What is the trace of a matrix?

The trace of an n × n matrix is the sum of its main diagonal entries: tr(A) = a₁₁ + a₂₂ + ⋯ + aₙₙ. It is defined only for square matrices. The trace of the identity matrix is n, and the trace of the zero matrix is 0.

How is the trace related to eigenvalues?

The trace of a matrix equals the sum of all its eigenvalues, counted with algebraic multiplicity. This connects a trivially computable quantity — adding diagonal entries — to spectral information that normally requires solving the characteristic polynomial. The determinant similarly equals the product of the eigenvalues.

What is the cyclic property of the trace?

The trace is invariant under cyclic permutations of a matrix product: tr(AB) = tr(BA), and more generally tr(ABC) = tr(BCA) = tr(CAB). Only cyclic reorderings are valid — swapping two adjacent factors changes the trace in general. This property also implies the trace is unchanged by similarity transformations.

What is the Frobenius inner product?

The Frobenius inner product of two n × n matrices A and B is tr(AᵀB), which equals the sum of all entry-by-entry products. It is the dot product of the matrices viewed as vectors of n² entries. The associated Frobenius norm is the square root of tr(AᵀA).

Trace of a Matrix: Properties & Identities

Q: Why is the trace a similarity invariant?

For any invertible matrix P, tr(P⁻¹AP) = tr(A). This follows directly from the cyclic property. Since similar matrices represent the same linear transformation in different bases, the trace is a property of the transformation itself, independent of the chosen coordinate system.

Definition

For an

n \times n

matrix

A

, the trace is the sum of the entries on the main diagonal:

\text{tr}(A) = \sum_{i=1}^{n} a_{ii} = a_{11} + a_{22} + \cdots + a_{nn}

The trace is defined only for square matrices — a rectangular matrix has no trace.

For example, if

A = \begin{pmatrix} 3 & 1 & 4 \\ 1 & 5 & 9 \\ 2 & 6 & 5 \end{pmatrix}

, then

\text{tr}(A) = 3 + 5 + 5 = 13

. The off-diagonal entries play no role.

The trace of the

n \times n

identity matrix is

\text{tr}(I_n) = n

, since there are

n

ones on the diagonal. The trace of the zero matrix is

0

.

Linearity

The trace is a linear function from the space of

n \times n

matrices to the real numbers. It satisfies additivity:

\text{tr}(A + B) = \text{tr}(A) + \text{tr}(B)

and scalar homogeneity:

\text{tr}(cA) = c \cdot \text{tr}(A)

Combined, these say that

\text{tr}(cA + dB) = c \cdot \text{tr}(A) + d \cdot \text{tr}(B)

for any scalars

c, d

and any

n \times n

matrices

A, B

. Both properties follow immediately from the definition — the sum of the diagonals of

A + B

is the sum of the individual diagonal sums, and scaling every entry by

c

scales each diagonal entry by

c

.

The transpose does not affect the trace:

\text{tr}(A^T) = \text{tr}(A)

, since transposition does not move the diagonal entries.

It is worth contrasting the trace with the determinant. The determinant is multiplicative (

\det(AB) = \det(A)\det(B)

) but not additive (

\det(A + B) \neq \det(A) + \det(B)

in general). The trace is additive but not multiplicative —

\text{tr}(AB)

generally has no relation to

\text{tr}(A) \cdot \text{tr}(B)

. Each captures different structural information about the matrix.

Property	Trace	Determinant
Additivity	tr(A + B) = tr(A) + tr(B)	det(A + B) ≠ det(A) + det(B) in general
Multiplicativity	tr(AB) has no simple relation to tr(A)·tr(B)	det(AB) = det(A) · det(B)
Transpose	tr(Aᵀ) = tr(A)	det(Aᵀ) = det(A)
Eigenvalue link	sum: λ₁ + λ₂ + ⋯ + λₙ	product: λ₁ · λ₂ · ⋯ · λₙ
Similarity invariance	tr(P⁻¹AP) = tr(A)	det(P⁻¹AP) = det(A)

Trace of Special Matrices

For a diagonal matrix

D = \text{diag}(d_1, \dots, d_n)

, the trace is simply

d_1 + d_2 + \cdots + d_n

— the entire matrix reduces to its diagonal, and the trace reads off everything. A scalar matrix

cI

has trace

cn

.

A skew-symmetric matrix has all diagonal entries equal to zero (since

a_{ii} = -a_{ii}

forces

a_{ii} = 0

), so

\text{tr}(A) = 0

for every real skew-symmetric matrix.

An idempotent matrix — one satisfying

A^2 = A

— has a striking property:

\text{tr}(A) = \text{rank}(A)

. The eigenvalues of an idempotent matrix are restricted to

0

and

1

, the trace counts the number of eigenvalues equal to

1

, and this count equals the dimension of the column space.

