2r
non-commuting variables S2r vanishes identically over the sub-algebra spanned
by A and A∗ . In particular, every A is n-normal, whereas A is 1-normal if and
only if it is normal.
Then A is r-normal if and only if there exists a unitary matrix U such that
U ∗ AU is block-diagonal with diagonal blocks of size m × m with m ≤ r.

The reader will prove easily that a matrix that is unitarily similar to such a block-diagonal
matrix is r-normal.

325. We consider here the linear equation AM = M B in M ∈ Mn×m (k), where A ∈ Mn (k)
and B ∈ Mm (k) are given. The solution set is a vector space denoted S(A, B).

(a) If R ∈ k[X], verify that R(A)M = M R(B). If the spectra of A and B are disjoint,
deduce that M = 0.
(b) When A = J(0; n) and B = J(0; m), compute the solutions, and verify that the
dimension of S(A, B) is min{m, n}.
(c) If A is conjugate to A0 and B conjugate to B 0 , prove that S(A0 , B 0 ) is obtained from
S(A, B) by applying an equivalence. In particular, their dimensions are equal.
(d) In this question, we assume that k is algebraically closed
i. Let {λ1 , . . . , λ` } be the union of the spectra of A and B. If i ≤ `, we denote
(X − λi )αij the elementary divisors of A, and (X − λi )βik those of B. Using a
canonical form A0 and B 0 , prove that the dimension of S(A, B) equals the sum
of the numbers X
Ni := min{αij , βik }.
j,k

ii. Deduce that the dimension of the solution set of the matrix equation AM = M B
equals X
deg[g.c.d.(pi , qj )],
where p1 , . . . , pn are the invariant factors of A and q1 , . . . , qm are those of B.
(e) Show that the result above persists when k is an arbitrary field, not necessarily al-
gebraically closed. This is the Cecioni–FrobeniusCecioniFrobenius Theorem. Hint:
S(A, B) is defined by a linear system in k nm . Its dimension remains the same when
one replaces k by a field K containing k.

326. A symmetric matrix S ∈ Symn (R) is said compatible if it is of the form abT + baT with
a, b ∈ Rn . Prove that S ∈ Symn (R) is compatible if and only if

• either S = 0n ,
• or S is rank-one and non-negative,
• or S has rank two and its non-zero eigenvalues have opposite signs.

175
327. (After R. A. HornHorn!Roger and C. R. JohnsonJohnson!Charles R. (I).) Let A ∈ Mn (C)
be given.

(a) Suppose that A is similar to a matrix B ∈ Mn (R). Prove that A is similar to A∗ .

Hint: B T is similar to B.
(b) Conversely, we assume that A is similar to A∗ .
i. Verify that A is similar to A. Deduce that the spectrum of A is invariant under
conjugation, and that when λ is a non-real eigenvalue, the JordanJordan!Camille
blocks corresponding to λ̄ have the same sizes as those corresponding to λ.
ii. Deduce that A is similar to a matrix B ∈ Mn (R).

328. (After R. A. HornHorn!Roger and C. R. JohnsonJohnson!Charles R. (II).) Let A ∈ Mn (C)

be given. We assume that A is similar to A∗ : A = T −1 A∗ T .

(a) Show that there exists θ ∈ R such that eiθ T + e−iθ T ∗ is non-singular.
(b) Deduce that there exists S ∈ Hn ∩ GLn (C) such that A = S −1 A∗ S.

329. (After R. A. HornHorn!Roger and C. R. JohnsonJohnson!Charles R. (III).) Let A ∈

Mn (C) be given.

(a) We assume that A is similar to A∗ via a Hermitian transformation: A = S −1 A∗ S

with S ∈ Hn . Verify that A∗ S is Hermitian. Deduce that A has the form HK where
H and K are Hermitian, one of them being non-singular.
(b) Conversely, assume that A has the form HK where H and K are Hermitian, one of
them (say H for definiteness) being non-singular. Verify that A is similar to A∗ .

So far, we have shown that A is similar to a matrix B ∈ Mn (R) if and only if it is of the
form HK where H and K are Hermitian, one of them being non-singular.

330. (After R. A. HornHorn!Roger and C. R. JohnsonJohnson!Charles R. (IV).) We now as-

sume that A = HK where H and K are Hermitian. We warn the reader that we allow
both H and K to be singular.
 0 
H 0
(a) To begin with, we assume that H = where H 0 ∈ Hp is non-singular. Let
 0  0 0
K ∗
be the block form of K matching that of H so that
∗ ∗
 0 0 
HK ∗
A= .
0 0

i. Let λ be a non-real eigenvalue of A. Show that λ is an eigenvalue of A0 := H 0 K 0

and that the Jordan blocks corresponding to λ in A or in A0 are the same.

176
 ii. Deduce that the Jordan blocks of A corresponding to λ̄ have the same sizes
as those corresponding to λ. Hint: According to the previous exercises, A0 is
similar to a matrix B 0 ∈ Mp (R).
iii. Deduce that A is similar to a matrix B ∈ Mn (R).
(b) Prove the same result in the general case. Hint: Diagonalize H.
Summary: The following properties are equivalent to each other for every A ∈ Mn (C).
• A is similar to a matrix B ∈ Mn (R),
• A is similar to A∗ ,
• A is similar to A∗ via a Hermitian transformation,
• There exist H, K ∈ Hn such that A = HK and one of them is non singular,
• There exist H, K ∈ Hn such that A = HK.
331. If A ∈ Mn (R) and x ∈ (0, +∞), verify that det(xIn + A2 ) ≥ 0. Deduce that if n is odd,
then −In cannot be written as A2 + B 2 with A, B ∈ Mn (R). Note: On the contrary, if
n is even, then every matrix M ∈ Mn (R) can be written as A2 + B 2 with A, B ∈ Mn (R).
332. If σ ∈ Sn , we denote by Pσ the permutation matrix associated with σ. A finite sum
of permutation matrices is obviously a matrix M ∈ Mn (N), whose sums of rows and
columns are equal. We shall prove the converse statement: If M ∈ Mn (N) has equal sum
S for rows and columns, then M is a finite sum of permutation matrices.
Let I, J be two sets of indices 1 ≤ i, j ≤ n. If the bloc MIJ is identically 0, prove that
the sum of the entries of the opposite bloc MI c J c equals (n − p − q)S. If S ≥ 1, deduce
that p + q ≤ n.
Let us recall (Exercise 9) that this property implies that there exists a permutation σ
such that miσ(i) 6= 0 for every 1 ≤ i ≤ n. Then argue by induction over S.
333. Let M ∈ Mn (R) be a given non-negative matrix. If σ ∈ Sn , let us denote
n
X
σ
m := miσ(i) .
i=1

Finally, we define
S := max mσ .
σ∈Sn

We assume that for every entry mij , there exist σ ∈ Sn such that σ(i) = j and mσ = S.
(a) Find a positive linear form f over Mn (R) such that mσ = f (Pσ ◦ M ) for every
σ ∈ Sn , where A ◦ B is the Hadamard product.
(b) Rephrase the assumption in the following way: There exists a subset X of Sn such
that mσ = S for every σ ∈ X, and
X
Q := Pσ > 0n .
σ∈X

177
 (c) Let θ ∈ Sn be given.
i. Verify that Q − Pθ is a sum of k − 1 permutation matrices, with k := |X|. Hint:
Use exercise 332.
ii. Deduce that mθ ≥ S, and therefore = S.

334. Recall that a numerical function f : (a, b) → R is operator monotone if whenever n ≥ 1,

A, B ∈ Hn are given with the spectra of A and B included in (a, b), we have

(A ≤ B) =⇒ (f (A) ≤ f (B)),

where H ≤ K is understood in the sense of Hermitian forms. More generally, f is

operator monotone of grade n if this holds true at fixed size n × n. We prove here the
Loewner’sLoewner Theorem, under some regularity assumption.

(a) Check that if m ≤ n, then operator monotonicity of grade n implies operator mono-
tonicity of grade m.
(b) We already know that if both H and K are ≥ 0n , their HadamardHadamard product
H ◦K is ≥ 0n too. Here is a converse: Let K ∈ Hn be given. If K ◦H ≥ 0n whenever
H ∈ Hn is ≥ 0n , prove that K ≥ 0n .
(c) We assume that f ∈ C(a, b) and recall the assumptions of Exercise 250: If D =
diag(d1 , . . . , dn } ∈ Mn (C), we define a matrix f [1] (D) ∈ Mn (C) by

f (dj ) − f (dk )
f [1] (D)jk = ,
dj − dk

where we identify
f (b) − f (a)
:= f 0 (a),
b−a
if b = a.
If f is operotor monotone of grade n, prove that f [1] (D) ≥ 0n whenever d1 , . . . , dn ∈
(a, b). Hint: Use Daletskiı̆–KreinDaletskiı̆Krein formula.
(d) If n ≥ 2, deduce that either f is constant, or it is strictly increasing.
(e) We assume that f ∈ C 3 (a, b) and n ≥ 2. We may take n = 2. Compute P T f [1] (D)P
when  1 
1 − d2 −d
P = 1
1
.
0 d2 −d 1

Deduce that 2f 0 f 000 ≥ 3f 002 . In other words, either f is constant or (f 0 )−1/2 is concave.

335. The relative gain array of a square matrix A ∈ GLn (k) is defined as Φ(A) := A ◦ A−T ,
where the product is that of HadamardHadamard (entrywise). It was studied by C. R.
JohnsonJohnson!Charles R. & H. ShapiroShapiro). Many questions about it remain open,
including that of the range of Φ.

178
 (a) If A is triangular, show that Φ(A) = In .
(b) Verify that 1 is an eigenvalue of Φ(A), associated with the eigenvector ~e whose
components are all ones. Open problem: Is this the only constraint ? In other
words, if M~e = ~e, does there exist a matrix A such that Φ(A) = M ?
(c) If A is not necessarily invertible, we may define Ψ(A) := A ◦ A,
b with A
b the cofactor
1
matrix. Therefore we have Φ(A) = det A Ψ(A) when A is non-singular.
Show that there exists a homogeneous polynomial ∆ in the entries of A, such that
det Ψ(A) ≡ ∆(A) det A.
(d) If n = 3 and A is symmetric,
 
a z y
A = z b x ,
y x c

verify that

∆(A) = (abc − xyz)2 + abx2 y 2 + bcy 2 z 2 + caz 2 x2 − (ax2 )2 − (by 2 )2 − (cz 2 )2 .

Deduce that ∆(A) = 0 when A is rank-one, or when a row of A vanishes. However

det A does not divide ∆(A).

336. (The symmetric positive definite case.) We continue with the study of the relative gain
array. We now restrict to matrices S ∈ SPDn .

(a) Show that Φ(S) ≥ PIn (FiedlerFiedler), with equality only if S = µIn for some µ.
Hint: Write S = λj vj vj in an orthonormal basis. Compute S −T and Φ(S). Then
T

use the inequality ab + ab ≥ 2.

(b) Following the same strategy, show that
 
1 1
Φ(S) ≤ κ(S) + In ,
2 κ(S)

where κ(S) is the condition number of S.

(c) Deduce the inequality  
1 1
κ(Φ(S)) ≤ κ(S) + .
2 κ(S)
(d) Draw the conclusion: if S ∈ SPDn , then

lim Φ(m) (S) = In .

m→+∞

Verify that this convergence has order 2 at least, like in a Newton’sNewton method.

179
337. (S. MaciejMaciej, R. StanleyStanley.)
Let us recall that the permanent of A ∈ Mn (k) is defined by the formula
n
XY
perA := aiσ(i) .
σ∈Sn i=1

It is the same formula as for the determinant, with the exception that the coefficient (σ)
has been replaced by +1. We consider matrices with real entries aij ∈ [0, 1]. We assume
that there are m zeroes among the entries, with m ≤ n. We wish to bound the permanent
of A.
(a) Verify that the maximum of the permanent is achieved, at some matrix whose n2 −m
other entries are 1’s.
(b) Prove the formula X
perA = perAjk · perB jk ,
1≤j<k≤n
jk
where A denotes the block obtained by retaining only the first two rows and the
j-th and k-th colums, whereas B jk is the block obtained by deleting the first two
rows and the j-th and k-th colums.
(c) Let us assume that A has m zeroes and n2 − m ones, and that two zeroes of A belong
to the same row.
i. Show that A has a row of ones.
ii. Wlog, we may assume that a11 = a12 = 0 and the second row is made of ones.
We define A0 form A by switching a11 and a21 . Thus the upper-left block of A0
is  
1 0
.
0 1
Using the formula above, show that perA < perA0 .
(d) If perA is maximal among the admissible matrices, deduce that the zeroes of A are
on m distinct rows and m distinct columns.
(e) We may therefore assume that aii = 0 for i = 1, . . . , m and aij = 0 otherwise. Prove
that m  
` m
X
perA = (−1) (n − `)! .
`
`=0
Deduce that  m
perA ≤ n! 1 − .
2n
338. Let δ ∈ Z be given. We assume that δ is not the square of an integer. We consider the
set Eδ of matrices A ∈ Mn (Z) whose characteristic polynomial is X 2 − δ. If a, b ∈ Z and
b|δ − a2 , we denote
δ − a2
 
a c
M(a,b) := ∈ Eδ , c := .
b −a b

180
 Finally, we say that two matrices A, B ∈ Mn (Z) are similar in Z if there exists P ∈
GL2 (Z) such that P A = BP .
(a) If (a, b) is above and λ ∈ Z, verify that M(a,b) , M(a,−b) , M(a+λb,b) , M(−a,(δ−a2 )/b) are
similar in Z.
(b) Let M ∈ Eδ be given. We define β(M ) as the minimal b > 0 such that M is similar
to M(a,b) . Prove that this definition makes sense.
(c) Show that there exists an a ∈ Z such that |a| ≤ 21 β(M ), such that M is similar to
M(a,b) .
(d) Compare |δ − a2 | with β(M )2 . Deduce that
 √
δ, if δ > 0,
β(M ) ≤ p
4|δ|/3, if δ < 0.

(e) Finally, show that Eδ is the union of finitely many conjugation classes in Z.
339. Let k be a field and P ∈ k[X] be a monic polynomial of degree n.
(a) When is the companion matrix BP diagonalizable ?
(b) Show that the Euclidian algorithm can be used to split P into factors having simple
roots, in finitely many elementary operations.
(c) Deduce an explicit construction of a diagonalizable matrix AP ∈ Mn (k) whose char-
acteristic polynomial is P (diagonalizable companion).
Nota: When k ∈ R and the roots of P are real, Exercise 92 gives an alternate
construction.
340. Elaborate a test which tells you in finite time whether a matrix A ∈ Mn (k) is diagonal-
izable or not. Hints: – A is diagonalizable if and only if P (A) = 0n for some polynomial
with simple roots, – One may construct explicitly the factor P of the characteristic poly-
nomial PA , whose roots are simple and are those of PA (see Exercise 339).
341. Let A ∈ Mn (C) be a strictly diagonally dominant matrix. We denote
|aii |
ri := P < 1, i = 1, . . . n.
j6=i |aij |

Let us recall that A is non-singular, and denote B := A−1 .

(a) Let i 6= k be indices. Show that

|bik | ≤ ri max{|bjk | ; j 6= i}.

(b) Deduce that for every i 6= k, we have |bik | ≤ ri |bkk |. Hint: Fix k and consider the
index j that maximizes |bjk |.
Remark that A−1 is not necessarily strictly diagonally dominant.

181
342. We come back to the relative gain array defined in Exercise 335:

Φ(A) := A ◦ A−T .

We consider the strictly diagonally dominant case and use the notations of Exercise 341.

(a) Verify that X

|Φ(A)ij | ≤ ri (max rj )|Φ(A)ii |.
j6=i
j6=i

(b) Deduce that Φ(A) is strictly diagonally dominant too. Denoting r(A) := maxi ri , we
have r(Φ(A)) ≤ r(A)2 .
(c) We now consider the iterates A(k) := Φ(k) (A). Show that A(k) is strictly diagonally
dominant and that r(A(k) ) → 0. Deduce that A(k) = D(k) (In + E (k) ), where D(k) is
diagonal and E (k) → 0n .

−1
(d) Show that A(k+1) = D(k) M (k) D(k) where M (k) → In . Deduce that D(k+1) → In .
(e) Finally prove (JohnsonJohnson!Charles R. & ShapiroShapiro)

lim Φ(k) (A) = In

k→+∞

for every strictly diagonally dominant matrix.

343. We continue the analysis of the relative gain array defined in Exercise 335, following
JohnsonJohnson!Charles R. & ShapiroShapiro. We consider permutation matrices and
linear combinations of two of them. We recall that permutation matrices are orthogonal:
P −T = P .

(a) Assume that C is the permutation matrix associated with an n-cycle. If z ∈ C is

such that z n 6= 1, verify that
1
(In − zC)−1 = (In + zC + · · · + z n−1 C n−1 ).
1 − zn
Deduce that
1
Φ(In − zC) = n
(In − z n C).
1−z
(b) Use the formula above to prove that
1 nk
Φ(k) (In − zC) = k (In − z C).
1 − zn
Deduce that if |z| < 1 (respectively |z| > 1) then Φ(k) (In − zC) converges towards
In (resp C) as k → +∞.
1
(c) If z n = z 6= 1, deduce that (I
1−z n
− zC) is a fixed point of Φ.
(d) If A ∈ GLn (C) and P is a permutation matrix, verify that Φ(P A) = P Φ(A).

