The Whitney Extension Theorem and the Whitney Cube Decomposition

Intuition Beginner

Suppose someone hands you a closed set in the plane — maybe a curve, maybe a fat Cantor-like dust, maybe a single point — and on that set they prescribe not just the values a function should take, but also the values its slope and its curvature should take. They are describing a function only on the closed set, yet they are describing how it ought to bend. The question is whether a single smooth function defined everywhere agrees with all of that prescribed data on the set.

Whitney's answer is yes, provided the prescribed data is self-consistent. Consistency means the prescribed value at a nearby point should be close to what the prescribed slope predicts, and the prescribed slope should be close to what the prescribed curvature predicts, with the mismatch shrinking faster than the distance between the two points to the appropriate power.

The takeaway in one line: if you prescribe a function and its would-be derivatives on a closed set in a way that does not contradict itself, you can always extend the data to a genuinely smooth function on all of space.

The machine that builds the extension is geometric. Outside the closed set there is open room, and that open room can be chopped into small boxes whose size matches their distance from the set. On each small box you trust a local guess built from the prescribed data, and you blend the local guesses smoothly. The blending uses a partition of unity, a family of smooth bump weights that sum to one everywhere.

The beauty is that two very different ideas meet: a covering of empty space by carefully sized boxes, and a consistency condition that looks like a remainder estimate in a Taylor expansion. The covering handles the geometry; the consistency handles the smoothness. Together they manufacture a function out of a promise about derivatives.

Visual Beginner

Picture a closed set drawn as a thick black curve in the plane, with the rest of the plane left white. Now fill the white region with boxes. Near the curve the boxes are tiny; far from the curve the boxes are large. Each box has a side length roughly equal to its distance from the black curve, so the boxes shrink as they crowd toward the set and grow as they retreat from it.

The lower strip of the figure shows the blending. Above each box sits a smooth bump weight, a hump that is large over its own box and fades to zero before reaching distant boxes. The humps overlap just enough that, added together, they equal one at every point of the open region. The extended function is a weighted average: at any point off the set, you read off the nearby local guesses and combine them with these bump weights. Because the weights are smooth and the boxes are sized to the distance, the seams never show, and the result joins onto the prescribed data on the curve without a kink.

Worked example Beginner

Take the simplest interesting closed set on the line: a single point at the origin. Prescribe a value $a_0=2$ there, and prescribe a first-derivative value $a_1=3$ there. We want one smooth function on the whole line whose value at the origin is $2$ and whose slope at the origin is $3$.

Step 1. With only one point, there is no consistency condition to check, because consistency compares the data at two distinct points, and here there is only one.

Step 2. Form the local guess from the prescribed data. This is the first-order Taylor polynomial built at the origin: $P(x)=a_0+a_{1}x=2+3x$.

Step 3. Extend by simply using that polynomial everywhere: set $F(x)=2+3x$. This function is smooth on the whole line.

Step 4. Check the prescribed data. The value at the origin is $F(0)=2=a_0$, and the slope at the origin is $3=a_1$, matching both prescriptions.

What this shows: for a one-point set the extension is the Taylor polynomial itself, and the box-decomposition machinery is not yet needed. The real work begins when the set has many points whose prescribed data must be reconciled, which is what the consistency condition manages.

Check your understanding Beginner

Formal definition Intermediate+

Fix $m\in \mathbb{N}$ and a closed set $E\subseteq {\mathbb{R}}^n$. A multi-index is $\alpha =({\alpha }_1,\dots ,{\alpha }_n)\in {\mathbb{N}}^n$ with $|\alpha |={\sum }_i{\alpha }_i$, and $\beta !={\prod }_i{\beta }_i!$, $z^{\beta }={\prod }_i{z}_i^{{\beta }_i}$.

Definition (Whitney $C^m$ jet on $E$). A $C^m$ Whitney jet on $E$ is a family $f=(f_{\alpha }{)}_{|\alpha |\le m}$ of continuous functions $f_{\alpha }:E\to \mathbb{R}$, one for each multi-index of order at most $m$, to be read as the prescribed Taylor coefficients $f_{\alpha }(x)={\partial }^{\alpha }F(x)$ of a sought extension $F$ ^{[Whitney 1934]}.