A nilpotent matrix has all eigenvalues equal to zero, so

\text{tr}(A) = 0

. More generally,

\text{tr}(A^k) = 0

for every positive integer

k

, since the eigenvalues of

A^k

are the

k

-th powers of the eigenvalues of

A

, and

0^k = 0

.

Matrix type	Defining condition	Trace
Diagonal D = diag(d₁,…,dₙ)	nonzero entries only on the diagonal	d₁ + d₂ + ⋯ + dₙ
Scalar cI	every diagonal entry equals c	cn
Identity Iₙ	diagonal with every diagonal entry equal to 1	n
Skew-symmetric	Aᵀ = −A (forces zero diagonal)	0
Idempotent	A² = A	rank(A)
Nilpotent	Aᵏ = O for some k ≥ 1	0

The Cyclic Property

The most distinctive algebraic property of the trace is its invariance under cyclic permutations of a product. For any two matrices

A

and

B

where both products

AB

and

BA

are defined:

\text{tr}(AB) = \text{tr}(BA)

Note that

AB

and

BA

need not even have the same dimensions — if

A

is

m \times n

and

B

is

n \times m

, then

AB

is

m \times m

and

BA

is

n \times n

. The traces of these differently-sized matrices are nevertheless equal.

The proof is a direct computation. The

(i,i)

entry of

AB

is

\sum_k a_{ik} b_{ki}

, so

\text{tr}(AB) = \sum_i \sum_k a_{ik} b_{ki}

. The

(k,k)

entry of

BA

is

\sum_i b_{ki} a_{ik}

, so

\text{tr}(BA) = \sum_k \sum_i b_{ki} a_{ik}

. Both double sums range over the same index pairs and contain the same terms.

For three matrices, the cyclic property extends to

\text{tr}(ABC) = \text{tr}(BCA) = \text{tr}(CAB)

Only cyclic reorderings are permitted. The rearrangement

\text{tr}(ABC) = \text{tr}(ACB)

is false in general — swapping two adjacent factors is not a cyclic permutation.

Trace and Eigenvalues

The trace of a matrix equals the sum of its eigenvalues, counted with algebraic multiplicity:

\text{tr}(A) = \lambda_1 + \lambda_2 + \cdots + \lambda_n

This identity connects a trivially computable quantity (add the diagonal entries) to eigenvalue information that ordinarily requires solving a degree-

n

polynomial.

The proof comes from the characteristic polynomial

p(\lambda) = \det(A - \lambda I)

. Expanding this determinant produces a polynomial of degree

n

whose leading term is

(-\lambda)^n

and whose

\lambda^{n-1}

coefficient is

(-1)^{n-1} \text{tr}(A)

. By Vieta's formulas, the sum of the roots of

p

equals

\text{tr}(A)

.

A companion identity links the determinant to the product of eigenvalues:

\det(A) = \lambda_1 \lambda_2 \cdots \lambda_n

. Together, the trace and the determinant capture the two simplest symmetric functions of the eigenvalue spectrum — their sum and their product.

For a

3 \times 3

matrix with eigenvalues

2, -1, 4

, the trace is

5

and the determinant is

-8

. Neither the trace nor the determinant individually determines the eigenvalues, but together they constrain them heavily. For a

2 \times 2

matrix, the trace and determinant determine the eigenvalues completely via the quadratic formula.

Trace and Similarity

Two matrices

A

and

B

are similar if

B = P^{-1}AP

for some invertible matrix

P

. Similar matrices represent the same linear transformation in different coordinate systems —

P

encodes the change of basis.

The trace is invariant under similarity:

\text{tr}(P^{-1}AP) = \text{tr}(A)

This follows in one step from the cyclic property:

\text{tr}(P^{-1}AP) = \text{tr}(APP^{-1}) = \text{tr}(A)

.

Invariance under similarity means the trace is a property of the transformation itself, not of any particular matrix representation. No matter which basis is chosen, the trace comes out the same. The eigenvalues share this invariance (similar matrices have the same eigenvalues), and indeed

\text{tr}(A) = \sum \lambda_i

is an eigenvalue identity, so trace invariance and eigenvalue invariance are two sides of the same coin.

The determinant is also a similarity invariant:

\det(P^{-1}AP) = \det(A)

. Together with the trace, it forms the beginning of a sequence of similarity invariants — the coefficients of the characteristic polynomial — that collectively determine the eigenvalue structure of the transformation.

The Frobenius Inner Product

The trace provides a natural inner product on the space of

n \times n

matrices. For two matrices

A

and

B

, the Frobenius inner product is

\langle A, B \rangle_F = \text{tr}(A^T B) = \sum_{i=1}^{n} \sum_{j=1}^{n} a_{ij} b_{ij}

This is the dot product of

A

and

B

viewed as vectors of

n^2

entries. It is symmetric (

\langle A, B \rangle_F = \langle B, A \rangle_F

), linear in each argument, and positive definite (

\langle A, A \rangle_F > 0

whenever

A \neq O

).

