EIGENVECTORS FROM EIGENVALUES: A SURVEY OF A

BASIC IDENTITY IN LINEAR ALGEBRA

PETER B. DENTON, STEPHEN J. PARKE, TERENCE TAO, AND XINING ZHANG

arXiv:1908.03795v2 [math.RA] 2 Dec 2019

Abstract. If A is an n×n Hermitian matrix with eigenvalues λ1 (A), . . . , λn (A)

and i, j = 1, . . . , n, then the j th component vi,j of a unit eigenvector vi associ-
ated to the eigenvalue λi (A) is related to the eigenvalues λ1 (Mj ), . . . , λn−1 (Mj )
of the minor Mj of A formed by removing the j th row and column by the for-
mula
Yn n−1
Y
|vi,j |2 (λi (A) − λk (A)) = (λi (A) − λk (Mj )) .
k=1;k6=i k=1
We refer to this identity as the eigenvector-eigenvalue identity. Despite the
simple nature of this identity and the extremely mature state of development
of linear algebra, this identity was not widely known until very recently. In this
survey we describe the many times that this identity, or variants thereof, have
been discovered and rediscovered in the literature (with the earliest precursor
we know of appearing in 1934). We also provide a number of proofs and
generalizations of the identity.

1. Introduction
If A is an n×n Hermitian matrix, we denote its n real eigenvalues by λ1 (A), . . . , λn (A).
The ordering of the eigenvalues will not be of importance in this survey, but for
sake of concreteness let us adopt the convention of non-decreasing eigenvalues:
λ1 (A) ≤ · · · ≤ λn (A).
If 1 ≤ j ≤ n, let Mj denote the n−1×n−1 minor formed from A by deleting the j th
row and column from A. This is again a Hermitian matrix, and thus has n − 1 real
eigenvalues λ1 (Mj ), . . . , λn−1 (Mj ), which for sake of concreteness we again arrange
in non-decreasing order. In particular we have the well known Cauchy interlacing
inequalities (see e.g., [Wil63, p. 103-104])
(1) λi (A) ≤ λi (Mj ) ≤ λi+1 (A)
for i = 1, . . . , n − 1.
By the spectral theorem, we can find an orthonormal basis of eigenvectors
v1 , . . . , vn of A associated to the eigenvalues λ1 (A), . . . , λn (A) respectively. For

Date: December 4, 2019.

PBD acknowledges the United States Department of Energy under Grant Contract desc0012704
and the Fermilab Neutrino Physics Center.
This manuscript has been authored by Fermi Research Alliance, LLC under Contract No. DE-
AC02-07CH11359 with the U.S. Department of Energy, Office of Science, Office of High Energy
Physics. FERMILAB-PUB-19-377-T.
TT was supported by a Simons Investigator grant, the James and Carol Collins Chair, the
Mathematical Analysis & Application Research Fund Endowment, and by NSF grant DMS-
1764034.
1
2 PETER B. DENTON, STEPHEN J. PARKE, TERENCE TAO, AND XINING ZHANG

any i, j = 1, . . . , n, let vi,j denote the j th component of vi . This survey paper is de-
voted to the following elegant relation, which we will call the eigenvector-eigenvalue
identity, relating this eigenvector component to the eigenvalues of A and Mj :
Theorem 1 (Eigenvector-eigenvalue identity). With the notation as above, we have
n
Y n−1
Y
(2) |vi,j |2 (λi (A) − λk (A)) = (λi (A) − λk (Mj )) .
k=1;k6=i k=1

If one lets pA : C → C denote the characteristic polynomial of A,

n
Y
(3) pA (λ) := det(λIn − A) = (λ − λk (A)),
k=1
where In denotes the n × n identity matrix, and similarly let pMj : C → C denote
the characteristic polynomial of Mj ,
n−1
Y
pMj (λ) := det(λIn−1 − Mj ) = (λ − λk (Mj ))
k=1
then the derivative p0A (λi (A)) of pA at λ = λi (A) is equal to
Y n
p0A (λi (A)) = (λi (A) − λk (A))
k=1;k6=i

and so (2) can be equivalently written in the characteristic polynomial form

(4) |vi,j |2 p0A (λi (A)) = pMj (λi (A)).
Example 2. If we set n = 3 and
 
1 1 −1
A= 1 3 1
−1 1 3
then the eigenvectors and eigenvalues are
 
2
1  
v1 = √ −1 ; λ1 (A) = 0
6 1
 
1
1
v2 = √  1  ; λ2 (A) = 3
3 −1
 
0
1  
v3 = √ 1 ; λ3 (A) = 4
2 1

with minors Mj and eigenvalues λi (Mj ) given by

 
3 1
M1 = ; λ1,2 (M1 ) = 2, 4
1 3
√
 
1 −1
M2 = ; λ1,2 (M2 ) = 2 ∓ 2 ≈ 0.59, 3.4
−1 3
√
 
1 1
M3 = ; λ1,2 (M3 ) = 2 ∓ 2 ≈ 0.59, 3.4;
1 3
 EIGENVECTORS FROM EIGENVALUES 3

one can observe the interlacing inequalities (1). One can then verify (2) for all
i, j = 1, 2, 3:
2 (0 − 2)(0 − 4)
= |v1,1 |2 =
3 (0 − 3)(0 − 4)
√ √
1 2 (0 − 2 − 2)(0 − 2 + 2)
= |v1,2 | =
6 (0 − 3)(0 − 4)
√ √
1 2 (0 − 2 − 2)(0 − 2 + 2)
= |v1,3 | =
6 (0 − 3)(0 − 4)

1 (3 − 2)(3 − 4)
= |v2,1 |2 =
3 (3 − 0)(3 − 4)
√ √
1 (3 − 2 − 2)(3 − 2 + 2)
= |v2,2 |2 =
3 (3 − 0)(3 − 4)
√ √
1 2 (3 − 2 − 2)(3 − 2 + 2)
= |v2,3 | =
3 (3 − 0)(3 − 4)

(4 − 2)(4 − 4)
0 = |v3,1 |2 =
(4 − 0)(4 − 3)
√ √
1 2 (4 − 2 − 2)(4 − 2 + 2)
= |v3,2 | =
2 (4 − 0)(4 − 3)
√ √
1 2 (4 − 2 − 2)(4 − 2 + 2)
= |v3,3 | = .
2 (4 − 0)(4 − 3)
One can also verify (4) for this example after computing
p0A (λ) = 3λ2 − 14λ + 12
pM1 (λ) = λ2 − 6λ + 8
pM2 (λ) = λ2 − 4λ + 2
pM3 (λ) = λ2 − 4λ + 2.
Numerical code to verify the identity can be found at [Den19].
Theorem 1 passes a number of basic consistency checks:
(i) (Dilation symmetry) If one multiplies the matrix A by a real scalar c, then
the eigenvalues of A and Mj also get multiplied by c, while the coefficients
vi,j remain unchanged, which does not affect the truth of (2). To put it
another way, if one assigns units to the entries of A, then the eigenvalues
of A, Mj acquire the same units, while vi,j remains dimensionless, and the
identity (2) is dimensionally consistent.
(ii) (Translation symmetry) If one adds a scalar multiple of the identity λIn to
A, then the eigenvalues of A and Mj are shifted by λ, while the coefficient
vi,j remains unchanged. Thus both sides of (2) remain unaffected by such
a transformation.
(iii) (Permutation symmetry) Permuting the eigenvalues of A or Mj does not
affect either side of (2) (provided one also permutes the index i accordingly).
Permuting the ordering of the rows (and colums), as well as the index j,
similarly has no effect on (2).
4 PETER B. DENTON, STEPHEN J. PARKE, TERENCE TAO, AND XINING ZHANG

(iv) (First degenerate case) If vi,j vanishes, then the eigenvector vi for A also
becomes an eigenvector for Mj with the same eigenvalue λi (A) after deleting
the j th coefficient. In this case, both sides of (2) vanish.
(v) (Second degenerate case) If the eigenvalue λi (A) of A occurs with multiplic-
ity greater than one, then by the interlacing inequalities (1) it also occurs
as an eigenvalue of Mj . Again in this case, both sides of (2) vanish.
(vi) (Compatibility with interlacing) More generally, the identity (2) is consis-
tent with the interlacing (1) because the component vi,j of the unit eigen-
vector vi has magnitude at most 1.
(vii) (Phase symmetry) One has the√ freedom to multiply each eigenvector vi by
an arbitrary complex phase e −1θi without affecting the matrix A or its
minors Mj . But both sides of (2) remain unchanged when one does so.
(viii) (Diagonal case) If A is a diagonal matrix with diagonal entries λ1 (A), . . . , λn (A),
then |vi,j | equals 1 when i = j and zero otherwise, while the eigenvalues of
Mj are formed from those of A by deleting one copy of λi (A). In this case
one can easily verify (2) by hand.
(ix) (Normalization) As the eigenvectors v1 , . . . , vn form Pn the columns of an
2
orthogonal matrix,Pone must have the identities |v
i=1 i,j | = 1 for all
n 2
j = 1, . . . , n and |v
j=1 i,j | = 1 for all i = 1, . . . , n. It is not imme-
diately obvious that these identities are consistent with (2), but we will
demonstrate the former identity in Remark 8. For the latter identity, we
use the translation symmetry P(ii) to normalize Pλni (A) = 0, and then observe
n
(e.g., from (7)) that (−1)n j=1 pMj (0) = j=1 det(Mj ) = tr adj(A) is
the (n − 1)th symmetricPn function of the eigenvalues and thus equal (since
λi (A) vanishes) to k=1;k6=i λk (A) = (−1)n p0A (0). Comparing this with
Pn
(4) we obtain j=1 |vi,j |2 = 1.

