Canonical: https://knowlpedia.com/real-analysis/frechet-derivative/

# Fréchet derivative

The derivative of a multivariable function as a best linear approximation at a point

Let \(E,F\) be real [normed vector spaces](https://knowlpedia.com/linear-algebra/normed-vector-space/), \(U\subset E\) open, and \(a\in U\). The **Fréchet derivative** of \(f:U\to F\) at \(a\) is a [bounded linear map](https://knowlpedia.com/functional-analysis/bounded-linear-operator/) \(Df(a):E\to F\) such that
\[
\lim_{h\to0,\ h\ne0}\frac{\|f(a+h)-f(a)-Df(a)h\|_F}{\|h\|_E}=0.
\]
It is the linear approximation with an error negligible compared with the size of the increment. Completeness of the spaces is not required.

## Uniqueness

If \(L_1,L_2\) both satisfy the definition, use increments \(h=tv\) for a fixed vector \(v\). Dividing the difference of their remainders by \(|t|\) and letting \(t\to0\) gives \((L_1-L_2)v=0\). Thus the derivative is unique.

## Euclidean spaces

For \(E=\mathbb R^n,F=\mathbb R^m\), the derivative is represented by the [Jacobian matrix](https://knowlpedia.com/real-analysis/jacobian-matrix/). Fréchet differentiability implies existence of the partial derivatives; their existence alone does not imply Fréchet differentiability. For \(f(x,y)=(x^2y,x+y)\),
\[
Df(a,b)=\begin{pmatrix}2ab&a^2\\1&1\end{pmatrix}.
\]
A bounded linear function \(f(x)=Ax\) has derivative \(Df(a)=A\) at every point.
