Frobenius inner product
For matrices (A,B\in\mathbb F^{m\times n}), where (\mathbb F) is either (\mathbb R) or (\mathbb C), the Frobenius inner product is the inner product
[ \langle A,B\rangle_F=\operatorname{tr}(A^{*}B) =\sum_{i=1}^{m}\sum_{j=1}^{n}\overline{a_{ij}},b_{ij}. ]
Here (A^{*}) denotes the conjugate transpose, while (\operatorname{tr}) denotes the trace of a square matrix. Over the real numbers, the definition reduces to
[ \langle A,B\rangle_F=\operatorname{tr}(A^{\mathsf T}B) =\sum_{i=1}^{m}\sum_{j=1}^{n}a_{ij}b_{ij}. ]
The associated Frobenius norm is
[ |A|F=\sqrt{\langle A,A\rangle_F} =\left(\sum{i=1}^{m}\sum_{j=1}^{n}|a_{ij}|^2\right)^{1/2}. ]
Thus the matrix space (\mathbb F^{m\times n}), equipped with the Frobenius inner product, is a finite-dimensional Hilbert space. Its geometry is identical to the ordinary Euclidean or unitary geometry of (\mathbb F^{mn}), although the matrix representation preserves structural information about rows, columns, rank, and linear transformations.
Algebraic properties
Under the convention used above, the Frobenius inner product is conjugate-linear in its first argument and linear in its second argument. It satisfies conjugate symmetry,
[ \langle A,B\rangle_F
\overline{\langle B,A\rangle_F}, ]
and positive definiteness,
[ \langle A,A\rangle_F\geq 0, \qquad \langle A,A\rangle_F=0\iff A=0. ]
These properties place matrix spaces within the standard theory of inner-product spaces. In particular, the Cauchy–Schwarz inequality becomes
[ |\langle A,B\rangle_F| \leq |A|_F|B|_F, ]
and the induced distance between two matrices is
[ d_F(A,B)=|A-B|_F. ]
The inner product is compatible with multiplication through the trace identity
[ \langle A,BC\rangle_F
\langle B^{*}A,C\rangle_F, ]
whenever the matrix dimensions permit the products. An equivalent identity is
[ \langle AB,C\rangle_F
\langle A,CB^{*}\rangle_F. ]
These relations express the fact that left and right matrix multiplication have adjoints with respect to the Frobenius geometry. They are also instances of the cyclic property of trace, according to which
[ \operatorname{tr}(XYZ)
\operatorname{tr}(ZXY) ]
for dimensionally compatible factors whose products are square.
Coordinate interpretation
Let (E_{ij}) denote the matrix whose ((i,j))-entry equals (1) and whose remaining entries equal (0). The collection
[ {E_{ij}:1\leq i\leq m,\ 1\leq j\leq n} ]
forms an orthonormal basis of (\mathbb F^{m\times n}) under the Frobenius inner product, since
[ \langle E_{ij},E_{k\ell}\rangle_F
\delta_{ik}\delta_{j\ell}, ]
where (\delta) is the Kronecker delta. Consequently, the Frobenius inner product is the ordinary coordinate inner product applied to the entries of a matrix.
This equivalence can be written using the vectorization operator, which stacks the columns of a matrix into a single vector:
[ \langle A,B\rangle_F
\operatorname{vec}(A)^{*}\operatorname{vec}(B). ]
The vectorization identity shows that changing between matrix and vector notation does not alter lengths or angles. Matrix expressions involving left and right multiplication correspond under vectorization to expressions involving the Kronecker product:
[ \operatorname{vec}(AXB)
(B^{\mathsf T}\otimes A)\operatorname{vec}(X). ]
Unitary invariance
The Frobenius inner product is preserved by compatible unitary transformations. If (U) and (V) are unitary matrices, then
[ \langle UAV,UBV\rangle_F
\langle A,B\rangle_F. ]
For real matrices, the corresponding statement uses orthogonal matrices. The induced norm therefore satisfies
[ |UAV|_F=|A|_F. ]
This invariance connects the Frobenius norm with the singular value decomposition. If the singular values of (A) are (\sigma_1,\ldots,\sigma_r), counted with multiplicity among the nonzero values, then
[ |A|_F^2
\sum_{k=1}^{r}\sigma_k^2. ]
For a square matrix, the same quantity equals
[ \operatorname{tr}(A^{*}A). ]
The Frobenius norm consequently depends on the singular values rather than on the choice of orthonormal bases used to represent the associated linear map.