The eigenvector-eigenvalue identity has a surprisingly complicated history in

the literature, having appeared in some form or another in over two dozen refer-
ences, and being independently rediscovered a half-dozen times, in fields as diverse
as numerical linear algebra, random matrix theory, inverse eigenvalue problems,
graph theory (including chemical graph theory, graph reconstruction, and walks
on graphs), and neutrino physics; see Figure 1. While the identity applies to all
Hermitian matrices, and extends in fact to normal matrices and more generally to
diagonalizable matrices, it has found particular application in the special case of
symmetric tridiagonal matrices (such as Jacobi matrices), which are of particular
importance in several fundamental algorithms in numerical linear algebra.
While the eigenvector-eigenvalue identity is moderately familiar to some math-
ematical communities, it is not as broadly well known as other standard identities
in linear algebra such as Cramer’s rule [Cra50] or the Cauchy determinant formula
[Cau41] (though, as we shall shortly see, it can be readily derived from either of
these identities). While several of the papers in which the identity was discovered
went on to be cited several times by subsequent work, the citation graph is only very
weakly connected; in particular, Figure 1 reveals that many of the citations com-
ing from the earliest work on the identity did not propagate to later works, which
instead were based on independent rediscoveries of the identity. In many cases,
the identity was not highlighted as a relation between eigenvectors and eigenvalues,
but was instead introduced in passing as a tool to establish some other application;
 EIGENVECTORS FROM EIGENVALUES 5

Figure 1. The citation graph of all the references in the litera-

ture we are aware of that mention some variant of the eigenvector-
eigenvalue identity. To reduce clutter, transitive references (e.g.,
a citation of a paper already cited by another paper in the bib-
liography) are omitted. Note the very weakly connected nature
of the graph, with many early initial references not being (transi-
tively) cited by many of the more recent references. This graph
was mostly crowdsourced from feedback received by the authors
after the publication of [Wol19].

also, the form of the identity and the notation used varied widely from appear-
ance to appearance, making it difficult to search for occurrences of the identity
by standard search engines. The situation changed after a popular science article
[Wol19] reporting on the most recent rediscovery [DPZ19, DPTZ19] of the identity
by ourselves; in the wake of the publicity generated by that article, we received
many notifications (see Section 6) of the disparate places in the literature where
the eigenvector-eigenvalue identity, or one closely related to it, was discovered. Ef-
fectively, this crowdsourced the task of collating all these references together. In
this paper, we survey all the appearances of the eigenvector-eigenvalue identity
that we are aware of as a consequence of these efforts, as well as provide several
proofs, generalizations, and applications of the identity. Finally, we speculate on
some reasons for the limited nature of the dissemination of this identity in prior
literature.
6 PETER B. DENTON, STEPHEN J. PARKE, TERENCE TAO, AND XINING ZHANG

2. Proofs of the identity

The identity (2) can be readily established from several existing standard iden-
tities in the linear algebra literature. We now give several such proofs.
2.1. The adjugate proof. We first give a proof using adjugate matrices, which is a
purely “polynomial” proof that avoids any invertibility, division, or non-degeneracy
hypotheses in the argument; in particular, as we remark below, it has an extension
to (diagonalizable) matrices that take values in arbitrary commutative rings. This
argument appears for instance in [Par80, Section 7.9].
Recall that if A is an n × n matrix, the adjugate matrix adj(A) is given by the
formula
adj(A) := (−1)i+j det(Mji ) 1≤i,j≤n


(5)
where Mji is the n − 1 × n − 1 matrix formed by deleting the j th row and ith column
from A. From Cramer’s rule we have the identity
adj(A)A = Aadj(A) = det(A)In .
If A is a diagonal matrix with (complex)
Qn entries λ1 (A), . . . , λn (A), then adj(A) is
also a diagonal matrix with ii entry k=1;k6=i λk (A). More generally, if A is a
normal matrix with diagonalization
X n
(6) A= λi (A)vi vi∗
i=1

where v1 , . . . , vn are an orthonormal basis of eigenfunctions of A, then adj(A) has

the same basis of eigenfunctions with diagonalization
n
X Yn
(7) adj(A) = ( λk (A))vi vi∗ .
i=1 k=1;k6=i

If one replaces A by λIn − A for any complex number λ, we therefore have1

Xn n
Y
adj(λIn − A) = ( (λ − λk (A)))vi vi∗ .
i=1 k=1;k6=i

If one specializes to the case λ = λi (A) for some i = 1, . . . , n, then all but one of
the summands on the right-hand side vanish, and the adjugate matrix becomes a
scalar multiple of the rank one projection vi vi∗ :
n
Y
(8) adj(λi (A)In − A) = ( (λi (A) − λk (A)))vi vi∗ .
k=1;k6=i

Extracting out the jj component of this identity using (5), we conclude that
n
Y
(9) det(λi (A)In−1 − Mj ) = ( (λi (A) − λk (A)))|vi,j |2
k=1;k6=i

which is equivalent to (2). In fact this shows that the eigenvector-eigenvalue identity
holds for normal matrices A as well as Hermitian matrices (despite the fact that
the minor Mj need not be Hermitian or normal in this case). Of course in this case

1To our knowledge, this identity first appears in [Hal42, p. 157].

 EIGENVECTORS FROM EIGENVALUES 7

the eigenvectors are not necessarily real and thus cannot be arranged in increasing
order, but the order of the eigenvalues plays no role in the identity (2).

Remark 3. The same argument also yields an off-diagonal variant

n
0 Y
(10) (−1)j+j det(λi (A)(In )j 0 j − Mj 0 j ) = ( (λi (A) − λk (A)))vi,j vi,j 0
k=1;k6=i

for any 1 ≤ j, j 0 ≤ n, where (In )j 0 j is the n − 1 × n − 1 minor of the identity matrix

In . When j = j 0 , this minor (In )j 0 j is simply equal to In−1 and the determinant
can be expressed in terms of the eigenvalues of the minor Mj ; however when j 6= j 0
there is no obvious way to express the left-hand side of (10) in terms of eigenvalues
of Mj 0 j (though one can still of course write the determinant as the product of the
eigenvalues of λi (A)(In )j 0 j − Mj 0 j ). Another way of viewing (10) is that for every
1 ≤ j 0 ≤ n, the vector with j th entry

(−1)j det(λi (A)(In )j 0 j − Mj 0 j )

is a non-normalized eigenvector associated to the eigenvalue λi (A); this observation

appears for instance in [Gan59, p. 85–86]. See [Van14] for some further identities
relating the components vi,j of the eigenvector vi to various determinants.

Remark 4. This remark is due to Vassilis Papanicolaou2. The above argument

also applies to non-normal matrices A, so long as they are diagonalizable with
some eigenvalues λ1 (A), . . . , λn (A) (not necessarily real or distinct). Indeed, if one
lets v1 , . . . , vn be a basis of right eigenvectors of A (so that Avi = λi (A)vi for all
i = 1, . . . , n), and let w1 , . . . , wn be the corresponding dual basis3 of left eigenvectors
(so wiT A = wiT λi (A), and wiT vj is equal to 1 when i = j and 0 otherwise) then one
has the diagonalization
n
X
A= λi (A)vi wiT
i=1

and one can generalize (8) to

n
Y
adj(λi (A)In − A) = ( (λi (A) − λk (A)))vi wiT
k=1;k6=i

leading to an extension
n
Y
(11) det(λi (A)In−1 − Mj ) = ( (λi (A) − λk (A)))vi,j wi,j
k=1;k6=i

of (9) to arbitrary diagonalizable matrices. We remark that this argument shows

that the identity (11) is in fact valid for any diagonalizable matrix taking values in
any commutative ring (not just the complex numbers). The identity (10) may be
generalized in a similar fashion; we leave the details to the interested reader.

2terrytao.wordpress.com/2019/08/13/eigenvectors-from-eigenvalues/#comment-519905
3In the case when A is a normal matrix and the v are unit eigenvectors, the dual eigenvector
i
wi would be the complex conjugate of vi .
8 PETER B. DENTON, STEPHEN J. PARKE, TERENCE TAO, AND XINING ZHANG

Remark 5. As observed in [Van14, Appendix A], one can obtain an equivalent

identity to (2) by working with the powers Am , m = 0, . . . , n − 1 in place of adj(A).
Indeed, from (6) we have
Xn
Am = λi (A)m vi vi∗
i=1

and hence on extracting the jj component

n
X
(Am )jj = λi (A)m |vi,j |2
i=1

for all j = 1, . . . , n and m = 0, . . . , n − 1. Using Vandermonde determinants (and

assuming for sake of argument that the eigenvalues λi (A) are distinct), one can
then solve for the |vi,j |2 in terms of the (Am )jj , eventually reaching an identity
[Van14, Theorem 2] equivalent to (2) (or (10)), which in the case when A is the
adjacency matrix of a graph can also be expressed in terms of counts of walks of
various lengths between pairs of vertices. We refer the reader to [Van14] for further
details.

2.2. The Cramer’s rule proof. Returning now to the case of Hermitian matrices,
we give a variant of the above proof of (2) that still relies primarily on Cramer’s
rule, but makes no explicit mention of the adjugate matrix; as discussed in the next
section, variants of this argument have appeared multiple times in the literature.
We first observe that to prove (2) for Hermitian matrices A, it suffices to do so
under the additional hypothesis that A has simple spectrum (all eigenvalues occur
with multiplicity one), or equivalently that

λ1 (A) < λ2 (A) < · · · < λn (A).