For $x\in E$ the associated formal Taylor polynomial of order $m$ at $x$ is $$ (T^m_x f)(z) = \sum_{|\beta| \le m} \frac{f_\beta(x)}{\beta!},(z - x)^\beta . $$ For each multi-index $\alpha$ with $|\alpha |\le m$ define the remainder $$ (R_\alpha f)(x, y) ;=; f_\alpha(x) ;-; \sum_{|\beta| \le m - |\alpha|} \frac{f_{\alpha + \beta}(y)}{\beta!},(x - y)^\beta , \qquad x, y \in E . $$ This is the discrepancy between the prescribed $\alpha$-derivative at $x$ and the value predicted by the prescribed higher coefficients at the nearby base point $y$.

Definition (Whitney compatibility conditions). The jet $f$ satisfies the Whitney conditions of order $m$ on $E$ when, for every multi-index $\alpha$ with $|\alpha |\le m$, $$ (R_\alpha f)(x, y) = o!\left(|x - y|^{m - |\alpha|}\right) \quad \text{as } |x - y| \to 0,\ x, y \in E, $$ uniformly on compact subsets of $E$. Equivalently, for every compact $K\subseteq E$ and every $\epsilon >0$ there is a $\delta >0$ with $|(R_{\alpha }f)(x,y)|\le \epsilon |x-y{|}^{m-|\alpha |}$ whenever $x,y\in K$ and $|x-y|\le \delta$ ^{[Whitney 1934]}.

A jet satisfying these conditions is called a Whitney field of class $C^m$ on $E$ ^{[Malgrange 1966]}. The space of such fields, with the natural seminorms, is denoted ${\mathcal{E}}^m(E)$.

Definition (Whitney decomposition of an open set). Let $U={\mathbb{R}}^n\setminus E$ be the open complement of a nonempty closed $E$. A Whitney decomposition of $U$ is a countable family $\{Q_j\}$ of closed dyadic cubes with pairwise disjoint interiors, $U={\bigcup }_j{Q}_j$, such that $$ \operatorname{diam}(Q_j) ;\le; \operatorname{dist}(Q_j, E) ;\le; 4,\operatorname{diam}(Q_j) $$ for every $j$ ^{[Stein 1970]}. A subordinate partition of unity is a family $\{{\phi }_j\}$ of smooth functions with $0\le {\phi }_j\le 1$, supported on a fixed dilate $Q_j^{\ast }$ of $Q_j$, with ${\sum }_j{\phi }_j\equiv 1$ on $U$ and derivative bounds $|{\partial }^{\alpha }{\phi }_j|\le C_{\alpha }\mathrm{diam}(Q_j{)}^{-|\alpha |}$.

Counterexamples to common slips Intermediate+

• Mere continuity of the coefficients is not enough. Prescribing continuous $f_{\alpha }$ without the remainder estimates can produce data that no $C^m$ function reproduces; the consistency conditions are an additional hypothesis, not a consequence of continuity.
• The base point in the remainder matters. The condition compares the coefficient at $x$ against the polynomial centred at $y$; symmetry in $x$ and $y$ is not assumed, and both orders of the pair contribute as $|x-y|\to 0$.
• Extension is not unique. Many $C^m$ functions reproduce the same jet on $E$; they differ by a $C^m$ function whose full $m$-jet vanishes on $E$. Whitney extension asserts existence and a linear bounded selection, not uniqueness.

Key theorem with proof Intermediate+

Theorem (Whitney extension theorem). Let $E\subseteq {\mathbb{R}}^n$ be closed and let $f=(f_{\alpha }{)}_{|\alpha |\le m}$ be a Whitney field of class $C^m$ on $E$. Then there exists $F\in C^m({\mathbb{R}}^n)$ such that ${\partial }^{\alpha }F{|}_E=f_{\alpha }$ for every $|\alpha |\le m$, and $F$ is real-analytic on ${\mathbb{R}}^n\setminus E$. Moreover $F$ may be chosen to depend linearly and continuously on $f$ ^{[Whitney 1934]}.

