CSIR NET Mathematical Sciences — Linear Algebra Compendium 1

CSIR NET Mathematical 1. Vector Spaces, Subspaces, Linear Combi-

nation, Span, Linear Independence, Ba-
Sciences sis, Dimension, Direct Sum, Quotient
Linear Algebra Complete Space, Ordered Basis, Isomorphism The-
LaTeX Compendium orems
Dense Revision Notebook Formula Handbook Concept Core Concept Skeleton
Map Quick Solver Guide
• field + abelian group
• vector addition
• scalar multiplication
• closure
• distributivity
• associativity
Contents • identity + inverse
• subspace criterion
• linear combination
• span / generated subspace
1 Vector Spaces, Subspaces, Linear Combina- • linear dependence relation
tion, Span, Linear Independence, Basis, Di- • uniqueness in basis expansion
mension, Direct Sum, Quotient Space, Or- • maximal independent set
dered Basis, Isomorphism Theorems 1 • minimal spanning set
• dimension invariance
• extension / reduction principles
2 Linear Transformations and Their Proper- • sum and intersection of subspaces
ties; Composition; Special Linear Operators; • direct sum uniqueness
Matrix Representation 4 • quotient by subspace
• cosets
• canonical projection
• rank-nullity viewpoint
3 Matrices and Their Properties: Terminology,
Special Matrices, Determinant, Adjoint, In- • ordered basis and coordinates
verse, Rank, Trace, Norm, Structural Matrix • isomorphism structure
Facts 7

4 System of Linear Equations 10

5 Eigenvalues, Eigenvectors, Characteristic

Polynomial, Minimal Polynomial, Compan-
ion Matrix, Gershgorin, Similarity 11

6 Diagonalizability and Canonical Forms: Di-

agonalization, Spectrum of Important Matri-
ces, Permutation Matrices, Jordan Canonical
Form 14

7 Inner Product Spaces, Orthogonality, Or-

thonormal Sets, Gram–Schmidt, Orthogonal
Complement, Adjoint, Unitary, Normal, Self-
Adjoint Operators 16

8 Bilinear Forms, Dual Space, Dual Map, An-

nihilator, Quadratic Forms, Reduction, Defi-
niteness, Congruence, Signature 18
CSIR NET Mathematical Sciences — Linear Algebra Compendium 2

Definitions Theorems / Results / Statements

Vector space. A set V over a field F with operations Subspace test. A nonempty subset W ⊆ V is a
+ and scalar multiplication such that: subspace iff αx + βy ∈ W for all x, y ∈ W and α, β ∈ F.
Intersection theorem. Arbitrary intersections of sub-
(V, +) is an abelian group, a(x+y) = ax+ay, (a+b)x =spaces
ax+bx,are subspaces.
Union criterion. W1 ∪ W2 is a subspace iff W1 ⊆ W2
(ab)x = a(bx), 1x = x.
or W2 ⊆ W1 .
Real / complex vector space. Over R / C respec- Span is smallest subspace. span(S) is the smallest
tively. subspace containing S.
Subspace. A nonempty subset W ⊆ V such that for Basic span facts.
all x, y ∈ W and α, β ∈ F,
span(X∪Y ) = span(X)+span(Y ), A ⊆ B ⇒ span(A) ⊆ span(
αx + βy ∈ W.
span(span(A)) = span(A), span(∅) = {0}.
Linear combination. Any vector of the form
Pm Dependence facts.
i=1 ai vi .
Span. For S ⊆ V , • Any set containing 0 is linearly dependent.
(m ) • ∅ is linearly independent.
X • Every subset of a linearly independent set is linearly
span(S) = ai si : m ∈ N, ai ∈ F, si ∈ S . independent.
i=1
• Every superset of a linearly dependent set is linearly
Linearly dependent. A set S is linearly dependent if dependent.
some finite distinct vectors x1 , . . . , xn ∈ S satisfy • A set is linearly independent iff every finite subset is
linearly independent.
c1 x1 + · · · + cn xn = 0 Two-vector dependence. For u ̸= 0, v ̸= 0,
for scalars not all zero. {u, v} dependent ⇐⇒ u, v are scalar multiples.
Linearly independent. Only trivial linear relation
gives 0. Finite-set dependence criterion. For finite
Basis. A linearly independent spanning set. S = {u1 , . . . , un }, S is dependent iff some uk ∈
Dimension. Number of vectors in any basis of a finite- span(u1 , . . . , uk−1 ) or some ui = 0.
dimensional vector space; denoted dim V . Basis characterizations.
Direct sum. V = W1 ⊕· · ·⊕Wm if V = W1 +· · ·+Wm • Basis = maximal linearly independent set.
and each vector has a unique decomposition; equiva- • Basis = minimal spanning set.
lently X • In dim V = n, any linearly independent set with n
Wk ∩ Wi = {0} ∀k. vectors is a basis.
i̸=k • In dim V = n, any spanning set with n vectors is a
Quotient space. For subspace W ≤ V , basis.
• Every linearly independent set extends to a basis.
V /W = {v + W : v ∈ V } , • Every spanning set contains a basis.
Dimension theorem. Any two bases of a finite-
with
dimensional vector space have the same cardinality.
(v+W )+(u+W ) = (v+u)+W, a(v+W ) = av+W. Zero space. dim{0} = 0.
Subspace dimension inequality. If W ≤ V , then
Ordered basis. A basis with a prescribed order. dim W ≤ dim V .
Coset. v + W = {v + w : w ∈ W }. Dimension formula. For finite-dimensional subspaces
Complementary subspaces. W1 , W2 are complemen- W1 , W2 ≤ V ,
tary if V = W1 ⊕ W2 .
dim(W1 + W2 ) = dim W1 + dim W2 − dim(W1 ∩ W2 ).

Direct sum criterion.

V = W1 ⊕ W2 ⇐⇒ V = W1 + W2 and W1 ∩ W2 = {0}.

Equivalently, each v ∈ V has a unique representation

v = w1 + w2 .
Dimension of direct sum.

dim(U ⊕ V ) = dim U + dim V.

Quotient dimension theorem. If W ≤ V and V is

finite-dimensional, then

dim(V /W ) = dim V − dim W.

First Isomorphism Theorem. If T : V → V ′ is

linear, then
V / ker T ∼
= Im T.
CSIR NET Mathematical Sciences — Linear Algebra Compendium 3

Direct Results / Formulae Examples

W1 + W2 = {w1 + w2 : w1 ∈ W1 , w2 ∈ W2 } Basic. In R3 , W = {(x, y, z) : x + y + z = 0} is

a subspace since it is kernel of the linear functional
W1 − W2 = {w1 − w2 : w1 ∈ W1 , w2 ∈ W2 } = W1 + W2 (x, y, z) 7→ x + y + z.
dim(W1 ∩ W2 ) ≤ min(dim W1 , dim W2 ) Exam-style. In R3 , let

max{0, dim W1 + dim W2 − dim V } ≤ dim(W1 ∩ W2 ) W1 = span{(1, 0, 0), (0, 1, 0)}, W2 = span{(1, 1, 0)}.
max(dim W1 , dim W2 ) ≤ dim(W1 +W2 ) ≤ min(dim V, dim W1 +dim Then W W22) ⊆ W1 , so W1 ∪ W2 = W1 is a subspace. If
If dim W1 + dim W2 > dim V , then W1 ∩ W2 ̸= {0}. instead W2 = span{(0, 0, 1)}, then W1 ∪ W2 is not a
For quotient: subspace.
Direct sum example. R2 = span{(1, 0)} ⊕
(v + W ) = (u + W ) ⇐⇒ v − u ∈ W. span{(0, 1)}.
Quotient example. In R2 , with W = span{(1, 0)},
Canonical projection: two vectors (a, b) and (c, d) lie in the same coset iff
b = d.
π : V → V /W, π(v) = v+W, ker π = W, Im π = V /W.
Related / Associated Concepts
Basis expansion uniqueness:
Related / Associated Concepts
n
X • coordinate systems
v= ai ei , unique if {ei } is a basis.
• linear maps
i=1
• kernel and image
Coordinate vector relative to ordered basis B = • matrices of transformations
(e1 , . . . , en ): • rank-nullity
[v]B = (a1 , . . . , an )T . • dual space
• affine subspaces
Alternative / Equivalent / Easier Forms • complements
• Subspace test can be used as: nonempty + closed • projections
under subtraction and scalar multiplication. • module analogy
• V = W1 ⊕ W2 iff V = W1 + W2 and decomposition • finite fields and Gaussian coefficients
is unique.
• S is a basis iff every v ∈ V has a unique linear Core Concept Skeleton for Related Concepts
combination in S. • coordinates in ordered basis
• In finite dimension: maximal independent ⇐⇒ basis • kernel-image decomposition
⇐⇒ minimal spanning. • coset partition
• Quotient equality: • induced map on quotient
• basis extension
x + W = x′ + W ⇐⇒ x − x′ ∈ W.
• dimension counting
• W1 ∪ W2 subspace only in nested case. • invariant decomposition
• W1 \ W2 is generally not a subspace. • projection idempotence