The associated norm is the Frobenius norm:

\|A\|_F = \sqrt{\text{tr}(A^T A)} = \sqrt{\sum_{i,j} a_{ij}^2}

This measures the "total size" of a matrix as the square root of the sum of squares of all entries — the matrix analogue of the Euclidean length of a vector.

The Frobenius inner product turns the space of

n \times n

matrices into an inner product space, bringing geometric concepts — angles, orthogonality, projections, distances — to bear on matrices themselves, not just on the vectors they act upon.

Trace of Commutators

The commutator of two

n \times n

matrices is the matrix

[A, B] = AB - BA

The commutator measures how far

A

and

B

are from commuting — it is zero if and only if

AB = BA

.

Regardless of what

A

and

B

are, the commutator always has trace zero:

\text{tr}([A, B]) = \text{tr}(AB - BA) = \text{tr}(AB) - \text{tr}(BA) = 0

The cancellation is a direct consequence of the cyclic property. This means the identity matrix

I

can never be a commutator, since

\text{tr}(I) = n \neq 0

. In particular, there exist no

n \times n

matrices

A, B

satisfying

AB - BA = I

when working over the real or complex numbers with finite-dimensional matrices.

The converse does not hold in general: a traceless matrix is not necessarily a commutator, though in the space of

n \times n

matrices over a field, every traceless matrix can in fact be written as a commutator — a result that requires proof beyond the trace identity itself.

Trace Identities

Several identities involving the trace appear frequently enough to be worth collecting.

If

S

is symmetric and

K

is skew-symmetric, then

\text{tr}(SK) = 0

. The proof:

\text{tr}(SK) = \text{tr}((SK)^T) = \text{tr}(K^T S^T) = \text{tr}(-KS) = -\text{tr}(KS) = -\text{tr}(SK)

, where the last step uses the cyclic property. The only number equal to its own negative is zero.

The trace can be written as a sum of quadratic forms against the standard basis:

\text{tr}(A) = \sum_{i=1}^{n} \mathbf{e}_i^T A \mathbf{e}_i

. Each term

\mathbf{e}_i^T A \mathbf{e}_i = a_{ii}

extracts one diagonal entry. This formula generalizes: for any orthonormal basis

\{\mathbf{q}_1, \dots, \mathbf{q}_n\}

,

\text{tr}(A) = \sum_{i=1}^{n} \mathbf{q}_i^T A \mathbf{q}_i

The result is independent of which orthonormal basis is used — another manifestation of the trace's invariance under orthogonal change of coordinates.

Trace in Differentiation

Many objective functions in optimization and statistics are expressed as traces of matrix products, and computing their gradients requires differentiating with respect to a matrix variable.

The simplest case is the linear function

f(X) = \text{tr}(AX)

, where

A

is a fixed matrix and

X

is the variable. The derivative with respect to

X

is

\frac{\partial}{\partial X} \text{tr}(AX) = A^T

For the quadratic form

f(X) = \text{tr}(X^T A X)

, the derivative is

\frac{\partial}{\partial X} \text{tr}(X^T A X) = (A + A^T)X

When

A

is symmetric, this simplifies to

2AX

.

These formulas are the matrix analogues of the scalar rules

\frac{d}{dx}(ax) = a

and

\frac{d}{dx}(ax^2) = 2ax

. They appear in deriving the normal equations for least squares, in the gradient descent updates for matrix factorization problems, and in the analysis of covariance estimators. The trace's linearity and cyclic property make these derivatives clean and systematic — full matrix calculus extends these patterns to products of arbitrary length and composition.

Trace Properties at a Glance

The properties developed across the preceding sections can be collected for quick reference. Each row pairs the defining identity with the structural fact it carries — the trace owes its usefulness to this short list of algebraic guarantees.

Property	Identity	Why it matters
Linearity	tr(cA + dB) = c · tr(A) + d · tr(B)	trace is a linear functional on n × n matrices
Transpose invariance	tr(Aᵀ) = tr(A)	transposition leaves the diagonal entries in place
Cyclic property	tr(AB) = tr(BA); tr(ABC) = tr(BCA) = tr(CAB)	foundation of similarity invariance and the zero-commutator-trace
Similarity invariance	tr(P⁻¹AP) = tr(A)	trace is a property of the transformation, not the chosen basis
Eigenvalue sum	tr(A) = λ₁ + λ₂ + ⋯ + λₙ	reads spectral information off the diagonal
Commutator trace	tr(AB − BA) = 0	the identity matrix is never a commutator
Frobenius inner product	⟨A, B⟩_F = tr(AᵀB)	gives n × n matrices a geometry — angles, lengths, projections

Trace of a Matrix

The Simplest Matrix Invariant