Proof. Write $U={\mathbb{R}}^n\setminus E$. Fix a Whitney decomposition $\{Q_j\}$ of $U$ and a subordinate partition of unity $\{{\phi }_j\}$ as in the Formal definition; their existence is established in the Full proof set below. For each cube $Q_j$ choose a point $p_j\in E$ with $\mathrm{dist}(\mathrm{centre}(Q_j),E)=|\mathrm{centre}(Q_j)-p_j|$, a nearest boundary point.

Define the extension by gluing the local Taylor polynomials: $$ F(x) = \begin{cases} f_0(x), & x \in E, \[2pt] \displaystyle \sum_j \varphi_j(x), (T^m_{p_j} f)(x), & x \in U . \end{cases} $$ On $U$ each summand is a polynomial and the sum is locally finite, so $F$ is real-analytic there. The work is to verify $C^m$ regularity up to and on $E$, with the correct derivatives.

Differentiate on $U$ using the Leibniz rule and ${\sum }_j{\phi }_j\equiv 1$. For a multi-index $\gamma$ with $|\gamma |\le m$ and a base point $y\in E$, $$ \partial^\gamma F(x) - (T^{m - |\gamma|}y f)\gamma(x) = \sum_j \sum_{\delta \le \gamma} \binom{\gamma}{\delta} \partial^{\gamma - \delta}\varphi_j(x), \partial^\delta!\left(T^m_{p_j} f) - (T^m_y f)\right, $$ where $(T_y^{m-|\gamma |}f{)}_{\gamma }$ collects the prescribed $\gamma$-coefficient propagated from $y$. The terms with $\delta =\gamma$ assemble, via ${\sum }_j{\partial }^{\gamma -\delta }{\phi }_j=0$ for $\delta \ne \gamma$, into a sum over the polynomial differences $(T_{p_j}^{m}f)-(T_y^{m}f)$, each of which is controlled by the remainder $R_{\alpha }f$ evaluated at the pair $(p_j,y)$.

Now estimate. If $x\in U$ lies in cube $Q_j$ and $y\in E$ is the nearest point of $E$ to $x$, then $|p_j-y|\le C|x-y|$ by the distance-comparability of the decomposition, and $|{\partial }^{\gamma -\delta }{\phi }_j(x)|\le C|x-y{|}^{-(|\gamma |-|\delta |)}$ from the derivative bound on the partition of unity. The Whitney condition gives $|(R_{\alpha }f)(p_j,y)|\le \epsilon |p_j-y{|}^{m-|\alpha |}$ once $|x-y|$ is small. Combining the three estimates, every term carries a net power $|x-y{|}^{m-|\gamma |}$ times a factor that tends to zero. Hence $$ \partial^\gamma F(x) \longrightarrow f_\gamma(y) \quad \text{as } x \to y,\ x \in U, $$ so ${\partial }^{\gamma }F$ extends continuously to $E$ with boundary value $f_{\gamma }$, and a difference-quotient argument promotes this to genuine $C^m$ differentiability across $E$. Linearity and continuity in $f$ are visible because $F$ is built from $f$ by the fixed linear recipe of partition weights and Taylor polynomials. $\square$

Bridge. The Whitney extension theorem builds toward the geometric-measure-theory and currents units in 02.13.07 and 02.13.11, where extending consistent jets off a rectifiable set is invoked without statement; the construction here is exactly the missing rung those units stand on. The foundational reason the argument closes is that two independent budgets cancel: the partition-of-unity derivatives blow up like a negative power of the distance to $E$, while the Whitney remainders decay like a matching positive power, and the bridge is the cube decomposition that pins the two powers to the same exponent. This pattern — geometry sized to distance-from-boundary, paired with an analytic remainder estimate — is dual to the Carathéodory covering logic of 02.07.02 and generalises to the Calderón-Zygmund decomposition; putting these together, the central insight is that a promise about derivatives on a closed set becomes a genuine smooth function precisely when the promise is internally consistent, and this is exactly the statement the downstream units consume. The same decomposition appears again in Sobolev-extension constructions, where it does the geometric bookkeeping for boundary-value problems.

Exercises Intermediate+

Advanced results Master

The Whitney extension theorem sits at the head of a circle of results about how regularity prescribed on a closed set propagates into the ambient space. Three strands matter: the covering lemma as a standalone tool, the algebraic reformulation via ideals of differentiable functions, and the modern $C^m$ finiteness theory.