This is because any Hermitian matrix with repeated eigenvalues can be approx-
imated to arbitrary accuracy by a Hermitian matrix with simple spectrum, and
both sides of (2) vary continuously with A (at least if we avoid the case when λi (A)
occurs with multiplicity greater than one, which is easy to handle anyway by the
second degenerate case (iv) noted in the introduction).
As before, we diagonalize A in the form (6). For any complex parameter λ not
equal to one of the eigenvalues λi (A), The resolvent (λIn − A)−1 can then also be
diagonalized as
n
−1
X vi vi∗
(12) (λIn − A) = .
i=1
λ − λi (A)

Extracting out the jj component of this matrix identity using Cramer’s rule [Cra50],
we conclude that
n
det(λIn−1 − Mj ) X |vi,j |2
=
det(λIn − A) i=1
λ − λi (A)
which we can express in terms of eigenvalues as
Qn−1 n
(λ − λk (Mj )) X |vi,j |2
(13) Qk=1
n = .
k=1 (λ − λk (A)) i=1
λ − λi (A)
 EIGENVECTORS FROM EIGENVALUES 9

Both sides of this identity are rational functions in λ, and have a pole at λ = λi (A)
for any given i = 1, . . . , n. Extracting the residue at this pole, we conclude that
Qn−1
(λ (A) − λk (Mj ))
Qnk=1 i = |vi,j |2
k=1:k6=i (λ i (A) − λ k (A))
which rearranges to give (2).
Remark 6. One can view the above derivation of (2) from (13) as a special case
of the partial fractions decomposition
P (t) X P (α) 1
=
Q(t) Q0 (α) t − α
Q(α)=0

whenever Q is a polynomial with distinct roots α1 , . . . , αn , and P is a polynomial of

degree less than that of Q. Equivalently, this derivation can be viewed as a special
case of the Lagrange interpolation formula (see e.g., [AS64, §25.2])
n n
X Y t − αj
P (t) = P (αi )
i=1
αi − αj
j=1;j6=i

whenever α1 , . . . , αn are distinct and P is a polynomial of degree less than n.

Remark 7. A slight variant of this proof was observed by Adam Harrow4, inspired
by the inverse power method for approximately computing eigenvectors numerically.
We again assume simple spectrum. Using the translation invariance noted in con-
sistency check (ii) of the introduction, we may assume without loss of generality that
λi (A) = 0. Applying the resolvent identity (12) with λ equal to a small non-zero
quantity ε, we conclude that
vi vi∗
(14) (εIn − A)−1 = + O(1).
ε
On the other hand, by Cramer’s rule, the jj component of the left-hand side is
det(εIn−1 − Mj ) pMj (ε) pM (0) + O(ε)
= = 0j .
det(εIn − A) pA (ε) εpA (ε) + O(ε2 )
Extracting out the top order terms in ε, one obtains (4) and hence (2). A variant
of this argument also gives the more general identity
X pMj (λ)(λ − λ∗ )
(15) |vi,j |2 = lim
λ→λ∗ pA (λ)
i:λi (A)=λ∗

whenever λ∗ is an eigenvalue of A of some multiplicity m ≥ 1. Note when m = 1

we can recover (4) thanks to L’Hôpital’s rule. The right-hand side of (15) can also
be interpreted as the residue of the rational function pMj /pA at λ∗ .
An alternate approach way to arrive at (2) from (13) is as follows. Assume
for the sake of this argument that the eigenvalues of Mj are all distinct from the
eigenvalues of A. Then we can substitute λ = λk (Mj ) in (13) and conclude that
n
X |vi,j |2
(16) =0
i=1
λk (Mj ) − λi (A)

4twitter.com/quantum aram/status/1195185551667847170
10 PETER B. DENTON, STEPHEN J. PARKE, TERENCE TAO, AND XINING ZHANG

for k = 1, . . . , n − 1. Also, since the vi form an orthonormal basis, we have from

expressing the unit vector ej in this basis that
n
X
(17) |vi,j |2 = 1
i=1

This is a system of n linear equations in n unknowns |vi,j |2 . For sake of notation

let use permutation symmetry to set i = n. From a further application of Cramer’s
rule, one can then write
det(S 0 )
|vn,j |2 =
det(S)
1
where S is the n×n matrix with ki entry equal to λk (Mj )−λi (A)
when k = 1, . . . , n−1,
0
and equal to 1 when k = n, and S is the minor of S formed by removing the nth
row and column. Using the well known Cauchy determinant identity [Cau41]
Q
1≤j<i≤n (xi − xj )(yi − yj )
 
1
(18) det = Qn Qn
xi − yj 1≤i,j≤n i=1 j=1 (xi − yj )

and inspecting the asymptotics as xn → ∞, we soon arrive at the identities

Q Q
1≤l<k≤n−1 (λk (Mj ) − λl (Mj )) 1≤l<k≤n (λk (A) − λl (A))
det(S) = Qn−1 Qn
l=1 k=1 (λl (Mj ) − λk (A))

and Q
0 1≤l<k≤n−1 (λk (Mj ) − λl (Mj ))(λk (A) − λl (A))
det(S ) = Qn−1 Qn−1
l=1 k=1 (λl (Mj ) − λk (A))

and the identity (2) then follows after a brief calculation.

Remark 8. The derivation of the eigenvector-eigenvalue identity (2) from (16), as
well as the obvious normalization (17), is reversible. Indeed, the identity (2) implies
that the rational functions on both sides of (13) have the same residues at each of
their (simple) poles, and these functions decay to zero at infinity, hence they must
agree by Liouville’s theorem. Specializing (13) to λ = λk (Mj ) then recovers (16),
while comparing the leading asymptotics of both sides of (13) as λ → ∞ recovers
(17) (thus completing half of the consistency check (ix) from the introduction). As
the identity (16) involves the same quantities vi,j , λk (Mj ), λi (A) as (2), one can
thus view (16) as an equivalent formulation of the eigenvector-eigenvalue identity,
at least in the case when all the eigenvalues of A are distinct; the identity (13)
(viewing λ as a free parameter) can also be interpreted in this fashion.
Remark 9. The above resolvent-based arguments have a good chance of being ex-
tended to certain classes of infinite matrices, or other Hermitian operators, partic-
ularly if they have good spectral properties (e.g., they are trace class). Certainly it
is well known that spectral projections of an operator to a single eigenvalue λ can
often be viewed as residues of the resolvent at that eigenvalue, in the spirit of (14),
under various spectral hypotheses of the operator in question. The main difficulty is
to find a suitable extension of Cramer’s rule to infinite-dimensional settings, which
would presumably require some sort of regularized determinant such as the Fredholm
determinant. We will not explore this question further here.
 EIGENVECTORS FROM EIGENVALUES 11

2.3. Coordinate-free proof. We now give a proof that largely avoids the use
of coordinates or matrices, essentially due to Bo Berndtsson5. For this proof we
assume familiarity with exterior algebra (see e.g., [BM41, Chapter XVI]). The key
identity is the following statement.
Lemma 10 (Coordinate-free eigenvector-eigenvalue identity). Let T : Cn → Cn be
a self-adjoint linear map that annihilates a unit vector v. For each unit vector f ∈
Cn , let ∆T (f ) be the determinant of the quadratic form w 7→ (T w, w)Cn restricted
to the orthogonal complement f ⊥ := {w ∈ Cn : (f, w)Cn = 0}, where (, )Cn denotes
the Hermitian inner product on Cn . Then one has
(19) |(v, f )Cn |2 ∆T (v) = ∆T (f )
for all unit vectors f ∈ Cn .
Proof. The determinant of a quadratic form w 7→ (T w, w)Cn on a k-dimensional
subspace V of Cn can be expressed as (T α, α)Vk Cn /(α, α)Vk Cn for any non-degenerate
Vk Vk n
element α of the k th exterior power V ⊂ C (equipped with the usual Her-
Vk n
mitian inner product (, ) k Cn ), where the operator T is extended to
V C in the
n−1 n
usual fashion. If f ∈ Cn is a unit vector, then the Hodge dual ∗f ∈
V
C is a
Vn−1 ⊥
unit vector in (f ), so that we have the identity
(20) ∆T (f ) = (T (∗f ), ∗f )Vn−1 Cn .
To prove (19), it thus suffices to establish the more general identity
(21) (f, v)Cn ∆T (v)(v, g)Cn = (T (∗f ), (∗g))Vn−1 Cn
for all f, g ∈ Cn . If f is orthogonal
Vn−2 ton v then ∗f can be expressed as a wedge
product of v with an element of C , and hence T (∗f ) vanishes, so that (21)
holds in this case. If g is orthogonal to v then we again obtain (21) thanks to the
self-adjoint nature of T . Finally, when f = g = v the claim follows from (20). Since
the identity (21) is sesquilinear in f, g, the claim follows. 
Now we can prove (2). Using translation symmetry we may normalize λi (A) = 0.
We apply Lemma 10 to the self-adjoint map T : w 7→ Aw, setting v to be the
null vector v = vi and f to be the standard basis vector Qn ej . Working in the
orthonormal eigenvector basis v1 , . . . , vn we have ∆(v) = k=1;k6=i λk (A); working
Qn−1
in the standard basis e1 , . . . , en we have ∆T (f ) = det(Mj ) = k=1 λk (Mj ). Finally
we have (v, f )Cn = vi,j . The claim follows.
Remark 11. In coordinates, the identity (20) may be rewritten as ∆T (f ) =
f ∗ adj(A)f . Thus we see that Lemma 10 is basically (8) in disguise.
2.4. Proof using perturbative analysis. Now we give a proof using perturba-
tion theory, which to our knowledge first appears in [MD89]. By the usual limiting
argument we may assume that A has simple eigenvalues. Let ε be a small param-
eter, and consider the rank one perturbation A + εej e∗j of A, where e1 , . . . , en is
the standard basis. From (3) and cofactor expansion, the characteristic polynomial
pA+εej e∗j (λ) of this perturbation may be expanded as
pA+εej e∗j (λ) = pA (λ) − εpMj (λ) + O(ε2 ).