Failure Cases / Exceptions / Warnings Formula Arsenal for Related Concepts

• Nonempty alone does not imply subspace. dim Im T + dim ker T = dim V
• Closure under addition only is insufficient.
V / ker T ∼
= Im T
• Closure under scalar multiplication only is insuffi-
cient. (W1 + W2 )/W1 ∼
= W2 /(W1 ∩ W2 )
• Union of two subspaces is usually not a subspace. (V /W1 )/(W2 /W1 ) ∼
= V /W2
• Infinite-dimensional spaces need not have finite bases.    
• Quotient space notation V /W makes sense only when n n
=
W is a subspace. r q n−r q
• Direct sum requires trivial intersection plus full sum; |ordered bases of Fnq | = | GL(n, q)|
either one alone is insufficient.
• In characteristic 2, some dependence tricks involving
u + v and u − v collapse.
• Wronskian nonzero implies independence for many
standard function families, but zero Wronskian need
not imply dependence without additional hypotheses.
CSIR NET Mathematical Sciences — Linear Algebra Compendium 4

CSIR NET Utility Layer 2. Linear Transformations and Their Prop-

• If you see “subset defined by homogeneous linear erties; Composition; Special Linear Op-
equations” think subspace. erators; Matrix Representation
• If you see “x + y + 1 = 0” think not a subspace:
Core Concept Skeleton
non-homogeneous.
• For union questions: immediately test nesting. • linearity
• For direct sum in MCQ: check both W1 + W2 = V • image / range
and W1 ∩ W2 = {0}. • kernel / null space
• In finite dimension n: independent set with n vectors • rank
or spanning set with n vectors is automatically a • nullity
basis. • injective / surjective / bijective
• Quotient dimension is a frequent one-line objective • invertibility
question: • composition
• operator on same space
dim(V /W ) = dim V − dim W.
• basis transport
• Over finite fields, ordered bases count = order of • matrix representation
GL(n, q). • change of basis
• Common trap: W1 − W2 is a subspace because it • idempotent operator
equals W1 + W2 . • involutory operator
• Common mistake: assuming every spanning set is a • invariant decomposition
basis; independence must also hold. • projection / reflection / rotation

Research-Based Enrichment Layer Definitions

• Every finite-dimensional nonzero vector space over Linear transformation. T : V → W is linear if
an infinite field has infinitely many distinct bases.
• For subspaces W1 , . . . , Wm , T (x + y) = T (x) + T (y), T (ax) = aT (x).

m
M Range / image.
V = Wi
i=1 R(T ) = Im T = {T (x) : x ∈ V } ⊆ W.
P
iff every 0 = i wi implies each wi = 0. Kernel / null space.
• If V is finite-dimensional, every quotient V /W is
finite-dimensional. N (T ) = ker T = {x ∈ V : T (x) = 0}.
• If dim V = n and T ⊆ V is linearly independent
while S ⊆ V spans, then Rank.
rank(T ) = dim(Im T ).
|T | ≤ n ≤ |S|. Nullity.
nullity(T ) = dim(ker T ).
• In any finite-dimensional vector space, all comple-
ments of a fixed subspace have the same dimension. Linear operator. A linear transformation T : V → V .
Invertible linear map. A linear map with linear
inverse T −1 .
Projection. T with T 2 = T .
Reflection / involution. T with T 2 = I.
Matrix representation. For ordered bases B =
(v1 , . . . , vn ) of V , C = (w1 , . . . , wm ) of W , the matrix
[T ]C←B = A = (aij ) defined by
m
X
T (vj ) = aij wi .
i=1
CSIR NET Mathematical Sciences — Linear Algebra Compendium 5

Theorems / Results / Statements Direct Results / Formulae

Kernel and image are subspaces. ker T ≤ V and For T : R2 → R2 ,
Im T ≤ W .
Rank–Nullity. If dim V < ∞, then T (a, b) = (αa + βb, γa + δb)

dim V = rank(T ) + nullity(T ). is invertible iff

αδ − βγ ̸= 0.
Finite-dimensional equal-dimension criterion. If
dim V = dim W < ∞ and T : V → W is linear, then Standard projections:
the following are equivalent:
Px (a, b) = (a, 0), Py (a, b) = (0, b).
• T is invertible
• T is injective 1
Projection on line y = mx along y = − m x:
• T is surjective  
• ker T = {0} a + mb m(a + mb)
P (a, b) = , .
• T sends a basis of V to a basis of W 1 + m2 1 + m2
Dimension obstruction. Reflections:
dim V < dim W ⇒ T cannot be onto, Rx (a, b) = (a, −b), Ry (a, b) = (−a, b), R0 (a, b) = (−a, −b)
dim V > dim W ⇒ T cannot be one-one. Reflection about line y = mx:
Composition. If T : V → W and S : W → U are  
linear, then S ◦ T : V → U is linear. a + mb m(a + mb)
R(a, b) = 2 − a, 2 − b .
Inverse linearity. If T is bijective linear, then T −1 is 1 + m2 1 + m2
linear.
One-sided inverse in finite dimension. If V is finite- Rotation by θ:
dimensional and T S = I, then both T, S are invertible
Tθ (a, b) = (a cos θ − b sin θ, a sin θ + b cos θ).
and S = T −1 .
Projection decomposition. If V = W ⊕ W ′ , define Rotation matrix:
T (w + w′ ) = w. Then:  
cos θ − sin θ
2
T = T, ker T = W , ′
Im T = W. [Tθ ] = .
sin θ cos θ

Reflection decomposition. If V = W ⊕ W ′ and If Tθ , Tϕ are rotations:

T (w + w′ ) = w − w′ , then:
Tθ Tϕ = Tθ+ϕ = Tϕ Tθ .
T 2 = I, ker T = {0}, Im T = V.
Matrix representation:
Rotation in R2 . The anticlockwise rotation by angle

θ is linear and invertible. [T ]C←B = [T (v1 )]C · · · [T (vn )]C .
Basis-image criterion. If β = {v1 , . . . , vn } is a basis
of V and T is invertible, then Alternative / Equivalent / Easier Forms

T (β) = {T (v1 ), . . . , T (vn )} • Injective ⇐⇒ ker T = {0}.

