0% found this document useful (0 votes)
1K views

Linear Algebra Cheat Sheet

1) The document provides formulas and explanations for concepts in linear algebra including vector spaces, subspaces, null spaces, row spaces, column spaces, ranks, linear transformations, inner products, orthonormal bases, and matrix exponentials. 2) It also summarizes methods for solving systems of equations including using the QR factorization to solve the least squares problem and using variation of parameters and the method of undetermined coefficients for solving non-homogeneous linear differential equations. 3) Eigenvectors and generalized eigenvectors are discussed as well as how to find them by solving related equations.

Uploaded by

NooraniManshad
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
1K views

Linear Algebra Cheat Sheet

1) The document provides formulas and explanations for concepts in linear algebra including vector spaces, subspaces, null spaces, row spaces, column spaces, ranks, linear transformations, inner products, orthonormal bases, and matrix exponentials. 2) It also summarizes methods for solving systems of equations including using the QR factorization to solve the least squares problem and using variation of parameters and the method of undetermined coefficients for solving non-homogeneous linear differential equations. 3) Eigenvectors and generalized eigenvectors are discussed as well as how to find them by solving related equations.

Uploaded by

NooraniManshad
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

Math 54 Cheat Sheet v1 = u1


u2 ·v1

Generalized eigenvectors If you only find one eigenvector v (even though there are supposed to be 2), then solve the
following equation for u: (A − λI)(u) = v (one solution is enough).
v2 = u2 − v1
v ·v
 
 1 1 Then: x(t) = Aeλt v + B teλt v + eλt u
Vector spaces
 
u3 ·v1 u3 ·v2
v3 = u3 − v1 − v2 Undetermined coefficients First find hom. solution. Then for xp , just like regular undetermined coefficients, except
v1 ·v1 v2 ·v2  
Subspace: If u and v are in W , then u + v are in W , and cu is in W vi a1
Then {v1 · · · vn } is an orthogonal basis for Span(B), and if wi = that instead of guessing xp (t) = aet + b cos(t), you guess aet + b cos(t), where a = is a
Nul(A): Solutions of Ax = 0. Row-reduce A. v , then {w1 · · · wn } is a2
i
Row(A): Space spanned by the rows of A: Row-reduce A and choose the rows that contain the pivots. an orthonormal basis for Span(B). vector. Then plug into x0 = Ax + f and solve for a etc.
Col(A): Space spanned by columns of A: Row-reduce A and choose the columns of A that contain the pivots Variation of parameters First hom. solution xh (t) = Ax1 (t) + Bx2 (t). Then sps
Rank(A): = Dim(Col(A)) = number of pivots QR-factorization: To find Q, apply G-S to columns of A. Then R = QT A
0
" #
Rank-Nullity theorem: Rank(A) + dim(N ul(A)) = n, where A is m × n Least-squares: To solve Ax = b in the least squares-way, solve AT Ax = AT b. xp (t) = v1 (t)x1 (t) + v2 (t)x2 (t), then solve W f (t) v1 = f , where
0
v2
Linear transformation: T (u + v) = T (u) + T (v), T (cu) = cT (u), where c is a number. Least squares solution makes kAx − bk smallest.
T is one-to-one if T (u) = 0 ⇒ u = 0 x̂ = R−1 QT b, where A = QR. f (t) −1 , integrate and solve for v (t),
   