182
 (e) Show that if P and Q are permutation matrices and P + Q is non-singular, then
1
2
(P + Q) is a fixed point of Φ. Hint: reduce to the case where P = In , then work
blockwise to deal only with cycles.

344. The QR method for the calculation of the eigenvalues of a complex matrix was designed in
1961–62 by J. FrancisFrancis. Not only he proved the convergence when the eigenvalues
have distinct moduli, but he recognized the necessity of shifts:

• shifts help to enhance the convergence by reducing the ratio λn /λn−1 , where λn is
the smallest eigenvalue,
• complex shifts help to discriminate pairs of distinct eigenvalues that have the same
modulus. This problem is likely to happen for matrices with real entries, because of
complex conjugate pairs.

We describe below a few basic facts about shifts. The algorithm works as follows. We
start with A, presumably a Hessenberg matrix. We choose a ρ0 ∈ C and make the QR
factorization
A − ρ0 In = Q0 R0 .
Then we re-combine A1 := R0 Q0 + ρ0 In . More generally, if Aj is an iterate, we choose
ρj ∈ C, decompose Aj − ρj In = Qj Rj and recompose Aj+1 := Rj Qj + ρj In . Thus the
standard QR algorithm (without shift) corresponds to choices ρj = 0.
We still denote Pj := Q0 · · · Qj−1 and Uj := Rj−1 · · · R0 .

(a) Verify that Aj+1 = Q∗j Aj Qj and then Ak = Pk∗ APk .

(b) Show that Pk Uk is the QR factorization of the product (A − ρk−1 In ) · · · (A − ρ0 In ).
(c) We consider the case where A ∈ Mn (R). We choose ρ1 = ρ̄0 . Show that P2 ∈ On (R)
and deduce that A2 ∈ Mn (R).
Nota: J. FrancisFrancis found a way to perform the two first iterations at once, by
using only calculations within real numbers. Therefore the shifted QR method does
not need to dive into the complex numbers when A has real entries. See D. Watkins,
American Math. Monthly, May 2011, pp 387–401.

345. A square matrix A is said to be non-derogatory if for every eigenvalue λ, one has dim ker(A−
λIn ) = 1.

(a) Let J = J(0; r) be the basic Jordan block. If B ∈ Mr (k) commutes with J, show
that B is a polynomial in J.
(b) More generally, if B commutes with a non-derogatory matrix A, show that there
exists p ∈ k[X] such that B = p(A). Hint: Jordanization, plus polynomial interpo-
lation.

346. (BezerraBezerra, R. HornHorn!Roger.) This is a follow-up of the previous exercise. Let

A ∈ Mn (C) be non-derogatory, and suppose that B ∈ Mn (C) commutes with both A

183
 and A∗ . Show that B is normal. Hint: Use Schur’s Theorem that a matrix is unitarily
trigonalizable.

347. If v1 , . . . , vr ∈ Rn have non-negative entries, then the matrix v1 v1T +· · ·+vr vrT is symmetric
positive semidefinite and has non-negative entries. A natural question is whether the
converse holds true: given a symmetric matrix S, positive semidefinite with non-negative
entries, do there exist vectors vj ≥ 0 such that S = v1 v1T + · · · + vr vrT ? According to
P. H. DianandaDiananda, and to M. HallHall!M. & M. NewmanNewman, this is true for
n ≤ 4. The following example, due to Hall, shows that it is false when n ≥ 5:
 
4 0 0 2 2
0 4 3 0 2 
 
S=  0 3 4 2 0 .

2 0 2 4 0 
2 2 0 0 4

(a) Verify that S is positive semidefinite. In particular, det S = 0, thus S has a non-
trivial kernel. Compute a generator of the kernel
(b) Suppose that S was a v1 v1T + · · · + vr vrT for some non-negative vectors vj .
i. Show that S can be written as a sum S1 + S2 with
   
4 0 0 2 2 0 0 0 0 0
0 0 0 0 0 0 4 3 0 2
   
0 0 0 0 0
S1 =  , S = 0 3 4 2 0
 2  ,
2 0 0 a 0 0 0 2 x 0
2 0 0 0 b 0 2 0 0 y

and a, b, x, y ≥ 0 and both matrices are positive semidefinite.

ii. Show that necessarily, a = b = 0 and x = y = 4. Hint: Use the kernel of S.
iii. Deduce a contradiction.
(c) Let Pn denote the cone of n × n symmetric matrices with non-negative entries. Let
Sym+ n denote the cone of positive semidefinite matrices. Finally Cn denotes the
cone of symetric n × n matrices S having the property that for v ∈ Rn ,

(v ≥ 0) =⇒ (v T Sv ≥ 0).

Show that if n ≥ 5, then Sym+

n + Pn ( Cn . Hint: argue by duality.

348. (a) Parametrization: Given (a, b, c, d) such that a2 + b2 + c2 + d2 = 1, verify that the
matrix  2 
a + b2 − c2 − d2 2bc − 2ad 2bd + 2ac
 2bc + 2ad a2 − b2 + c2 − d2 2cd − 2ab 
2 2 2 2
2bd − 2ac 2cd + 2ab a −b −c +d
is orthogonal.

184
 (b) Interpretation: Let H be the skew field of quaternions, whose basis is (1, i, j, k). If
q = a + bi + cj + dk ∈ H, we denote q̄ := a − bi − cj − dk its conjugate. We identify
the Euclidian space R3 with the imaginary quaternions q = bi + cj + dk. In other
words, v ∈ R3 iff v̄ = −v. We have qr = r̄q̄. Finally, the norm over H is
√ √
kqk := q q̄ = a2 + b2 + c2 + d2 .

One has kqq 0 k = kqk kq 0 k.

Suppose that q ∈ H has unit norm. Verify that q −1 = q̄. We consider the linear map
r 7→ Lq r := qrq −1 . Show that Lq (R3 ) ⊂ R3 . Verify that the restriction Rq of Lq to
R3 is an isometry.
Prove that q 7→ Rq is a continuous group homomorphism from the unit sphere of H
into O3 (R). Deduce that its image is included into SO3 (R).
(c) Conversely, let R be a rotation of R3 with axis ~u and angle α. Show that R = Rq
for q := cos α2 + ~u sin α2 . The morphism q 7→ Rq is thus onto.
349. Let M ∈ Mn (k) be given. We denote M (j) the j-th principal minor,
 
(j) 1 ... j
M := M .
1 ... j
We assume that these principal minors are all nonzero. Recall that this ensures a factor-
ization M = LU .
(a) For any fixed pair (i, j), use the Desnanot–JacobiDesnanotJacobi formula (Dodg-
sonDodgson (see Lewis C.) condensation formula, exercise 24) to establish the iden-
tity    
1 ... k − 1 i 1 ... ... k
M M
1 ... ... k 1 ... k − 1 j
(k) (k−1)
= Ak−1 − Ak ,
M M
where  
1 ... k i
M
1 ... k j
Ak := .
M (k)
(b) Deduce the formula
   
1 ... k − 1 i 1 ... ... k
n M M
X 1 ... ... k 1 ... k − 1 j
= mij .
k=1
M (k) M (k−1)

(c) Deduce that in the factorization M = LU , we have

   
1 ... j − 1 i 1 ... ... i
M M
1 ... ... j 1 ... i − 1 j
`ij = (j)
, uij = .
M M (i−1)
185
  
i1 . . . ir
(d) In particular, if M is totally positive, meaning that all minors M are
j1 . . . jr
strictly positive whenever i1 < · · · < ir and j1 < · · · < jr , then the non-trivial entries
of L and U are strictly positive. Comment: Actually, all the minors
   
i1 . . . ir i1 . . . i r
L , respectively U ,
j1 . . . jr j1 . . . jr

with j1 < · · · < jr ≤ i1 < · · · < ir (resp. i1 < · · · < ir ≤ j1 < · · · < jr ) are strictly
positive.

350. Let Tn ⊂ Mn (R) be the set of n × n tridiagonal bi-stochastic matrices.

(a) Verify that every M ∈ Tn is of the form

 
1 − c1 c1 0 ··· 0
 ... 
 c1 1 − c 1 − c2 c 2 

. .. . .. .
.. 

 0 c2 1 − c2 − c3

.

 . . . .
 .. .. .. .. c 0 
 n−1 
 ... 
 cn−1 1 − cn−1 − cn cn 
0 ··· 0 cn 1 − cn

In particular, M is symmetric.
(b) Verify that Tn is a convex compact subset of Mn (R), defined by the inequalities

0 ≤ ci ≤ 1, ci + ci+1 ≤ 1.

(c) Prove that the extremal points of Tn are those matrices M for which (c1 , . . . , cn ) is
a sequence of 0s and 1s, in which two 1s are always separated by one or several 0s.
We denote Fn the set of those sequences.
(d) Find a bijection between Fn ∪ Fn−1 and Fn+1 . Deduce that the cardinal of Fn is the
n-th Fibonacci number.

351. (After Y. BenoistBenoist and B. JohnsonJohnson!Bill.) In Exercise 330, we proved that

every real matrix M ∈ Mn (R) can be written as the product HK of (possibly complex)
Hermitian matrices. Of course, one has kM k ≤ kHk · kKk. We show here that if n = 3
(and therefore also if n > 3), there does not exist a finite number cn such that (H, K)
can always be chosen so that kHk · kKk ≤ cn kM k. Thus this factorization is unstable.
So let us assume that for some finite c3 , the following property holds true: for every
M ∈ M3 (R), there exist H, K ∈ H3 such that M = HK and kHk · kKk ≤ c3 kM k.
We recall that the condition number of a non-singular matrix P is

κ(P ) := kP k · kP −1 k ≥ 1.

186
 (a) Let M ∈ GLn (R) be given. Show that χ(M ) ≤ c3 κ(M ), where χ(M ) is defined as
the infimum of κ(P ) where P −1 M P = M T .
(b) Let B := (v1 , v2 , v3 ) be a basis of R3 and let (w1 , w2 , w3 ) be the dual basis. If
M vi = λi vi for all i, verify that M T wi = λi wi .
(c) Chosing pairwise distinct eigenvalues λi , deduce that Φ(B) ≤ c3 κ(M ), where Φ(B)
is the infimum of κ(P ) for which P vi k wi .
(d) Deduce that for every basis B, one has Φ(B) ≤ c3 .
(e) We choose the basis B given by
     
1 1 1
v1, = 0 ,
 v2, =  ,
 v3, = 0 .

0 0 

Compute the dual basis. If the inequality above is true, prove that there exists a
subsequence (P )m →0 that converges to some P0 ∈ GL3 (C), such that P vi, k wi, .
Conclude.

352. Let H ∈ Hn be a Hermitan matrix, with indices of inertia (n− , 0, n+ ). Hence H is

non-degenerate.

(a) Let us write H blockwise as  

H− X
,
X ∗ H+
where H− has size n− × n− . If H− is definite negative, prove that its Schur comple-
ment H+ − X ∗ H−−1 X is positive definite.
(b) Deduce that if E ⊂ Cn is a subspace of dimension n− on which the form x 7→ x∗ Hx
is negative definite (E is a maximal negative subspace for H), then the form y 7→
y ∗ H −1 y is positive definite on E ⊥ (E ⊥ is a maximal positive subspace for H −1 ).

353. We consider the method of JacobiJacobi for the approximate calculation of the spectrum
of a Hermitian matrix H. We take the notations of Section 13.4 of the second edition.

(a) Recall that the equation t2 + 2tσ − 1 = 0 admits two roots t, t0 , with t = tan θ and
θ ∈ [− π4 , π4 ). Verify that the other root corresponds to t0 = tan θ0 with θ0 = θ + π2 .
(b) We call θ the inner angle and θ0 the outer angle. Show that if the choice of angle θ
or θ0 leads to an iterate K or K 0 , then K 0 is conjugated to K by a rotation of angle
π
2
in the (p, q)-plane.
(c) Deduce that if we fix the list of positions (pk , qk ) to be set to zero at step k, the
choice of the angles θk is irrelevant because, if A(k) and B (k) are two possible iterates
at step k, then they are conjugated by a sign-permutation matrix. Such a matrix is
a permutation matrix in which some 1s have been replaced by −1s.

187
354. (After X. TuniTuni.) We show here that it is not possible to define a continuous square
root map over GL2 (C).
(a) Let A ∈ GL2 (C) be given. If A has two distinct eigenvalues, prove that there are
exactly four matrices X such that X 2 = A.
(b) If instead A is not semi-simple, verify that there are exactly two matrices X such
that X 2 = A.
(c) We suppose that there exists a square root map A 7→ A1/2 over GL2 (C), which is
continuous. Without loss of generality, me may assume that
 1/2  1 
1 1 1 2
= .
0 1 0 1
For x ∈ [0, π), prove the formula
 2ix 1/2  ix
(1 + eix )−1

e 1 e
= .
0 1 0 1
Hint: Use a continuity argument.
(d) By letting x → π, conclude.
355. Let A ∈ Symn be a positive semi-
definite matrix. We assume that its
diagonal part D is positive definite.
If b ∈ R(A), we consider the Gauss–
SeidelGaussSeidel method to solve the
system Ax = b:

(D − E)x(m+1) = E T x(m) + b.

Remark that D−E is non-singular from

Gauß, the assumption. The iteration matrix
Carl Friedrich (East Germany) is G = (D − E)−1 E T .
Carl Friedrich Gauß.

(a) Define y (m) = x(m) − x̄, where x̄ is some solution of Ax̄ = b. Verify that y (m+1) =
Gy (m) . In order that y (m) converges for every choice of the initial data y (0) , prove
that it is necessary and sufficient that
• if λ = 1 is an eigenvalue of G, then it is semi-simple,
• the rest of the spectrum of G is of modulus < 1.
(b) Verify that ker(G − In ) = ker A.
(c) Show that G commutes with (D − E)−1 A. Deduce that (D − E)−1 R(A) is a G-
invariant subspace.

188
 (d) Prove that Rn = ker A ⊕ (D − E)−1 R(A). Hint: if v T (D − E)v = 0, then v T (D +
A)v = 0, hence v = 0.
(e) Prove that the spectrum of the restriction of G to (D − E)−1 R(A) has modulus < 1.
Conclude. Hint: follow the proof of Lemma 20 in Chapter 12 of the 2nd edition.

356. (After C. HillarHillar.) A matrix A ∈ Mn (C) is completely invertible if its numerical

range W (A) does not contain 0. In particular, it is invertible, because W (A) contains the
spectrum of A.
We suppose that A is completely invertible.

(a) Show that there exists θ ∈ R and  > 0 such that W (eiθ A) is contained in the
half-plane <z ≥ .
(b) Let us denote B := eiθ A. Verify that B + B ∗ ≥ 2In . Deduce that B −∗ + B −1 is
positive definite.
(c) Prove that there exists α > 0 such that W (B −1 ) is contained in the half-plane
<z ≥ α.
(d) Conclude that A−1 is completely invertible. Therefore a matrix is completely invert-
ible if and only if its inverse is so.

357. I shall not comment the title of this exercise, but it is related to the following fact: if
M ∈ SOn (R) is given in block form
 
A B
M=
C D

where A and D are square matrices, not necessarily of the same size, prove that det A =
det D. Hint: Find block-triangular matrices L and U such that M L = U .
The same identity holds true if M ∈ SO(p, q).

358. Let us say that a lattice

L = ⊕dj=1 Zvj
of Rd has a five-fold symmetry if there exists a matrix A ∈ Md (R) such that A5 = Id ,
A 6= Id and AL = L.

(a) Verify that the lattice Z5 has a five-fold symmetry.

(b) If AL = L, verify that A is similar to a matrix B ∈ Md (Z). Deduce that its
eigenvalues are algebraic integers of degree ≤ d.
(c) Suppose d = 3 and L has a five-fold symmetry A. Show that A is diagonalizable in
M3 (C), with eigenvalues 1, ω and ω̄, where ω is a primitive root of unity of order 5.
Deduce that there does not exist a 3-dimensional lattice with a five-fold symmetry.
(d) In the same vein, if L is a 4-dimensional lattice with a five-fold symmetry A, show
that A4 = −I4 − A − A2 − A3 .

189
 (e) Let α 6= β be the roots of X 2 − X − 1. Let us form the matrix
 
−1 −α −β 0
1 1 0 α −β 
A :=  .
2 1 β 0 −α
−1 1 1 −1

Define  
1
0
v1 = 
0 ,
 v2 = Av1 , v3 = Av2 , v4 = Av3 .
0
Show that the vj ’s span a lattice that has a five-fold symmetry.

359. (After O. Taussky and H. ZassenhausTausskyZassenhaus.) We show here that if A ∈

Mn (k), then there exists a symmetric matrix S ∈ GLn (k) such that AT = SAS −1 (com-
pare with Exercises 327–330). Of course, we already know that there exists a (possibly
non-symmetric) matrix R ∈ GLn (k) such that AT = RAR−1 .

(a) Let us begin with the case where the characteristic and minimal polynomials of A
coincide. We recall that then, the subspace Com(A) of matrices commuting with A
equals
{P (A) | P ∈ k[X]},
and that its dimension equals n.
i. Define the subspace CT(A) ⊂ Mn (k) by the equation M A = AT M . Define also
the subspace ST(A) ⊂ Mn (k) by the equations

SA = AT S and S = S T .

