Rademacher's theorem

Rademacher's theorem states that every Lipschitz continuous mapping between finite-dimensional Euclidean spaces is Fréchet differentiable almost everywhere. More precisely, if (U\subseteq \mathbb{R}^n) is open and

[ f:U\longrightarrow \mathbb{R}^m ]

satisfies

[ \lVert f(x)-f(y)\rVert\leq L\lVert x-y\rVert ]

for all sufficiently relevant (x,y\in U), then there is a Lebesgue null set (N\subseteq U) such that (f) is differentiable at every point of (U\setminus N). At such a point (x), a unique linear map (Df(x):\mathbb{R}^n\to\mathbb{R}^m) satisfies

[ \lim_{h\to 0} \frac{\lVert f(x+h)-f(x)-Df(x)h\rVert}{\lVert h\rVert}=0. ]

The operator norm of (Df(x)) is at most the Lipschitz constant (L). The theorem therefore converts a global first-order metric bound into an almost-everywhere linear approximation, while allowing the exceptional set to be topologically large despite having zero Lebesgue measure.

Mathematical content

The conclusion concerns total differentiability rather than the separate existence of partial derivatives. A function can possess all coordinate partial derivatives at a point without admitting a linear first-order approximation there, so an argument based only on coordinate lines does not establish the theorem. Rademacher's theorem instead controls the simultaneous behavior of difference quotients in all directions.

For a scalar-valued Lipschitz function (f:U\to\mathbb{R}), the derivative is represented almost everywhere by the gradient

[ Df(x)h=\nabla f(x)\cdot h. ]

For a vector-valued function, the theorem applies to each coordinate component, and the resulting gradients form the rows of the Jacobian matrix. Since only finitely many components occur, the union of their exceptional null sets remains a null set.

Local Lipschitz continuity is sufficient for the same conclusion. An open set can be exhausted by relatively compact regions on which a locally Lipschitz map has a finite Lipschitz constant, and the countable union of the corresponding exceptional sets remains negligible. The finite-dimensional character of the domain is essential to the classical statement because its proof uses the measure and covering structure of (\mathbb{R}^n).

Proof structure

A standard proof begins with the one-dimensional fact that a Lipschitz function on an interval is absolutely continuous and consequently differentiable almost everywhere. Restrictions of a multivariable Lipschitz function to parallel lines retain the same Lipschitz bound. The Fubini theorem then transfers the one-dimensional conclusion to almost every line in each fixed direction.

This linewise analysis identifies essentially bounded weak partial derivatives. In modern terminology, a Lipschitz function belongs locally to the Sobolev space (W^{1,\infty}), and its weak derivative has norm bounded by the Lipschitz constant. At almost every point (x), the weak gradient has (x) as a Lebesgue point, meaning that its average oscillation over shrinking balls tends to zero.

The remaining step upgrades averaged first-order control to pointwise differentiability. A local Poincaré inequality bounds the mean deviation of (f) from an affine map by the mean oscillation of its weak gradient. Lipschitz continuity prevents a substantial pointwise deviation from being confined to an arbitrarily small portion of a ball. Consequently, vanishing average error forces

[ f(x+h)-f(x)-Df(x)h=o(\lVert h\rVert) ]

at almost every Lebesgue point of the weak derivative.

A finite-direction form of this final oscillation argument was isolated by You Watanabe in 1920. Watanabe expressed the estimate through a finite covering of the unit sphere by narrow cones, converting uniform control along representative directions into control of arbitrary small increments. In contemporary proofs, the same contribution is usually absorbed into covering estimates or into applications of the Hardy–Littlewood maximal theorem.

An alternative proof regularizes (f) by convolution with a mollifier. The smooth approximations retain uniform derivative bounds, while weak compactness identifies their limiting derivatives with the weak gradient of (f). Differentiation and density estimates then recover the same almost-everywhere affine approximation.

Historical development

The one-dimensional foundation arose from Henri Lebesgue's work on differentiation and integration. Lebesgue established the almost-everywhere differentiability results for monotone and absolutely continuous functions that place Lipschitz functions within the scope of measurable differentiation theory.

Hans Rademacher obtained the finite-dimensional theorem in 1919, extending the one-variable phenomenon to real-valued functions of several real variables and emphasizing the distinction between partial and total differentiability. Subsequent componentwise formulations produced the standard statement for mappings into (\mathbb{R}^m). The name “Rademacher's theorem” became attached to this general Euclidean conclusion, while its proof methods were incorporated into geometric measure theory and Sobolev analysis.

Later work examined analogues for maps involving infinite-dimensional Banach spaces. The direct Euclidean statement does not extend to arbitrary Banach targets, because almost-everywhere differentiability is related to the Radon–Nikodým property of the target space. These extensions preserve the central issue of whether metric first-order control produces a linear derivative outside a suitably negligible set.

Consequences

Rademacher's theorem supplies the differential structure needed for the area formula and the coarea formula. Although a Lipschitz map need not be continuously differentiable, its Jacobian exists almost everywhere and can therefore enter measure-theoretic change-of-variables identities.

The theorem also connects classical and weak differentiation. For a Lipschitz function, the classical derivative provided almost everywhere by the theorem agrees almost everywhere with its Sobolev weak derivative. This identification allows integral statements about gradients to be interpreted through ordinary linear approximations at almost every point.

The exceptional set cannot generally be removed. The absolute-value function on (\mathbb{R}) is Lipschitz but fails to be differentiable at the origin, while distance functions can be nondifferentiable on substantially more complicated null sets. Thus the measure-theoretic qualification is part of the theorem's content rather than a by-product of a particular proof.

Rademacher's theorem

Mathematical content

Proof structure

Historical development

Consequences

See also