5terrytao.wordpress.com/2019/08/13/eigenvectors-from-eigenvalues/#comment-519914
12 PETER B. DENTON, STEPHEN J. PARKE, TERENCE TAO, AND XINING ZHANG

On the other hand, from perturbation theory the eigenvalue λi (A + εej e∗j ) may be
expanded as
λi (A + εej e∗j ) = λi (A) + ε|vi,j |2 + O(ε2 ).
If we then Taylor expand the identity
pA+εej e∗j (λi (A + εej e∗j )) = 0
and extract the terms that are linear in ε, we conclude that
ε|vi,j |2 p0A (λi (A)) − εpMj (λi (A)) = 0
which gives (4) and hence (2).

2.5. Proof using a Cauchy-Binet type formula. Now we give a proof based on
a Cauchy-Binet type formula, which is also related to Lemma 10. This argument
first appeared in [DPTZ19].
Lemma 12 (Cauchy-Binet type formula). Let A be an n × n Hermitian matrix
with a zero eigenvalue λi (A) = 0. Then for any n × n − 1 matrix B, one has

2
det(B ∗ AB) = (−1)n−1 p0A (0) det B vi 


where B vi denotes the n × n matrix with right column vi and all remaining
columns given by B.
Proof. We use a perturbative argument related to that in Section 2.4. Since Avi =
0, vi∗ A = 0, and vi∗ vi = 1, we easily confirm the identity
 ∗
−B ∗ AB + O(ε) O(ε)
 
B

(εIn − A) B v i =
vi∗ O(ε) ε
for any parameter ε, where the matrix on the right-hand side is given in block form,
with the top left block being an n − 1 × n − 1 matrix and the bottom right entry
being a scalar. Taking determinants of both sides, we conclude that

2
pA (ε) det B vi  = (−1)n−1 det(B ∗ AB ∗ )ε + O(ε2 ).


Extracting out the ε coefficient of both sides, we obtain the claim. 

Remark 13. In the case when  vi is the basis vector en , we may write A in block
Mn 0n−1×1
form as A = , where 0i×j denotes the i × j zero matrix, and
 0 01×n−1 0
B
write B = for some n − 1 × n − 1 matrix B 0 and n − 1-dimensional vector x,
x∗
in which case one can calculate
det(B ∗ AB) = det((B 0 )∗ Mn B 0 ) = det(Mn )|det(B 0 )|2
and
vi = det(B 0 ).


det B
Since p0A (0) = (−1)n−1 det(Mn ) in this case, this establishes (12) in the case vi =
en . The general case can then be established from this by replacing A by U AU ∗ and
B by U B, where U is any unitary matrix that maps vi to en .
 EIGENVECTORS FROM EIGENVALUES 13

We now prove (4) and hence (2). Using the permutation and translation sym-
metries
 we may  normalize λi (A) = 0 and j = 1. If we then apply Lemma 12 with
01×n−1
B= , in which case
In−1
det(B ∗ AB) = det(M1 ) = (−1)n−1 pM1 (0)
and


det B vi = vi,1 .
Applying Lemma 12, we obtain (4).

2.6. Proof using an alternate expression for eigenvector component mag-

nitudes. There is an alternate formula for the square |vi,j |2 of the eigenvector
component vi,j that was first introduced in the 2009 paper [ESY09, (5.8)] of Erdős,
Schlein and Yau in the context of random matrix theory, and then highlighted fur-
ther in a 2011 paper [TV11, Lemma 41] of the third author and Vu; it differs from
the eigenvector-eigenvalue formula in that it involves the actual coefficients of A
and M1 , rather than just their eigenvalues. For sake of notation we just give the
formula in the j = 1 case.
Lemma 14 (Alternate expression for vi,1 ). Let A be an n × n Hermitian matrix
written in block matrix form as
a11 X ∗
 
A=
X M1
where X is an n − 1-dimensional column vector and a11 is a scalar. Let i =
1, . . . , n, and suppose that λi (A) is not an eigenvalue of M1 . Let u1 , . . . , un−1
denote an orthonormal basis of eigenvectors of M1 , associated to the eigenvalues
λ1 (M1 ), . . . , λn−1 (M1 ). Then
1
(22) |vi,1 |2 = Pn−1 .
1+ j=1 X ∗ (λ i (A)In−1 − M1 )−2 X
This lemma is useful in random matrix theory for proving delocalization of eigen-
vectors of random matrices, which roughly speaking amounts to proving upper
bounds on the quantity sup1≤j≤n |vi,j |.
Proof. One can verify that this result enjoys the same translation symmetry as
Theorem 1 (see consistency check (ii) from the introduction),
 sowithout loss of
vi,1
generality we may normalize λi (A) = 0. If we write vi = for an n − 1-
wi
dimensional column vector wi , then the eigenvector equation Avi = λi (A)vi = 0
can be written as the system
a11 vi,1 + X ∗ wi = 0
Xvi,1 + M1 wi = 0.
By hypothesis, 0 is not an eigenvalue of M1 , so we may invert M1 and conclude
that
wi = −M1−1 Xvi,1 .
Since vi is a unit vector, we have |wi |2 + |vi,1 |2 = 1. Combining these two formulae
and using some algebra, we obtain the claim. 
14 PETER B. DENTON, STEPHEN J. PARKE, TERENCE TAO, AND XINING ZHANG

Now we can give an alternate proof of (4) and hence (2). By permutation
symmetry (iii) it suffices to establish the j = 1 case. Using limiting arguments as
before we may assume that A has distinct eigenvalues; by further limiting arguments
we may also assume that the eigenvalues of M1 are distinct from those of A. By
translation symmetry (ii) we may normalize λi (A) = 0. Comparing (4) with (22),
our task reduces to establishing the identity
p0A (0) = pM1 (0)(1 + X ∗ M1−2 X).
However, for any complex number λ not equal to an eigenvalue of M1 , we may
apply Schur complementation [Cot74] to the matrix
−X ∗
 
λ − a11
λIn − A =
−X λIn−1 − M1
to obtain the formula
det(λIn − A) = det(λIn−1 − M1 )(λ − a11 − X ∗ (λIn−1 − M1 )−1 X)
or equivalently
pA (λ) = pM1 (λ)(λ − a11 − X ∗ (λIn−1 − M1 )−1 X)
which on Taylor expansion around λ = 0 using pA (0) = 0 gives
p0A (0)λ + O(λ2 ) = (pM1 (0) + O(λ))(λ − a11 + X ∗ M1−1 X + λX ∗ M1−2 X + O(λ2 )).
Setting λ = 0 and using pM1 (0) 6= 0, we conclude that a11 + X ∗ M1−1 X vanishes. If
we then extract the λ coefficient, we obtain the claim.
Remark 15. The same calculations also give the well known fact that the minor
eigenvalues λ1 (M1 ), . . . , λn−1 (M1 ) are precisely the roots for the equation
λ − a11 − X ∗ (λIn−1 − M1 )−1 X = 0.
Among other things, this can be used to establish the interlacing inequalities (1).
2.7. A generalization. The following generalization of the eigenvector-eigenvalue
identity was recently observed by Yu Qing Tang (private communication), relying
primarily on the Cauchy-Binet formula and a duality relationship (23) between the
various minors of a unitary matrix. If A is an n × n matrices and I, J are subsets of
{1, . . . , n} of the same cardinality m, let MI,J (A) denote the n − m × n − m minor
formed by removing the m rows indexed by I and the m columns indexed by J.
Proposition 16 (Generalized eigenvector-eigenvalue identity). Let A be a normal
n × n matrix diagonalized as A = U DU ∗ for some unitary U and diagonal D =
diag(λ1 , . . . , λn ), let 1 ≤ m < n, and let I, J, K ⊂ {1, . . . , n} have cardinality m.
Then
Y Y
detMJ c ,I c (U )(detMK c ,I c (U )) (λj − λi ) = detMJ,K ( (A − λi In ))
i∈I,j∈I c i∈I

where I denotes the complement of I in {1, . . . , n}, and similarly for J c , K c .

c

Note that if we set m = 1, I = {i}, and J = K = {j}, we recover (2). The

identity (10) can be interpreted as the remaining m = 1 cases of this proposition.
 EIGENVECTORS FROM EIGENVALUES 15

Proof. We have Y Y
(A − λi In ) = U (D − λi In )U ∗
i∈I i∈I
and hence by the Cauchy-Binet formula
Y X Y
detMJ,K ( (A−λi In )) = (detMJ,L (U ))(detML,L0 ( (D−λi In )))(detML0 ,K (U ∗ ))
i∈I L,L0 i∈I
0
Qof {1, . . . , n} of cardinality m. A computation
where L, L range over subsets
0
reveals
that the quantity detML,L 0(
Q i∈I (D − λ i In )) vanishes unless L = L = I, in which
case the quantity equals i∈I,j∈I c (λj − λi ). Thus it remains to show that
detMJ c ,I c (U )(detMK c ,I c (U )) = detMJ,I (U )detMI,K (U ∗ ).
Since detMI,K (U ∗ ) = detMK,I (U ), it will suffice to show that
(23) detMJ,I (U ) = detMJ c ,I c (U )detU
for any J, I ⊂ {1, . . . , n} of cardinality m. By permuting rows and columns we may
assume that J = I = {1, . . . , m}. If we split the identity matrixIn into the
 left
I m 0 m×n−m
m columns In1 := and the right n − m columns In2 := and
0n−m×m In−m
take determinants of both sides of the identity
U U ∗ In1 In2 = In1 U In2

we conclude that
det(U )detMI c ,J c (U ∗ ) = detMJ,I (U )
giving the claim. 