Verify that ST(A) ⊂ CT(A) ⊂ R · Com(A).

ii. Show that dim ST(A) ≥ n.
iii. Deduce that every matrix M such that M A = AT M is symmetric.
(b) We now drop the condition about the characteristic and minimal polynomials of
A. Using the case above, show that there exists a symmetric and non-singular S
such that SA = AT S. Hint: Apply the case above to the diagonal blocks of the
FrobeniusFrobenius form of A.

Remark. This situation is interesting in infinite dimension too. For instance, let us
take a differential operator L = D2 + D ◦ a, where a is a C 1 -bounded function, that is
Lu = u00 + (au)0 . As an unbounded operator over L2 (R), L has an adjoint L∗ = D2 − aD,
that is Lv = v 00 − av 0 . There are a lot of self-adjoint operators S satisfying SL = L∗ S.
For instance,
Sz = (αz 0 )0 + γz

190
 works whenever α(x) and γ(x) solve the linear differential equations

α0 = aα and γ = αa0 + cst α.

More generally, the analysis above suggests that for every P ∈ R[X], S := αP (L) is a
self-adjoint operator satisfying SL = L∗ S.
However, the situation can be very different from the finite-dimensional one, just because
it may happen that an operator M is not conjugated to M ∗ . This happens to the deriva-
d
tion D = dx within the algebra of differential operators. We recover conjugacy by leaving
the realm of differential operators: the symmetry

S0 : (x 7→ f (x)) 7−→ (x 7→ f (−x))

satisfies S0 D = D∗ S0 = −DS0 . Then, as above, every S = S0 P (D) with P ∈ R[X] is

self-adjoint and satisfies SD = D∗ S.

360. (Continuation.) We investigate now the real case: k = R. We ask under which condition
it is possible to choose S symmetric positive definite such that SA = AT S.

(a) Show that a necessary condition is that A has a real spectrum.

(b) Prove that if A is similar to B, and if there exists Σ ∈ SPDn such that ΣB = B T Σ,
then there exists S ∈ SPDn such that SA = AT S.
(c) If A is a Jordan block, describe explicitly the solutions of SA = AT S. Verify that
some of them are positive definite.
(d) Deduce that a necessary and sufficient condition is that A has a real spectrum.

361. Let n ≥ 2 be a given integer. We identify below the convex cone K spanned by matrices
of the form −A2 , with A ∈ Mn (R) running over skew-symmetric matrices.

(a) If A is skew-symmetric with real entries, verify that −A2 is symmetric, positive
semi-definite, and that its non-zero eigenvalues have even multiplicities.
(b) Show that K is contained in the set
 
Tr S
S ∈ Symn | 0n ≤ S ≤ In .
2

(c) Let us define

( )
X 1
C1 := a ∈ Rn | aj = 1 and 0 ≤ aj ≤ , ∀ j .
j
2

i. Show that C1 is a compact, convex subset of Rn , and that its extremal points
have the form 21 (ei + ej ) for some i < j.

191
 ii. Deduce that C1 is the convex hull of
 
1
C1 := (ei + ej ) | 1 ≤ i < j ≤ n .
2

iii. Finally, show that

( )
1X
a ∈ R n | 0 ≤ aj ≤ ak , ∀ j = conv ({λ(ei + ej ) | λ ≥ 0, 1 ≤ i < j ≤ n}) .
2 k

(d) Using this last result, show that actually,

 
Tr S
K = S ∈ Symn | 0n ≤ S ≤ In .
2

362. (After FeitFeit & HigmanHigman.) Let M ∈ Mn (k) be given, and p ∈ k[X] a non-
zero polynomial such that p(M ) = 0 (that is, a multiple of the minimum polynomial
of M ). Let m be the multiplicity of λ ∈ k as a root of p, and define the polynomial
q(X) = p(X)/(X − λ)m . Prove that the multiplicity of λ as an eigenvalue of M equals

Tr q(M )
.
q(λ)

Hint: decompose k n into R((M − λ)m ) and ker(M − λ)m .

363. Let φ : R → R be a continuous convex function. If x, y ∈ Rn are such that x ≺ y, prove

that n n
X X
φ(yj ) ≤ φ(xj ).
j=1 j=1

Hint: Use Proposition 6.4 in the 2nd edition.

Application: Let M ∈ Mn (C) be given, with eigenvalues λj and singular values σj . If s
is a positive real number, deduce from above and from Exercise 69 the inequality
n
X n
X
s
|λj | ≤ σjs .
j=1 j=1

364. (Golden, Wasin So, ThompsonGoldenSo, WasinThompson.)

(a) Let H, K ∈ HPDn be given. Using the previous exercise, prove that for every
integer m ≥ 1,
Tr (HK)2m ≤ Tr (KH 2 K)m .

(b) If m = 2, prove that the equality holds above if, and only if [H, K] = 0n .

192
 (c) Let A, B be Hermitian matrices. Show that the sequence (uk )k≥1 defined by
 
A/2k B/2k 2k
uk = Tr (e e )

is non-increasing.
(d) Deduce that
Tr(eA eB ) ≥ Tr eA+B .
Hint: UseTrotter’s formula

lim (eA/m eB/m )m = eA+B .

m→+∞

(e) In the equality case, that is if Tr(eA eB ) = Tr eA+B , show that Tr((eA eB )2 ) =
Tr(eA e2B eA ). Then deduce that [eA , eB ] = 0n , and actually that [A, B] = 0n .

365. Let t 7→ A(t) be a continuous map with values in the cone of real n × n matrices with
non-negative entries. We denote by E the vector space of solutions of the differential
equation
dx
= −A(t)x.
dt
We also define e = (1, . . . , 1)T .

(a) Verify that the map x 7→ kx(0)k1 is a norm over E.

(b) If τ > 0 is given, we denote xτ ∈ E the solution satisfying xτ (τ ) = e. Verify that
xτ (t) ≥ 0 for every t ∈ [0, τ ].
(c) Let us define y τ = xτ /kxτ (0)k1 . Show that the family (y τ )τ >0 is relatively compact
in E, and that it has a cluster point as τ → +∞.
(d) Deduce Hartman–Wintner’sHartmanWintner Theorem: There exist a non-zero solu-
tion y(t) of the ODE such that y(t) ≥ 0 and y 0 (t) ≤ 0 for every t ≥ 0. In particular,
y(t) admits a limit as t → +∞.
(e) When A is constant, give such a solution in close form.

366. Let K be a non-void compact convex subset in finite dimension. If x ∈ ∂K, the Hahn–
BanachHahnBanach Theorem ensures that there exists at least one convex cone (actually
a half-space) with apex x, containing K. The set of all such convex cones admits a smaller
one, namely the intersection of all of them. We call it the supporting cone of K at x, and
denote it CK (x). If there is no ambiguity, we just write C(x).
We admit the following properties, which are classical in convex analysis:

• K is the intersection of its supporting cones C(x) when x runs over the extremal
points of K,
• If x ∈ ∂K, C(x) is the smallest cone with apex at x, containing all the extremal
points of K.

193
 In what follows, we determine CK (In ) when K is the convex hull of SOn (R) in Mn (R).
We assume that n ≥ 2.

(a) We look for those matrices Q ∈ Mn (R), such that Tr Q(R − In ) ≤ 0 for every
R ∈ SOn (R).
• Using the one-parameters subgroups of SOn (R), show that for every skew-
symmetric matrix A, one has

Tr(QA) = 0, Tr(QA2 ) ≤ 0.

• Show that Q is symmetric.

• We denote q1 ≤ q2 ≤ · · · ≤ qn the eigenvalues of Q. Using Exercise 361, show
that q1 + q2 ≥ 0. In other words,

|q1 | ≤ q2 ≤ · · · ≤ qn .

Hint: Use the extremal points of the cone C1 .

(b) Conversely, let Q ∈ Symn (R) be such that q1 +q2 ≥ 0. Using Dacorogna–MaréchalDacorognaMaréch
Inequality (17), prove that Tr(QR) ≤ Tr Q for every R ∈ SOn (R).
(c) Deduce that C(In ) is the set of all matrices such that Tr(QM ) ≤ Tr Q for every
symmetric Q whose least eigenvalues satisfy q1 + q2 ≥ 0.
(d) Show that C(In ) is the set of matrices M whose symmetric part S = 21 (M + M T )
satisfies  
Tr S − n
1+ In ≤ S ≤ In .
2
Hint: Use again Exercise 361.

367. (From Zhiqin LuLu, Zhiqin.) In this exercise and in the next one, one can replace the
scalar field R by C, to the price that X T be replaced by X ∗ .
 
0n X
Let X ∈ Mn (R) be given. We form the matrix P = ∈ M2n (R). We define the
0n 0n
linear map
TX : V ∈ Mn (R) 7→ [P T , [P, V ]].

(a) Show that TX is self-adjoint over M2n (R), for the standard Euclidian product hV, W i =
Tr(W T V ).
 
B 0n
(b) Verify that the set of block-diagonal matrices is an invariant subspace for
0n C
TX .
The restriction of TX to this subspace will be denoted UX .

194
 (c) Let s1 ≥ · · · ≥ sn be the singular values of X. Show that the spectrum of UX
consists in the numbers µ = s2i + s2j , with multiplicities equal to the number of pairs
(i, j) such that this equality holds.
For instance, if all the numbers s2i and s2j + s2k are pairwise distinct (except for the
trivial s2j + s2k = s2k + s2j ) then s2i is simple and s2j + s2k has multiplicity two for j 6= k.

368. (Continuation of the previous one.) We present below the proof by Zhiqin LuLu, Zhiqin
of the Böttcher–Wenzelbottcher@BöttcherWenzel Inequality

k[X, Y ]k2F ≤ 2kXk2F kY k2F , ∀X, Y ∈ Mn (R).

For X ∈ Mn (R), one defines SX : M 7→ [X T , [X, Y ]].

(a) Show that SX is self-adjoint, positive semi-definite.

k[X,Y ]k2
(b) Verify that the ratio kY k2 F is maximal if and only if Y 6= 0 belongs to the eigenspace
F
E associated with the largest eigenvalue of SX .
(c) If Y ∈ E, show that [X T , Y T ] ∈ E. Show also that [X T , Y T ] is not colinear to Y ,
unless X = 0 or Y = 0. Deduce that dim E ≥ 2.
 
Z 0n
(d) Show that there exists Z 6= 0 in E such that Z1 = is orthogonal to the
0n Z
the main eigenvector of UX (that associated with s21 ).
(e) Show that
k[X, Y ]k2F
sup = hUX Z1 , Z1 i
Y 6=0 kY k2F
and deduce that
k[X, Y ]k2F
sup 2
≤ (s21 + s22 )kY k2F .
Y 6=0 kY kF
Then conclude.
Remark:pThis proof gives a little more when n ≥ 3, because we now that X 7→
φ(X) = s21√+ s22 is a norm (use Exercise 162), with φ ≤ k · kF . We do have
k[X, Y ]kF ≤ 2 φ(X)kY kF .

369. (With the help of P. MigdolMigdol.) Let n ≥ 2 be given. If M ∈ Mn (C), we denote

r(M ) the numerical radius
r(M ) = sup |x∗ M x|.
kxk2 =1

Recall that r is norm over Mn (C). We also define the real and imaginary parts of M by
1 1
<M = (M + M ∗ ) ∈ Hn =M = (M − M ∗ ).
2 2i
(a) Prove that
r(M ) = sup k<(e−iθ M )k2 .
θ∈R/2πZ

195
 (b) Let A, B ∈ Mn (C) be given. We apply the previous question to M = [A, B]. Let θ
be such that r(M ) = k<(e−iθ M )k2 .
i. Let us denote X = <(e−iθ A), Y = =(e−iθ A), Z = <B and T = =B. Check that
r([A, B]) = k[X, T ] + [Y, Z]k2 and deduce
r([A, B]) ≤ 2(kXk2 · kT k2 + kY k2 · kZk2 ).
ii. Conclude that
(37) r([A, B]) ≤ 4r(A)r(B), ∀A, B ∈ Mn (C).
(c) By picking up a convenient pair A, B ∈ Mn (C), show that the constant 4 in (37) is
the best possible. In other words,
r([A, B])
sup = 4.
A,B6=0 r(A)r(B)

370. Let Σ be Hermitian positive definite, with CholeskyCholesky factorization LL∗ .

√
(a) Show that in the polar factorization L = HU , one has H = Σ.
√
(b) Show that in the QR-factorization Σ = QR, one has R = L∗ .
371. We consider real symmetric matrices; the Hermitian case could be treated the same way.
We recall that the eigenvalues and eigenvectors of a matrix are smooth functions of its
entries so long as the eigenvalues are simple (see Theorem 5.3 of the second edition). We
also recall that a functional calculus is available over Hermitian matrices with continuous
functions: if f : I → R is continuous and H = U ∗ DU where U is unitary and D =
diag(a, b), then f (H) = U ∗ f (D)U with f (D) = diag(f (a), f (b)), whenever the spectrum
of H is contained in I. This construction does not depend upon the way we diagonalize
H.
Loewner’sLoewner theory is the study of operator monotone functions. A numerical func-
tion f over an interval I is operator monotone if whenever the real symmetric matrices
A, B have their spectrum included in I, then A ≤ B implies f (A) ≤ f (B). This no-
tion depends upon the size n of the matrices under consideration, the class of operator
monotone functions getting narrower as n increases.
Finally, we recall that if M ∈ GLn (R), then A ≤ B if and only if M AM T ≤ M BM T .
(a) If f is operator monotone, show that f is monotone in the classical sense. Hint:
Take A = λI2 .
(b) Let θ 7→ S(θ) be a smooth curve such that S(θ) has simple eigenvalues. Up to a
unitary conjugation, we may assume that S(θ0 ) = diag(a1 , . . . , an ). Writing S(θ) =
P (θ)diag(λ1 (θ), λ2 (θ))P (θ)T with λj (θ0 ) = aj , P (θ0 ) = In and P (θ) orthogonal,
compute the derivatives of λj and of P at θ0 . Deduce that
d
f (S(θ)) = H ◦ S 0 (θ0 ),
dθ θ0

196
 is the Hadamard product of S 0 (θ0 ) with the matrix H = H(~a) whose entries are
(
f 0 (aj ), if i = j,
hij = f (aj )−f (ai )
aj −ai
, if i 6= j.

(c) If f is operator monotone over n × n matrices whose spectrum belong to I, deduce

from above that H(~a) ≥ 0n for every a1 , . . . , an ∈ I.
(d) Conversely, we suppose that the matrices H(~a) above are non-negative. Let A, B ∈
Symn (R) have their spectra in I, with B − A positive definite. We admit that there
exists a smooth curve θ 7→ S(θ) such that S(0) = A, S(A) = B, S(θ) has simple
eigenvalues for every θ ∈ (0, 1) and S 0 (θ) ≥ 0n . This follows from the fact that
the set of symmetric matrices C such that A ≤ C ≤ B is open, and the subset of
matrices with a multiple eigenvalue has codimension 2 in Symn (R).
d
Prove that dθ f (S(θ)) ≥ 02 for all θ ∈ (0, 1). Deduce that f (A) ≤ f (B).
(e) By continuity, extends this result to the case where B − A ≥ 0n .
Hence a C 1 -function f is operator monotone over n × n matrices (property (OMn))
if and only if the matrices H(~a) are non-negative for every a1 , . . . , an in the domain
of f .
(f) If k ≤ n verify that (OMn) implies (OMk).
(g) Show that (OMn) amounts to saying that for all 1 ≤ k ≤ n, and for every a1 , . . . , ak ∈
I, then
det H(~a) ≥ 0.
(h) Show that (OM2) amounts to saying that f 0 ≥ 0 and √1 0
f
is concave.
(i) If I = R and f is operator monotone, show that f is affine.
(j) Consider the case I = (0, +∞) and f (t) = tα . Show that f is operator monotone if
and only if α ∈ [−1, 1].
372. Let k be a field and E a linear subspace of Mn (k). We assume that every element of E
is singular. We wish to prove that there exist P, Q ∈ GLn (k) such that P EQ =: E 0 has
the property that every M ∈ E 0 has a zero entry mnn .
(a) This is true if n = 1.
(b) Suppose that M 6= {0n }. Show that there exist non-singular P1 , Q1 and 1 ≤ r < n
such that  
Ir 0
Jr = ∈ P1 EQ1 =: E1 .
0 0n−r
(c) Let F ⊂ Mn−r (k) be the subpace of matrices N such that there exists a matrix
 
· ·
A= ∈ E1 .
· N
Prove that every element of F is singular. Hint: Apply Schur’sSchur complement
formula to tA + Jr .

197
 (d) Then argue by induction over n.

373. (From SmyrlisSmyrlis.) We address the following ‘practical’ problems:

• Let n = 2p + 1 be an odd integer, and let n coins be given. The facial value of each
coin is an integer (in cents, say). Suppose that, whenever we remove one coin, the
n − 1 remaining ones can be arranged into two sets of p coins, the total value of the
sets being equal to each other. Prove that all coins have the same value.
• Let n = 2p + 1 be an odd integer, and let n complex numbers zj be given. Suppose
that, whenever we remove one number, the n−1 remaining ones can be arranged into
two sets of p numbers, which share the same isobarycenter. Prove that all numbers
coincide.

Of course the first problem is a special case of the second, a natural integer being a
complex number. Yet, we shall prove the general case after proving the special case.