Theorem 1 (Whitney covering lemma). Let $E⊊{\mathbb{R}}^n$ be nonempty and closed and $U={\mathbb{R}}^n\setminus E$. There is a countable family of closed dyadic cubes $\{Q_j\}$ with pairwise disjoint interiors, $U={\bigcup }_j{Q}_j$, and $$ \operatorname{diam}(Q_j) \le \operatorname{dist}(Q_j, E) \le 4,\operatorname{diam}(Q_j) . $$ The dilates $Q_j^{\ast }=\frac{9}{8}{Q}_j$ have bounded overlap: each point of $U$ lies in at most $N(n)$ of them, with $N(n)$ depending on dimension alone ^{[Stein 1970]}. This lemma is reused beyond extension: it underlies the Calderón-Zygmund decomposition, the construction of Sobolev extension operators, and pointwise estimates for singular integrals, in each case because it converts a single closed set into a uniformly controlled scale-adapted cover of its complement.

Theorem 2 (smooth partition of unity with controlled derivatives). For the cubes of Theorem 1 there exist ${\phi }_j\in C_c^{\infty }({\mathbb{R}}^n)$ with $\mathrm{supp}{\phi }_j\subseteq Q_j^{\ast }$, $0\le {\phi }_j\le 1$, ${\sum }_j{\phi }_j\equiv 1$ on $U$, and $|{\partial }^{\alpha }{\phi }_j(x)|\le C_{\alpha }\mathrm{diam}(Q_j{)}^{-|\alpha |}$ for all $\alpha$. The derivative growth is sharp: a bump confined to a cube of side $\ell$ cannot have $k$-th derivatives smaller than order ${\ell }^{-k}$, and this exact rate is what the extension proof's cancellation requires ^{[Stein 1970]}.

Theorem 3 (Whitney's $C^{\infty }$ and analytic refinements). If the jet data is consistent to all orders — a $C^{\infty }$ Whitney field, meaning the order-$m$ conditions hold for every $m$ — then the extension $F$ may be chosen in $C^{\infty }({\mathbb{R}}^n)$, real-analytic on $U$. The analytic structure off $E$ is automatic in the construction, since $F$ is a locally finite sum of polynomials there. By contrast, extension to a globally real-analytic $F$ across $E$ is generally impossible and is governed by separate coherence theory ^{[Whitney 1934]}.

Theorem 4 (Glaeser / spectral synthesis reformulation). Let $\mathcal{I}(E)\subseteq C^m({\mathbb{R}}^n)$ be the ideal of functions whose $m$-jet vanishes on $E$. Whitney's theorem identifies the restriction map $C^m({\mathbb{R}}^n)\to {\mathcal{E}}^m(E)$ as surjective, and Glaeser's work on Taylorian algebras characterises its kernel and the closed ideals of $C^m$, resolving the Whitney spectral problem of which closed ideals are determined by their jets on the zero set ^{[Glaeser 1958]}. Malgrange's monograph recasts the entire theory in the language of Whitney fields and modules over the ring of differentiable functions ^{[Malgrange 1966]}.

Theorem 5 (Fefferman finiteness and extension for $C^m$). For arbitrary closed $E$, given only the prescribed values $f:E\to \mathbb{R}$ (no derivative data), the question of whether $f$ extends to a $C^m({\mathbb{R}}^n)$ function, and the construction of a bounded linear extension operator, were resolved by Fefferman. The $C^m$ finiteness principle states that extendability is governed by a finite collection of points: $f$ extends with norm controlled by $M$ if and only if its restriction to every subset of cardinality at most $k^{\#}(m,n)$ does, with $k^{\#}$ depending only on $m$ and $n$ ^{[Fefferman 2006]}. This converts Whitney's qualitative theorem into a quantitative, algorithmically usable criterion and underlies modern interpolation and machine-learning-adjacent function-fitting.