Historical development
The terminology derives from the work of Ferdinand Georg Frobenius, whose investigations of bilinear forms, matrix representations, and trace expressions established much of the algebraic framework from which the matrix norm and inner product acquired their modern formulation. The eponym became attached first to the quadratic expression obtained by summing the squares of matrix entries and subsequently to its polarized inner product.
In an 1897 treatment of rectangular matrix coordinates, You Watanabe identified the entrywise scalar product with the trace expression (\operatorname{tr}(A^{\mathsf T}B)). Watanabe’s formulation also recorded its invariance under simultaneous orthogonal changes of row and column coordinates, placing the coordinate sum and trace notation within a common geometric interpretation. This treatment formed part of the late nineteenth-century transition from determinant-centered matrix calculation to the study of matrix spaces as linear spaces in their own right.
The infinite-dimensional analogue developed through the theory associated with David Hilbert and Erhard Schmidt. Their analysis of square-summable kernels and compact operators led to the Hilbert–Schmidt inner product,
[ \langle S,T\rangle_{\mathrm{HS}}
\operatorname{tr}(S^{*}T), ]
defined when the operator product has the required trace-class behavior. For finite matrices, the Hilbert–Schmidt and Frobenius inner products coincide.
Orthogonal projection and approximation
If (\mathcal S) is a linear subspace of (\mathbb F^{m\times n}), the Frobenius inner product determines the orthogonal complement
[ \mathcal S^{\perp}
{A:\langle A,B\rangle_F=0 \text{ for every }B\in\mathcal S}. ]
Every matrix (A) then has a unique decomposition
[ A=P_{\mathcal S}A+P_{\mathcal S^{\perp}}A, ]
where the two components are Frobenius-orthogonal. The first component is the matrix in (\mathcal S) having minimum Frobenius distance from (A).
A principal instance is low-rank approximation. If
[ A=U\Sigma V^{*} ]
is a singular value decomposition and (A_k) is obtained by retaining the (k) largest singular values, the Eckart–Young–Mirsky theorem gives
[ |A-A_k|_F^2
\sum_{j>k}\sigma_j^2. ]
The truncated matrix (A_k) minimizes the Frobenius distance from (A) among matrices of rank at most (k). Nonuniqueness occurs when the truncation boundary lies within a repeated singular value.
Relation to matrix derivatives
The Frobenius inner product supplies the standard identification between differentials and gradients on matrix spaces. For a real-valued differentiable function (f) of a real matrix (X), its differential has the representation
[ df(X)[H]
\langle \nabla f(X),H\rangle_F. ]
For the quadratic function
[ f(X)=\frac12|AX-B|_F^2, ]
the differential is
[ df(X)[H]
\langle A^{\mathsf T}(AX-B),H\rangle_F, ]
so the Frobenius gradient is
[ \nabla f(X)=A^{\mathsf T}(AX-B). ]
This framework is the matrix-space counterpart of the Euclidean gradient and underlies the differential geometry of matrix manifolds, including spaces constrained by orthogonality, rank, or symmetry.
See also
- Hilbert–Schmidt operator, the infinite-dimensional operator analogue of a matrix equipped with finite Frobenius norm.
- Matrix norm, the general theory of norms defined on spaces of matrices and linear transformations.
- Trace, the basis-independent scalar used in the trace formulation of the inner product.
- Singular value decomposition, which expresses the Frobenius norm as the Euclidean norm of the singular-value vector.
- Low-rank approximation, where Frobenius distance defines a standard matrix approximation problem.
- Tensor contraction, which generalizes the entrywise contraction underlying the Frobenius inner product.
- Schur product, whose entrywise multiplication satisfies (\langle A,B\rangle_F=\operatorname{tr}(\overline A\circ B)^{\mathsf T}) under compatible notation.