(a) We begin with the first problem. Show that there exists a matrix A ∈ Mn (Z) and
x ∈ Zn with x > 0 such that Ax = 0, the diagonal entries of A vanish, and the other
entries of any ith line are ±1, summing up to 0.
(b) Prove that the coordinates of x all have the same parity.
(c) Let us define y = x − x1 (1, . . . , 1)T . Verify that Ay = 0 and y1 = 0. Prove that the
coordinates of y have to be even. Finally, prove y = 0 and conclude.
(d) One turns towards the second problem and denote z = (z1 , . . . , zn )T . Then there
exists a matrix A with the same properties as above, and Az = 0. In particular, 0
is an eigenvalue of A.
(e) Using the answer to the first problem, prove that the kernel of A, when acting
over Qn , is one-dimensional. Deduce that its kernel, when acting over Cn , is one-
dimensional. Conclude.

374. This gives an alternate argument about the rank of A in the previous exercise.
Let A ∈ Mn (Z) (n needs not be odd) be such that its diagonal entries are even, whereas
its off-diagonal entries are odd. Prove that rk A ≥ n − 1. Hint: After deleting the first
row and the last column, compute the determinant modulo 2.

375. (After A. BezerraBezerra & H.-J. WernerWerner.) Recall that an n × n matrix M is

idempotent if M 2 = M . In particular, it satisfies

k n = R(M ) ⊕ ker M, ker(In − M ) = R(M ), ker M = R(In − M ).

Let A and B be n × n idempotent matrices

(a) Show that

ker(In − AB) = R(A) ∩ (R(B) ⊕ (ker A ∩ ker B)).

198
 (b) Deduce that

dim ker(In − AB) = dim(R(A) ∩ R(B)) + dim((R(A) + R(B)) ∩ ker A ∩ ker B).

Hint: Use twice the identity dim(X + Y ) + dim(X ∩ Y ) = dim X + dim Y .

376. This is to show that when H = A + A∗ , then the spectrum of A does not tell us much
about that of H, apart from TrH = 2<(TrA), and the obvious fact that the spectrum of
H is real.
Thus, let H be a Hermitian n×n matrix. Show that there exists a matrix A = n1 TrH In +

N with N nilpotent, such that A + A∗ = H. Therefore the spectrum of A is a singleton.

Hint: Use Exercise 131 above, which is Exercise 9 of Chapter 5 in the 2nd edition.

377. Let A and B ∈ Mn (C) be given. We assume that they span a two-dimensional subspace.
We wish to prove that there exist non-trivial factors sj A + tj B (1 ≤ j ≤ N = 2n − 1)
such that
YN
(sj A + tj B) = 0n .
j=1

Here non-trivial means that (sj , tj ) 6= (0, 0).

(a) Let M, P ∈ Mn (k) be given, with r = rkR. Using the rank decomposition, we write
r
X
R= xi aTi .
i=1

Show that rk(RM R) ≤ r, and that rk(RM R) < r if and only if the r × r matrix P
defined by pij := aTi Rxj is singular.
(b) Show, by induction, that for every 1 ≤ r ≤ n, there exists a product of 2n−r − 1
factors whose rank is at most r.

378. (Thanks to F. Brunault.)Brunault We prove here that if n ≥ 2 and P ∈ C[X] is such

that P (A) is diagonalizable for every A ∈ Mn (C), then P is a constant polynomial.

(a) If P ∈ C[X], z ∈ C and B ∈ Mn (C), verify

1 00
P (zIn + B) = P (z)In + P 0 (z)B + P (z)B 2 + · · · ,
2!
where the sum is finite. Hint: this is Taylor expansion for a polynomial in a com-
mutative algebra.
(b) If w ∈ C, M is nilpotent and wIn + M is diagonalizable, prove that M = 0n .
(c) Given P ∈ C[X], suppose that P (A) is diagonalizable for every A ∈ Mn (C) for some
n ≥ 2.

199
 i. If N is nilpotent, show that for every z ∈ C,
1 00
P 0 (z)N + P (z)N 2 + · · · = 0n .
2!
ii. Chosing N 6= 0n above, deduce that det(P 0 (z)In + N 0 ) = 0 for some nilpotent
N 0.
iii. Conclude.
Remark: In this analysis, C can be replaced by any field of characteristic 0.

379. (Continuing.) On the contrary, consider some finite field k = Fpm .

(a) Given n ≥ 2, prove that there exists a Pn ∈ k[X] that is divisible by every polynomial
p ∈ k[X] of degree n.
(b) Show that for all A ∈ Mn (k), one has Pn (A) = 0n ; hence the matrix Pn (A) is
diagonalizable!
(c) We consider the case m = 1, that is k = Fp . Show that we may take
n
m
Y
Pn (X) = (X p − X).
m=1

n−1
Its degree is p pp−1 .

380. We begin with some geometry over closed convex cones in Rn . Let K be such a cone.
We assume in addition that K ∩ (−K) = {0}. By Hahn–Banach Theorem, it is not hard
to see that there exists a compact convex section K0 of K such that K = R+ · K0 . For
instance, if K = (R+ )n , then the unit simplex works.
If x, y ∈ Rn , we write x ≤ y when y − x ∈ K.

(a) If x, y ∈ K, we define

α(x, y) = sup{λ ≥ 0 | λx ≤ y}, β(x, y) = inf{µ ≥ 0 | µx ≥ y}.

We may have α = 0 or β = +∞. We set

β(x, y)
θ(x, y) = log .
α(x, y)

i. Verify that θ(x, y) ∈ [0, +∞], θ(x, y) = θ(y, x) and θ(x, z) ≤ θ(x, y) + θ(y, z).
ii. Suppose that the interior U of K is non-void. Show that θ is a distance over
the (projective) quotient of U by the following relation: y ∼ x if there exists
t ∈ (0, +∞) such that y = tx. This is called the HilbertHilbert distance.
(b) Suppose now that a matrix A ∈ Mn (R) is given, such that AK ⊂ K. Let us define

∆ := sup{θ(Ax, Ay) | x, y ∈ K}.

200
 i. Let x, y ∈ K be given, and α = α(x, y), β = β(x, y). Show that α ≤ α(Ax, Ay)
and β ≥ β(Ax, Ay), and therefore θ(Ax, Ay) ≤ θ(x, y). In summary, A induces
a non-expansive map over U/ ∼.
ii. Define (remark that µ ≥ λ ≥ 0

λ = α(A(y − αx), A(βx − y)), µ = β(A(y − αx), A(βx − y)).

Show that
µα + β λα + β
≤ α(Ax, Ay), ≥ β(Ax, Ay).
µ+1 λ+1
Deduce that
θ(x,y)
λ + et
Z
λα + β µ + 1
θ(Ax, Ay) ≤ log = f 0 (t) dt, f (t) := log .
λ + 1 µα + β 0 µ + et
iii. Verify that
λ+s
s ∈ (0, +∞) 7→ log
µ+s
√
is maximal at s = λµ.
iv. Deduce that
θ(Ax, Ay) ≤ k · θ(x, y), k := tanh e∆/4 .
In particular, if ∆ < +∞, then A induces a contraction over U/ ∼.
(c) Deduce an other proof of a part of the Perron–Frobenius Theorem: if A is strictly
positive, then it has one and only one positive eigenvector.
It is possible to recover the full Perron–Frobenius Theorem with arguments in the
same vein. This proof is due to G. BirkhoffBirkhoff.

381. Let R be an abelian ring and A ∈ Mn (R) be given. The left annihilator Ann` (A) is
the set of B ∈ Mn (R) such that BA = 0n ; it is a left-submodule. Likewise, the right
annihilator Annr (A) is the set of B ∈ Mn (R) such that AB = 0n .

(a) If R is a principal ideal domain, show that there exists a non-singular matrix Q
(depending on A) such that

Ann` (A)T = Q · Annr (A).

Deduce that Ann` (A) and Annr (A) have the same cardinality.
(b) (After K. ArdakovArdakov.) We choose instead R = k[X, Y ]/(X 2 , XY, Y 2 ), where
k is a finite field.
i. Verify that R = k ⊕ m, where m = kx ⊕ ky is a maximal ideal (the unique one)
satisfies m2 = (0). We have |R| = |k|3 and |m| = |k|2 .
ii. Let a, b ∈ R be such that ax + by = 0. Show that a, b ∈ m.

201
 iii. Let us define  
0 x
A= .
0 y
Describe its left- and right-annihilators. Verify that

|k|8 = Ann` (A) 6= Annr (A) = |k|10 .

382. Let Symd be the space of d × d real symmetric matrices, where d ≥ 2 is given. When
endowed with the scalar product hS, T i = Tr(ST ), this becomes a Euclidian space of
dimension N = d(d+1)
2
, isomorphic to RN . The space Q of quadratic forms over Symd is
therefore isomorphic to SymN (!!). If q ∈ Q and V is a subspace of Symd , the trace of q
over V is well-defined, and it equals
X
q(Si ),
i

where S1 , . . . , is any orthonormal basis of V .

By duality, the space of linear forms over Q is isomorphic to Q itself, through L(q) =
Tr(Σσq ), where Σ ∈ SymN and σq ∈ Symn is the matrix associated with q.
(a) Let L be a linear form over Q. Show that there exist a unitary basis S1 , . . . , SN and
numbers αi such that
X
L(q) = αi q(Si ), ∀q ∈ Q.
i

In particular, the numbers αi are unique up to a permutation, and given a number

α, the subspace Vα of Symd spanned by the Si ’s such that αi = α is unique. Hint:
Diagonalize the matrix Σ associated with L.
Finally, show that X

L(q) = α Tr q|Vα , ∀q ∈ Q.
α

(b) We consider the linear representation of Od over Symd by

(U, S) 7→ U T SU.

It induces a representation of Od over Q by

(U, q) 7→ q U , q U (S) = q(U T SU ).

We suppose that a linear form L over Q is invariant under this action:

L(q) = L(q U ), ∀q ∈ Q, U ∈ Od .

Show that there exist numbers α, β such that

L(q) = αq(Id ) + β Tr q, ∀q ∈ Q.

202
 (c) Let K be the (convex) cone of semi-positive quadratic forms, K ⊂ Q. Show that
the extremal rays of K are spanned by the forms qH : S 7→ (Tr(HS))2 .
(d) Let L(q) = 0 be the equation of a supporting hyperplane of K at qId . By convention,
L ≥ 0 over K. We thus have L(qId ) = 0. Let us define
Z
L(q) := L(q U ) dµ(U ),

where µ is the HaarHaar measure over Od . Verify that L(q) = 0 is the equation of
a supporting hyperplane to K at qId , and also that L is invariant under the action
of Od .
(e) Deduce that
L(q) = γ(d Tr q − q(Id )), ∀q ∈ Q,
for some positive constant γ.

v α ⊗v α
P
383. Let Pn be the set of n×n symmetric real matrices that can be written as a sum α
where the vectors v α ∈ Rn are non-negative.

(a) Show that Pn is a convex cone, stable under the HadamardHadamard product.
(b) As an example, let ~a = (a1 , . . . , an ) and ~b = (a1 , . . . , an ) be two sequences of real
numbers, with 0 < a1 < · · · < an and 0 < bn < · · · < b1 . We form the symmetric
matrix S(~a, ~b) whose entries sij are given by amin(i,j) bmax(i,j) . Show that S(~a, ~b) ∈ Pn

384. Let us order Sn , for instance by lexicographic order. Given a matrix A ∈ Mn (k), we
form a matrix P ∈ Mn! (k), whose rows and columns are indexed by permutations, in the
following way:
Yn
pσρ := aσ(i)ρ(i) .
i=1

Notice that pσρ depends only upon σ −1 ρ. Thus P is a kind of circulant matrix.

(a) Show that det A and X

perA := aiσ(i)
σ∈Sn

are eigenvalues of P , and exhibit the corresponding eigenvectors.

(b) If k = R and A is entrywise non-negative, prove that perA is the Perron eigenvalue
of P , that is P ≥ 0n and perA is the spectral radius of P .

385. This exercise is at the foundation by FrobeniusFrobenius of the representation theory.

Let G be a finite group, with n = |G|. If (Xg )g∈G are indeterminates, DedekindDedekind
formed the matrix A ∈ Mn (C[Xe , · · · ]) whose entries are

agh = Xgh−1 , g, h ∈ G.

203
 Remark that if G is cyclic, then A is a circulant matrix.
Let V := C[G] and ρ be the regular representation, where a basis of V as a linear space
is G, and we have ρ(g)h = gh (G acts on itself by left-multiplications).

(a) Show that X

L := Xg ρ(g)
g∈G

acts linearly on C[Xe , · · · ] ⊗ V =: W , and its matrix in the basis G is A.

(b) By decomposing V into irreducible representations V1 , · · · , show that W decomposes
as the direct sum of W1 , · · · , in which each factor is stable under L.
(c) Deduce that the polynomial P (Xe , · · · ) := det A splits as
Y
P = Pini
i

where each Pi is a homogeneous polynomial of degree ni = dim Vi .

Remark: One may prove that each Pi is irreducible, and that they are pairwise
prime.

386. Let H ∈ HPDn be given. Using CholeskyCholesky factorization, find an other proof of
HadamardHadamard Inequality
Yn
det H ≤ hii .
i=1

This gives also the equality case (H must be diagonal).

387. We show here that if a1 , . . . , an are integers, then

Y aj − ai
1≤i<j≤n
j−i

is an integer.

(a) Define pj (X) := X(X − 1) · · · (X − j + 1) ∈ Z[X], where p0 (X) = 1 (empty product)

and p1 (X) = X. Show that there exists an infinite triangular matrix T = (tij )0≤i,j
with 1’s on the diagonal, such that the basis {p0 , p1 , p2 , . . .} is the image of the basis
{1, X, X 2 , . . .} under T .
(b) Deduce that  
1 1 ···
p1 (x1 ) p1 (x2 ) · · · Y
det p (x ) p (x ) · · · = (xj − xi ).
 
 2 1 2 2  1≤i<j≤n
.. ..
. .
Hint: There is a VandermondeVandermonde determinant.

204
 (c) Let us denote  
a a(a − 1) · · · (a − k + 1) pk (a)
= = .
k k! k!
Prove that  
1
1
···
x1 x2

1
1

· · ·
det 
 x1 x2


 2 2 
.. ..
. .
is an integer. Conclude.

388. (After E. StarlingStarling.) Let k be a field whose characteristic is not 2. Let A, B ∈

Mn (k) be such that A2 = B 2 = In . If A + B is non-singular, define

A ? B = (A + B)−1 (A − B + 2In ).

(a) Show that

k n = E+ (A) ⊕ E− (A) = E+ (B) ⊕ E− (B),
where E± (M ) denotes the eigenspace of M associated with the eigenvalue ±1.
(b) If A + B is non-singular, show that dim E± (A) = dim E± (B). In other words, A and
B have the same spectrum.
(c) Show that
E+ (A ? B) = E+ (B), E− (A ? B) = E− (A).
(d) Deduce that (A ? B)2 = In .

389. Let S ∈ Mn (R) be symmetric, with non-negative entries. Concerning the diagonal entries,
we even assume sii > 0 for every i = 1, . . . , n.

(a) Let us define the subset of Rn

n
Y
K := {x ≥ 0 | xi = 1}.
i=1

We consider a minimizing sequence (xk )k∈N of x 7→ 21 xT Sx over K. Show that each

coordinate sequence (xkj )k∈N is bounded below by some  > 0. Deduce that the
sequence is bounded, and therefore it has a cluster point in K.
(b) Deduce that x 7→ 21 xT Sx achieves its infimum over K, at some point x∗ ∈ K.
(c) Using x∗ , show that there exists a vector X ∈ Rn such that X ◦ (SX) = 1, where ◦
is the Hadamard product and 1 = (1, . . . , 1)T .
(d) Deduce that there exists a diagonal matrix D > 0n such that DSD is bi-stochastic.

390. Recall that an idempotent matrix M ∈ Mn (k) represents a projector: M 2 = M . Let

A1 , . . . , Ar be idempotent matrices and define A = A1 + · · · + Ar .

205
 (a) If Ai Aj = 0n for every pair i 6= j, verify that A is idempotent.
(b) Conversely, we suppose that A is idempotent. Using the trace, show that R(A) =
⊕ri=1 R(Ai ). Deduce that Ai Aj = 0n for every i 6= j.

391. This a third proof (due to M. RosenblumRosenblum) of the Putnam–Fuglede’sPutnamFuglede

Theorem that if M, N ∈ Mn (C) are normal matrices and if BN = M B, then BN ∗ =
M ∗ B. The first proof is given in Exercise 255 and the second one in Exercise 297.

(a) Verify that if f : C → C is an entire function, then Bf (N ) = f (M )B.

(b) Deduce that the holomorphic function
∗ ∗
F (z) := eizM Be−izN : C → Mn (C)
∗
is bounded. Hint: Remark that eizM eiz̄M is unitary, hence bounded. This is where
the assumption of normality enter in the play.
(c) Deduce that F is constant. Conclude by calculating F 0 (0).