f (t) = x (t) |
W x2 (t) . Multiply both sides by W
T is onto if Col(T ) = Rm . Rb 1 1
Inner product spaces f · g = a f (t)g(t)dt. G-S applies with this inner product as well. v2 (t), and plug back into xp . Finally, x = xh + xp
Linearly independence: a1 v1 + a2 v2 + · · · + an vn = 0 ⇒ a1 = a2 = · · · = an = 0.
Cauchy-Schwarz: |u · v| ≤ kuk kvk P∞ An tn . To calculate eAt , either diagonalize:
To show lin. ind, form the matrix of the vectors, and show that N ul(A) = {0}
Triangle inequality: ku + vk ≤ kuk + kvk Matrix exponential eAt = n=0
Linear dependence: a1 v1 + a2 v2 + · · · + an vn = 0 for a1 , a2 , · · · , an , not all zero. n!
Span: Set of linear combinations of v1 , · · · vn A = P DP −1 ⇒ eAt = P eDt P −1 , where eDt is a diagonal matrix with diag. entries eλi t . Or
Basis B for V : A linearly independent set such that Span (B) = V
To show sthg is a basis, show it is linearly independent and spans.
Symmetric matrices (A = AT ) if A only has one eigenvalue λ with multiplicity m, use eAt = eλt
Pm−1 (A−λI)n tn
n=0 n!
. Solution
0 At
To find a basis from a collection of vectors, form the matrix A of the vectors, and find Col(A). Has n real eigenvalues, always diagonalizable, orthogonally diagonalizable (A = P DP T , P is an orthogonal of x = Ax is then x(t) = e c, where c is a constant vector.
To find a basis for a vector space, take any element of that v.s. and express it as a linear combination of ’simpler’ matrix, equivalent to symmetry!).
vectors. Then show those vectors form a basis.
Dimension: Number of elements in a basis.
Theorem: If A is symmetric, then any two eigenvectors from different eigenspaces are orthogonal.
How to orthogonally diagonalize: First diagonalize, then apply G-S on each eigenspace and normalize. Then P =
Coupled mass-spring system
To find dim, find a basis and find num. elts. matrix of (orthonormal) eigenvectors, D = matrix of eigenvalues. Case N = 2
 
Theorem: If V has a basis of vectors, then every basis of V must have n vectors. Quadratic forms: To find the matrix, put the x2
i -coefficients on the diagonal, and evenly distribute the other terms. Equation: x00 = Ax, A =
−2 1
Basis theorem: If V is an n−dim v.s., then any lin. ind. set with n elements is a basis, and any set of n elts. 1 −2
For example, if the x1 x2 −term is 6, then the (1, 2)th and (2, 1)th entry of A is 3.
which spans V is a basis. Matrix of a lin. transf T with respect to bases B and C: For every vector v in B, √
Then orthogonally diagonalize A = P DP T . Proper frequencies: Eigenvalues of A are: λ = −1, −3, then proper frequencies ±i, ± 3i (± square
evaluate T (v), and express T (v) as a linear combination of vectors in C. Put the coefficients in a column vector,
and then form the matrix of the column vectors you found! Then let y = P T x, then the quadratic form becomes λ1 y1
2 + · · · + λ y 2 , where λ are the
n n i roots of eigenvalues)
eigenvalues. √   √ 
Coordinates: To find [x]B , express x in terms of the vectors in B.
     
sin π 3 sin 2 π 3
x = PB [x]B , where PB is the matrix whole columns are the vectors in B. Spectral decomposition: λ1 u1 u1 T + λ2 u2 u2 T + · · · + λn un un T Proper modes: v1 =   3  =  √2 , v = 
2
 3  =  √2 
Invertible matrix theorem: If A is invertible, then: A is row-equivalent to I, A has n pivots, T (x) = Ax is sin 2 π 3 sin 4 π − 3
3 2 3 2
one-to-one and onto, Ax = b has a unique solution for every b, AT is invertible, det(A) 6= 0, the columns
of A form a basis for Rn , N ul(A) = {0}, Rank(A) = n
Second-order and Higher-order differential Case N = 3
b −1
equations
     
a 1 d −b −2 1 0
=
c d ad−bc −c a Equation: x00 = Ax, A =  1 −2 1 
Homogeneous solutions: Auxiliary equation: Replace equation by polynomial, so y 000 becomes r 3 etc. Then find 0 1 −2
h i
A−1
 