• Surjective ⇐⇒ rank(T ) = dim W in finite dimension.
is a basis of W . • Invertible ⇐⇒ bijective ⇐⇒ non-singular.
Dependence transfer. • Projection = idempotent operator.
• If x1 , . . . , xk are dependent, then T (x1 ), . . . , T (xk ) • Reflection = involutory operator.
are dependent. • Matrix columns of [T ]C←B are coordinate vectors of
• If T (x1 ), . . . , T (xk ) are independent, then x1 , . . . , xk T (vj ) in basis C.
are independent. • If V = W ⊕ W ′ , then projection onto W along W ′
depends on the chosen complement.
CSIR NET Mathematical Sciences — Linear Algebra Compendium 6

Failure Cases / Exceptions / Warnings Formula Arsenal for Related Concepts

• If images are dependent, preimages may still be inde- dim V = rank(T ) + nullity(T )
pendent.
• If vectors are independent, their images need not be ker(S ◦ T ) ⊇ ker T
independent unless T is injective. Im(S ◦ T ) ⊆ Im S
• In infinite-dimensional spaces, T S = I need not imply
ST = I. [T ◦ S]D←B = [T ]D←C [S]C←B
• Projection is not necessarily orthogonal projection [T −1 ]B←C = [T ]−1
C←B
unless an inner product and orthogonal complement
are specified. CSIR NET Utility Layer
• Reflection formulas about a line require the line to
• For injective/surjective MCQs in finite dimension,
pass through the origin for linearity.
use dimensions first, not computation.
• Rotation by an angle in affine geometry about a
• If domain and codomain dimensions are equal, prove
non-origin point is not linear.
either injective or onto; the other follows.
Examples • For linearity check, test T (0) = 0 immediately. If
not, reject.
Basic. T : R2 → R2 , T (x, y) = (x + y, x − y) is linear • If T 2 = T , think eigenvalues 0, 1 and projection.
with matrix   • If T 2 = I, think eigenvalues ±1 and reflec-
1 1
, tion/involution.
1 −1
• Rotation matrix recognition:
so T is invertible since determinant = −2 ̸= 0.  
Exam-style. Let T : R3 → R2 , T (x, y, z) = (x + y, y + cos θ − sin θ
.
z). Then sin θ cos θ

ker T = {(−t, t, −t) : t ∈ R}, • Common trap: a geometric transformation may fail
to be linear because it does not preserve origin.
so nullity(T ) = 1, hence rank(T ) = 3 − 1 = 2. Therefore
T is onto. Research-Based Enrichment Layer
Projection example. If V = R2 , T (x, y) = (x, 0), then
• If rank(T ) = rank(T 2 ) on finite-dimensional V , then
T 2 = T , ker T = span{(0, 1)}, Im T = span{(1, 0)}.
Rotation trick. A nonzero real vector can be an V = Im T ⊕ ker T.
eigenvector of a real rotation iff θ = 0 or π modulo 2π.
• If T is invertible from V to W , then V finite-
Related / Associated Concepts dimensional iff W finite-dimensional, and dimensions
Related / Associated Concepts coincide.
• matrices • The set L(V, W ) is itself a vector space.
• change of basis • If dim V = n, dim W = m, then
• kernel-image decomposition dim L(V, W ) = mn.
• isomorphism theorems
• invariant subspaces • For T ∈ L(V ), the minimal polynomial of T controls
• eigenvalues decomposition phenomena later.
• projection matrices
• orthogonal transformations

Core Concept Skeleton for Related Concepts

• column-space viewpoint
• basis-image map
• decomposition into kernel + complement
• coordinate transform
• operator equations
• idempotent / involutive algebra
CSIR NET Mathematical Sciences — Linear Algebra Compendium 7

3. Matrices and Their Properties: Termi- Definitions

nology, Special Matrices, Determinant, Row matrix / column matrix. Single row / single
Adjoint, Inverse, Rank, Trace, Norm, column matrices.
Structural Matrix Facts Square matrix. A ∈ Mn (F).
Diagonal matrix. aij = 0 for i ̸= j.
Core Concept Skeleton Upper triangular / lower triangular. aij = 0 for
• matrix as coordinate form of linear map i > j / i < j.
• rows / columns Strictly triangular. Triangular with all diagonal en-
tries 0.
• square vs rectangular
Transpose. AT = (aji ).
• transpose Conjugate. A = (aij ).
• conjugation T
Conjugate transpose. A∗ = A .
• adjoint matrix
Trace.
• determinant n
X
• cofactors tr(A) = aii .
• inverse i=1

• rank Determinant. det(A) for square A.

• null space Adjugate / adjoint matrix. adj(A) = transpose of
• row space / column space cofactor matrix.
• trace Singular / non-singular. det(A) = 0 / det(A) ̸= 0.
• special matrix classes Symmetric. AT = A.
Skew-symmetric. AT = −A.
• commutation
Hermitian. A∗ = A.
• block structure
Skew-Hermitian. A∗ = −A.
• factorization Orthogonal. AT A = AAT = I over R.
• spectral cues Unitary. A∗ A = AA∗ = I.
Idempotent. A2 = A.
Involutory. A2 = I.
Nilpotent. Ak = 0 for some k ∈ N.
Block matrix.
 
C D
A= .
P Q

Boolean matrix. Entries only 0 or 1.

Periodic matrix. An+r = Ar for some r; in the source
style often An+1 = A for least such n.
Tridiagonal matrix. Nonzero entries only on main
diagonal and first super/sub-diagonals.
Circulant matrix. Each row is cyclic right shift of the
previous row.
CSIR NET Mathematical Sciences — Linear Algebra Compendium 8

Theorems / Results / Statements Direct Results / Formulae

Transpose and conjugate laws. Cofactor expansion. If Cij is cofactor of aij , then

(AT )T = A, (AB)T = B T AT , AB = A B, (AB)∗ = B ∗ A∗ . n

X n
X
det(A) = aij Cij = aij Cij .
Determinant laws. j=1 i=1

det(AB) = det(A) det(B), det(AT ) = det(A), det(A∗Determinant

) = det(A). row/column operations.
• Swap two rows ⇒ determinant changes sign.
If A triangular, then • Multiply a row by k ⇒ determinant multiplied by k.
n
Y • Add multiple of one row to another ⇒ determinant
det(A) = aii . unchanged.
i=1 Adjugate formulas.
Adjugate identity. adj(AB) = adj(B) adj(A), adj(I) = I, adj(0) = 0.
A adj(A) = adj(A) A = det(A)I. Block trace.
 
Inverse criterion. A B
tr = tr(A) + tr(D).
C D
−1
A exists ⇐⇒ det(A) ̸= 0.
Norm of real matrix.
Inverse formula.  1/2
1
q X
−1
A = adj(A) (det A ̸= 0). ∥A∥F = tr(AT A) =  a2ij  .
det(A)
i,j

Uniqueness of inverse. Inverse is unique.

Inverse product rule. Circulant matrix. If first row is (a0 , a1 , . . . , an−1 ) and

(AB)−1 = B −1 A−1 . f (x) = a0 + a1 x + · · · + an−1 xn−1 ,

Rank facts. then eigenvalues are

• rank(A) = rank(AT ). λj = f (ωj ), ωj = e2πij/n , j = 0, . . . , n − 1.
• 0 ≤ rank(A) ≤ min(m, n) for A ∈ Mm×n .
• rank(A) = 0 ⇐⇒ A = 0. Fibonacci matrix.
• If A ∈ Mn (F) is invertible, then rank(A) = n.    
1 1 Fn+1 Fn
• For echelon form, rank = number of nonzero rows. A= , An = .
1 0 Fn Fn−1
Rank inequalities.
Alternative / Equivalent / Easier Forms
rank(A + B) ≤ rank(A) + rank(B),
• Orthogonal over R = unitary over R.
rank(A − B) ≥ rank(A) − rank(B) , • Hermitian = self-adjoint matrix.
rank(AB) ≤ min(rank A, rank B). • Skew-Hermitian = anti-self-adjoint matrix.
• Idempotent = projection matrix.
Sylvester inequality. If A is m × p and B is p × n,
• Involutory = reflection-type matrix.
then
rank(AB) ≥ rank(A) + rank(B) − p. • For invertible A, AB = AC ⇐⇒ B = C and
BA = CA ⇐⇒ B = C.
Frobenius inequality. If ABC exists, then • If A and B commute, many scalar-like identities be-
come valid, e.g.
rank(AB) + rank(BC) ≤ rank(B) + rank(ABC).
A2 − B 2 = (A − B)(A + B).
Trace facts.

tr(A+B) = tr(A)+tr(B), tr(λA) = λ tr(A), tr(AB) = tr(BA).

Also
tr(ABC) = tr(BCA) = tr(CAB).
Real positivity trace identity.
X
tr(AT A) = a2ij ≥ 0,
i,j

with equality iff A = 0.