392. This exercise gives a version of Lemma 20, Section 13.3 (2nd edition), where the as-
sumption that matrices are “Hermitian positive definite” is replaced by “entrywise pois-
itive”. We borrow it from Nonnegative matrices in the mathematical sciences, by A.
BermanBerman & R. J. PlemmonsPlemmons, chapter 5.
Let A, M ∈ Mn (R) be non-singular. We form N = M − A and H = M −1 N , so that H
is the iteration matrix of the scheme

M xk+1 = N xk + b

in the resolution of Ax = b.
We assume that H ≥ 0 (which could be verified by a discrete maximum principle). Show
that is ρ(H) < 1, if and only if A−1 N ≥ 0n . Hint: If
that the iteration is convergent, P
ρ(H) < 1, prove that A−1 N = k≥1 H k . Conversely, if A−1 N is non-negative, apply
Perron–FrobeniusPerronFrobenius and give a relation between the eigenvalues of A−1 N
and those of H.

393. (After K. CostelloCostello & B. YoungYoung.)

Let F be a finite set of cardinal n ≥ 2 and denote P(F ) its Boolean algebra, made of all
subsets of F . Let us define a square matrix B, whose rows and columns are indexed by
the non-void subsets of F :

1 if I ∩ J = ∅,
bij =
0 if I ∩ J 6= ∅.

We shall prove that B is invertible and compute its inverse matrix.

(a) Show that there exists only one permutation of P(F ) satisfying σ(I) ∩ I = ∅ for
every I ⊂ F .

206
 (b) Let M be the square matrix, whose rows and columns are indexed by all the subsets
of F , defined by 
0 if I ∩ J = ∅,
mij =
1 if I ∩ J 6= ∅.
Deduce that det M = 1.
(c) Let A be the square matrix, whose rows and columns are indexed by all the subsets
of F , defined by 
1 if I ∩ J = ∅,
aij =
0 if I ∩ J 6= ∅.
Show that A has rank ≥ 2n − 1. Deduce that B is invertible. Hint: M = eeT − A
where e is defined below.
(d) We chose an order in P(F ) such that the last element is F itself. Denote
   
0 1
 ..   .. 
v = . , e = . .
   
0 1
1 1

Verify Bv = e.
(e) We identify B −1 through an induction over F . If G = F ∩ {a} with a 6∈ F , and if an
order has been given in P(F ), then we order P(G) as follows: first list the non-void
subsets of F , then {a}, then the non-void subsets of F augmented of a. If B and C
denote the matrices associated with F and G respectively, verify that
 
B 0 B
C = 0T 1 eT  .
B e eeT

(f) If B is invertible, show that C is too, with inverse

 
0 −v B−1
C −1 = −v T 0 vT  .
B −1 v −B −1

Conclude.

394. Let M ∈ Mn (R) be bistochastic, with singular values σ1 ≥ · · · ≥ σn .

(a) Show that σ1 = 1.

(b) If J denotes the matrix whose all entries equal 1, what are the singular values of
M − n1 J ?

207
395. Let T ∈ Mn (C) be given. Prove that T ∗ T and T T ∗ are unitarily similar. Hint: Use the
singular value decomposition.
More generally, if T ∈ Mn×m (C) with n ≥ m, prove that there exists an isometry W ∈
Mn×m (C) (that is W ∗ W = Im ) such that T ∗ T = W ∗ (T T ∗ )W .

396. (After J.-C. BourinBourin & E.-Y. LeeLee, Eun-Yung.)

Let H ∈ H+
n+m a Hermitian matrix, positive semi-definite, given blockwise
 
A X
H= , A ∈ H+ +
n , B ∈ Hm .
X∗ B

(a) Show that there exists a decomposition

   
∗ ∗ C Y 0 0
H = T T + S S, T = ,S= .
0 0 Y∗ D

Hint: Use the square root of H.

(b) Deduce that there exist unitary matrices U, V , such that
   
A 0 ∗ 0 0
H=U U +V V ∗.
0 0 0 B

Hint: Use Exercise 395.

Remark: This decomposition implies a lot of inequalities between sums of eigen-
values of A, B and H, respectively, following A. HornHorn!Alfred.
(c) Arguing by induction, show that there exist vectors x1 , . . . , xn ∈ Sn−1 (the unit
sphere) such that
X n
H= hjj xj x∗j .
j=1

397. (After J.-C. BourinBourin.)

Usually, one uses the convexity of the numerical range to prove that for a given ma-
trix M ∈ Mn (C), there exists a unitarily similar U ∗ M U that has a constant diagonal.
However, one may prove directly the latter property.

(a) Show that M is unitarily similar to a matrix R such that i 6= j implies rji = −rij .
Hint: Consider the so-called real part 12 (M + M ∗ ).
(b) If x ∈ Cn is such that |xj | = n−1/2 , show that x∗ Rx = 1
n
Tr M + i=φ(x, x) where φ
is a sesquilinear form to be determined.
(c) Deduce that there exists a unit vector y such that y ∗ M y = 1
n
Tr M .
(d) Show that M is unitarily similar to a matrix of constant diagonal.

208
398. (After D. R. RichmanRichman.)
Let k be a field and n ≥ 2 an integer.
(a) Consider a matrix M ∈ Mn (k) of Hessenberg form
 
· 1
 .. . .
. 1

. O 
. .
 
 .. .. 
M = .
 
 . .. 1 
 
 .. .. 
. . 1
a1 · · · · · · an

If X is an indeterminate, we form M + XJ ∈ Mn (k[X]), where J = diag(1, . . . , 1, 0).

Prove that det(M + XJ) = 0 if and only if a1 = · · · = an = 0. Hint: Induction over
n.
(b) Let p ∈ k[X] be a given monic polynomial. Let M be given as above. Deduce that
there exists a vector z ∈ k n such that the characteristic polynomial of M + e~n z T
equal p. Hint: Linearity of the determinant as a function of a row.
399. (After D. R. RichmanRichman.)

(a) Let S be an integral ring and p a prime number such that pS = {0}. If B ∈ Mn (S),
prove that Tr(B p ) = (Tr B)p . Hint: Work in a splitting field of the characteristic
polynomial.
(b) Let R be an integral ring and n, k ≥ 2 integers. If M ∈ Mn (R) is the sum of k-th
powers of matrices Bj ∈ Mn (R), prove that for every prime factor p of k, there exists
an x ∈ R such that Tr M ≡ xp mod pR.
Nota: RichmanRichman gives also a sufficient condition for the Waring problem to
have a solution. For instance, when p is prime and p ≤ n, the fact that Tr M is a
k-th power mod pR implies that M is a sum of p-th powers. As an example, every
M ∈ Mn (Z) with n ≥ 2 is a sum of squares of integral matrices.

400. Here is another proof of the concavity of f : S 7→ log det S and g : S 7→ (det S)1/n over
SDPn . Of course, it works in the Hermitian case too.
(a) Show that the differential of f at S is T 7→ Tr(S −1 T ).
(b) Verify that the Hessian of f at S is T 7→ − Tr ((S −1 T )2 ).
(c) Conclude that f is concave. Hint: Use the fact that if Σ ∈ SPDn and T is sym-
metric, then the spectrum of ΣT is real.
(d) Follow the same strategy to prove that g is concave. Hint: At the end, you have
to apply Cauchy–Schwarz inequality to the vector of all ones and the vector of
eigenvalues of S −1 T .

209
401. Here is an iterative method for the calculation of the factors in the polar decomposition
U H of a given matrix A ∈ GLn (C). Define a sequence of matrices by
1
A0 = A, Ak+1 = (Ak + A−∗
k ).
2
If Uk Hk is the polar decomposition of Ak , prove that Uk = U and
1
Hk+1 = (Hk + Hk−1 ).
2
Deduce that the sequence is defined for every k ≥ 0, and

lim Ak = U.
k→+∞

Verify that the convergence is quadratic.

DOUBLON avec l’Exercice 262 !!!

402. Let A ∈ Mn (k) and w, z ∈ k n be given.

(a) If det A = 0, show that

z T Aw
b A b TA
b − Awz b = 0n ,

where A
b is the cofactor matrix. Hint: Ab is a rank-one matrix (see Exercise 56).
(b) Suppose that A is diagonalisable, that is
X
A= λj xj yjT , where yiT xj = δij , ∀i, j.
j

Prove the formula

1 X
(z T Aw
b A b T A)
b − Awz b = λ
d T
j,k [(z · yj )yk − (z · yk )yj )] [(w · xj )xk − (w · xk )xj )] ,
det A j<k

where Y
λ
dj,k = λm .
m6=j,k

2
(c) Show actually that there is a polynomial mapping Pw,z : Mn (k) ∼ k n → Mn (k),
such that
1
(z T Aw
b A b T A)
b − Awz b = Pw,z (A).
det A
Hint: Apply Desnanot–JacobiDesnanotJacobi formula (see Exercise 24).
Remark that Pw,z (A)w = 0 and z T Pw,z (A) = 0.
 
0 0T 0 z2
(d) If n = 2, verify that Pw,z (A) ≡: z w , where z := .
−z1

210
403. (After R. MarsliMarsli & F. J. HallHall!F. J..)

(a) Let E be a subspace of Cn , of dimension r. Show that there exists a basis {v 1 , . . . , v r }

of E and pairwise distinct indices i1 , . . . , ir such that kv j k∞ = |vijj | for every j =
1, . . . , r. Hint: Argue by induction over the dimension r.
(b) Let A ∈ Mn (C) be given. Deduce that, if λ is an eigenvalue of A of geometric
multiplicity r, then λ belongs to at least r GershgorinGershgorin disks D(aii ; ri ),
where we recall that X
ri = |aij |.
j6=i

404. Let A, B ∈ Mn×p (C) be given.

(a) Show that  ∗ 

AA A
≥ 0n+p .
A∗ Ip
(b) Deduce that
(AA∗ ) ◦ (BB ∗ ) A ◦ B
 
≥ 0n+p .
A∗ ◦ B ∗ Ip
(c) Finally, show that
(AA∗ ) ◦ (BB ∗ ) ≥ (A ◦ B)(A ◦ B)∗ .
(d) As an application, let H, K be Hermitian non-negative, of same size. Show that

(H ◦ K)1/2 ≥ H 1/2 ◦ K 1/2 .

405. Let A, B be Hermitian matrices of same size, with B positive definite and A positive
semi-definite. We already know that A ◦ B is positive semidefinite.

(a) If aii > 0 for all A, show that A ◦ B is positive semi-definite (SchurSchur). Hint: go
back to the proof that it is positive semidefinite.
(b) In general, show that the rank of A◦B equals the number of positive diagonal entries
of A (BallantineBallantine).
(c) Finally, prove that a positive semidefinite Hermitian matrix C can be factorized
A ◦ B with A, B Hermitian, B positive definite and A positive semi-definite, if an
only if the rank of C equals its number of positive diagonal entries.
(d) The positive definite case is easier and more explicit. Let J be the matrix with all
entries equal to 1, and denote Kα = (1 − α)In + αK. If C is Hermitian positive
definite, verify that C ◦ Kα is positive definite for some α > 1. Check that C =
(C ◦ Kα ) ◦ K1/α and conclude (DjokovicDjokovic).

211
406. If f : RN → R ∩ {+∞} is a proper (f 6≡ +∞) convex function, and h is a positive
parameter, the Yosida approximationYosida of f is
 
h 2
fh (u) = inf |u − v| + f (v) .
v∈RN 2

On another hand, we know that f : Symn (R) → R ∩ {+∞}, defined by


− log det S if S > 0n ,
f (S) =
+∞ otherwise,

is convex.
Compute its Yosida approximation fh .

407. Let A ∈ GLn (R) be such that aij ≤ 0 for every pair i 6= j, and A−1 ≥ 0n (entrywise).
Prove that aii > 0 for every i = 0, . . . , n. Such a matrix is called an M -matrix.

(a) If M ≥ 0n entrywise, show that ρ(M ) < 1 if and only if In − M is non-singular with
(In − M )−1 ≥ 0n . Hint: Sufficiency comes from Perron–FrobeniusPerronFrobenius
Theorem, while necessity involves a series.
(b) Let M ∈ Mn (R) be such that mii > 0, while mij ≤ 0 otherwise. Show that M is an
M -matrix if and only if ρ(In − D−1 M ) < 1, where D = diag{m11 , . . . , mnn }.
(c) Let A be an M -matrix, and B ∈ Mn (R) be such that A ≤ B entrywise, and bij ≤ 0
−1 −1
for every pair i 6= j. We denote DA , DB their diagonals. Verify that DA A ≤ DB B
and deduce that B is an M -matrix.

408. We recall that the function H 7→ φ(H) := −(det H)1/n is convex over HPDn . We extend
φ to the whole of Hn by posing φ(H) = +∞ otherwise. This extension preserves the
convexity of φ. We define as usual the LegendreLegendre transform

φ∗ (K) := sup {Tr(HK) − φ(H)}.

H∈Hn

Show that 
∗ 0 if K ∈ E,
φ (K) =
+∞ otherwise,
where E denotes the set of matrices K ∈ Hn that are non-positive and satisfy
1
(det(−K))1/n ≥ .
n

409. (After W. MascarenhasMascarenhas.) We continue our analysis of the JacobiJacobi al-

gorithm.

212
 (a) In an iteration, compute (kpp − kqq )2 − (hpp − hqq )2 . Deduce that
|kpp − kqq | ≥ |hpp − hqq |.
This means that an iteration tends to separate the relevant diagonal terms. This is
reminiscent to the well-known repulsion phenomenon in quantum mechanics between
two energy levels.
(b) We define X X
Σ := |hjj − hii |, Σ0 := |kjj − kii |.
i,j i,j
i. If x ≤ y ≤ w ≤ z are such that x + z = w + y, verify that the function
a 7→ |a − x| − |a − y| − |a − w| + |a − z| is non-negative. Deduce that
|hii − kpp | + |hii − kqq | − |hii − hpp | − |hii − hqq | ≥ 0.
ii. Show that
Σ0 − Σ ≥ 2|kpp − hpp |.
(c) Let ~δk be the diagonal of A(k) . We define also ∆k to be |kpp − hpp | if H = A(k) ,
K = A(k+1) and (p, q) = (pk , qk ). Finally, Σ and Σ0 above are denoted Σk and Σk+1 .
i. Verify that k~δk+1 − ~δk k∞ ≤ ∆k .
ii. Deduce that k k~δk+1 − ~δk k∞ is finite, and that the diagonal of A(k) converges
P
as k → +∞.
Remark. This result is independent of the choice of the sequence (pk , qk ). It
does not say that A(k) converges to a diagonal matrix. Therefore we don’t claim
that the limit of the diagonal is the spectrum of the initial matrix A.
410. Let A ∈ Symn (Z/2Z) be such that aii = 1 for all diagonal entries. Show that the vector
 
1
 .. 
1 = .
1
belongs to the range of A. Hint: using xT Ax, show that ker A⊥1.
Interpretation. A is the adjacency matrix of a graph, an electric network whose vertices
are light bulbs. At the beginning, all bulbs are turned off. If you switch (off or on) a
bulb, its neighbours are switched simultaneously (but their resulting states depend on
their original states). The problem is to act so that all bulbs are switched on at the end.
This is equivalent to finding an x ∈ (Z/2Z)n such that Ax = 1.
411. Recall that a circulant matrix has the form
   
a1 a2 · · · · · · an 0 1 0 ··· 0
.. ..   .. . . . . . . .. 
 an a1 . . . . . . .

. . .. .. .  = P (J), . .. .. 
A=  .. .. . ..  J = .
. . . . 0,
  
. ... ... .
 .. .
  
a2  0 . 1
a2 · · · · · · an a1 1 0 ··· ··· 0

213
 where P is a polynomial. We consider complex circulant matrices.

(a) Find a permutation matrix P such that every circulant matrix satisfies M T =
P −1 M P .
(b) Deduce that if p ∈ [1, ∞] and p0 is the conjugate exponent, then circulant matrices
satisfy kM kp0 = kM kp .
(c) Show that the map p 7→ kM kp is non-increasing over [1, 2], non-decreasing over
[2, ∞]. Hint: remember that the map
1
7→ log kM kp
p
enjoys a nice property.
(d) Compute kM kp for p = 1, 2, ∞.
(e) For p ∈ [1, ∞), deduce the inequality
n n p n
!p p
X X X X
aj+k xj ≤ |aj | |xj |p .
k=1 j=1 j=1 j=1

Examine the equality case.

412. Given S 1 , . . . , S r ∈ Symn (R), prove that the following statements are equivalent to each
other:

• None of the matrices x1 S 1 + · · · + xr S r is positive definite, when x runs over Rr ,

• There exists a matrix Σ ∈ Symn (R) such that Σ ≥ 0n and Tr(ΣS j ) = 0 for every
j = 1, . . . , r.

Hint: Apply Hahn-Banach.

413. This is graph theory. A graph is a pair (V, E) where V is a finite set (the vertices) and
E ⊂ V × V (the edges) is symmetric: (s, t) ∈ E implies (t, s) ∈ E. The adjacency matrix
A is indexed by V ×V , with ast = 1 if (s, t) ∈ E and ast = 0 otherwise ; this is a symmetric
matrix. The group of the graph is the subgroup G of Bij(V ) formed by elements g such
that (gs, gt) ∈ E if and only if (s, t) ∈ E.
A unit distance representation of the graph is a map from V in some Euclidian space Rd ,
where (after denoting u1 , . . . , un the images of the vertices) |uj − ui | = 1 whenever (ui , uj )
is an edge. For instance, the complete graph (E = V × V ) with n vertices has a UDR in
dimension d = n − 1, a regular simplex. A graph admits a UDR if and only if it does not
contains an edge (s, s).
Given a graph (V, E) without edges (s, s) (the diagonal of A is zero), we look for the
smallest R ≥ 0 such that there is a UDR contained in a ball of radius R. We denote n
the number of vertices.