A| I → I| √ √
the zeros (use the rational roots theorem and long division, see the ‘Diagonalization-section). ’Simple zeros’ give Proper frequencies: Eigenvalues of A: λ = −2, −2 − 2, −2 + 2, then proper frequencies
Change of basis: [x]C = PC←B [x]B (think of C as the new, cool basis)
  you ert , Repeated zeros (multiplicity m) give you Aert + Btert + · · · Ztm−1 ert , Complex zeros √
q

 q


[C | B] → I | PC←B
PC←B is the matrix whose columns are [b]C , where b is in B r = a + bi give you Aeat cos(bt) + Beat sin(bt). ± 2i, ± 2 + 2 i, ± 2− 2 i
Undetermined coefficients: y(t) = y0 (t) + yp (t), where y0 solves the hom. eqn. (equation = 0), and    √    
yp is a particular solution. To find yp : sin π sin 2 π
Diagonalization If the inhom. term is Ctm ert , then: yp = ts (Am tm · · · + A1 t + 1)ert , where if r is a root of
  4 
Proper modes: v1 = sin 2 π  = 
 2
2
1 
   4 

1

sin 4 π  =  0  , v3 =
 , v2 = 
   
 4  √  4 
Diagonalizability: A is diagonalizable if A = P DP −1 for some diagonal D and invertible P . aux with multiplicity m, then s = m, and if r is not a root, then s = 0. 
2
 −1
If the inhom term is Ctm eat sin(βt), then: sin 3 π 2 sin 6 π
A and B are similar if A = P BP −1 for P invertible √ 
4 4
yp = ts (Am tm · · · + A1 t + 1)eat cos(βt) + ts (Bm tm · · · + B1 t + 1)ert sin(βt),
  
Theorem: A is diagonalizable ⇔ A has n linearly independent eigenvectors sin 3 π 2
where s = m, if a + bi is also a root of aux with multiplicity m (s = 0 if not). cos always goes with sin  4   2 
Theorem: IF A has n distinct eigenvalues, THEN A is diagonalizable, but the opposite is not always true!!!!

sin 6 π  =  −1 
 

0 0
 and vice-versa, also, you have to look at a + bi as one entity.  4  √ 
Variation of parameters: First, make sure the leading coefficient (usually the coeff. of y 00 ) is = 1.. Then 2

Notes: A can be diagonalizable even if it’s not invertible (Ex: A =
0 0
). Not all matrices are sin 9 π 2
4

1 1
 y = y0 + yp as above. Now suppose yp (t) = v1 (t)y1 (t) + v2 (t)y2 (t), where y1 and y2 are General case (just in case!)
diagonalizable (Ex: ) " 0# −2 ···
  
0 1 y1 y2 v1 0 0 0 1 0 0 
your hom. solutions. Then 0 0 0 = f (t) . Invert the matrix and solve for v1 and v2 , and −2 ···
Consequence: A = P DP −1 ⇒ An = P D n P −1 y1 y2 v2  1 1 0 0 
How to diagonalize: To find the eigenvalues, calculate det(A − λI), and find the roots of that.
 0 1 −2 1 0 · · ·
integrate to get v1 and v2 , and finally use: yp (t) = v1 (t)y1 (t) + v2 (t)y2 (t).  
Equation: x00 = Ax, A = 
 
 . . . . . . 
To find the eigenvectors, for each λ find a basis for N ul(A − λI), which you do by row-reducing b −1
   
a 1 d −b  . . . . . . 
Rational roots theorem: If p(λ) = 0 has a rational root r = a , then a divides the constant term of p, and b Useful formulas:
c d
=
ad−bc −c a  . . . . . . 

b
divides the leading coefficient. 0 0 ··· 1 −2 1
 
R R R 2
R sec(t) = ln |sec(t) + tan(t)|, tan(t) = ln |sec(t)|, tan (t) = tan(x) − x,
Use this to guess zeros of p. Once you have a zero that works, use long division! Then A = P DP −1 , where ln(t) = t ln(t) − t 0 0 0 0 1 −2
D= diagonal matrix of eigenvalues, P = matrix of eigenvectors Linear independence: f, g, h are linearly independent if  