Symmetric matrix facts.
• A + AT is symmetric.
• AT A and AAT are symmetric.
• If A, B symmetric, then AB symmetric iff AB = BA.
CSIR NET Mathematical Sciences — Linear Algebra Compendium 9

Failure Cases / Exceptions / Warnings Formula Arsenal for Related Concepts

• AB = 0 does not imply A = 0 or B = 0. det(A−1 ) = (det A)−1
• AB = 0 does not imply BA = 0.
• AB symmetric need not hold even if A, B are sym- (A−1 )T = (AT )−1 , (A−1 )∗ = (A∗ )−1
metric. rank(AT A) = rank(A) = rank(AAT )
• tr(ABC) = tr(ACB) is false in general. X
• Rankof AB need not equal rank of BA. tr(A∗ A) = |aij |2
i,j
A B
• rank ≤ rank(A) + rank(D) is false in gen-
C D AB = I ⇐⇒ BA = I for square matrices
eral.
• Skew-symmetric determinant odd-order zero requires CSIR NET Utility Layer
characteristic ̸= 2. • For determinant of triangular/block triangular ma-
• For complex matrices, use A∗ , not AT , in uni- trices: multiply diagonal/block determinants.
tary/Hermitian questions. • For rank questions, row-reduce first; for theoretical
MCQs, use inequalities.
Examples
• If matrix is idempotent, instantly think eigenvalues
Basic. For 0, 1 and tr(A) = rank(A).
    • If matrix is involutory, think eigenvalues ±1.
1 2 1 4 −2
A= , det A = −2, A−1 = . • If matrix is orthogonal, determinant is only ±1.
3 4 −2 −3 1
• Skew-symmetric odd order ⇒ determinant 0 is a
Exam-style. If A is 3 × 3 skew-symmetric over R, then favorite trap.
• AT A = 0 ⇒ A = 0 over R; useful for elimination.
det(A) = det(AT ) = det(−A) = (−1)3 det(A) = − det(A), • For circulant matrices, row sum is an eigenvalue.
• Common trap: “all row sums equal α” gives eigen-
hence det(A) = 0.   vector (1, . . . , 1)T .
1 0
Idempotent example. A = satisfies A2 = A,
0 0 Research-Based Enrichment Layer
rank A = tr A = 1.
Orthogonal example. • If A has every entry equal to 1, then rank(A) = 1
  and eigenvalues are n, 0, . . . , 0.
1 1 −1 • A matrix both unitary and upper triangular is diago-
Q= √
2 1 1 nal.
• Any complex matrix decomposes as
has QT Q = I.
A + A∗ A − A∗
A= + ,
Related / Associated Concepts 2 2
Related / Associated Concepts Hermitian + skew-Hermitian; also
• linear operators
• determinant tests A = H + iK, H, K Hermitian.
• eigenvalues • For real matrices,
• similarity
X
• diagonalization tr(AT B) = aij bij
• quadratic forms i,j
• row reduction
• singular values gives Frobenius inner product.
• normal matrices • If A, B are similar, determinant, trace, rank, nullity,
characteristic polynomial and minimal polynomial
Core Concept Skeleton for Related Concepts all coincide.
• matrix = linear map in coordinates
• elementary operations
• row space / column space
• determinant multiplicativity
• trace cyclicity
• orthogonality constraints
• spectral classes
CSIR NET Mathematical Sciences — Linear Algebra Compendium 10

4. System of Linear Equations Theorems / Results / Statements

Core Concept Skeleton Consistency criterion.
• matrix equation AX = b AX = b is consistent ⇐⇒ rank(A) = rank([A|b]).
• homogeneous / non-homogeneous
• consistency Equivalently,
• augmented matrix b ∈ Im A.
• rank criterion Homogeneous system always consistent. Since
• null space X = 0 is always a solution.
• solution space Homogeneous solution set.
• affine translate of kernel
• unique / infinite / no solution {X : AX = 0} = ker A.
• row-reduction logic
Dimension of homogeneous solution space.
• pivot / free variable count
• convex combination of solutions dim ker A = n − rank(A).
Definitions Homogeneous nontrivial solution criterion.
A system of m linear equations in n variables can be
written as AX = 0 has a nonzero solution ⇐⇒ nullity(A) > 0 ⇐⇒ rank(A)
AX = b,
If m < n over an infinite field, then homogeneous
where A ∈ Mm×n (F), X ∈ Fn , b ∈ Fm . system has a nontrivial solution.
Homogeneous system. Non-homogeneous solution structure. If x0 is one
solution of AX = b, then all solutions are
AX = 0.
x = x0 + z, z ∈ ker A.
Non-homogeneous system.
Equal cardinality of nonempty affine solution
AX = b, b ̸= 0. sets. If AX = b1 and AX = b2 are both consistent,
then their solution sets have equal cardinality and same
Consistent system. At least one solution exists.
dimension structure.
Inconsistent system. No solution exists.
Unique solution criterion for square system. For
Augmented matrix. [A | b].
A ∈ Mn (F):
Solution space of homogeneous system. ker A.
AX = b has a unique solution for every b ⇐⇒ A invertible.

Convex combinationP y1 , . . . , yn are

statement. If P
solutions of Ax = b and i αi = 1, then i αi yi is also
a solution.

Direct Results / Formulae

No. of free variables = n − rank(A)

(
rank(A) = n ⇒ only trivial solution,
AX = 0 :
rank(A) < n ⇒ infinitely many solutions over infi
For non-homogeneous AX = b:

consistent ⇐⇒ b ∈ Im A ⇐⇒ rank(A) = rank([A|b]).

If consistent:

dim(solution set) = n − rank(A).

Square case m = n:

rank(A) = n ⇒ unique solution, rank(A) < n ⇒ either none o

If A invertible:
X = A−1 b.
CSIR NET Mathematical Sciences — Linear Algebra Compendium 11

Alternative / Equivalent / Easier Forms Core Concept Skeleton for Related Concepts
• Consistency can be checked either by augmented rank • echelon form
or by asking whether b lies in the column space. • pivot positions
• Non-homogeneous solution set = affine space = one • free parameters
particular solution + null space. • image membership
• For homogeneous systems, “one nontrivial solution” • nullity count
already means infinitely many over infinite fields. • affine translation
• Unique solution iff no free variable.
Formula Arsenal for Related Concepts
Failure Cases / Exceptions / Warnings
dim ker A = n − rank(A)
• rank(A) < n does not guarantee a non-homogeneous (
system is consistent. ∅, b∈/ Im A,
solution set of AX = b =
• m > n does not imply inconsistency; overdetermined x0 + ker A, b ∈ Im A.
systems may still be consistent.
• For homogeneous systems, consistency is automatic. If A ∈ Mn (F), det(A) ̸= 0 ⇒ unique solution for all b
• “More variables than equations” guarantees nontriv-
CSIR NET Utility Layer
ial homogeneous solution, not non-homogeneous so-
lution. • First question in MCQ: homogeneous or not?
• Convex combination language P is natural over R; the • Second question: compare ranks.
affine-combination condition αi = 1 is the field- • Third question: count free variables = n − rank(A).
general statement. • If square and determinant nonzero, do not row-reduce
further; answer is unique solution.
Examples • If two equations are scalar multiples but constants
Basic. Solve are not, instantly inconsistent.
     • If you see m < n in homogeneous case, mark nontriv-
1 1 x 2 ial solution.
= .
1 −1 y 0
Research-Based Enrichment Layer
Then x = y = 1; unique solution since determinant ̸= 0.
Exam-style. For • Any two nonempty solution sets of the same linear
system matrix A are translates of ker A.
 