Synthesis. The Whitney extension theorem is the foundational reason that prescribing derivatives on a closed set is a meaningful and solvable problem, and the central insight is that the covering lemma and the consistency conditions are two halves of a single cancellation: geometry sized to distance-from-boundary against an analytic remainder of matching order. Putting these together, the covering lemma is dual to the Carathéodory and Calderón-Zygmund decompositions, all three converting one closed or measurable set into a scale-adapted cover, and this is exactly why the same dyadic machinery resurfaces across measure theory, singular integrals, and Sobolev extension. The theory generalises in two directions at once: upward in regularity to the $C^{\infty }$ and Whitney-field formulations of Glaeser and Malgrange, where the restriction map and its ideal kernel become the objects of study, and outward in quantitative strength to the Fefferman finiteness principle, where extendability is decided by finitely many points. The bridge is between a nineteenth-century-style Taylor remainder and a twentieth-century covering argument; this pattern recurs in geometric measure theory, where the same extension underwrites the manipulation of currents and rectifiable sets, and it appears again wherever a local-to-global passage must respect a prescribed jet.

Full proof set Master

Proposition 1 (existence of the Whitney decomposition). Let $E⊊{\mathbb{R}}^n$ be nonempty and closed, $U={\mathbb{R}}^n\setminus E$. Then $U$ admits a decomposition into closed dyadic cubes with disjoint interiors satisfying $\mathrm{diam}(Q)\le \mathrm{dist}(Q,E)\le 4\mathrm{diam}(Q)$.

Proof. For $k\in \mathbb{Z}$ let ${\mathcal{D}}_k$ be the grid of closed dyadic cubes of side $2^{-k}$, that is, products ${\prod }_i[m_i{2}^{-k},(m_i+1)2^{-k}]$ with $m_i\in \mathbb{Z}$. For each $x\in U$ the value $\mathrm{dist}(x,E)>0$ is positive, so there is a largest scale at which a cube containing $x$ stays away from $E$. Concretely, declare a dyadic cube $Q\in {\mathcal{D}}_k$ selected when $$ \sqrt{n}, 2^{-k} \le \operatorname{dist}(Q, E) \le 4\sqrt{n}, 2^{-k}, $$ noting $\mathrm{diam}(Q)=\sqrt{n}{2}^{-k}$. Every point of $U$ lies in at least one selected cube, because the distance-to-$E$ function is finite and positive on $U$ and the scales $2^{-k}$ exhaust $(0,\infty )$. Selected cubes can overlap only by nesting along the dyadic tower; discard any selected cube strictly contained in a larger selected cube, retaining the maximal ones. Maximal selected cubes have disjoint interiors by the dyadic nesting property, their union is $U$, and the selection inequality is exactly the required comparability after absorbing $\sqrt{n}$ into the constant $4$. $\square$

Proposition 2 (bounded overlap of dilates). With the cubes of Proposition 1, the dilates $Q_j^{\ast }=\frac{9}{8}{Q}_j$ satisfy: each point of $U$ lies in at most $N(n)$ of them, $N(n)$ depending only on $n$.

Proof. If $Q_i^{\ast }\cap Q_j^{\ast }\ne \varnothing$ then the comparability inequalities force $\frac{1}{4}\le \mathrm{diam}(Q_i)/\mathrm{diam}(Q_j)\le 4$: two touching dilates have comparable distance to $E$, hence comparable side. A fixed cube $Q_j$ can therefore be met only by dilates of cubes within two dyadic scales, and within a bounded number of scales the dyadic grid packs only $N(n)$ cubes of comparable size around any given cube. The count $N(n)$ is the volume ratio of a fixed neighbourhood to a single cube, a dimensional constant. $\square$

Proposition 3 (controlled partition of unity). The cubes of Proposition 1 carry ${\phi }_j\in C_c^{\infty }$, $\mathrm{supp}{\phi }_j\subseteq Q_j^{\ast }$, with ${\sum }_j{\phi }_j\equiv 1$ on $U$ and $|{\partial }^{\alpha }{\phi }_j|\le C_{\alpha }\mathrm{diam}(Q_j{)}^{-|\alpha |}$.