214
(a) Show that R ≤ 1 and that the minimal UDR can be taken in a space of dimension
d ≤ n − 1.
(b) Show that (V, E) admits a UDR in a ball of radius ρ if and only if there exists a
matrix S ∈ Sym+n such that

Spp ≤ ρ2 ∀p ∈ V,
Spp − 2Spq + Sqq = 1 ∀(p, q) ∈ E.
Show that such a UDR exists in a space whose dimension equals the rank of S.
Hint: Consider a GramGram matrix.
(c) Let R be the infimum of those ρ ≥ 0 for which the graph has a UDR in a ball of
radius ρ. Show that a UDR exists in a ball of radius R (minimal UDR).
(d) Show that among the MUDR’s, there is at least one that is invariant under the group
of the graph:
PgT SPg = S, ∀g ∈ G,
where Pg denotes the permutation matrix associated with g. We call it a symmetric
minimal UDR.
(e) Consider the complete graph with n vertices. For a SMUDR, show that the matrix
S above is of the form 21 In + yJn , where Jn is the matrix of entries 1 everywhere.
Prove that R2 = n−1
2n
. Prove that it can be realized in dimension n − 1 but not in
smaller dimension.
(f) Consider PetersenPetersen graph:

The invariance group of Petersen graph

has the property that it is transitive on
edges: if (st) and (uv) are two edges,
there exists a g ∈ G such that gs = u
and gt = v. It is also transitive on anti-
edges (!): if both (st) and (uv) are not
edges, there exists a g ∈ G such that
gs = u and gt = v. Remark that if
two vertices are not neighbours, there
exists a unique path of length two from
one to the other.
i. Let S be the matrix associated with an SMUDR of Petersen graph. Show that
it has the form
S = S(x, y, z) = xI10 + yJ10 + zA, x + y = R2 .
ii. Verify that 3 is a simple eigenvalue of A. For which eigenvector?
iii. We admit that the other eigenvalues of A are 1 and −2. Compute the eigenvalues
of S(x, y, z). Determine the triples for which S(x, y, z) is positive semi-definite.

215
 3
iv. Finally, show that R2 = 10
and that the SMUDR of Petersen graph can be
realized in dimension 4.
(g) Consider the
p graph of an octahedron (the dual of a cube). Show that its SMUDR
has radius 1/3 and dimension 2. Of course, this dimension is pathologic: the map
V → R2 is not injective !
(h) What is the SMUDR of a square, of a pentagon, more generally of an n-agon ?

414. Consider the product Y

P (z1 , . . . , zn ) := |zk − zj |.
1≤j<k≤n

Prove that the maximum of P , as z1 , . . . , zn run over the unit disk D, equals nn/2 , and
that it is achieved when the points form a regular n-agon.
Hint: Consider a VandermondeVandermonde determinant. Apply the HadamardHadamard
inequality and use the equality case.

415. Denote An (F ) the space of alternate matrices over a field F . Recall that Sr is the standard
non commutative polynomial in r variables (see exercise 289).

(a) Prove that S4 ≡ 0 over A3 (F ). Hint: Invoque the dimension of A3 .

(b) With a similar argument, prove that there exists a matrix J ∈ A4 (F ), and an
alternate 6-form φ over A4 (F ) such that

S6 (A1 , . . . , A6 ) = φ(A1 , . . . , A6 ) · J, ∀ A1 , . . . , A6 ∈ A4 (F ).

(c) Let Ωij = ei eTj − ej eTi be the elements standard basis of A4 . Verify that the products
of all the Ωij ’s in any order, is trivial. Deduce J = 04 . Hence S6 ≡ 0 over A4 .
Comment: B. KostantKostant and L. H. RowenRowen proved that S2n−2 ≡ 0 over
An ; the case n = 2 being trivial.
(d) Likewise, show that

S3 (A, B, C) = φ(A, B, C) · I3 , ∀ A, B, C ∈ A3 (F ),

where φ is a non-zero alternate 3-form over A3 (F ).

416. Let k be a field, n = 2m and A ∈ Mn (k) be alternate. Define the polynomial

 
f 0m Im
PA (X) := Pf(XJn + A), Jn = .
−Im 0m

The matrix XJn + A ∈ Mn (k[X]) is alternate. Its alternate adjoint XJ

\n + A is defined
in Exercise 11.

(a) Show that the entries of XJn + A have degrees less or equal to m − 1.
\

216
 (b) Mimicking the proof of Cayley–HamiltonCayleyHamilton Theorem, deduce that

PAf (JA) = 0n .

Comment. It is remarquable that if A is alternate, then the minimal polynomial

of JA has degree less than or equal to n2 = m.
(c) Let K, B ∈ Mn (k) be alternate matrices, with K invertible. Define Q(X) =
Pf(XK − B). Deduce from above the identity

Q(K −1 B) = 0n .

417. We consider elements M of the orthogonal group O(p, q; R). Wlog, we assume p ≥ q. We
use the block decomposition
 
A B
M= , A ∈ Mp (R).
C D

(a) Write the equations satisfied by the blocks. Deduce that A and D are non-singular.
(b) Show the identity
t
AT A = AT B(Iq + B T B)−1 B T A + Ip = AT f (BB T )A + Ip , f (t) = ,
1+t
equivalently
A−T A−1 + f (BB T ) = Ip .
(c) Let us denote a1 ≥ · · · ≥ ap (> 0) the singular values of A. Likewise, bj , cj , dj are
those of B, C, D, in decreasing order (with 1 ≤ j ≤ q). Prove the relations

a2i = 1 + b2i , ai = di , bi = ci , i = 1, . . . , q

and ai = 1 for q < i ≤ p.

In particular, one has  
kAk2 kBk2
∈ O(1, 1).
kCk2 kDk2
(d) Prove that the image of O(p, q; R) under the projection M 7→ A is precisely the set
of matrices A ∈ GLp (R) satisfying ap ≥ 1 (that is kA−1 k2 ≤ 1) and

(q < i ≤ p) =⇒ (ai = 1).

Hint: First, construct a matrix B, then C and D.

418. (Thanks to P.-L. LionsLions, P.-L..) Let us say that a matrix A ∈ Mn (C) is monotone
if for every vector x ∈ Cn , one has <x∗ Ax ≥ 0. It is strictly monotone if <x∗ Ax > 0 for
every non-zero vector. This amounts to saying that the numerical range of A is contained
in the right half-space (closed or open, respectively) of the complex plane.

217
 (a) Verify that A is monotone if and only if A∗ is so. If P ∈ Mn (C), then P ∗ AP is
monotone too.
(b) Show that A is strictly monotone if and only if A−1 is so.
(c) Let A be monotone and B strictly monotone. Prove that the spectrum of AB (or
BA as well) is contained in C \ (−∞, 0). If both A, B are strictly monotone, then
the spectrum of AB avoids (−∞, 0].
(d) Conversely, let M ∈ Mn (C) be given, its spectrum being contained in C \ (−∞, 0].
We define D a diagonal matrix with the same spectrum.
i. Obviously, D = D02 with D0 strictly monotone. Prove that there exists a neigh-
bourhood V of D such that if N ∈ V, then N = A2 with A strictly monotone.
Hint: The square root of N makes sense thanks to DunfordDunford calculus.
ii. Prove that M is the product of two strictly monotone matrices. Hint: M is
similar to an N ∈ V.
(e) Adapt the previous proof to the real case: if the spectrum of M ∈ Mn (R) avoids
(−∞, 0], then M is the product of two real strictly monotone matrices. Hint:
Consider first the block diagonal case, where the diagonal blocks are either scalar or
rotation matrices.
(f) Show that the solution of the ODE

dX
+ X 2 = 0n , X(0) = M
dt
has a global solution over R+ if and only if the spectrum of M avoids (−∞, 0).

419. (After R. BhatiaBhatia & R. SharmaSharma.) Let A ∈ Mn (C) be normal, with eigen-
values λj for j = 1, . . . , n. Prove

max |aii − ajj | ≤ max |λi − λj |.

i,j i,j

In other words, the spread of the diagonal part of A is not greater than that of A itself
see also Exercise 259).
Hint: All aii and λi belong to the numerical range, but some of them are vertices.

420. (Lewis’ TheoremLewis.)

Let N be a norm over Mn (C). Recall that the dual norm is defined as

N ∗ (M ) = sup | Tr(M T )|.

N (T )≤1

(a) Verify that N (M )N ∗ (M −1 ) ≥ n for all M ∈ GLn (C).

(b) Prove that M 7→ | det M | reaches a maximum over the unit ball of N , say at a matrix
P . Show that N (P ) = 1 and det P 6= 0.

218
 (c) Prove that for every T ∈ Mn (C), we have

| det(In + P −1 T )| ≤ N (P + T )n ≤ (1 + N (T ))n .

(d) Derive the inequality N ∗ (P −1 ) ≤ n.

(e) Deduce Lewis’ TheoremLewis: There exists P ∈ GLn (C) such that

N (P ) = 1 and N ∗ (P −1 ) = n.

421. Recall (see Exercise 165) that a non-negative n×n matrix A is primitive if it is irreducible
and ρ(A) > 0 is the only eigenvalue of maximal modulus. Equivalently, Am is positive
for some m ≥ 1. We prove here Wielandt’sWielandt Theorem: Ap is positive for p =
n2 − 2n + 2 = (n − 1)2 + 1, and this p is sharp in the sense that there is an example for
which Ap−1 is only non-negative.
(a) If Am has a positive column, verify that the same column of Am+1 is positive.
(b) If M ≥ 0 is irreducible and m11 > 0, prove that the first row and column of M q
are positive for every q ≥ n − 1. Hint: it is enough to find positive products
miα mαβ · · · mγτ mτ 1 and m1α mαβ · · · mγτ mτ i of length n − 1, for every i.
(c) Suppose that for every i = 1, . . . , n, there exists an exponent ` = `i ∈ {1, . . . , n − 1}
(`)
such that the matrix A` has a positive diagonal entry aii . Deduce from above that
Ap > 0.
(d) There remains the case where one of the diagonal entries of all of A, A2 , . . . , An−1
(n−1) (n)
vanishes. Say the upper-left entry : a11 = · · · = a11 = 0. Prove that a11 > 0.
Hint: Cayley–HamiltonCayleyHamilton. Deduce that, after a possible reordering,
one has a12 a23 · · · an1 > 0, that is
 
0 1 0 ... 0
 . . . . . . .. 
0 . . . .
. .. .. 
A ≥  .
. . . 0,
  > 0.

0 0 ... 
1
1 0 ... ... 0

(e) Then, if for some q ≥ 1, Aq has a positive column, prove that Aq+1 , . . . , Aq+n−1 also
have a positive column, a different one at each step, and therefore Aq+n−1 is positive.
(f) If Tr Aλ > 0 for some λ ∈ {1, . . . , n − 2}, conclude.
(g) There remains the case where Tr A = Tr A2 = · · · = Tr An−2 = 0. Show that A
satisfies an equation
Tr An−1 Tr An
An = A+ In .
n−1 n
Hint: Apply Newton’sNewton relations.
Deduce a, b > 0 from the fact that A is primitive.

219
 (h) Let c = min(a, b). Show that Ap ≥ cn−2 (A2 + · · · An ), whence Ap ≥ d(In + A + · · · +
An−1 for some d > 0. Deduce Ap > 0n .
422. We prove sharpness of Wielandt’s Theorem. Define
 
0 1 0 ... 0
 .. . . . . .. 
0 . . . .
. .. .. 
A =  ..
 . .
. 0
 .. 
0 0 . 1
1 1 0 ... 0
Show that An = A + In . Deduce that Ap−1 ≤ g(A + · · · + An−1 ) for some g > 0. Conclude
that the upper-left entry of Ap−1 is zero.

423. Given S ∈ SPDn and x ∈ Rn , show that S\ + xxT ≥ Ŝ, where Ŝ = (det S)S −1 denotes
the matrix of cofactors. Hint: Apply ShermanSherman–MorrisonMorrision Formula.
Deduce that the map S 7→ Ŝ is monotonous over SPDn . If S ≤ T and the rank of T − S
is ≥ 2, show that actually Ŝ < T̂ .
Of course, these results have a counterpart in the realm of positive Hermitian matrices.
424. Let B, C ∈ SPDn be given. Let us define a function
φ
SPDn −→ (0, +∞)
A 7→ det(A + B + C) + det C − det(A + C) − det(B + C).
(a) Show that the differential of φ is
dA φ = A +
\ B+C −A
\ + C.
(b) Prove that φ is monotonous. Hint: Use Exercise 423.
(c) Deduce the inequality
det(A + B + C) + det C ≥ det(A + C) + det(B + C).
425. Let F0 = 0, F1 = F2 = 1, etc ... be the FibonacciFibonacci sequence. Prove the identity
 n  
1 1 Fn+1 Fn
= .
1 0 Fn Fn−1
Deduce Fn+1 Fn−1 − Fn2 = (−1)n .
426. (From Z. BradyBrady.) In Sym3 (R), define the set
C = {S | Tr(SO) ≤ 3, ∀O ∈ O3 }.
Show that C is a convex subset, with −C = C (symmetric subset), such that
(S ∈ C) =⇒ (| det S| ≤ 1).
Hint: Use Schur’sSchur triangularization.

220
427. Let K be a field with characteristic 0 and V be a finite-dimensional vector space over K.

(a) Let {e1 , . . . , en } be a basis of V . For k ∈ Z, define

vk = e1 + ke2 + · · · + k n−1 en .

If W is a proper subspace, show that W contains at most n − 1 vectors of the form

vk .
(b) Deduce that V cannot be the finite union of proper subspaces.

428. Let x1 , . . . , xd ∈ Cn be vectors. If m ∈ N, let us form the Hermitian matrix

H = ((mx∗j xi In − xi x∗j ))1≤i,j≤n .

Denote also q(ξ) := ξ ∗ Hξ the form associated with H.

(a) If ξ = (ξ1 , . . . , ξd ) is given blockwise, with ξj ∈ Cn , develop q(ξ) and show that it
equals m Tr F F ∗ − | Tr F |2 for some matrix F (ξ, x).
(b) If m ≥ min(n, d), deduce that H is positive semi-definite. Hint: Recall (see Eexer-
cise 49) the inequality
|TrM |2 ≤ rkM · TrM M ∗ .
2 2
  A ∈ Mn (R) acts upon the space R by X 7→ AX. Let us identify R ∼ C : if
429. A matrix
x
X= , then X ∼ z = x + iy. We therefore may write Az instead of AX.
y

(a) Show that there exist uniquely defined complex numbers a± such that

Az = a+ z + a− z̄ for all z ∈ C.

Write the entries of A in terms of the real and imaginary parts of a+ and a− .
(b) Prove the formulæ

det A = |a+ |2 − |a− |2 , kAk2F = 2|a+ |2 + 2|a− |2 .

(c) Compute the singular values of A in terms of a± . Deduce kAk2 = |a+ | + |a− |. If A
is non-singular, compute kA−1 k2 .
(d) At which condition does A have two real distinct eigenvalues ?

430. Recall that Fq denotes the finite field with q elements (q is a power of a prime number).
We consider matrices chosen randomly in Mn (Fq ), when the entries are independent and
uniformly distributed over Fq .

(a) Show that GLn (Fq ) is of order

(q n − 1)(q n − q) · · · (q n − q n−1 ).

221
 (b) Show that the cardinal of the class

Da := {M ∈ Mn (Fq ) | det M = a}

does not depend on a 6= 0. Deduce the value of this cardinal.

(c) Show that the probability pn (q) that det M = 1 for M chosen randomly in Mn (Fq )
is stricly less than 1q . What is the probability that det M = 0 ?
(d) Show that limn→+∞ pn (q) =: p(q) exists and is non-zero.
Comment: p(q) is the probability that a large random matrix with entries in Fq
have determinant 1 (or any other number a 6= 0). This probability can be expressed
in terms of Dedekind’sDedekind eta function η.

431. If m ∈ N∗ and λ ∈ k, we denote Jm (λ) the Jordan block of size m with eigenvalue λ.

(a) If λ 6= 0, prove that the minimal polynomial of Jm (λ)2 is (X − λ2 )m . Deduce that

Jm (λ)2 is similar to Jm (λ2 ).
(b) On the contrary, verify that Jm (0)2 is permutation-similar to

J[ m+1 ] (0) ⊕ J[ m ] (0).

2 2

432. Let k be a field of characteristic 0 (that is Q ⊂ k). If n ≥ 1, we denote φn the nth

cyclotomic polynomial: Y
Xn − 1 = φd (X).
d|n

Remark that if n 6= m, then φn ∧ φm = 1, because the roots of X ` − 1 are simple and

φn φm divides X ` − 1, where ` = lcm(m, n). With σ ∈ Sn , we associate the permutation
matrix Pσ as usual.

(a) If σ and ρ are conjugated in Sn , prove that Pσ and Pρ are similar.