1 2 3 • For consistent AX = b, the set of solutions is a coset
A= , in Fn / ker A viewpoint.
2 4 6
• In computational problems, row rank = column rank
rank(A) = 1 < 3. Hence AX = 0 has nontrivial solu- justifies rank checks using either rows or columns.
tions and in fact a 2-dimensional solution space.
Consistency example. 5. Eigenvalues, Eigenvectors, Characteristic
( Polynomial, Minimal Polynomial, Com-
x+y =1
panion Matrix, Gershgorin, Similarity
2x + 2y = 3
Core Concept Skeleton
has rank(A) = 1, rank([A|b]) = 2, so inconsistent.
• eigen equation
Related / Associated Concepts • invariant direction
• characteristic polynomial
Related / Associated Concepts • roots and multiplicity
• row reduction • eigenspace
• column space • algebraic vs geometric multiplicity
• null space • annihilating polynomial
• rank-nullity • minimal polynomial
• inverse matrix • Cayley–Hamilton
• linear transformations • companion matrix
• affine subspaces • spectral mapping
• least squares • trace and determinant as symmetric sums
• commuting matrices
• Gershgorin discs
• row-sum eigenvalue trick
• similarity invariants
CSIR NET Mathematical Sciences — Linear Algebra Compendium 12

Definitions Theorems / Results / Statements

Characteristic polynomial. For A ∈ Mn (F), Cayley–Hamilton Theorem.
χA (x) = det(xI − A). χA (A) = 0.
Characteristic equation. Eigenvalue criterion.
χA (λ) = 0. λ is eigenvalue ⇐⇒ det(A − λI) = 0
Eigenvalue / eigenvector. λ ∈ F is an eigenvalue of for finite-dimensional matrix representation.
T : V → V if there exists 0 ̸= x ∈ V such that Eigenspace criterion.
T (x) = λx. Eλ = ker(A − λI).
Such x is an eigenvector corresponding to λ. Multiplicity inequality.
Eigenspace.
1 ≤ gm(λ) ≤ am(λ).
Eλ = ker(T − λI) = {x ∈ V : T (x) = λx}.
Distinct eigenvalues give independence. Eigenvec-
Geometric multiplicity. tors corresponding to distinct eigenvalues are linearly
independent.
gm(λ) = dim Eλ . Spectral mapping for polynomials. If λ is an eigen-
value of A, then for any polynomial f ,
Algebraic multiplicity. Multiplicity of λ as a root of
χA (x). f (λ) is an eigenvalue of f (A).
Annihilating polynomial. A polynomial f with
f (A) = 0. Determinant and trace via eigenvalues. Over an
Minimal polynomial. The unique monic polynomial algebraically closed extension:
mA (x) of least degree such that
n
X n
Y
mA (A) = 0. tr(A) = λi , det(A) = λi
i=1 i=1
Companion matrix. For
counting algebraic multiplicities.
p(x) = xn + cn−1 xn−1 + · · · + c1 x + c0 , Inverse eigenvalue rule. If A is invertible and λ is
an eigenvalue, then
the companion matrix is
  λ−1 is an eigenvalue of A−1 .
0 0 ··· 0 −c0
1 0 ··· 0 −c1  Adjugate eigenvalue rule. If A is invertible and
λ ̸= 0 is an eigenvalue, then
 
C(p) = 0
 1 ··· 0 −c2 .

 .. .. .. .. .. 
. . . . .  det(A)
0 0 ··· 1 −cn−1 λ

Similar matrices. A, B ∈ Mn (F) are similar if is an eigenvalue of adj(A).

Minimal polynomial properties.
B = P −1 AP • mA divides every annihilating polynomial.
• mA divides χA .
for some invertible P .
• Every eigenvalue of A is a root of mA .
Orthogonally similar / unitarily similar.
• If χA has distinct roots only, then mA = χA .
B = P T AP or B = P ∗ AP Companion matrix property. For companion matrix
C(p),
with P orthogonal or unitary. χC(p) (x) = mC(p) (x) = p(x).
Gershgorin Circle Theorem. If A = (aij ) ∈ Mn (C)
and X
Ri = |aij |,
j̸=i

then every eigenvalue λ of A lies in at least one disc

|λ − aii | ≤ Ri .

Similarity invariants. If A and B are similar, then

they have the same:

det, tr, rank, nullity, χ(x), m(x), eigenvalues.

AB and BA. If A, B are square, then AB and BA have

the same nonzero eigenvalues with the same algebraic
multiplicities; if both invertible, they are similar.
CSIR NET Mathematical Sciences — Linear Algebra Compendium 13

Direct Results / Formulae Examples

Basic. For
χA (x) = det(xI − A) 
2 0

A= ,
λ ∈ spec(A) ⇐⇒ rank(A − λI) < n 0 3

gm(λ) = nullity(A − λI) eigenvalues are 2, 3; eigenspaces are span{(1, 0)} and
span{(0, 1)}.
mA (x) | χA (x) Exam-style. For
A−1 = q(A) for some polynomial q of degree ≤ n−1 if A invertible  
1 1
by Cayley–Hamilton. A = ,
0 1
If each row sum of A is α, then
χA (x) = (x − 1)2 , mA (x) = (x − 1)2 .
A1 = α1, 1 = (1, . . . , 1)T ,
Only one eigenvalue 1, and gm(1) = 1 since
so α is an eigenvalue.  
If each row sum of A is α ̸= 0 and A invertible, then 0 1
A − I = .
each row sum of A−1 is α−1 . 0 0
For circulant matrix:
Row-sum trick. For
λj = f (ωj ), ωj = e2πij/n .  
2 1 0
A = 1 1 1 ,
Alternative / Equivalent / Easier Forms
0 2 1
• Eigenvalue test:
row sums are not equal, so (1, 1, 1)T is not directly an
(A−λI)x = 0 has nonzero solution ⇐⇒ λ eigenvalue. eigenvector.
Gershgorin example. For
• Similar matrices are the same linear operator under  
different bases. 5 1
A= ,
• For matrix questions, “characteristic roots” = eigen- 2 4
values.
• Over R, a real matrix may have no real eigenvalues. eigenvalues lie in discs |λ − 5| ≤ 1 and |λ − 4| ≤ 2.

Failure Cases / Exceptions / Warnings Related / Associated Concepts

• Number of eigenvalues depends on the field; over R, Related / Associated Concepts

x2 + 1 has no roots. • diagonalization
• Same eigenvalues do not imply similarity in general. • Jordan form
• Same characteristic polynomial does not imply same • invariant subspaces
minimal polynomial. • normal matrices
• gm(λ) = am(λ) need not hold automatically. • companion matrices
• Distinct eigenvalues imply independence of eigenvec- • matrix polynomials
tors; converse is false. • trace/determinant relations
• AB and BA need not be similar if singular. • commuting matrices
• Gershgorin gives bounds/regions, not exact eigenval-
ues in general. Core Concept Skeleton for Related Concepts
• polynomial functional calculus
• root multiplicity
• eigenspace dimension
• basis of eigenvectors
• similarity orbit
• decomposition by generalized eigenspaces

Formula Arsenal for Related Concepts

χP −1 AP (x) = χA (x), mP −1 AP (x) = mA (x)

det(λI − A) = (−1)n det(A − λI)

χA (0) = (−1)n det(A)
X Y
tr(A) = λi , det(A) = λi
CSIR NET Mathematical Sciences — Linear Algebra Compendium 14

CSIR NET Utility Layer Definitions

• If every row sum equals α, immediately mark α as Diagonalizable matrix. A is diagonalizable over F if
eigenvalue.
• For 2 × 2 matrix, use P −1 AP = D

χA (x) = x2 − (tr A)x + det A. for some invertible P and diagonal D over F.
Orthogonally / unitarily diagonalizable. Same
• Distinct eigenvalues ⇒ diagonalizable; high-yield one- with P orthogonal / unitary.
liner. Jordan block. An m × m matrix
• Minimal polynomial divides characteristic polyno- 
λ 1 0 ··· 0

mial: useful for elimination in options. 0 λ 1 · · · 0
• Companion matrix question? Characteristic and min- 
 .. .. ..

..  .
imal polynomial are the defining polynomial. Jm (λ) =  . . . .
 
• If matrix is triangular, eigenvalues are diagonal en- 0 · · · 0 λ 1
tries. 0 ··· ··· 0 λ
• Common trap: algebraic multiplicity is from char-
Jordan canonical form. A block diagonal matrix
acteristic polynomial, geometric multiplicity from
with Jordan blocks on the diagonal, similar to A over a
eigenspace dimension.
splitting field.
Research-Based Enrichment Layer
• If A is real skew-symmetric and I + A invertible, the
Cayley transform

Q = (I − A)(I + A)−1

is orthogonal.
• If A and B commute and are diagonalizable over
a splitting field, simultaneous triangularization is
common; simultaneous diagonalization needs stronger
hypotheses.
• The roots of the minimal polynomial are exactly the
eigenvalues.
• The exponent of each factor (x − λ) in mA equals the
size of the largest Jordan block for λ.