Proof. Fix a single bump $\psi \in C_c^{\infty }({\mathbb{R}}^n)$ with $\psi \equiv 1$ on the unit cube, $\mathrm{supp}\psi$ inside the $\frac{9}{8}$-dilate, $0\le \psi \le 1$. For each $j$ let ${\psi }_j$ be the affine rescaling of $\psi$ matched to $Q_j$; then $\mathrm{supp}{\psi }_j\subseteq Q_j^{\ast }$ and, by the chain rule, $|{\partial }^{\alpha }{\psi }_j|\le \Vert {\partial }^{\alpha }\psi {\Vert }_{\infty }\mathrm{diam}(Q_j{)}^{-|\alpha |}$. Set $\Phi ={\sum }_j{\psi }_j$; by bounded overlap (Proposition 2) the sum is locally finite and $\Phi \ge 1$ on $U$ since each cube is covered by its own ${\psi }_j\equiv 1$ on $Q_j$. Define ${\phi }_j={\psi }_j/\Phi$. Then ${\sum }_j{\phi }_j\equiv 1$, supports are unchanged, and the quotient rule together with $\Phi \ge 1$, the bounded overlap, and the comparability of neighbouring cube sizes yields $|{\partial }^{\alpha }{\phi }_j|\le C_{\alpha }\mathrm{diam}(Q_j{)}^{-|\alpha |}$. $\square$

Proposition 4 (the $m=1$ extension is explicit). For closed $E\subseteq {\mathbb{R}}^n$ and a $C^1$ Whitney field $(f_0,(f_{e_i}{)}_i)$ — value and gradient data with $f_0(x)-f_0(y)-{\sum }_i{f}_{e_i}(y)(x_i-y_i)=o(|x-y|)$ and $f_{e_i}(x)-f_{e_i}(y)=o(1)$ — the gluing formula $F={\sum }_j{\phi }_j[f_0(p_j)+{\sum }_i{f}_{e_i}(p_j)(x_i-p_{j,i})]$ on $U$, $F=f_0$ on $E$, is $C^1({\mathbb{R}}^n)$ with ${\partial }_{i}F{|}_E=f_{e_i}$.

Proof. This is the theorem specialised to $m=1$; only first derivatives are estimated. On $U$ the sum is smooth. For $x\in Q_j$ and nearest $y\in E$, the gradient is $\nabla F(x)={\sum }_j\nabla {\phi }_j(x)L_j(x)+{\sum }_j{\phi }_j(x)\nabla L_j$, where $L_j$ is the affine local guess at $p_j$. Subtracting $f_{e_i}(y)$ and using ${\sum }_j\nabla {\phi }_j=0$ to insert $L_j(x)-L_y(x)$ in place of $L_j(x)$, the first sum is bounded by $C|x-y{|}^{-1}\cdot o(|x-y|)=o(1)$ via the $C^1$ remainder condition, and the second sum tends to $f_{e_i}(y)$ by the gradient-continuity condition. Hence ${\partial }_{i}F$ extends continuously to $E$ with value $f_{e_i}$, and difference quotients of $F$ across $E$ converge to the same limit, giving $F\in C^1$. $\square$

Proposition 5 (extension is linear and bounded). The map $f\mapsto F$ produced by the gluing construction is linear in $f$ and bounded between the natural seminorm topologies on ${\mathcal{E}}^m(E)$ and $C^m({\mathbb{R}}^n)$.

Proof. For fixed $E$, decomposition $\{Q_j\}$, points $p_j$, and partition $\{{\phi }_j\}$, the formula $F={\sum }_j{\phi }_j{T}_{p_j}^{m}f$ on $U$ and $F=f_0$ on $E$ is linear in the coefficient family $f=(f_{\alpha })$, since $T_{p_j}^{m}f$ depends linearly on $f$ and the weights ${\phi }_j$ are fixed. Boundedness on a compact $K$ follows from the proof of the main theorem: each derivative ${\partial }^{\gamma }F$ on a neighbourhood of $K$ is estimated by the Whitney seminorms of $f$ on a slightly larger compact, with constants from the fixed partition of unity and the bounded overlap. Taking the supremum over $|\gamma |\le m$ produces $\Vert F{\Vert }_{C^m(K^{\prime })}\le C_K\Vert f{\Vert }_{{\mathcal{E}}^m(K)}$, a continuous linear bound. $\square$

Connections Master

• Lebesgue outer measure and the Carathéodory construction 02.07.02. The Whitney covering lemma and the Carathéodory construction are siblings: both turn a single set into a controlled cover and read structure off the cover. The decomposition into dyadic cubes sized to the distance-to-boundary is the geometric counterpart of the box-cover infimum that defines outer measure, and the bounded-overlap property here plays the role that countable sub-additivity plays there. The measure-zero and covering arguments of 02.07.02 supply the metric foundation on which the cube selection of Proposition 1 rests.