3. History of the identity

In this section we present, roughly in chronological order, all the references to
the eigenvector-eigenvalue identity (2) (or closely related results) that we are aware
of in the literature. For the primary references, we shall present the identity in
the notation of that reference in order to highlight the diversity of contexts and
notational conventions in which this identity has appeared.
An identity related to (2) appears in a paper of Löwner [L3̈4, (7)]. In this paper,
a diagonal quadratic form
Xn
An (x, x) = λi x2i
i=1
is considered, as well as a rank one perturbation
n
!2
X
Bn (x, x) = An (x, x) + αi xi
i=1
for some real numbers λ1 , . . . , λn , α1 , . . . , αn . If the eigenvalues of the quadratic
form Bn are denoted µ1 , . . . , µn . If the eigenvalues are arranged in non-decreasing
order, one has the interlacing inequalities
λ1 ≤ µ1 ≤ λ2 ≤ · · · ≤ λn ≤ µn
(compare with (1)). Under the non-degeneracy hypothesis
λ1 < µ1 < λ2 < · · · < λn < µn
16 PETER B. DENTON, STEPHEN J. PARKE, TERENCE TAO, AND XINING ZHANG

the identity
(µ1 − λk )(µ2 − λk ) . . . (µn − λk )
(24) αk2 =
(λ1 − λk )(λ2 − λk ) . . . (λk−1 − λk )(λk+1 − λk ) . . . (λn − λk )
is established for k = 1, . . . , n, which closely resembles (2). The identity (24) is
obtained via “Eine einfache Rechnung” (an easy calculation) from the standard
relations
n
X αk2
=1
µi − λk
k=1
for i = 1, . . . , n (compare with (16)), after applying Cramer’s rule and the Cauchy
determinant identity (18); as such, it is very similar to the proof of (2) in Section
2.2 that is also based on (18). The identity (24) was used in [L3̈4] to help classify
monotone functions of matrices. It can be related to (2) as follows. For sake of
notation let us just consider the j = n case of (2). Let ε be a small parameter
and consider the perturbation en e∗n + εA of the rank one matrix en e∗n . Standard
perturbative analysis reveals that the eigenvalues of this perturbation consist of
n − 1 eigenvalues of the form ελi (Mn ) + O(ε2 ) for i = 1, . . . , n − 1, plus an outlier
eigenvalue at 1 + O(ε). Rescaling, we see that the rank one perturbation A + 1ε en e∗n
of A has eigenvalues of the form λi (Mn ) + O(ε) for i = 1, . . . , n − 1, plus an outlier
eigenvalue at 1ε + O(1). If we let An , Bn be the quadratic forms associated to
A, A + 1ε en e∗n expressed using the eigenvector basis v1 , . . . , vn , the identity (24)
becomes Qn
1 (λi (A + 1ε en e∗n ) − λk (A))
|vk,n |2 = i=1
Qn .
ε i=1;i6=k (λi (A) − λk (A))
Extracting the 1/ε component of both sides of this identity using the aforementioned
perturbative analysis, we recover (2) after a brief calculation.
A more complicated variant of (24) involving various quantities related to the
Rayleigh-Ritz method of bounding the eigenvalues of a symmetric linear operator
was stated by Weinberger [Wei60, (2.29)], where it was noted that it can be proven
much the same method as in [L3̈4].
The first appearance of the eigenvector-eigenvalue identity in essentially the form
presented here that we are aware of was by Thompson [Tho66, (15)], which does not
reference the prior work of Löwner or Weinberger. In the notation of Thompson’s
paper, A is a normal n × n matrix, and µ1 , . . . , µs are the distinct eigenvalues of A,
with each µi occurring with multiplicity ei . To avoid non-degeneracy it is assumed
that s ≥ 2. One then diagonalizes A = U DU −1 for a unitary U and diagonal
D = diag(λ1 , . . . , λn ), and then sets
X
θiβ = |Uij |2
j:λj =µβ

for i = 1, . . . , n and β = 1, . . . , n − 1, where Uij are the coefficients of U . The

minor formed by removing the ith row and column from A is denoted A(i|i); it has
“trivial” eigenvalues in which each µi with ei > 1 occurs with multiplicity ei − 1,
as well as additionally some “non-trivial” eigenvalues ξi1 , . . . , ξi,s−1 . The equation
[Tho66, (15)] then reads
s−1
Y s
Y
(25) θiα = (µα − ξij ) (µα − µj )−1
j=1 j=1,j6=α
 EIGENVECTORS FROM EIGENVALUES 17

for 1 ≤ α ≤ s and 1 ≤ i ≤ n. If one specializes to the case when A is Hermitian

with simple spectrum, so that all the multiplicities ei are equal to 1, and set s = n
and µi = λi , it is then not difficult to verify that this identity is equivalent to
the eigenvector-eigenvalue identity (2) in this simple spectrum case. In the case
of repeated eigenvalues, the eigenvector-eigenvalue identity (2) may degenerate (in
that the left and right hand sides both vanish), but the identity (25) remains non-
trivial in this case. The proof of (25) given in [Tho66] is written using a rather
complicated notation (in part because much of the paper was concerned with more
general k × k minors rather than the n − 1 × n − 1 minors A(i|i)), but is essentially
the adjugate proof from Section 2.1 (where the adjugate matrix is replaced by the
closely related (n − 1)th compound matrix). In [Tho66], the identity (25) was not
highlighted as a result of primary interest in its own right, but was instead employed
to establish a large number of inequalities between the eigenvalues µ1 , . . . , µs and
the minor eigenvalues ξi1 , . . . , ξi,s−1 in the Hermitian case; see [Tho66, Section 5].
In followup paper [TM68] by Thompson and McEnteggert, the analysis from
[Tho66] was revisited, restricting attention specifically to the case of an n × n
Hermitian matrix H with simple eigenvalues λ1 < · · · < λn , and with the minor
H(i|i) formed by deleting the ith row and column having eigenvalues ξi1 ≤ · · · ≤
ξi,n−1 . In this paper the inequalities
n
X λj − ξi,j−1 ξij − λj
(26) ≥1
i=1
λj − λj−1 λj+1 − λj

and
n
X λj − ξi,j−1 ξij − λj
(27) ≤1
i=1
λj − λ1 λn − λj

for 1 ≤ j ≤ n were proved (with most cases of these inequalities already established
in [Tho66]), with a key input being the identity
      
2 λj − ξi1 λj − ξi,j−1 ξij − λj ξi,n−1 − λj
(28) |uij | = ... ...
λj − λ1 λj − λj−1 λj+1 − λj λn − λj
where uij are the components of the unitary matrix U used in the diagonalization
H = U DU −1 of H. Note from the Cauchy interlacing inequalities that each of the
expressions in braces takes values between 0 and 1. It is not difficult to see that
this identity is equivalent to (2) (or (25)) in the case of Hermitian matrices with
simple eigenvalues, and the hypothesis of simple eigenvalues can then be removed
by the usual limiting argument. As in [Tho66], the identity is established using
adjugate matrices, essentially by the argument given in the previous section. How-
ever, the identity (28) is only derived as an intermediate step towards establishing
the inequalities (26), (27), and is not highlighted as of interest in its own right. The
identity (25) was then reproduced in a further followup paper [Tho69], in which the
identity (13) was also noted; this latter identity was also independently observed
in [DH78].
In the text of Šilov [Š69, Section 10.27], the identity (2) is established, essentially
by the Cramer rule method. Namely, if A(x, x) is a diagonal real quadratic form
on Rn with eigenvalues λ1 ≥ · · · ≥ λn , and Rn−1 is a hyperplane in Rn with unit
normal vector (α1 , . . . , αn ), and µ1 ≥ · · · ≥ µn−1 are the eigenvalues of A on Rn−1 ,
18 PETER B. DENTON, STEPHEN J. PARKE, TERENCE TAO, AND XINING ZHANG