6. Diagonalizability and Canonical Forms:

Diagonalization, Spectrum of Important
Matrices, Permutation Matrices, Jordan
Canonical Form
Core Concept Skeleton
• similarity to diagonal matrix
• eigenbasis
• splitting of characteristic polynomial
• equality of AM and GM
• minimal polynomial square-free condition
• special spectra
• permutation cycles
• roots of unity
• Jordan block
• generalized eigenspace
• block sizes
• canonical form uniqueness up to order
CSIR NET Mathematical Sciences — Linear Algebra Compendium 15

Theorems / Results / Statements Direct Results / Formulae

Diagonalizability criteria. For A ∈ Mn (F):
k
• If A has n distinct eigenvalues in F, then A is diago- Y
nalizable over F. A diagonalizable ⇐⇒ mA (x) = (x−λi ) with distinct λi
i=1
• A is diagonalizable over F iff there exists a basis of
eigenvectors. when it splits over the field.
• A is diagonalizable over F iff for each eigenvalue, Idempotent:

gm(λ) = am(λ), A2 = A ⇒ mA (x) | x(x−1), χA (x) = xn−r (x−1)r , r = rank(A

provided χA splits over F. Involutory:

• A is diagonalizable over F iff mA (x) splits into distinct
linear factors over F. A2 = I ⇒ mA (x) | (x − 1)(x + 1).
Jordan form existence. If χA splits over F, then A
Nilpotent:
is similar over F to a Jordan matrix.
Jordan form uniqueness. Unique up to permutation Ak = 0 ⇒ mA (x) = xs , χA (x) = xn .
of Jordan blocks.
Jordan block polynomials. For J = Jm (λ), All-ones-type matrix:
m m
χJ (x) = (x − λ) , mJ (x) = (x − λ) . A = αJn , χA (x) = xn−1 (x−nα), mA (x) = x(x−nα) (α ̸=
Jordan structure from multiplicities. If Permutation matrix Pσ :
Y Y
χA (x) = (x − λi )ni , mA (x) = (x − λi )mi , det(Pσ ) = sgn(σ), Pσ−1 = PσT .
i i
Jordan block:
then:
• largest Jordan block size for λi is mi , rank(Jm (λ) − λI) = m − 1.
• sum of sizes of Jordan blocks for λi is ni ,
• number of Jordan blocks for λi is gm(λi ). Alternative / Equivalent / Easier Forms
Spectra of standard classes. • Distinct eigenvalues criterion is sufficient, not neces-
• Real symmetric / Hermitian: eigenvalues real. sary.
• Real skew-symmetric / skew-Hermitian: eigenvalues • Diagonalizable ⇐⇒ sum of eigenspace dimensions
purely imaginary or 0. equals n when characteristic polynomial splits.
• Orthogonal / unitary: eigenvalues have modulus 1. • For idempotent and involutory matrices, minimal
• Idempotent: eigenvalues 0 or 1. polynomial has no repeated factors; hence diagonal-
• Involutory: eigenvalues ±1. izable.
• Nilpotent: only eigenvalue 0. • A permutation matrix is a direct sum of cycle per-
Idempotent diagonalizability. Every idempotent mutation matrices up to permutation similarity.
matrix is diagonalizable.
Involutory diagonalizability. Over char(F) ̸= 2, Failure Cases / Exceptions / Warnings
every involutory matrix is diagonalizable. • A matrix may have all eigenvalues in F and still fail
Nilpotent diagonalizability. A nilpotent matrix is to be diagonalizable.
diagonalizable iff it is the zero matrix. • Idempotent with same eigenvalues as A2 does not
Permutation matrices. imply idempotent.
• A permutation matrix is orthogonal (and unitary). • Over R, permutation matrices with cycles of length >
• If corresponding permutation decomposes into dis- 2 are not diagonalizable over R though diagonalizable
joint cycles of lengths r1 , . . . , rk , then over C.
k
• In characteristic 2, involutory diagonalization needs
χA (x) =
Y
(xrj − 1). care because x2 − 1 = (x − 1)2 .
j=1
• Jordan form requires characteristic polynomial to
split in the field.
• Eigenvalues are roots of unity.
• Over C, every permutation matrix is diagonalizable.
• Over R, diagonalizable iff all cycles have lengths 1 or
2 only.
All-entries-equal matrix. If A = αJn where Jn is
all-ones matrix, then eigenvalues are nα and 0 with
multiplicity n − 1; A is diagonalizable.
CSIR NET Mathematical Sciences — Linear Algebra Compendium 16

Examples Research-Based Enrichment Layer

Basic.   • A matrix is diagonalizable iff each primary component
2 0 is diagonalizable iff no Jordan block has size > 1.
A=
0 3 • For nilpotent matrix, the Jordan form is a direct sum
is diagonalizable with eigenbasis the standard basis. of nilpotent Jordan blocks Jm (0).
Exam-style. • The companion matrix gives a canonical cyclic model
 
1 1 for matrices with a cyclic vector.
A=
0 1
has only one eigenvector direction, so not diagonalizable;
7. Inner Product Spaces, Orthogonality, Or-
its Jordan form is itself. thonormal Sets, Gram–Schmidt, Orthog-
Idempotent example. onal Complement, Adjoint, Unitary, Nor-
  mal, Self-Adjoint Operators
1 0
A=
0 0 Core Concept Skeleton

has eigenvalues 1, 0 and is diagonalizable. • inner product axioms

Permutation example. The 3-cycle permutation ma- • norm from inner product
trix has characteristic polynomial • orthogonality
• orthonormality
x3 − 1 = (x − 1)(x2 + x + 1), • Pythagoras
• Cauchy–Schwarz
so diagonalizable over C but not over R.
• triangle inequality
Related / Associated Concepts • Gram–Schmidt
• orthogonal complement
Related / Associated Concepts • direct sum decomposition
• eigenvalues and eigenvectors • orthogonal projection
• minimal polynomial • isometry
• similarity • adjoint operator
• normal matrices • unitary operator
• spectral theorem • normal operator
• generalized eigenvectors • self-adjoint operator
• cyclic decomposition • spectral theorem

Core Concept Skeleton for Related Concepts Definitions

• eigenspace dimension count Inner product. On a vector space V over R or C, a
• square-free minimal polynomial map ⟨·, ·⟩ : V × V → F satisfying:
• block decomposition
⟨au + bv, w⟩ = a⟨u, w⟩ + b⟨v, w⟩,
• generalized eigenspace
• root structure of annihilators ⟨v, u⟩ = ⟨u, v⟩,
Formula Arsenal for Related Concepts ⟨v, v⟩ > 0 (v ̸= 0), ⟨v, v⟩ = 0 ⇐⇒ v = 0.
X X
gm(λ) ≤ n, am(λ) = n Norm. p
λ λ ∥x∥ = ⟨x, x⟩.
largest Jordan block for λ = exponent of (x−λ) in mA (x) Orthogonal vectors. ⟨u, v⟩ = 0.
Orthonormal set. ⟨ui , uj ⟩ = δij .
No. of Jordan blocks for λ = gm(λ) Orthogonal complement.
CSIR NET Utility Layer W ⊥ = {v ∈ V : ⟨v, w⟩ = 0 ∀w ∈ W }.
• Idempotent? Immediately mark diagonalizable and
eigenvalues 0, 1. Adjoint of linear map. For T : V → W between
finite-dimensional inner product spaces, T ∗ : W → V is
• Nilpotent? If nonzero, never diagonalizable.
defined by
• Distinct eigenvalues? Instantly diagonalizable.
• For permutation matrices, use cycle decomposition, ⟨T v, w⟩ = ⟨v, T ∗ w⟩ ∀v ∈ V, w ∈ W.
not determinant expansion.
• If matrix is triangular, diagonal entries are eigenval- Unitary operator. T ∗ T = T T ∗ = I.
ues; then compare eigenspace dimensions. Normal operator. T ∗ T = T T ∗ .
• Jordan block of size > 1 means not diagonalizable. Self-adjoint operator. T = T ∗ .
Isometry. ∥T x∥ = ∥x∥ for all x.
CSIR NET Mathematical Sciences — Linear Algebra Compendium 17

Theorems / Results / Statements Direct Results / Formulae

Norm properties. Angle formula in Rn :

∥x∥ ≥ 0, ∥αx∥ = |α| ∥x∥ , ∥x + y∥ ≤ ∥x∥+∥y∥ . ⟨u, v⟩ = ∥u∥ ∥v∥ cos θ.