Complex eigenvalues If λ = a + bi, and v is an eigenvector, then A = P CP −1 , where af (t) + bg(t) + ch(t) = 0 ⇒ a = b = c = 0. To show linear dependence, do it directly. To show Proper frequencies: ±2i sin
2(N +1)
, k = 1, 2, · · · N
   
a b f (t) = f0(t) g(t)
 
P = Re(v) Im(v) , C = linear independence, form the Wronskian: W (for 2 functions),
−b a f (t) g 0 (t) 



sin
 N +1 
p  
C is a scaling of det(A) followed by a rotation by θ, where: f (t) g(t) h(t)
sin 2kπ 


cos(θ) sin(θ)

Wf (t) =  f 0 (t) g 0 (t) h0 (t)  (for 3 functions). Then pick a point t0 where det(W f (t ))
p 1 C = 0  N +1 
f 00 (t) g 00 (t) h00 (t)
 
det(A) − sin(θ) cos(θ) Proper modes: vk = 
 .


is easy to evaluate. If det 6= 0, then f, g, h are linearly independent! Try to look for simplifications before you .
Orthogonality
 
 .
 
differentiate.  
u, v orthogonal if u · v = 0. Fundamental solution set: If f, g, h are solutions and linearly independent. sin N kπ
√ N +1
kuk = u·u Largest interval of existence: First make sure the leading coefficient equals to 1. Then look at the domain of each
{u1 · · · un } is orthogonal if ui · uj = 0 if i 6= j, orthonormal if ui · ui = 1 term. For each domain, consider the part of the interval which contains the initial condition. Finally, intersect the

W ⊥ : Set of v which are orthogonal to every w in W .


intervals and change any brackets to parentheses. Harmonic oscillator: my 00 + by 0 + ky = 0 (m = inertia,
b = damping, k = stiffness)
Partial differential equations
If {u1 · · · un } is an orthogonal basis, then: Full Fourier series: f defined on (−T , T ):
y·uj P∞   
πmx + b

πmx

f (x) f m=0 am cos m sin , where:
y = c1 u1 + · · · cn un ⇒ cj =
uj ·uj Systems of differential equations a0 = 1 RT
f (x)dx
T T

Orthogonal matrix Q has orthonormal columns! Consequence:QT Q = I, QQT = Orthogonal projection 2T −T


To solve x0 = Ax: x(t) = Aeλ1 t v1 + Beλ2 t v2 + eλ3 t v3 (λi are your eigenvalues, vi are RT  
on Col(Q). your eigenvectors) am = 1 −T f (x) cos πmx
T T
kQxk = kxk Fundamental matrix: Matrix whose columns are the solutions, without the constants (the columns are solutions and b0 = 0
(Qx) · (Qy) = x ·y RT  
linearly independent) bm = 1 −T f (x) sin πmx
Orthogonal projection: If u1 · · · uk is a basis for W , then orthogonal projection of y on W is: Complex eigenvalues If λ = α + iβ, and v = a + ib. Then: T T
    P∞  
πmx , where:
y·u1 y·u1
    Cosine series: f defined on (0, T ): f (x) f m=0 am cos
ŷ = u1 + · · · + uk x(t) = A eαt cos(βt)a − eαt sin(βt)b + B eαt sin(βt)a + eαt cos(βt)b T
u1 u1 uk uk T f (x)dx (not a typo)
a0 = 2
R
y − ŷ is orthogonal to ŷ, shortest distance btw y and W is ky − ŷk Notes: You only need to consider one complex eigenvalue. For real eigenvalues, use the formula above. Also, 2T 0
= a−bi
 
1 am = 2 0T f (x) cos πmx
R
Gram-Schmidt: Start with B = {u1 , · · · un }. Let:
a+bi a2 +b2 T T
∂u = β ∂ 2 u + P (x)

P∞
Step 2: Deal with X 00 = λX. Use boundary conditions to find X(0) etc. (if you have ∂u , you might have
 
πmx , where:

Sine series: f defined on (0, T ): f (x) f m=0 bm sin ∂t ∂x2

T ∂x Nonhomogeneous heat equation: u(0, t) = U1 , u(L, t) = U 2
b0 = 0 X 0 (0) instead of X(0)). 

  u(x, 0) = f (x)
bm = 2 0T f (x) sin πmx Step 3: Case 1: λ = ω 2 , then X(x) = Aeωx + Be−ωx , then find ω = 0, contradiction. Case 2:
R
T T Then u(x,h t) = v(x) + w(x, t), where:
λ = 0, then X(x) = Ax + B, then eihter find X(x) = 0 (contradiction), or find X(x) = A. Case 3:
Tabular integration: (IBP: f 0 g = f g − f g 0 ) To integrate f (t)g(t)dt where f is a polynomial,
i
v(x) = U2 − U1 + 0L 0z 1 P (s)dsdz x + U1 − 0x 0z 1 P (s)dsdz and w(x, t)
R R R R R R R
λ = −ω 2 , then X(x) = A cos(ωx) + B sin(ωx). Then solve for ω, usually ω = πm . Also, if β L β
make a table whose first row is f (t) and g(t). Then differentiate f as many times until you get 0, and T  2
antidifferentiate as many times until it aligns with the 0 for f . Then multiply the diagonal terms and do + first term case 2 works, should find cos, if case 2 doesn’t work, should find sin.  ∂w = β ∂ w
∂t 2

− second term etc. Finally, λ = −ω 2 , and X(x) = whatever you found in 2) w/o the constant. solves the hom. eqn: ∂x
w(0, t) = 0, w(L, t) = 0
RT     Step 4: Solve for T (t) with the λ you found. Remember that for the heat equation: 
cos πmx sin πnx dx = 0

Orthogonality formulas: −T u(x, 0) = f (x) − v(x)
RT    T T T 0 = λT ⇒ T (t) = ] Am eλt . And for the wave equation: D’Alembert’s formula: ONLY works for wave equation and −∞ < x < ∞:
πmx πnx
−T cos  T  cos T  dx = 0 if m 6= n T 00 = λT ⇒ T (t) = PA
] m cos(ωt) + ] Bm sin(ωt). u(x, t) = 1 (f (x + αt) + f (x − αt)) + 1
R x+αt
g(s)ds, where
RT πmx sin πnx dx = 0 if m 6= n Step 5: Then u(x, t) = ∞ P∞ 2 2α x−αt
−T sin T T m=0 T (t)X(x) (if case 2 works), u(x, t) = m=1 T (t)X(x) (if 2 ∂u
case 2 doesn’t work!) utt = α uxx , u(x, 0) = f (x), u(x, 0) = g(x). The integral just means ‘antidifferentiate
Convergence: Fourier series F goes to f (x) is f is continuous at x, and if f has a jump at x, F goes to the ∂t
Step 6: Use u(x, 0), and plug in t = 0. Then use Fourier cosine or sine series or just ‘compare’, i.e. if and plug in’.
average of the jumps. Finally, at the endpoints, F goes to average of the left/right endpoints.
Heat/Wave equations: u(x, 0) = 4 sin(2πx) + 3 sin(3πx), then A g = 4, A
2
g = 3, and ]
3 Am = 0 if m 6= 2, 3.
Step 1: Suppose u(x, t) = X(x)T (t), plug this into PDE, and group X-terms and T -terms. Then Step 7: (only for wave equation): Use ∂u u(x, 0): Differentiate Step 5 with respect to t and set t = 0. Then Laplace equation:
∂t
X 00 (x) use Fourier cosine or series or ‘compare’ Same as for Heat/Wave, but T (t) becomes Y (y), and we get Y 00 (y) = −λY (y). Also, instead of writing
= λ, so X 00 = λX. Then find a differential equation for T . Note: If you have an α-term, put it
X(x) Y (y) = ] Am eωy + ] Bm e−ωy , write Y (y) = ] Am cosh(ωy) + ] Bm sinh(ωy).
with T . Remember cosh(0) = 1, sinh(0) = 0

You might also like