then it is observed that

Qn−1
k=1 (µk − λ)
(29) αk2 =
(λk − λ1 ) . . . (λk − λk−1 )(λk − λk+1 ) . . . (λk − λn )
for k = 1, . . . , n, which is (2) after changing to the eigenvector basis; identities
equivalent to (16) and (4) are also established. The text [Š69] gives no references,
but given the similarity of notation with [L3̈4] (compare (29) with (24)), one could
speculate that Šilov was influenced by Löwner’s work.
In a section [Pai71, Section 8.2] of the PhD thesis of Paige entitled “A Useful
Theorem on Cofactors”, the identity (28) is cited as “a fascinating theorem ... that
relates the elements of the eigenvectors of a symmetric to its eigenvalues and the
eigenvalues of its principal submatrices”, with a version of the adjugate proof given.
In the notation of that thesis, one considers a k×k real symmetric tridiagonal matrix
C with distinct eigenvalues µ1 > · · · > µk with an orthonormal of eigenvectors
y1 , . . . , yk . For any 0 ≤ r < j ≤ k, let Cr,j denote the j − r × j − r minor of
C defined by taking the rows and columns indexed between r + 1 and j, and let
pi,j (µ) := det(µIj−r − Cr,j ) denote the associated trigonometric polynomial. The
identity
2
(30) yri = p0,r−1 (µi )pr,k (µi )/f (i)
is then established for i = 1, . . . , n, where yri is the ith component of yr and
Qk
f (i) := r=1;r6=i (µi − µr ). This is easily seen to be equivalent to (2) in the case of
real symmetric tridiagonal matrices with distinct eigenvalues. One can then use this
to derive (2) for more general real symmetric matrices by a version of the Lanczos
algorithm for tridiagonalizing an arbitrary real symmetric matrix, followed by the
usual limiting argument to remove the hypothesis of distinct eigenvalues; we leave
the details to the interested reader. Returning to the case of tridiagonal matrices,
Paige also notes that the method also gives the companion identity
(31) f (i)yri ysi = δr+1 . . . δs p0,r−1 (µi )ps,k (µi )
for 1 ≤ r < s ≤ k, where δ2 , . . . , δk are the upper diagonal entries of the tridiagonal
matrix C; this can be viewed as a special case of (10). These identities were then
used in [Pai71, Section 8] as a tool to bound the behavior of errors in the symmetric
Lanczos process.
Paige’s identities (30), (31) for tridiagonal matrices are reproduced in the text-
book of Parlett [Par80, Theorem 7.9.2], with slightly different notation. Namely,
one starts with an n×n real symmetric tridiagonal matrix T , decomposed spectrally
as SΘS ∗ where S = (s1 , . . . , sn ) is orthogonal and Θ = diag(θ1 , . . . , θn ). Then for
1 ≤ µ ≤ ν ≤ n and 1 ≤ j ≤ n, the j th component sµj of sµ is observed to obey the
formula
s2µj = χ1:µ−1 (θj )χµ+1:n (θj )/χ01:n (θj )
when θj is a simple eigenvalue, where χi:j is the characteristic polynomial of the
j − i + 1 × j − i + 1 minor of T formed by taking the rows and columns between i
and j. This identity is essentially equivalent to (30). The identity (30) is similarly
reproduced in this notation; much as in [Pai71], these identities are then used to
analyze various iterative methods for computing eigenvectors. The proof of the
theorem is left as an exercise in [Par80], with the adjugate method given as a
hint. Essentially the same result is also stated in the text of Golub and van Loan
[GVL83, p. 432–433] (equation (8.4.12) on page 474 in the 2013 edition), proven
 EIGENVECTORS FROM EIGENVALUES 19

using a version of the Cramer rule arguments in Section 2.2; they cite as reference
the earlier paper [Gol73, (3.6)], which also uses essentially the same proof. A
similar result was stated without proof by Galais, Kneller, and Volpe [GKV12,
Equations (6), (7)]. They provided expressions for both |vi,j |2 and the off-diagonal
eigenvectors as a function of cofactors in place of adjugate matrices. Their work
was in the context of neutrino oscillations.
The identities of Parlett and Golub-van Loan are cited in the thesis of Knyazev
[Kny86, (2.2.27)], again to analyze methods for computing eigenvalues and eigen-
vectors; the identities of Golub-van Loan and Šilov are similarly cited in the paper
of Knyazev and Skorokhodov [KS91] for similar purposes. Parlett’s result is also
reproduced in the text of Xu [Xu95, (3.19)]. In the survey [CG02, (4.9)] of Chu
and Golub on structured inverse eigenvalue problems, the eigenvector-eigenvalue
identity is derived via the adjugate method from the results of [TM68], and used
to solve the inverse eigenvalue problem for Jacobi matrices; the text [Par80] is also
cited.
In the paper [DLNT86, page 210] of Deift, Li, Nanda, and Tomei, the eigenvector-
eigenvalue identity (2) is derived by the Cramer’s rule method, and used to con-
struct action-angle variables for the Toda flow. The paper cites [BG78], which also
reproduces (2) as equation (1.5) of that paper, and in turn cites [Gol73].
In the paper of Mukherjee and Datta [MD89] the eigenvector-eigenvalue identity
was rediscovered, in the context of computing eigenvectors of graphs that arise in
chemistry. If G is a graph on n vertices v1 , . . . , vn , and G − vr is the graph on
n − 1 vertices formed by deleting a vertex vr , r = 1, . . . , n, then in [MD89, (4)] the
identity
(32) P (G − vr ; xj ) = P 0 (G; xj )Crj
2

is established for j, r = 1, . . . , n, where P (G; x) denotes the characteristic polyno-

mial of the adjacency matrix of G evaluated at x, and Crj is the coefficient at the
rth vertex of the eigenvector corresponding to the j th eigenvalue, and one assumes
that all the eigenvalues of G are distinct. This is equivalent to (4) in the case that
A is an adjacency matrix of a graph. The identity is proven using the perturba-
tive method in Section 2.4, and appears to have been discovered independently. A
similar identity was also noted in the earlier work of Li and Feng [LF79], at least
in the case j = 1 of the largest eigenvalue. In a later paper of Hagos [Hag02], it
is noted that the identity (32) “is probably not as well known as it should be”,
and also carefully generalizes (32) to an identity (essentially the same as (15)) that
holds when some of the eigenvalues are repeated. An alternate proof of (32) was
given in the paper of Cvetkovic, Rowlinson, and Simic [CRS07, Theorem 3.1], es-
sentially using the Cramer rule type methods in Section 2.2. The identity (13) is
also essentially noted at several other locations in the graph theory literature, such
as [God93, Chapter 4], [GM81, Lemma 2.1], [God12, Lemma 7.1, Corollary 7.2],
[GGKL17, (2)] in relation to the generating functions for walks on a graph, though
in those references no direct link to the eigenvector-eigenvalue identity in the form
(2) is asserted.
In [NTU93, Section 2] the identity (2) is derived for normal matrices by the
Cramer rule method, citing [Tho69], [DH78] as the source for the key identity (13);
the papers [Tho66], [TM68] also appear in the bibliography but were not directly
cited in this section. An extension to the case of eigenvalue multiplicity, essentially
corresponding to (15), is also given. This identity is then used to give a complete
20 PETER B. DENTON, STEPHEN J. PARKE, TERENCE TAO, AND XINING ZHANG

description of the relations between the eigenvalues of A and of a given minor Mj

when A is assumed to be normal. In [BFdP11] a generalization of these results
was given to the case of J-normal matrices for some diagonal sign matrix J; this
corresponds to a special case of (11) in the case where each left eigenvector wi is
the complex conjugate of Jvi .
The paper of Baryshnikov [Bar01] marks the first appearance of this identity
in random matrix theory. Let H be a Hermitian form on CM with eigenvalues
λ1 ≥ · · · ≥ λM , and let L be a hyperplane of CM orthogonal to some unit vector
l. Let li be the component of l with respect to an eigenvector vi associated to λi ,
set wi := |li |2 , and let µ1 ≥ · · · ≥ µM −1 be the eigenvalues of the Hermitian form
arising from restricting H to L. Then after [Bar01, (4.5.2)] (and correcting some
typos) the identity
Q
1≤j≤M −1 (λi − µj )
wi = Q
1≤j≤M ;j6=i (λi − λj )
is established, by an argument based on Cramer’s rule and the Cauchy determinant
formula (18), similar to the arguments at the end of Section 2.2, and appears to have
been discovered independently. If one specializes to the case when l is a standard
basis vector ej then li is also the ej component of vi , and we recover (2) after a brief
calculation. This identity was employed in [Bar01] to study the situation in which
the hyperplane normal l was chosen uniformly at random on the unit sphere. This
formula was rederived (using a version of the Cramer rule method in Section 2.2) in
the May 2019 paper of Forrester and Zhang [FZ19, (2.7)], who recover some of the
other results in [Bar01] as well, and study the spectrum of the sum of a Hermitian
matrix and a random rank one matrix.
In the paper [DE02, Lemma 2.7] of Dumitriu and Edelman, the identity (30) of
Paige (as reproduced in [Par80, Theorem 7.9.2]) is used to give a clean expression
for the Vandermonde determinant of the eigenvalues of a tridiagonal matrix, which
is used in turn to construct tridiagonal models for the widely studied β-ensembles
in random matrix theory.
In the unpublished preprint [Van14] of Van Mieghem, the identity (4) is promi-
nently displayed as the main result, though in the notation of that preprint it is
expressed instead as
1
(xk )2j = − det(A\{j} − λk In )
c0A (λk )
for any j, k = 1, . . . , n, where A is a real symmetric matrix with distinct eigenvalues
λ1 , . . . , λn and unit eigenvectors x1 , . . . , xn , A\{j} is the minor formed by removing
the j th row and column from A, and c0A is the derivative of the (sign-reversed)
characteristic polynomial cA (λ) = det(A − λIn ) = (−1)n pA (λ). Two proofs of this
identity are given, one being essentially the Cramer’s rule proof from Section 2.2 and
attributed to the previous reference [CRS07]; the other proof is based on Cramer’s
rule and the Desnanot-Jacobi identity (Dodgson condensation); this identity is used
to quantify the effect of removing a node from a graph on the spectral properties
of that graph. The related identity (22) from [TV11] is also noted in this preprint.
Some alternate formulae from [VM11] for quantities such as (xk )2j in terms of walks
of graphs are also noted, with the earlier texts [God93], [GVL83] also cited.
The identity (2) was independently rediscovered and then generalized by Kausel
[Kau18], as a technique to extract information about components of a generalized
 EIGENVECTORS FROM EIGENVALUES 21