(b) Let c be an n-cycle.
i. Show that the similarity invariants of Pc are 1, . . . , 1, X n − 1. Hint: there is a
basis in which Pσ becomes a companion matrix.
ii. Let d1 , . . . , dt = n be the divisors of n. Show that XIn − Pc is equivalent, in
Mn (k[X]), to diag(1, . . . , 1, φd1 , . . . , φdt ).
(c) Let σ ∈ Sn be given, and n1 ≤ · · · ≤ nr be the cardinals of the σ-orbits. If d ≥ 1,
we denote m(d) the number of lengths nj that are multiple of d. Prove that the
similarity invariants of Pσ are pn , . . . , p1 defined by
Y
ps = φd .
d ; m(d)≥s

222
 (d) Show that the map (n1 , . . . , nr ) 7→ (pn , . . . , p1 ) which, to a partition of n, associates
the similarity invariants of Pσ , is injective. Hint: The list (pn , . . . , p1 ) determines
the list of φd ’s with multiplicities. If d is maximal for division, then it is one of the
nj . Argue by induction over r.
(e) Deduce Brauer’sBrauer Theorem: if Pσ and Pρ are similar, then ρ and σ are con-
jugated. Hint: It amounts to proving that the σ-orbits and the ρ-orbits have the
same cardinals.

433. Let A ∈ Mn×m (R) be a matrix with non-negative entries: aij ≥ 0 for every i, j. The
positive rank of A is the minimal number ` such that A can be written as the sum of `
rank-one matrices with non-negative entries:
`
X
A= xα (y α )T , xαi ≥ 0, yjα ≥ 0, ∀α, i, j.
α=1

(a) Verify that rk(A) ≤ ` ≤ min(n, m).

(b) For  
1 1 0 0
1 0 1 0
A=
0
,
1 0 1
0 0 1 1
verify that rk(A) < `.

434. We consider n × n matrices A and view their entries aij as indeterminates. Thus det A is
a homogeneous polynomial of degree n. Let us define the CayleyCayley operator
 ∂ ∂

∂a11
. . . ∂a1n
Ω = det  ... ..  ,

. 
∂ ∂
∂an1
... ∂ann

which is a differential operator.

When s ∈ N , the CapelliCapelli identity says that

Ω[(det A)s ] = s(s + 1) · · · (s + n − 1)(det A)s−1 .

Prove the cases n ≤ 2 (and, why not ?, n = 3).

Nota: The polynomial b(s) = s(s + 1) · · · (s + n − 1) is called the BernsteinBernstein–
SatoSato polynomial of the determinant.

435. Let A, B ∈ Mn (k) be such that σ(A) ∩ σ(B) = ∅. Let C ∈ Mn (k) be commuting with
both A + B and AB.
Show that AC − CA belongs to the kernel of X 7→ AX − XB. Deduce that C commutes
with both A and B (Embry’sCapelli Theorem). Hint: see Exercise 167.

223
436. Let q1 and q2 be two non-degenerate quadratic forms over Rn . Prove that the orthogonal
groups O(q1 ) and O(q2 ) are isomorphic if and only if q1 and q2 are isomorphic, that is, if
there exist u ∈ GL(Rn ) such that q2 = q1 ◦ u.
437. We recall the numerical radius of a complex n × n matrix :
r(A) = sup{|x∗ Ax| | kxk2 = 1}.
(a) If U ∈ Un is unitary, verify that r(U −1 ) = r(U ) = 1.
(b) Let T be triangular, with unitary diagonal : |tjj | = 1 for every j ≤ n. Show that
r(T ) > 1 unless T is diagonal. Hint: Start with the 2 × 2 case.
(c) Let A ∈ Mn (C) be such that r(A) ≤ 1 and r(A−1 ) ≤ 1.
i. Prove that the spectrum of A is contained in the unit circle.
ii. Show the converse property : A is unitary. Hint: Start with the triangular case.
438. This is an other proof of the fact that if A, B and AB are normal, then BA is normal too
(Exercise 281). We use the Schur–Frobenius norm over Mn (C),
!1/2
√ X
kM k = Tr M ∗ M = |mij |2 .
i,j

We denote λ1 (M ), . . . , λn (M ) the eigenvalues of M (the order is not important).

(a) Show that kABk = kBAk.
(b) Deduce that X
|λi (BA)|2 = kBAk2 .
i

(c) Conclude.
439. Let A ∈ Mn (C) be given. We suppose that an eigenvalue λ of (algebraic) multiplicity `
belongs to the boundary of the numerical range H(A).
(a) In the case where A is triangular with a11 = · · · = a`` = λ, show that actually A is
block diagonal, A = diag(λI` , A0 ).
(b) Deduce that in the general case, λ is semi-simple (dim ker(A − λIn ) = `), and the
orthogonal of ker(A−λIn ) is stable under A. In other words, λ is a normal eigenvalue,
in the sense of normal matrices.
440. (From L. LessardLessard.) Let A ∈ Mn (k) be given, and X n − a1 X n−1 + · · · + (−1)n an
be its characteristic polynomial. Denote adjA the adjugate matrix (the transpose of the
matrix of cofactors).
Prove
adjA = an−1 In − an−2 A · · · + (−1)n−1 An−1 .
Hint: When A is non-singular, Cayley–HamiltonCayleyHamilton. Then it is nothing but
a polynomial identity in n2 indeterminates.

224
441. If X ∈ Mn (k), we denote LX ∈ L(Symn (k)) the linear operator

S 7→ X T M X.

(a) If λ, µ are eigenvalues of X, show that λµ is an eigenvalue of LX .

(b) If LX is diagonalisable with eigenvalues µ1 , . . . , µn , and if the products µi µj are
pairwise distinct for 1 ≤ i ≤ j ≤ n, prove that the characteristic polynomial ΠX (t)
equals
Y
(38) (t − µi µj ).
1≤i≤j≤n

(c) Remarking that the expression in (38) is a polynomial in the entries of X, with
integer coefficients, deduce that the formula holds true for every field k and every
X ∈ Mn (k).

442. Let A, B ∈ Un be unitary matrices. If Y := (AB ∗ )1/2 ∈ Un is a solution of Y 2 = AB ∗ (a

square root), show that X := A∗ Y is a solution of the quadratic equation

XAXB = In .

443. This exercise shows that O(n3 ) operations suffice to compute the characteristic polynomial
of a real or complex square matrix.

(a) Let R be a commutative ring and M ∈ Mn (R) be an upper Hessenberg matrix.

Denote a1 , . . . , an−1 the sub-diagonal entries, x1 , . . . , xn those of the last column and
M1 , . . . , Mn = M the principal square submatrices ; Mk is obtained by keeping only
the kth first rows and columns of M . Prove the formula

det M = xn det Mn−1 − an−1 xn−1 det Mn−2 + an−1 an−2 xn−2 det Mn−3 − · · · .

(b) If A ∈ Mn (k), with A1 , . . . , An = A its principal submatrices, and if one knows the
characteristic polynomials of A1 , . . . , An−1 , show that PA can be deduced in n2 +O(n)
elementary operations in k.
(c) Deduce that the calculation of PA (X) can be done in 21 n3 + O(n2 ) elementary oper-
ations.
(d) For a general real or complex n × n matrix B, show that that there is a calculation
of PB in O(n3 ) operations.

444. We say that A, B ∈ Mn (k) are orthogonally similar if there exists an O ∈ On (k) such
that AO = OB.

(a) Suppose that A and B are orthogonally similar, say OT AO = B. Verify that for
every word w(X, Y ) in two letters, one has OT w(A, AT )O = w(B, B T ).

225
 (b) Deduce the necessary condition for A and B to be orthogonally similar:

(39) Tr w(A, AT ) = Tr w(B, B T ), for all word w(X, Y ).

(c) If n = 2, show that (39) is equivalent to

Tr A2 = Tr B 2 , Tr(AAT )k = Tr(BB T )k , Tr((AAT )k A) = Tr((BB T )k B), ∀ k ≥ 0.

Hint: Use Cayley–Hamilton.

(d) Still, when n = 2, deduce that for every A ∈ M2 (k), the pair (A, B := AT ) satifies
(39).
(e) Yet, prove that if k = Fp (p an odd prime), and
 
a b
A= ,
c d

where (d − a)2 + (b + c)2 is not a square, then A and AT are not orthogonaly similar.
(f) On the contrary, show that if k = R and n = 2, then A and AT are always orthogo-
nally similar.

445. We endow Mn (R) with the standard scalar product hM, N i = Tr(M T N ).

(a) Let E be the vector space spanned by the bistochastic matrices. Show that E ⊥
consists of matrices for which every sum
n
X
aiσ(i) , ∀σ ∈ Sn
i=1

vanishes.
(b) Verify: The entries a1j (2 ≤ j ≤ n) and aii (1 ≤ i ≤ n − 1) form a coordinate system
in E ⊥ . Actually, the other entries are determined through

aij = aii + a1j − a1i , if i ≤ n − 1

and
n−1
X
anj = a1j − a1n − akk .
k=1

(c) Let K be set of matrices M ∈ Mn (K), entrywise non-negative, whose diagonals sum
to one: n
X
aiσ(i) = 1, ∀σ ∈ Sn .
i=1

Show that K is convex and compact. Prove that its extremal points are the matrices
1 ⊗ ~ej and ~ei ⊗ 1.

226
446. We begin by completing one entry of a positive definite symmetric matrix, and then
consider a more general completing problem.

(a) Let A ∈ Symn (R) be given with a1n = 0. We denote S(x) the symmetric matrix
obtained from A by replacing a1n by x ∈ R. We ask whether it is possible to find x
such that S(x) be positive definite.
i. A necessary condition is that the sub-matrices
   
1 ... n − 1 2 ... n
A and A
1 ... n − 1 2 ... n

are positive definite.

ii. Express det S(x) as a polynomial in x.
iii. We suppose that the necessary condition above is fulfilled. Show that S(x) can
be made positive definite if, and only if
    2
2 ... n − 1 1 ... n − 1
(det A) det A + det A > 0.
2 ... n − 1 2 ... n

iv. Conclude, with the help of the Desnanot–JacobiDesnanotJacobi formula (Exer-

cise 24).
(b) Let T be an n × n tri-diagonal symmetric matrix. Assume that

aii ai+1,i+1 − a2i,i+1 > 0, i = 1, . . . , n − 1.

Prove that the remaining diagonals can be completed in such a way that the new
matrix is positive definite.

447. (G. StrangStrang; thanks to B. SévennecSévennec.) This is a sequel of the previous

exercise. Let T ∈ SPDn be tridiagonal.

(a) Prove that there exists a unique S ∈ SPDn maximizing the determinant among
those matrices whose three main diagonals coincide with those of T :

(|i − j| ≤ 1) =⇒ (sij = tij ).

(b) Show that the inverse T 0 := S −1 is tridiagonal.

Hint: the differential of the map S 7→ det S is the cofactor matrix Ŝ.
(c) If T is diagonal, show that T 0 is diagonal too.
Hint: Hadamard’sHadamard inequality.
(d) Likewise, if ti,i+1 = 0 for some i < n, prove that t0i,i+1 = 0.

227
448. We know that the function x 7→ x2 is not operator monotone over (0, +∞): there exist
Hermitian matrices 0n ≤ A ≤ B such that B 2 − A2 is not positive semidefinite. This does
not tell what happens for projections. Therefore, let P and Q be orthogonal projections
over Cn :
P = P ∗ , Q = Q∗ , P 2 = P, Q2 = Q.
Let us point out that P and Q are positive, and therefore P + Q ≥ P ≥ 0n . We look for
a necessary and sufficient condition in order that (P + Q)2 ≥ P 2 . So suppose that this
inequality holds true.

(a) Show that

(P x = 0, Qy = 0) =⇒ (<hP y, Qxi = 0).
(b) Deduce that (In − Q)P Q(In − P ) = 0n .
(c) Show that X = P Q satisfies X 3 − 2X 2 + X = 0n , that is (X − In )2 X = 0.
(d) Prove that the eigenvalue 1 of X is semi-simple. Hint: kXk2 ≤ 1. Deduce X 2 = X.
(e) Deduce that R(X)⊥ ker(X) and therefore X is Hermitian. Conclude that [P, Q] =
0n .
(f) Conversely, verify that [P, Q] = 0n is sufficient in order that (P + Q)2 ≥ P 2 .
(g) What does the property [P, Q] = 0n mean geometrically, for orthogonal projections ?

449. (After Ilya BogdanovBogdanov.) Let p be a prime number and n ≥ 2 be an integer. If

M ∈ Mn (Fp ), we denote A(M ) the algebra spanned by M .

(a) Show that dim A(M ) ≤ n. If M ∈ GLn (Fp ), deduce that the order of M is at most
pn − 1.
(b) Conversely, let P ∈ Fp [X] be a monic polynomial such that Fpn ∼ Fp [X]/(P ) (we
know that Fpn is an extension of degree n of Fp ; actually, any irreducible polynomial
of degree n works). Let CP be its companion matrix. Show that CP has order pn − 1
in GLn (Fp ).

450. (After L. BorisovBorisov, A. FischleFischle & P. NeffNeff.) We are going to prove the
following assertion:

If M ∈ Mn (R) is such that M 2 is symmetric, then M is orthogonally similar

to a block-diagonal matrix whose diagonal blocks have sizes 1 × 1 and/or 2 × 2.

(a) Show that it is enough to treat the case where M 2 = λIn .

This identity is supposed from now on.
(b) Suppose first that λ 6= 0. We denote M = U SV the singular value decomposition
and N := SQ, where Q = V U ∈ On (R).
i. Show that SQS = λQT , then S 2 QS 2 = λ2 Q. Deduce that either si sj = |λ| or
qij = 0.

228
 ii. Show that S and Q can be (up to a permutation of indices) written simulat-
neously in block-diagonal form, where the diagonal blocks are respectively
   
sI` 0 0` P
, with P ∈ O` (R), if s < t, st = |λ|,
0 tI` P T 0`
or √
λ Im , P ∈ Om (R), but only if λ > 0.
iii. In the first case above (s < t and st = |λ|), show that there exist R ∈ O2` (R)
such that     
T sI` 0 0` P 0` sI`
R R=
0 tI` P T 0` tI` 0`
and conclude the case λ 6= 0.
(c) There remains the case where M 2 = 0n .
i. Proceeding as before, show that
   
Σ 0 0n−k P
S= , Q= , k = dim ker M, P 0T P 0 = In−k
0 0k P0 X

(mind that P, P 0 need not be square matrices).

ii. Deduce that N is block-triangular,
 
0n−k K
N= , K ∈ Mn−k×k (R).
0 0k

iii. Finally, show that N is orthogonally equivalent to a block-diagonal matrix with

diagonal blocks of sizes 1×1 or 2×2. Hint: Use the singular value decomposition
of K.

451. Prove the identity

 
A B
∀A, B ∈ Mn (R), det = | det(A + iB)|2 .
−B A

Hint: the matrix in the left-hand side is conjugated to diag(A + iB, A − iB).

452. Here is the rather simple proof of Toeplitz–HausdorffToeplitzHausdorff Theorem found

by Chandler DavisDavis. Recall that it is enough to prove the theorem for 2 × 2 matrix.
Therefore, let A ∈ M2 (C) be given, and denote W (A) = {x∗ Ax ; x∗ x = 1} its numerical
range.

(a) Verify that W (A) is the image of the set

S := {xx∗ ; x∗ x = 1}

under a real-linear map fA : H2 → C.

229
 (b) Verify that S is a sphere in {H ∈ H2 ; Tr H = 1} : it is parametrized by
   2  2
a z 1 1 1
, a− + b− + 2|z|2 = .
z̄ b 2 2 4

(c) Conclude.

Remark: We infer that in dimension n, the set Sn := {xx∗ ; x ∈ Cn and x∗ x = 1} can

be viewed as a rather complicated union of 2-spheres.

453. Let A ∈ Mn (R) be an alternate matrix. We form the 2n × 2n alternate matrix

 
A −In
B= .
In A

(a) Verify that det B = det(A2 + In ).

(b) Prove det B = (det(In + iA))2 .
(c) Show that the polynomial A 7→ det(In + iA) is a polynomial with real coefficients.
We denote this polynomial Qf.
(d) Conclude that Pf(B) = Qf(A).
(e) Show that the degree of Qf is n if n is even, but n − 1 if n is odd. Actually, Qf is
an even polynomial.

454. Let K ⊂ Mn (R) be the convex cone formed by the matrices S + A where S is symmetric
positive semi-definite, and A is alternate.

(a) Show that with S and A as above, det(S + A) ≥ det S. Hint: Start with the case
where S is positive definite, and use the square root.
(b) Deduce that the determinant takes only non-negative values over K.
(c) Show that K is maximal with these properties: if K is contained in a convex cone
C, and if det ≥ 0 over C, then C = K.

455. (After J. BochiBochi.) Let A ∈ Mn (C) be given, ρ(A) its spectral radius and

r(A) := lim kAm k1/m ,

m→+∞

where k · k is some operator norm over Mn (C). Because of ρ(B) ≤ kBk, we have easily
ρ(A) ≤ r(A). We now prove the converse (Banach’sBanach Formula), using the Cayley–
HamiltonCayleyHamilton’s Theorem. We may assume n ≥ 2.

(a) Show that there exists a finite constant Cn such that kAn k ≤ Cn ρ(A)kAkn−1 . Hint:
CH and, again, ρ(B) ≤ kBk.
(b) Apply the previous question to Am and deduce r(A)n ≤ ρ(A)r(A)n−1 . Conclude.

230
456. (after HörmanderHörmander & MelinMelin.)

