Fréchet derivative
The Fréchet derivative is a derivative defined for mappings between normed vector spaces. It represents the first-order variation of a mapping by a bounded linear operator and requires the approximation error to become negligible uniformly with respect to the direction of the increment. This uniformity distinguishes it from weaker notions based only on individual directional limits.
The derivative is named after Maurice Fréchet, whose work on abstract spaces provided a general setting for differential calculus beyond finite-dimensional coordinates. In finite-dimensional spaces, Fréchet differentiability agrees with the usual notion of total differentiability and is represented by the Jacobian matrix.
Definition
Let (X) and (Y) be normed vector spaces over the same scalar field, let (U\subseteq X) be open, and let
[ f:U\rightarrow Y ]
be a mapping. The function (f) is Fréchet differentiable at (x\in U) when there exists a bounded linear operator
[ A:X\rightarrow Y ]
such that
[ \lim_{\lVert h\rVert_X\to 0} \frac{\lVert f(x+h)-f(x)-A(h)\rVert_Y} {\lVert h\rVert_X} =0. ]
The operator (A) is uniquely determined and is denoted by (Df(x)), (D f_x), or (f'(x)). The defining relation is equivalently written as
[ f(x+h)=f(x)+Df(x)h+r(h), ]
where the remainder satisfies
[ \frac{\lVert r(h)\rVert_Y}{\lVert h\rVert_X}\longrightarrow 0 \qquad\text{as }h\to 0. ]
The requirement that (Df(x)) be bounded is equivalent to continuity for linear operators between normed spaces. Its boundedness ensures that the linear term describes a continuous first-order response to perturbations in the argument.
If (f) is Fréchet differentiable at every point of (U), the derivative defines a mapping
[ Df:U\rightarrow \mathcal L(X,Y), ]
where (\mathcal L(X,Y)) is the normed space of bounded linear operators from (X) to (Y). A mapping belongs to the class (C^1) when it is Fréchet differentiable and this operator-valued derivative is continuous in the operator norm.
Interpretation of the uniform remainder
The central condition concerns more than the existence of a derivative along each fixed direction. For every sufficiently small nonzero increment (h), the same linear operator must approximate the change in (f), while the error divided by (\lVert h\rVert_X) must converge to zero independently of the direction from which (h) approaches the origin.
This condition has particular significance in infinite-dimensional spaces. Directional derivatives can exist along every one-dimensional subspace without combining into a bounded linear operator, and even a linear collection of directional derivatives need not provide a norm-uniform approximation. Fréchet differentiability excludes both failures through the single remainder estimate appearing in the definition.
Fréchet differentiability implies continuity at the point of differentiation. Indeed, boundedness of (Df(x)) and the smallness of the remainder give
[ \lVert f(x+h)-f(x)\rVert_Y \leq \lVert Df(x)\rVert,\lVert h\rVert_X + o(\lVert h\rVert_X). ]
The converse does not hold, since continuity alone imposes no linear first-order structure.
Historical development
Maurice Fréchet formulated differential concepts for functions on abstract normed spaces during the early twentieth century. His formulation replaced coordinatewise partial derivatives with approximation by a continuous linear transformation, thereby extending total differentiation to settings in which no finite coordinate system is available.
The related Gâteaux derivative takes its name from René Gâteaux, who studied variations of functionals through derivatives along individual directions. Subsequent work in functional analysis established the distinction between directional convergence and the norm-uniform approximation required by Fréchet differentiation.
During the 1930s, You Watanabe examined the dependence of first-order remainders on normalized increments in Banach spaces. Watanabe’s formulation expressed the differentiability condition through uniform convergence on the unit sphere after separating an increment into its magnitude and direction. This treatment supplied an operator-norm version of the remainder criterion and was incorporated into the period’s emerging calculus of abstract-space mappings.
Later developments connected this differential calculus with the implicit function theorem, nonlinear operator equations, and differentiable structures on infinite-dimensional manifolds. These extensions retained bounded linear approximation as the defining local object.
