I know that,

A function $f: \mathbb{R}^n \to \mathbb{R}$ is said to be differentiable at $x$ if there exists a vector $v$ such that, $$ \lim_{h \to 0} \frac{f(x+h) - f(x) - v^Th} {\|h\|} = 0. $$ When $v$ exists, it is given by the "gradient" $\nabla f(x) = \left(\frac{\partial f}{\partial x_1}, ..., \frac{\partial f}{\partial x_n}\right)(x)$

Does there exist a similar definition for "matrix derivative"


  • $\begingroup$ As far as physics is concerned, it is defined in the same way as in systems of ODE's. For example, the equation for an open system follows a master equation of the form $\dot{\rho}=\mathcal{L}(\rho)$, where $\rho$ is a matrix (a density matrix) and $\mathcal{L}$ is the "Lindbladian" which is a superoperator that takes operators into operators (in finite Hilbert spaces, $\rho$ can be represented as a matrix). In order to solve this equation, say numerically, one can consider that every $\rho_{ij}$ fulfills an ordinary ODE. Overall, one ends up with a system of ODEs to be solved. $\endgroup$ – user2820579 Jan 21 at 21:28
  • $\begingroup$ Replace $v^T$ with a linear map $L\colon \Bbb R^n \to \Bbb R^k$. $\endgroup$ – Ivo Terek Jan 21 at 21:54
  • $\begingroup$ Just reiterating what has been said here, but what you're looking for is the Frechet derivative. See my answer here which later also provides a sample calculation. Also, you may want to take a look at this for the motivation of the general definition. $\endgroup$ – peek-a-boo Jan 22 at 8:14

Matrix derivation is just a particular case of Fréchet derivative between two Banach spaces. Which by the way is very similar in term of definition to the definition of the derivative of a function $f : \mathbb R^n \to \mathbb R$ provided in the question.

Applied to matrix derivatives, you just have to consider a map $f : V \to M$ where $M$ is a linear space of matrices endowed with the norm of your choice and $V$ a Banach space that can be (or not) of finite dimension.


Yes, you can use a very similar definition. First of all, the map $h \mapsto v^\top h$ encodes an arbitrary linear map on $\mathbb R^n$. For matrices, you can substitute it by the Frobenius inner product, e.g., $$A \mapsto (A,B)_F := \sum_{i,j = 1}^n A_{ij} B_{ij},$$ where $B \in \mathbb R^{n \times n}$ is fixed.

Thus, for $f \colon \mathbb R^{n \times n} \to \mathbb R$ you can define $\nabla f(A)$ to be the (unique) matrix $B \in \mathbb R^{n \times n}$ (if it exists), that satisfies $$ \lim_{H \to 0} \frac{f(A + H) - f(A) - (B,H)_F}{\|H\|} = 0.$$

This definition directly extends to Hilbert spaces (in which you use the inner product determined by the Hilbert space structure). In Banach spaces, one uses the duality product and this yields a derivative which belongs to the dual space.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.