eigenmode without having to compute the entire eigenmode. Here the generalized
eigenvalue problem
Kψj = λj Mψj
for j = 1, . . . , N is considered, where K is a positive semi-definite N × N real
symmetric matrix, M is a positive definite N × N real symmetric matrix, and the
matrix Ψ = (ψ1 . . . ψN ) of eigenfunctions is normalized so that ΨT M Ψ = I. For
any 1 ≤ α ≤ N , one also solves the constrained system
(α) (α)
Kα ψj = λαj Mα ψj
where Kα , Mα are the N − 1 × N − 1 minors of K, M respectively formed by
removing the αth row and column. Then in [Kau18, (18)] the Cramer rule method
is used to establish the identity
s v
u QN −1
ψαj = ± t Q k=1 (λj − λαk )
|Mα | u
|M| N −1
k=1;k6=j (λj − λk )

for the α component ψαj of ψj , where |M| is the notation in [Kau18] for the
determinant of M. Specializing to the case when M is the identity matrix, we
recover (2).
The eigenvector-eigenvalue identity was discovered by three of us [DPZ19] in
July 2019, initially in the case of 3 × 3 matrices, in the context of trying to find
a simple and numerically stable formula for the eigenvectors of the neutrino os-
cillation Hamiltonian, which form a separate matrix known as the PMNS lepton
mixing matrix. This identity was established in the 3 × 3 case by direct calcula-
tion. Despite being aware of the related identity (22), the four of us were unable
to locate this identity in past literature and wrote a preprint [DPTZ19] in August
2019 highlighting this identity and providing two proofs (the adjugate proof from
Section 2.1, and the Cauchy-Binet proof from Section 2.5). The release of this
preprint generated some online discussion6, and we were notified by Jiyuan Zhang
(private communication) of the prior appearance of the identity earlier in the year
in [FZ19]. However, the numerous other places in the literature in which some form
of this identity appeared did not become revealed until a popular science article
[Wol19] by Wolchover was written in November 2019. This article spread awareness
of the eigenvector-eigenvalue identity to a vastly larger audience, and generated a
large number of reports of previous occurrences of the identity, as well as other
interesting related observations, which we have attempted to incorporate into this
survey.

4. Further discussion
The eigenvector-eigenvalue identity (2) only yields information about the magni-
tude |vi,j | of the components of a given eigenvector vi , but does not directly reveal
the phase of these components. On one hand, this is to be expected, since (as al-
ready noted in the consistency check (vii) in the introduction) one has the freedom
to multiply vi by a phase; for instance, even if one restricts attention to real sym-
metric matrices A and requires the eigenvectors to be real vi , one has the freedom
to replace vi by its negation −vi , so the sign of each component vi,j is ambiguous.

6terrytao.wordpress.com/2019/08/13, www.reddit.com/r/math/comments/cq3en0
22 PETER B. DENTON, STEPHEN J. PARKE, TERENCE TAO, AND XINING ZHANG

However, relative phases, such as the phase of vi,j vi,j 0 are not subject to this ambi-
guity. There are several ways to try to recover these relative phases. One way is to
employ the off-diagonal analogue (10) of (2), although the determinants in that for-
mula may be difficult to compute in general. For small matrices, it was suggested
in [MD89] that the signs of the eigenvectors could often be recovered by direct
inspection of the components of the eigenvector equation Avi = λi (A)vi . In the ap-
plication in [DPZ19], the additional phase could be recovered by a further neutrino
specific identity [Tos91]. For more general matrices, one way to retrieve such phase
information is to apply (2) in multiple bases. For instance, suppose A was real sym-
metric and the vi,j were all real. If one were to apply the eigenvector-eigenvalue
identity after changing to a basis that involved the unit vector √12 (ej + ej 0 ), then
one could use the identity to evaluate the magnitude of √12 (vi,j + vi,j 0 ). Two fur-
ther applications of the identity in the original basis would give the magnitude of
vi,j , vi,j 0 , and this is sufficient information to determine the relative sign of vi,j and
vi,j 0 .
For large unstructured matrices, it does not seem at present that the identity
(2) provides a competitive algorithm to compute eigenvectors. Indeed, to use this
identity to compute all the eigenvector component magnitudes |vi,j |, one would
need to compute all n − 1 eigenvalues of each of the n minors M1 , . . . , Mn , which
would be a computationally intensive task in general; and furthermore, an addi-
tional method would then be needed to also calculate the signs or phases of these
components. However, if the matrix is of a special form (such as a tridiagonal form),
then the identity could be of more practical use, as witnessed by the uses of this
identity (together with variants such as (31)) in the literature to control the rate
of convergence for various algorithms to compute eigenvalues and eigenvectors of
tridiagonal matrices. Also, as noted recently in [Kau18], if one has an application
that requires only the component magnitudes |v1,j |, . . . , |vn,j | at a single location j,
then one only needs to compute the the characteristic polynomial of a single minor
Mj of A at a single value λi (A), and this may be more computationally feasible.

5. Sociology of science issues

As one sees from Section 3 and Figure 1, there was some partial dissemination
of the eigenvector-eigenvalue identity amongst some mathematical communities,
to the point where it was regarded as “folklore” by several of these communities.
However, this process was unable to raise broader awareness of this identity, result-
ing in the remarkable phenomenon of multiple trees of references sprouting from
independent roots, and only loosely interacting with each other. For instance,
as discussed in the previous section, for two months after the release of our own
preprint [DPTZ19], we only received a single report of another reference [FZ19]
containing a form of the identity, despite some substantial online discussion and
the dozens of extant papers on the identity. It was only in response to the popular
science article [Wol19] that awareness of the identity finally “went viral”, leading
to what was effectively an ad hoc crowdsourced effort to gather all the prior refer-
ences to the identity in the literature. While we do not know for certain why this
particular identity was not sufficiently well known prior to these recent events, we
can propose the following possible explanations:
• The identity was mostly used as an auxiliary tool for other purposes. In al-
most all of the references discussed here, the eigenvector-eigenvalue identity
 EIGENVECTORS FROM EIGENVALUES 23

was established only in order to calculate or bound some other quantity; it

was rarely formalized as a theorem or even as a lemma. In particular, with
a few notable exceptions such as the preprint [Van14], this identity would
not be mentioned in the title, abstract, or even the introduction. In a few
cases, the identity was reproven by authors who did not seem to be fully
aware that it was already established in one of the references in their own
bibliography.
• The identity does not have a standard name, form, or notation, and does
not involve uncommon keywords. As one can see from the previous section,
the identity comes in many variants and can be rearranged in a large num-
ber of ways; furthermore, the notation used for the various mathematical
objects involved in the identity vary greatly depending on the intended ap-
plication, or on the authors involved. Also, none of the previous references
attempted to give the identity a formal name, and the keywords used to de-
scribe the identity (such as “eigenvector” or “eigenvalue”) are in extremely
common use in mathematics. As such, there are no obvious ways to use
modern search engines to locate other instances of this identity, other than
by manually exploring the citation graph around known references to that
identity. Perhaps a “fingerprint database” for identities [BT13] would be
needed before such automated searches could become possible.
• The field of linear algebra is too mature, and its domain of applicability is
too broad. The vast majority of consumers of linear algebra are not domain
experts in linear algebra itself, but rather use it as a tool for a very diverse
array of other applications. As such, the diffusion of linear algebra knowl-
edge is not guided primarily by a central core of living experts in the field,
but instead relies on more mature sources of authority such as textbooks
and lectures. Unfortunately, only a small handful of linear algebra text-
books mention the eigenvector-eigenvalue identity, thus preventing wider
dissemination of this identity.
It is not fully clear to us how best to attribute authorship for the eigenvector-
eigenvalue identity (2). The earliest reference that contains an identity that implies
(2) is by Löwner [L3̈4], but the implication is not immediate, and this reference
had only a modest impact on the subsequent literature. The paper of Thompson
[Tho66] is the first explicit place we know of in which the identity appears, and
it was propagated through citations into several further papers in the literature;
but this did not prevent the identity from then being independently rediscovered
several further times; furthermore, we are not able to guarantee that there is an
even earlier place in the literature where some form of this identity has appeared.
We propose the name “eigenvector-eigenvalue identity” for (2) on the grounds that
it is descriptive, and hopefully is a term that can be detected through search engines
by researchers looking for for identities of this form.

6. Acknowledgments
We thank Carlo Beenakker, Percy Deift, Laurent Demanet, Alan Edelman, Chris
Godsil, Aram Harrow, James Kneller, Andrew Knysaev, Manjari Narayan, Michael
Nielsen, Karl Svozil, Gang Tian, Carlos Tomei, Piet Van Mieghem, Fu Zhang,
Jiyuan Zhang, and Zhenzhong Zhang for pointing out a number of references where
24 PETER B. DENTON, STEPHEN J. PARKE, TERENCE TAO, AND XINING ZHANG

some variant of the eigenvector-eigenvalue identity has appeared in the literature,

or suggesting various extensions and alternate proofs of the identity.