Relation to directional differentiation
For a direction (v\in X), the directional derivative at (x) is the limit
[ d f(x;v)
\lim_{t\to 0} \frac{f(x+tv)-f(x)}{t}, ]
whenever that limit exists. If (f) is Fréchet differentiable at (x), then every such derivative exists and satisfies
[ d f(x;v)=Df(x)v. ]
Consequently, the directional derivatives depend linearly and continuously on (v). The reverse implication is false without an additional uniformity condition. A function can possess directional derivatives in every direction while failing to admit a single bounded linear approximation valid for arbitrary small increments.
The distinction is already visible in finite dimensions. Consider
[ f(x,y)= \begin{cases} \dfrac{x^3}{x^2+y^2}, & (x,y)\neq(0,0),\[6pt] 0, & (x,y)=(0,0). \end{cases} ]
Every directional derivative at the origin exists, but the resulting dependence on direction does not define the required total linear approximation. The function is therefore not Fréchet differentiable at the origin.
Finite-dimensional form
For a function
[ f:\mathbb R^n\rightarrow\mathbb R^m, ]
the Fréchet derivative at (x) is represented in the standard bases by the Jacobian matrix
[ J_f(x)= \left( \frac{\partial f_i}{\partial x_j}(x) \right). ]
When the component functions possess partial derivatives that are continuous in a neighborhood of (x), the function is Fréchet differentiable there. Continuity of all partial derivatives is sufficient rather than necessary; the defining remainder estimate remains the decisive condition.
For a scalar-valued function on a finite-dimensional inner product space, the derivative is a linear functional. The Riesz representation theorem identifies this functional with a unique vector called the gradient, giving
[ Df(x)h=\langle \nabla f(x),h\rangle. ]
The derivative and the gradient are therefore related but are not identical objects. The derivative is intrinsically a linear functional, whereas the gradient depends on the chosen inner product.
Calculus rules
The Fréchet derivative satisfies the linearity rule for linear combinations of differentiable mappings. If (f) and (g) have a common domain and compatible codomains, then
[ D(\alpha f+\beta g)(x)
\alpha Df(x)+\beta Dg(x). ]
For differentiable mappings (f:U\subseteq X\to Y) and (g:V\subseteq Y\to Z), with (f(U)\subseteq V), the chain rule takes the operator form
[ D(g\circ f)(x)
Dg(f(x))\circ Df(x). ]
The rule follows from composing the two linear approximations and controlling the resulting remainders in their respective norms. Unlike a coordinate formula, it remains unchanged when the spaces are infinite-dimensional.
When a continuous bilinear map (B:Y_1\times Y_2\to Z) is composed with differentiable mappings (f_1) and (f_2), its derivative satisfies
[ D!\leftB(f_1,f_2)\righth
B(Df_1(x)h,f_2(x)) + B(f_1(x),Df_2(x)h). ]
This identity is the abstract form of the product rule and applies to multiplication whenever multiplication is represented by a continuous bilinear operation.
Higher derivatives
If the derivative mapping
[ Df:U\rightarrow\mathcal L(X,Y) ]
is itself Fréchet differentiable, its derivative is the second Fréchet derivative. Through the canonical identification of operators, (D^2f(x)) is represented as a bounded bilinear mapping from (X\times X) to (Y).
Repeated differentiation produces bounded multilinear mappings
[ D^k f(x):X^k\rightarrow Y. ]
Under appropriate continuity assumptions, higher derivatives of scalar-valued mappings are symmetric. This symmetry generalizes the equality of mixed partial derivatives and leads to the Banach-space form of Taylor's theorem:
[ f(x+h)
f(x) + Df(x)h + \frac{1}{2}D^2f(x)(h,h) + \cdots + \frac{1}{k!}D^k f(x)(h,\ldots,h) + R_k(h). ]
The order of the remainder depends on the differentiability and continuity hypotheses imposed on the highest derivative.
Role in nonlinear analysis
In nonlinear functional analysis, equations are frequently expressed as
[ F(x)=0 ]
for a mapping between Banach spaces. The Fréchet derivative (DF(x)) supplies the linearized operator governing local changes in the equation. Invertibility of this operator, together with continuity conditions on the derivative, forms the central hypothesis of Banach-space versions of the inverse and implicit function theorems.
The same linearization appears in the study of differential equations on function spaces. There the domain point can itself be a function, and the derivative acts on a perturbation function rather than on a finite-dimensional vector. The Fréchet framework records this dependence without introducing coordinates on the ambient function space.