## What Is a Fréchet Derivative?

Let $U$ and $V$ be Banach spaces (complete normed vector spaces). The Fréchet derivative of a function $f:U \to V$ at $X\in U$ is a linear mapping $L:U\to V$ such that $\notag f(X+E) - f(X) - L(X,E) = o(\|E\|)$

for all $E\in U$. The notation $L(X,E)$ should be read as “the Fréchet derivative of $f$ at $X$ in the direction $E$”. The Fréchet derivative may not exist, but if it does exist then it is unique. When $U = V = \mathbb{R}$, the Fréchet derivative is just the usual derivative of a scalar function: $L(x,e) = f'(x)e$.

As a simple example, consider $U = V = \mathbb{R}^{n\times n}$ and $f(X) = X^2$. From the expansion $f(X+E) - f(X) = XE + EX + E^2$

we deduce that $L(X,E) = XE + EX$, the first order part of the expansion. If $X$ commutes with $E$ then $L_X(E) = 2XE = 2EX$.

More generally, it can be shown that if $f$ has the power series expansion $f(x) = \sum_{i=0}^\infty a_i x^i$ with radius of convergence $r$ then for $X,E\in\mathbb{R}^{n\times n}$ with $\|X\| < r$, the Fréchet derivative is $L(X,E) = \displaystyle\sum_{i=1}^\infty a_i \displaystyle\sum_{j=1}^i X^{j-1} E X^{i-j}.$

An explicit formula for the Fréchet derivative of the matrix exponential, $f(A) = \mathrm{e}^A$, is $L(A,E) = \displaystyle\int_0^1 \mathrm{e}^{A(1-s)} E \mathrm{e}^{As} \, ds.$

Like the scalar derivative, the Fréchet derivative satisfies sum and product rules: if $g$ and $h$ are Fréchet differentiable at $A$ then \notag \begin{alignedat}{2} f &= \alpha g + \beta h &&\;\Rightarrow\; L_f(A,E) = \alpha L_g(A,E) + \beta L_h(A,E),\\ f &= gh &&\;\Rightarrow\; L_f(A,E) = L_g(A,E) h(A) + g(A) L_h(A,E). \end{alignedat}

A key requirement of the definition of Fréchet derivative is that $L(X,E)$ must satisfy the defining equation for all $E$. This is what makes the Fréchet derivative different from the Gâteaux derivative (or directional derivative), which is the mapping $G:U \to V$ given by $\notag G(X,E) = \lim_{t\to0} \displaystyle\frac{f(X+t E)-f(X)}{t} = \frac{\mathrm{d}}{\mathrm{dt}}\Bigm|_{t=0} f(X + tE).$

Here, the limit only needs to exist in the particular direction $E$. If the Fréchet derivative exists at $X$ then it is equal to the Gâteaux derivative, but the converse is not true.

A natural definition of condition number of $f$ is $\mathrm{cond}(f,X) = \lim_{\epsilon\to0} \displaystyle\sup_{\|\Delta X\| \le \epsilon \|X\|} \displaystyle\frac{\|f(X+\Delta X) - f(X)\|}{\epsilon\|f(X)\|},$

and it can be shown that $\mathrm{cond}$ is given in terms of the Fréchet derivative by $\mathrm{cond}(f,X) = \displaystyle\frac{\|L(X)\| \|X\|}{\|f(X)\|},$

where $\|L(X)\| = \sup_{Z\ne0}\displaystyle\frac{\|L(X,Z)\|}{\|Z\|}.$

For matrix functions, the Fréchet derivative has a number of interesting properties, one of which is that the eigenvalues of $L(X)$ are the divided differences $\notag f[\lambda_i,\lambda_j] = \begin{cases} \dfrac{ f(\lambda_i)-f(\lambda_j) }{\lambda_i - \lambda_j}, & \lambda_i\ne\lambda_j, \\ f'(\lambda_i), & \lambda_i=\lambda_j, \end{cases}$

for $1\le i,j \le n$, where the $\lambda_i$ are the eigenvalues of $X$. We can check this formula in the case $F(X) = X^2$. Let $(\lambda,u)$ be an eigenpair of $X$ and $(\mu,v)$ an eigenpair of $X^T$, so that $Xu = \lambda u$ and $X^Tv = \mu v$, and let $E = uv^T$. Then $L(X,E) = XE + EX = Xuv^T + uv^TX = (\lambda + \mu) uv^T.$

So $uv^T$ is an eigenvector of $L(X)$ with eigenvalue $\lambda + \mu$. But $f[\lambda,\mu] = (\lambda^2-\mu^2)/(\lambda - \mu) = \lambda + \mu$ (whether or not $\lambda$ and $\mu$ are distinct).

## References

This is a minimal set of references, which contain further useful references within.