References
[AS64] Milton Abramowitz and Irene A. Stegun. Handbook of mathematical functions with
formulas, graphs, and mathematical tables, volume 55 of National Bureau of Stan-
dards Applied Mathematics Series. For sale by the Superintendent of Documents, U.S.
Government Printing Office, Washington, D.C., 1964.
[Bar01] Yu. Baryshnikov. GUEs and queues. Probab. Theory Related Fields, 119(2):256–274,
2001.
[BFdP11] N. Bebiano, S. Furtado, and J. da Providência. On the eigenvalues of principal sub-
matrices of J-normal matrices. Linear Algebra Appl., 435(12):3101–3114, 2011.
[BG78] D. Boley and G. H. Golub. Inverse eigenvalue problems for band matrices. In Numer-
ical analysis (Proc. 7th Biennial Conf., Univ. Dundee, Dundee, 1977), pages 23–31.
Lecture Notes in Math., Vol. 630, 1978.
[BM41] Garrett Birkhoff and Saunders MacLane. A Survey of Modern Algebra. Macmillan
Company, New York, 1941.
[BT13] Sara C. Billey and Bridget E. Tenner. Fingerprint databases for theorems. Notices
Amer. Math. Soc., 60(8):1034–1039, 2013.
[Cau41] Augustin Louis Cauchy. Exercices d’analyse et de physique mathématique. vol. 2.
Bachelier, page 154, 1841.
[CG02] Moody T. Chu and Gene H. Golub. Structured inverse eigenvalue problems. Acta
Numer., 11:1–71, 2002.
[Cot74] Richard W. Cottle. Manifestations of the Schur complement. Linear Algebra Appl.,
8:189–211, 1974.
[Cra50] Gabriel Cramer. Introduction à l’analyse des lignes courbes algébriques. Geneva, pages
657–659, 1750.
[CRS07] D. Cvetković, P. Rowlinson, and S.K. Simić. Star complements and exceptional graphs.
Linear Algebra and its Applications, 423(1):146 – 154, 2007. Special Issue devoted to
papers presented at the Aveiro Workshop on Graph Spectra.
[DE02] Ioana Dumitriu and Alan Edelman. Matrix models for beta ensembles. J. Math. Phys.,
43(11):5830–5847, 2002.
[Den19] Peter B. Denton. Eigenvalue-eigenvector identity code. https://github.com/
PeterDenton/Eigenvector-Eigenvalue-Identity, 2019.
[DH78] Emeric Deutsch and Harry Hochstadt. On Cauchy’s inequalities for Hermitian matri-
ces. Amer. Math. Monthly, 85(6):486–487, 1978.
[DLNT86] P. Deift, L. C. Li, T. Nanda, and C. Tomei. The Toda flow on a generic orbit is
integrable. Comm. Pure Appl. Math., 39(2):183–232, 1986.
[DPTZ19] Peter B. Denton, Stephen J. Parke, Terence Tao, and Xining Zhang. Eigenvectors from
Eigenvalues. arXiv e-prints, page arXiv:1908.03795v1, Aug 2019.
[DPZ19] Peter B Denton, Stephen J Parke, and Xining Zhang. Eigenvalues: the Rosetta Stone
for Neutrino Oscillations in Matter. arXiv:1907.02534. 2019.
[ESY09] László Erdős, Benjamin Schlein, and Horng-Tzer Yau. Semicircle law on short
scales and delocalization of eigenvectors for Wigner random matrices. Ann. Probab.,
37(3):815–852, 2009.
[FZ19] Peter J. Forrester and Jiyuan Zhang. Co-rank 1 projections and the randomised Horn
problem. arXiv e-prints, page arXiv:1905.05314, May 2019.
[Gan59] F. R. Gantmacher. The theory of matrices. Vol I. Translated by K. A. Hirsch. Chelsea
Publishing Co., New York, 1959.
[GGKL17] Chris Godsil, Krystal Guo, Mark Kempton, and Gabor Lippner. State transfer in
strongly regular graphs with an edge perturbation, 2017.
[GKV12] Sbastien Galais, James Kneller, and Cristina Volpe. The neutrino-neutrino interaction
effects in supernovae: the point of view from the matter basis. J. Phys., G39:035201,
2012.
[GM81] C. D. Godsil and B. D. McKay. Spectral conditions for the reconstructibility of a
graph. J. Combin. Theory Ser. B, 30(3):285–289, 1981.
 EIGENVECTORS FROM EIGENVALUES 25

[God93] C. D. Godsil. Algebraic combinatorics. Chapman and Hall Mathematics Series. Chap-
man & Hall, New York, 1993.
[God12] Chris Godsil. When can perfect state transfer occur? Electron. J. Linear Algebra,
23:877–890, 2012.
[Gol73] Gene H. Golub. Some modified matrix eigenvalue problems. SIAM Rev., 15:318–334,
1973.
[GVL83] Gene H. Golub and Charles F. Van Loan. Matrix computations, volume 3 of Johns
Hopkins Series in the Mathematical Sciences. Johns Hopkins University Press, Balti-
more, MD, 1983.
[Hag02] Elias M. Hagos. Some results on graph spectra. Linear Algebra and its Applications,
356(1):103 – 111, 2002.
[Hal42] Paul R. Halmos. Finite Dimensional Vector Spaces. Annals of Mathematics Studies,
no. 7. Princeton University Press, Princeton, N.J., 1942.
[Kau18] Eduardo Kausel. Normalized modes at selected points without normalization. Journal
of Sound and Vibration, 420:261–268, April 2018.
[Kny86] Andrew Knyazev. Computation of eigenvalues and eigenvectors for mesh problems:
algorithms and error estimates (in Russian). PhD thesis, Department of Numerical
Mathematics, USSR Academy of sciences, Moscow, 1986.
[KS91] A. V. Knyazev and A. L. Skorokhodov. On exact estimates of the convergence rate
of the steepest ascent method in the symmetric eigenvalue problem. Linear Algebra
Appl., 154/156:245–257, 1991.
[L3̈4] Karl Löwner. Über monotone Matrixfunktionen. Math. Z., 38(1):177–216, 1934.
[LF79] Qiao Li and Ke Qin Feng. On the largest eigenvalue of a graph. Acta Math. Appl.
Sinica, 2(2):167–175, 1979.
[MD89] Asok K. Mukherjee and Kali Kinkar Datta. Two new graph-theoretical methods for
generation of eigenvectors of chemical graphs. Proceedings of the Indian Academy of
Sciences - Chemical Sciences, 101(6):499–517, Dec 1989.
[NTU93] Peter Nylen, Tin Yau Tam, and Frank Uhlig. On the eigenvalues of principal subma-
trices of normal, Hermitian and symmetric matrices. Linear and Multilinear Algebra,
36(1):69–78, 1993.
[Pai71] Christopher Conway Paige. The computation of eigenvalues and eigenvectors of very
large sparse matrices. PhD thesis, London University Institute of Computer Science,
1971.
[Par80] Beresford N. Parlett. The symmetric eigenvalue problem. Prentice-Hall, Inc., Engle-
wood Cliffs, N.J., 1980. Prentice-Hall Series in Computational Mathematics.
[Tho66] R. C. Thompson. Principal submatrices of normal and Hermitian matrices. Illinois J.
Math., 10:296–308, 1966.
[Tho69] R. C. Thompson. Principal submatrices. IV. On the independence of the eigenvalues
of different principal submatrices. Linear Algebra and Appl., 2:355–374, 1969.
[TM68] R.C. Thompson and P. McEnteggert. Principal submatrices ii: the upper and lower
quadratic inequalities. Linear Algebra and its Applications, 1:211–243, 1968.
[Tos91] S. Toshev. On T violation in matter neutrino oscillations. Mod. Phys. Lett., A6:455–
460, 1991.
[TV11] Terence Tao and Van Vu. Random matrices: Universality of local eigenvalue statistics.
Acta Math., 206(1):127–204, 2011.
[Van14] Piet Van Mieghem. Graph eigenvectors, fundamental weights and centrality metrics
for nodes in networks. arXiv e-prints, page arXiv:1401.4580, Jan 2014.
[VM11] Piet Van Mieghem. Graph spectra for complex networks. Cambridge University Press,
Cambridge, 2011.
[Š69] G. E. Šilov. Matematicheskiĭ analiz. Konechnomernye lineĭ nye prostranstva. Izdat.
“Nauka”, Moscow, 1969.
[Wei60] H. F. Weinberger. Error bounds in the Rayleigh-Ritz approximation of eigenvectors.
J. Res. Nat. Bur. Standards Sect. B, 64B:217–225, 1960.
[Wil63] J. H. Wilkinson. Rounding errors in algebraic processes. Prentice-Hall, Inc., Englewood
Cliffs, N.J., 1963.
[Wol19] Natalie Wolcholver. Neutrinos lead to unexpected discovery in basic math. Quanta
Magazine, Nov 2019.
26 PETER B. DENTON, STEPHEN J. PARKE, TERENCE TAO, AND XINING ZHANG

[Xu95] Shufang Xu. Theories and Methods of Matrix Calculations (in Chinese). Peking Uni-
versity Press, Beijing, 1995.

Department of Physics, Brookhaven National Laboratory, Upton, NY 11973, USA

E-mail address: [email protected]

Theoretical Physics Department, Fermi National Accelerator Laboratory, Batavia,

IL 60510, USA
E-mail address: [email protected]

Department of Mathematics, UCLA, Los Angeles CA 90095-1555

E-mail address: [email protected]

Enrico Fermi Institute & Department of Physics, University of Chicago, Chicago,

IL 60637, USA
E-mail address: [email protected]