Cauchy–Schwarz inequality. Gram–Schmidt:

k−1
|⟨x, y⟩| ≤ ∥x∥ ∥y∥ . X ⟨uk , vj ⟩
v1 = u1 , vk = uk − vj .
⟨vj , vj ⟩
Pythagorean theorem. If x ⊥ y, then j=1

2 2
∥x + y∥ = ∥x∥ + ∥y∥ .
2 Orthonormal version:
vk
Orthogonal nonzero sets are linearly indepen- ek = .
∥vk ∥
dent.
Gram–Schmidt process. Any linearly independent Orthogonal projection of x onto span of orthonormal
finite set can be converted into an orthogonal basis of set {e1 , . . . , en }:
the same span; normalizing gives an orthonormal basis.
n
Finite-dimensional orthogonal decomposition. If X
PW x = ⟨x, ei ⟩ei .
W ≤ V and V finite-dimensional, then
i=1

V = W ⊕ W ⊥. Parseval/Bessel form for complete orthonormal basis:

Double orthogonal complement. For finite- X 2
X
x= ⟨x, ei ⟩ei , ∥x∥ = |⟨x, ei ⟩|2 .
dimensional W ,
i i
W ⊥⊥ = W.
Complement identities. Adjoint matrix in orthonormal basis:

(W1 + W2 )⊥ = W1⊥ ∩ W2⊥ , [T ∗ ] = [T ]∗ .

(W1 ∩ W2 )⊥ = W1⊥ + W2⊥ Unitary matrix criterion:

in finite dimension. A∗ A = I ⇐⇒ columns of A form an orthonormal basis.

Row/column orthogonality.
Alternative / Equivalent / Easier Forms
(Row A)⊥ = Null(A), (Col A)⊥ = Null(AT )
• Orthogonal projection exists uniquely in finite-
over standard real inner product. dimensional inner product spaces.
Equivalent isometry conditions. For linear T be- • “Preserves norms” and “preserves inner products” are
tween inner product spaces, the following are equivalent: equivalent for linear maps.
• T preserves inner products • Unitary matrix = matrix with orthonormal columns
= matrix with orthonormal rows.
• T preserves norms
• Self-adjoint = Hermitian matrix in orthonormal co-
• T preserves distances
ordinates.
Adjoint properties.
• Normal = unitarily diagonalizable over C.
(S+T )∗ = S ∗ +T ∗ , (λS)∗ = λS ∗ , (ST )∗ = T ∗ S ∗ ,
Failure Cases / Exceptions / Warnings
−1 ∗ ∗ −1
(T ) = (T ) if T invertible. • Orthogonal complement decomposition V = W ⊕W ⊥
Kernel-image relations. is guaranteed in finite dimension; in infinite dimension
topological closure issues appear.
ker T ∗ = (Im T )⊥ , Im T ∗ = (ker T )⊥ . • Projection onto a subspace is not necessarily orthog-
onal projection.
Hence • T U self-adjoint need not hold for self-adjoint T, U
unless they commute.
V = ker T ⊕ Im T ∗ , W = Im T ⊕ ker T ∗ .
• Over R, a normal matrix need not be orthogonally
Unitary characterizations. For T ∈ L(V ) finite- diagonalizable unless symmetric; it is orthogonally
dimensional, the following are equivalent: block diagonalizable.
• T is unitary • In non-orthonormal bases, matrix of adjoint is not
simply conjugate transpose.
• ⟨T u, T v⟩ = ⟨u, v⟩
• ∥T u∥ = ∥u∥ for all u, v
Unitary spectrum. Every eigenvalue of a unitary
operator has modulus 1.
Normal operator facts.
• Eigenvectors corresponding to distinct eigenvalues
are orthogonal.
• If T x = λx and T normal, then T ∗ x = λx.
Spectral theorem for normal operators. On a
CSIR NET Mathematical Sciences — Linear Algebra Compendium 18

Examples CSIR NET Utility Layer

Basic. In R3 with standard inner product, vectors • If columns are orthonormal, matrix is uni-
(1, 1, 0) and (1, −1, 0) are orthogonal. tary/orthogonal.
Exam-style. Apply Gram–Schmidt to • For projection on orthonormal basis, coefficients are
just inner products.
u1 = (1, 1, 0), u2 = (1, 0, 1). • In Gram–Schmidt, choose simple vectors first to avoid
algebra.
Then
• Self-adjoint ⇒ real eigenvalues; unitary ⇒ |λ| = 1;
1 1skew-Hermitian ⇒ purely imaginary spectrum.
 
⟨u2 , v1 ⟩ 1
v1 = u1 , v2 = u2 − v1 = (1, 0, 1)− (1, 1, 0) = ,− , 1 .
⟨v1 , v1 ⟩ 2 2 • 2Common trap: orthogonal set need not be orthonor-
mal.
Projection example. Projection of (1, 2) onto • Common trap: preserving norm for a nonlinear map
span{(1, 1)}: does not imply unitary; linearity is essential.

⟨(1, 2), (1, 1)⟩ 3 Research-Based Enrichment Layer

(1, 1) = (1, 1).
⟨(1, 1), (1, 1)⟩ 2
• In finite-dimensional spaces, every subspace has an
Unitary example. orthonormal basis.
  • W is T -invariant for normal T iff W ⊥ is also T -
1 1 1 invariant.
U=√
2 −1 1 • For complete orthonormal sets, Fourier coefficients
give best approximation and projection formulas.
is orthogonal/unitary. • QR decomposition emerges from Gram–Schmidt:
Related / Associated Concepts A = QR
Related / Associated Concepts
for full-column-rank A.
• projections
• orthogonal matrices 8. Bilinear Forms, Dual Space, Dual Map,
• Hermitian matrices Annihilator, Quadratic Forms, Reduc-
• normal matrices tion, Definiteness, Congruence, Signa-
• spectral theorem
ture
• least squares
• QR decomposition Core Concept Skeleton
• duality and adjoints • linear functional
Core Concept Skeleton for Related Concepts • dual space
• dual basis
• orthogonality relations • transpose / dual map
• basis normalization • annihilator
• decomposition into orthogonal parts • bilinear pairing
• adjoint via bilinear pairing • matrix of bilinear form
• unitary invariance • symmetric / skew / alternating
• spectral decomposition • quadratic form
• polarization viewpoint
Formula Arsenal for Related Concepts
• diagonal reduction
∥x + y∥2 = ∥x∥2 + ∥y∥2 + 2ℜ⟨x, y⟩ • congruence
• rank / index / signature
∥x − y∥2 = ∥x∥2 + ∥y∥2 − 2ℜ⟨x, y⟩ • definiteness
ker(T ∗ T ) = ker T, Im(T ∗ T ) = Im T ∗ • Sylvester criterion
• positive semidefinite structure
T unitary ⇐⇒ T −1 = T ∗
T self-adjoint ⇐⇒ ⟨T x, x⟩ ∈ R ∀x
CSIR NET Mathematical Sciences — Linear Algebra Compendium 19