(a) Let M be a square matrix, bloc-diagonal:

 
B 0
M= .
0 C

Show that its cofactor matrix is given by the formula

 
(det C)B̂ 0
M̂ = .
0 (det B)Ĉ

Generalize this formula to a larger number of diagonal blocs.

(b) Let D ∈ Mn (k) be Jordan matrix: it is block-diagonal and the diagonal blocks are
Jordan blocks. Suppose that 0 is an eigenvalue, of algebraic multiplicity m. Let also
r be the size of the larger Jordan bloc of D, associated with this eigenvalue λ = 0.
Let uv T ∈ Mn (k) be a rank-one matrix. Show that X m−r divides the characteristic
polynomial of D + uv T .
(c) More generally, Let A ∈ Mn (k) have an eigenvalue λ of algebraic multiplicity µ.
Let r be the multiplicity of λ as a root of the minimal polynomial. Finally, let
uv T ∈ Mn (k) be a rank-one matrix. If r < m, show that λ is still an eigenvalue of
A + uv T , of algebraic multiplicity ≥ m − r.
Remark. One can be much more precise. For each eigenvalue, only one of the
Jordan blocks (one of the largest size) disappears after a rank-one pertubation.
Under a rank-` pertubation, up to ` blocks (of the largest sizes) may disappear.
(d) Let µ be an eigenvalue of A + uv T , which is not an eigenvalue of A. Show that µ is
geometrically simple. Hint: Prove instead that rk(A + uv T − µ) ≥ n − 1.

457. (After an anonymous answer to an MO question.) Given integers ai ∈ Z, we form the

matrix  
a1 −1 · · · · · · −1
.. .. 
−1 a2 . . 

 . . . . . 
M =  ..
 .. .. . . ..  ∈ Mn (Z).
 . ... ...
 ..

−1
−1 · · · · · · −1 an
Let d1 , . . . , dn be its elementary divisors. For ` ≤ n − 1, prove that d` is the gcd of all the
products of ` factors of the form ai + 1.

458. This is about the spread of a matrix A ∈ Mn (C). Recall that it is the number

s(A) = max |λj − λi |

i,j

where the λi ’s are the eigenvalues of A.

231
 (a) Verify that the quadratic form
n n
!2
X 1 X 1
qn (x) = x2i − xi − (x2 − x1 )2
i=1
n 1
2

is positive semi-definite. Hint: It depends only upon x1 + x2 , x23 + · · · + x2n and

x3 + · · · + xn .
(b) Deduce that if D ∈ Mn (C) is diagonal, then
 1/2
2 2 2
s(D) ≤ 2kDkF − (Tr D) .
n
(c) Deduce that
 1/2
2
s(A) ≤ − (Tr A)2
2kAk2F .
n
Hint: Use Schur’s trigonalisation Theorem.
459. Let x, y ∈ Rn−1 be given vectors. Show that kxk2 = kyk2 is a necessary and sufficient
condition in order that there exist a ∈ R and N ∈ Mn (R) such that the n × n matrix
 
a yT
M= ,
x N
is normal.
460. For a matrix M ∈ Mn×m (R), we recall the norm
kM kp→q := max{kM xkq | kxkp ≤ 1} = max{|y T M x|q | kxkp , kykq0 ≤ 1}.
We also define its cut-norm
X
kM kc := max mij .
S,T
i∈S,j∈T

where S (resp. T ) runs over the subsets of [[1, n]] (resp. [[1, m]]).
(a) Let M be such that M 1 = 0 and M T 1 = 0 (the row-wise and column-wise sums all
vanish). Prove that
kM k∞→1 = 4kM kc .
(b) More generally, let M ∈ Mn×m (R) be given, and let A be the (uniquely defined)
completed matrix  
M X
A= ∈ M(n+1)×(m+1) (R)
YT a
such that each row/column sums up to zero. Show that
1
kM kc = kAk∞→1 .
4
232
461. Let S be a simplex in the Euclidian space Rn , whose vertices are v0 , . . . , vn . We recall
that the volume of S is given by the formula
1
V (S) = det(v1 − v0 , . . . , vn − v0 ).
n!
(a) Verify that  
1 1 ··· 1
V (S) = det .
n! v 0 ··· vn
(b) Deduce that
 
1 1 ··· 1
0 1 + v0 · v0 ··· 
(n! V (S))2 = det((1 + vi · vj ))0≤i,j≤n = det  ..
 
.. .. 
. . 1 + vi · vj . 
0 ··· 1 + vn · vn
 
1 1 ··· 1
−1 v0 · v0 · · · 
= det  .. ..  .
 
..
 . . vi · vj . 
−1 · · · vn · vn

(c) Show that det((vi · vj ))0≤i,j≤n = 0. Deduce that

 
0 1 ··· 1
−1 v0 · v0 · · · 
2
(n! V (S)) = det  ..
 
.. .. 
 . . vi · vj . 
−1 · · · vn · vn
as well.
(d) Show that
 
0 1 ··· 1
1 −2v0 · v0 ··· 
n+1 n 2
(−1) 2 (n! V (S)) = det  ..
 
.. .. 
. . −2vi · vj . 
1 ··· −2vn · vn

(e) Deduce the Cayley–Menger Formula:

(−1)n+1
 
2 0 1T
V (S) = n det , E(S) := ((kvi − vj k2 ))0≤i,j≤n .
2 (n!)2 1 E(S)

(f) When n = 2 (case of a triangle), recover Heron’s Formula for the area in terms of
the lengths p, q, r of the edges:
1p
(p + q + r)(p + q − r)(q + r − p)(r + p − q) .
4
233
462. For h : (a, b) → R a smooth function, we form the HankelHankel matrix
 
h h0 · · · h(n)
 h0 h00 · · · h(n+1) 
Mn,h :=  .. ,
 
..
 . . 
h(n) h(n+1) · · · h(2n)

which involves derivatives.

(a) If h satisfies a constant-coefficient ODE of order n, show that det Mn,h ≡ 0.

(b) Conversely, let us assume that det Mn,h ≡ 0. We suppose that n is minimal for this
property, w.l.o.g. det Mn−1,h 6= 0 over (a, b).
i. Show the existence of smooth functions g0 , . . . , gn−1 such that
n−1
X
h(j+n) = gi h(j+i) , j = 0, . . . n.
i=0

ii. Prove
n−1
X
gi0 h(k+i) = 0, k = 0, . . . n − 1.
i=0

iii. Deduce that h satisfies a constant-coefficient ODE of order n.

463. (From G. Zaimi.) Consider the n × n matrix
 
c1 a . . . a
. . . .. 
 b c2 .

A=. . . .
 .. .. .. a 
b . . . b cn

(a) Let 1 denote the n × n matrix whose entries are 1’s. Compute det(A − b1) and
det(A − c1).
(b) Show that X 7→ det(A − X1) is affine.
(c) Deduce the value of det A.

464. Prove that the function

SDPn × Rn −→ R+
(S, x) 7−→ xT S −1 x

is convex.

465. We improve Proposition 6.1 by relaxing the assumption of positive definiteness.

Let H ∈ Symn (R) and K ∈ Sym+
n be given.

234
 √ √
(a) Verify that ker K = ker K and R( K) = R(K).
(b) Let λ 6= 0 be a real number. Show that x 7→ K 1/2 x is an isomorphism from ker(HK −
λIn ) to ker(K 1/2 HK 1/2 − λIn ).
(c) We now suppose that H ∈ Sym+
n as well. Show that

dim ker(HK) = dim ker(K 1/2 HK 1/2 ) = dim ker K + dim(R(K) ∩ ker H).

(d) Deduce that if both H and K are positive semi-definite, then HK is diagonalizable
with non-negative real eigenvalues.
(e) In particular, the zero eigenvalue is semi-simple: ker((HK)2 ) = ker(HK). Find a
direct proof of that point.
(f) Find a pair H ∈ Sym2 and K ∈ Sym+
2 such that HK is not diagonalizable.

466. This is a complement to Proposition 6.1.

(a) Let H, K ∈ Sym+ n (R) be given, prove an arithmetico-geometric inequality for the
spectrum of the matrix product:
1 1
(det(HK)) n ≤ Tr(HK).
n
(b) Conversely, let M ∈ Mn (R) be such that det M > 0. Suppose that
11
(det(M K) n ≤Tr(M K), ∀K ∈ SPDn .
n
Prove that M is symmetric. Hint: use the polar factorization.
467. These are two examples taken from either gas dynamics or wave equation, where the
determinant can be computed using Schur’s complement formula.
(a) Let ρ, p ∈ R and w ∈ Rn be given. Show that
 
ρ ρwT
det = ρpn .
ρw ρw ⊗ w + pIn
Mind that the matrix belongs to Mn+1 (R).
(b) Let a ∈ R and w ∈ Rn be given. Show that
 2 
a + |w|2 −2awT
det = (a2 − |w|2 )n+1 .
−2aw 2w ⊗ w + (a2 − |w|2 )In

468. (H. R. Suleimanova.) The Inverse Eigenvalue Problem consists in characterizing the
possible spectra (n-tuples of complex numbers) of non-negative matrices. We say that
V = (z1 , . . . , zn ) is realizable if there exists a matrix M ∈ Mn (R) with non-negative
entries, whose spectrum is V . An obvious necessary condition is that V be stable under
complex conjugacy. An other one, given by the Perron–Frobenius Theorem is that the
maximal modulus of V belongs to V : z1 is real and z1 ≥ |zj | for every j.

235
 (a) For k ∈ N, we define the Newton sums sk = z1k + · · · + znk . If V is realizable, prove
that sk ≥ 0 for every k ≥ 1.
Remark. Actually, R. Loewy & D. London, and independently C. R. Johnson, proved
that nk−1 skm ≥ skm for k, m ≥ 1.
In the sequel, we denote σk the elementary symmetric polynomial of degree k.
(b) From now on, we assume that V is real and only z1 is positive:

V = (x1 , −x2 , −x3 , . . . , −xn ), every xj ≥ 0.

From the previous question, a necessary condition for realizability is

x1 ≥ σ1 (x2 , . . . , xn ),

which we assume from now on.

i. If k ≥ 1, show that σk (x2 , . . . , xn ) ≤ x1 σk−1 (x2 , . . . , xn ). Deduce the sign of the
coefficients of the polynomial
n
Y
P (X) = (X − zj ).
j=1

ii. Show that the necessary condition is also sufficient (Suleimanova): if V is real
and all but one element are non-positive, then V is realizable if and only if
s1 ≥ 0.

469. (After Borisov, Fischle & Neff.) Let M ∈ Mn (R) be such that M 2 is symmetric.

(a) Define S := [M, M T ]. Verify that S is symmetric and M S + SM = 0.

(b) Prove that there exists an orthogonal matrix Q such that, if N = QT M Q and
R = [N, N T ], then R is block-diagonal, with diagonal blocs of the form 0p (the first
one), or dIp dIq for some distinct real numbers d > 0.
(c) Show that N is bloc-diagonal as well, the first diagonal block being normal, and the
other ones of the form  
0p B
.
C 0q
(d) For such blocks, prove that q = p.
(e) Conclude that M is orthogonally similar to a bloc-diagonal matrix whose diagonal
blocks are either scalars or 2 × 2 matrices.
(f) Show also that these 2 × 2 blocks are either symmetric, or of zero trace.

470. Let A, B ∈ Mn (k) be two commuting matrices: AB = BA. We assume that the minimal
polynomial πA equals its characteristic polynomial PA .

236
 (a) Prove that there exists a vector v1 ∈ k n such that (v1 , Av1 , . . . , An−1 v1 ) forms a basis
of k n .
(b) Show that there exists a polynomial Q of degree < n such that Q(A)v1 = Bv1 .
(c) Verify that BAk v1 = Q(A)Ak v1 .
(d) Deduce that B = Q(A), that is, B is a polynomial in A.

471. Let A ∈ GLn (C) be given. Use the QR factorization to prove Hadamard’s Inequality, plus
the fact that the equality holds if and only if the columns of A are pairwise orthogonal.
Hint: Express the columns A(j) in terms of the columns Q(k) .

472. According to Exercise 26, Chapter 4 (#23, Chap. 7 in the second edition ; the exercise
does not exist in the French edition), the unit sphere Σ of M2 (R) for the induced norm
k · k2 is parametrized by SO2 × O− 2 × [0, 1]. Using the maps
   
cos α sin α cos β sin β
α 7→ R(α) := , β 7→ S(β) := ,
− sin α cos α sin β − cos β

the sphere is actually parametrized by

(α, β, t) 7→ tR(α) + (1 − t)S(β), (α, β, t) ∈ [0, 2π) × [0, 2π) × [0, 1].

(See also exercise 153).

We equip M2 (R) with its standard Euclidian structure, that associated with the Frobenius
norm kAkF = (Tr(AT A))1/2 . For a piecewise smooth hypersurface, it induces an area
element, with which we may define the 3-dimensional volume. Show that
8 2
vol3 (Σ) = π .
3
2
Remark: The same volume of the unit √ sphere of k · kF is 2π . But Σ is clamped between
the Frobenius spheres of radii 1 and 2.

473. Denote    
1 0 1 1
L= , U= .
1 1 0 1

√ spectral radius (see Exercise 140) of F = {L, U } equals∗the golden

Show that the joint
1
ratio φ = 2 (1 + 5 ). More generally, the joint spectral radius of F = {M, M } is kM k2 .

474. (After T. Tao and others.) Let A ∈ Mn (C) be normal, with spectral decomposition
n
X
A= λk (A)vk vk∗ .
k=1

In particular, (v1 , . . . , vn ) is a unitary basis. The coordinates of vk be will denoted

(vk,1 , . . . , vk,n ) below.

237
 (a) Write the spectral decomposition of (λIn − A)−1 , when λ is not an eigenvalue.
(b) From the formula above, prove the formula
n
X |vk,j |2 det(λIn−1 − Mj )
= ,
k=1
λ − λk (A) det(λIn − A)

where Mj is the matrix obtained from A be deleting the jth row and column.
(c) Suppose that λi (A) is a simple eigenvalue. Deduce the identity
Y
|vi,j |2 (λi (A) − λk (A)) = PMj (λi (A)).
k6=i

(d) Verify that this identity is trivial when λi (A) has multiplicity ≥ 2.
475. Let k be a field and a, b, c, d ∈ k be such that n = a2 + b2 + c2 + d2 6= 0.
(a) Verify that the matrix
 2 
a + b2 − c 2 − d 2 2(−ad + bc) 2(ac + bd)
1
R :=  2(ad + bc) a2 − b 2 + c 2 − d 2 2(−ab + cd) 
n
2(−ac + bd) 2(ab + cd) a − b2 − c2 + d2
2

is orthogonal with determinant +1.

(b) Give an interpretation in terms of quaternions. Hint: think of a rotation in the
subspace of pure imaginary quaternions.
(c) If k = Q, verify that every rotation is of the form above with a, b, c, d ∈ Z.
476. For A ∈ Mn (C), prove the identity
1
min{ρ(QA) | Q ∈ Un } = | det A| n .

477. All matrices below belong to SPDn (HPDn might work as well).
(a) Show that the maps S 7→ Ŝ (cofactor matrix) is invertible from SPDn into itself.
We denote A 7→ Ǎ the inverse map. Express Ǎ in terms of A−1 and det A.
(b) Let D be a diagonal matrix, with positive diagonal. Show that
 
1 1
In + D ≤ (det (In + D)) In + D .
2 det D

(c) Deduce that for two matrices B1 , B2 ∈ SPDn , one has

1 1 1
B1 +B2
(B1 + B2 ) ≤ B1 + B2 .
det 2
det B1 det B2
1
In other words, the map B 7→ det B
B is operator convex over SPDn .

238
(d) Given A1 , A2 ∈ SPDn , we define a mean-by-cofactors:
1
mC (A1 , A2 ) = Ĉ, C := (Ǎ1 + Ǎ2 ).
2
Verify that mC (A1 , A2 ) ∈ SPDn . Show that

mH (A1 , A2 ) ≤ mC (A1 , A2 ),

where mH denotes the harmonic mean.

239
Index
author, 5–25, 27, 28, 30–36, 38–41, 43–47, 49–
56, 58–63, 65–68, 70–79, 81–83, 85–89,
91–93, 95–98, 100–122, 124–126, 128–
138, 141–143, 145–149, 151–164, 166,
167, 169–175, 177–179, 181–189, 191–
200, 202–205, 207–211, 214–219, 221–
223, 226–230

stp, 30, 39, 50, 51, 55, 58, 62, 77, 81, 82, 91, 97,
98, 101, 106, 107, 114, 117–121, 124,
128, 135, 141, 148, 160, 164, 170, 187

240
authorIndex of scientists (mostly mathematicians)

241
Index
author, 5–25, 27, 28, 30–36, 38–41, 43–47, 49–
56, 58–63, 65–68, 70–79, 81–83, 85–89,
91–93, 95–98, 100–122, 124–126, 128–
138, 141–143, 145–149, 151–164, 166,
167, 169–175, 177–179, 181–189, 191–
200, 202–205, 207–211, 214–219, 221–
223, 226–230

stp, 30, 39, 50, 51, 55, 58, 62, 77, 81, 82, 91, 97,
98, 101, 106, 107, 114, 117–121, 124,
128, 135, 141, 148, 160, 164, 170, 187

242
stpIndex of stamps

243