Definitions Theorems / Results / Statements

Linear functional. A linear map f : V → F. Dual space dimension. If dim V = n < ∞, then
Dual space.
V ∗ = L(V, F). dim V ∗ = n.
Dual basis. If (v1 , . . . , vn ) is a basis of V , then
Matrix of dual map. Relative to a basis and its dual
(f1 , . . . , fn ) in V ∗ with
basis,
fi (vj ) = δij [T ∗ ] = [T ]T .
Annihilator identities. For finite-dimensional V and
is the dual basis. subspaces X, Y ,
Dual map / transpose map. For T : V → W linear,
(X + Y )◦ = X ◦ ∩ Y ◦ ,
T ∗ : W ∗ → V ∗, T ∗ (g) = g ◦ T.
(X ∩ Y )◦ = X ◦ + Y ◦ ,
Annihilator. For U ⊆ V ,
(X ◦ )◦ = X.
U ◦ = {f ∈ V ∗ : f (u) = 0 ∀u ∈ U }.
Dual map kernel-image relations.
Bilinear form. A map B : V × V → F linear in each
ker T ∗ = (Im T )◦ , Im T ∗ = (ker T )◦ , rank T ∗ = rank T.
variable.
Symmetric bilinear form. Invariant annihilator. If X is T -invariant, then X ◦
is T ∗ -invariant.
B(u, v) = B(v, u).
Alternating vs skew-symmetric.
Skew-symmetric bilinear form. • Every alternating bilinear form is skew-symmetric.
• If char(F) ̸= 2, skew-symmetric ⇐⇒ alternating.
B(u, v) = −B(v, u). Decomposition of bilinear forms. If char(F) ̸= 2,
every bilinear form B decomposes uniquely as
Alternating bilinear form.
B = Bs + Ba ,
B(v, v) = 0 ∀v.
where Bs is symmetric and Ba is alternating, with
Quadratic form. A function
B(u, v) + B(v, u) B(u, v) − B(v, u)
q(x) = xT Ax Bs (u, v) = , Ba (u, v) = .
2 2
for a suitable matrix A (usually with A taken symmetric Matrix representation of bilinear form. Relative
over R). to basis (e1 , . . . , en ), there exists unique M such that
Congruent matrices. A, B ∈ Mn (F) are congruent if
B(u, v) = [u]T M [v].
T
B = P AP
Quadratic form and symmetry.
for some invertible P . • For real A, xT Ax depends only on the symmetric
Rank of quadratic form. Rank of its associated part:
symmetric matrix. A + AT
 
T T
Index of a real quadratic form. Number of positive x Ax = x x.
2
squares in its normal form.
Signature. Number of positive squares minus number • xT Ax = 0 for all real x iff A is skew-symmetric.
of negative squares. • If A, B are symmetric and xT Ax = xT Bx for all x,
then A = B.
Diagonal reduction of quadratic forms. Over
char(F) ̸= 2, every quadratic form is reducible by invert-
ible linear change of variables to

c1 y12 + · · · + cn yn2 .

Sylvester criterion. A real symmetric matrix A is

positive definite iff all leading principal minors are posi-
tive.
Negative definite criterion. A real symmetric is
negative definite iff

(−1)k ∆k > 0, k = 1, . . . , n,

where ∆k are leading principal minors.

Positive semidefinite matrix characterization.

A ⪰ 0 ⇐⇒ xT Ax ≥ 0 ∀x.

Positive definite implies nonsingular.

CSIR NET Mathematical Sciences — Linear Algebra Compendium 20

Direct Results / Formulae Failure Cases / Exceptions / Warnings

Dual basis:   • Bilinear form and inner product are not the same;
n
X positivity/conjugate symmetry may fail.
fi  aj vj  = ai . • In characteristic 2, symmetric/skew-
j=1 symmetric/alternating distinctions behave dif-
Quadratic form in variables: ferently.
• Sylvester criterion applies to real symmetric matrices,
q(x) = xT Ax not arbitrary nonsymmetric matrices.
• xT Ax cannot detect skew-symmetric part over R.
with symmetric reduction • Similar matrices need not define equivalent quadratic
forms; congruence is the right equivalence.
A + AT
As = , q(x) = xT As x.
2 Examples
Positive definite: Basic. On R2 , the bilinear form
xT Ax > 0 ∀x ̸= 0. B((x1 , x2 ), (y1 , y2 )) = x1 y1 + 2x2 y2
Positive semidefinite: has matrix  
1 0
T
x Ax ≥ 0 ∀x. M= .
0 2
Negative definite: Exam-style. For

xT Ax < 0 ∀x ̸= 0. q(x, y) = x2 + 4xy + 3y 2 ,

Indefinite: takes both positive and negative values. associated symmetric matrix is
Signature: If normal form has p positive and r − p  
negative squares, then 1 2
A= .
2 3
signature = p − (r − p) = 2p − r.
Leading principal minors are 1 > 0, det A = −1 < 0, so
Index: q is indefinite.
index = p. Positive definite example.
Hadamard inequality for positive semidefinite A = (aij ):
 
2 1
A=
n
1 2
Y
| det A| ≤ aii .
has leading minors 2, 3 > 0, hence positive definite.
i=1
Annihilator example. If V = R3 and X =
Positive semidefinite block: span{(1, 0, 0), (0, 1, 0)}, then

X ◦ = span{f },
 
I A f (x, y, z) = z.
⪰ 0.
AT AT A
Related / Associated Concepts
Alternative / Equivalent / Easier Forms
Related / Associated Concepts
• Bilinear form matrix changes by congruence, not • dual space
similarity. • adjoint / transpose
• Quadratic form is determined by symmetric part of • symmetric matrices
A. • orthogonal diagonalization
• For real quadratic forms, diagonalization under or- • spectral theorem
thogonal change needs symmetry; under general in-
• inertia law
vertible change, use congruence.
• positive definiteness tests
• Over R, positive definite ⇐⇒ all eigenvalues positive
• optimization and Hessians
for symmetric matrix.
Core Concept Skeleton for Related Concepts
• basis-dual basis pairing
• matrix by coordinates
• symmetric part extraction
• congruence transformation
• inertia invariance
• principal minors
• eigenvalue-sign relation
CSIR NET Mathematical Sciences — Linear Algebra Compendium 21

Formula Arsenal for Related Concepts Compact Summary Lines

M (T ∗ ) = M (T )T Ultra-Compact High-Yield Summary

B + BT B − BT • Subspace test: nonempty + αx + βy closure.

Bs = , Ba = • Basis = independent + spanning; in dim n, n inde-
2 2
 T
 pendent or n spanning vectors already give basis.
A+A
xT Ax = xT x • dim(W1 + W2 ) = dim W1 + dim W2 − dim(W1 ∩ W2 ).
2 • V / ker T ∼
= Im T , and dim V = rank T + nullity T .
A ≻ 0 ⇐⇒ ∆k > 0 ∀k • T invertible ⇐⇒ injective ⇐⇒ surjective in equal
finite dimension.
A ≺ 0 ⇐⇒ (−1)k ∆k > 0 ∀k
• det(AB) = det A det B, tr(AB) = tr(BA),
CSIR NET Utility Layer rank(A) = rank(AT ).
• Orthogonal/unitary ⇒ A−1 = AT /A∗ , eigenvalues
• For quadratic form questions, first symmetrize the have modulus 1.
matrix. • Idempotent ⇒ eigenvalues 0, 1; involutory ⇒ eigen-
• In 2 × 2 case, positive definite iff a11 > 0 and deter- values ±1; nilpotent ⇒ only 0.
minant > 0. • λ eigenvalue ⇐⇒ det(A − λI) = 0.
• If real symmetric matrix has one positive and one • A diagonalizable ⇐⇒ eigenbasis exists ⇐⇒ mA
negative eigenvalue, form is indefinite. splits into distinct linear factors.
• Signature/index are congruence invariants, not simi- • Real symmetric/Hermitian ⇒ orthogonally/unitarily
larity invariants. diagonalizable with real eigenvalues.
• For dual basis questions, just read off coefficients in • AX = b consistent ⇐⇒ rank(A) = rank([A|b]);
the given basis. homogeneous solution dimension = n − rank(A).
• Common trap: using determinant alone to test posi- • In inner product spaces: Cauchy–Schwarz, Gram–
tive definiteness in dimension > 2. Schmidt, V = W ⊕ W ⊥ .
Research-Based Enrichment Layer • T unitary ⇐⇒ T ∗ T = I; self-adjoint ⇒ real spec-
trum; normal ⇐⇒ orthonormal eigenbasis over C.
• Sylvester’s law of inertia: rank, index, signature are • Quadratic form xT Ax depends only on A+A
T
; pos-
2
invariant under real congruence. itive definite iff symmetric and Sylvester criterion
• For symmetric positive semidefinite A, there exists holds.
B such that A = B T B.
• A real matrix A is positive definite in quadratic-form
sense iff its symmetric part is positive definite.
• Hessian-based second derivative tests in multivariable
calculus are direct applications of quadratic form
definiteness.