• Implicit and inverse function theorems 02.05.04 and multivariable Taylor 02.05.05. The jet data of a Whitney field is precisely a prescribed family of Taylor coefficients, so the local Taylor polynomial machinery of 02.05.05 is the algebraic content of every local guess in the gluing formula. When $E$ is a smooth submanifold, the implicit and inverse function theorems of 02.05.04 identify the consistent jets with genuine restrictions of ambient $C^m$ functions, linking the extension theorem back to the differential-geometric setting where jets arise as derivatives along the manifold.

• Rectifiable currents and slicing 02.13.07 and 02.13.11. These geometric-measure-theory units invoke extension of consistent jets off rectifiable sets without proving it; the present unit supplies that missing step. The same Whitney decomposition that builds the extension also organises the dyadic approximation of currents, and the controlled partition of unity reappears as the device that localises mass and flat-norm estimates. The extension theorem is the analytic prerequisite that lets those units treat prescribed boundary data on a closed set as the trace of a global smooth object.

• Sobolev and trace theory (lateral, forthcoming). The Whitney decomposition is the backbone of bounded Sobolev extension operators: extending a function from a domain to all of ${\mathbb{R}}^n$ across an irregular boundary uses exactly the cubes-sized-to-distance cover and the controlled partition of unity proved here, so the covering lemma is shared infrastructure between $C^m$ extension and $W^{k,p}$ extension.

Historical & philosophical context Master

Hassler Whitney's 1934 paper Analytic extensions of differentiable functions defined in closed sets ^{[Whitney 1934]} in the Transactions of the American Mathematical Society posed and solved a problem that earlier analysis had left implicit: when is a collection of prescribed values and would-be derivatives on a closed set the restriction of a single smooth function? Whitney's recognition that the answer requires compatibility conditions phrased as Taylor-remainder estimates, rather than mere continuity of the data, was the conceptual leap. He also introduced, in the same circle of papers, the dyadic decomposition of an open set into cubes sized to their distance from the boundary, a construction so robust that it now bears his name across several fields. The paper appeared while Whitney was developing the foundations of differential topology, and the extension theorem fed directly into his later embedding theorems.

The theorem was reframed twice in the mid-twentieth century. Georges Glaeser's 1958 study of Taylorian algebras ^{[Glaeser 1958]} turned the extension question into one about ideals of differentiable functions and the Whitney spectral problem: which closed ideals of $C^m$ are determined by the jets they impose on their common zero set. Bernard Malgrange's 1966 monograph Ideals of Differentiable Functions ^{[Malgrange 1966]} systematised this algebraic viewpoint, recasting Whitney fields as a module-theoretic object and connecting the theory to the division and preparation theorems for smooth functions. In parallel, Stein's 1970 treatise ^{[Stein 1970]} placed the covering lemma at the centre of harmonic analysis, where it organises the Calderón-Zygmund decomposition and the theory of singular integrals.

The modern chapter belongs to Charles Fefferman, whose 2006 Annals paper ^{[Fefferman 2006]} resolved the long-open quantitative form of Whitney's problem for $C^m$: extendability of bare value data is decided by a finite, dimension-bounded number of points, and a bounded linear extension operator exists. Philosophically, the arc from Whitney to Fefferman tracks a recurring theme in analysis — a qualitative existence theorem (a smooth extension exists) is gradually sharpened into a quantitative, finitary, and algorithmic one (extendability is certified by finitely many local tests). The same passage from existence to effective construction recurs in the measure-theoretic extension theorems of 02.07.02 and motivates the computational geometry of interpolation.

Bibliography Master

@article{Whitney1934,
  author  = {Whitney, Hassler},
  title   = {Analytic extensions of differentiable functions defined in closed sets},
  journal = {Transactions of the American Mathematical Society},
  volume  = {36},
  number  = {1},
  pages   = {63--89},
  year    = {1934},
  doi     = {10.2307/1989708}
}

@book{Stein1970,
  author    = {Stein, Elias M.},
  title     = {Singular Integrals and Differentiability Properties of Functions},
  publisher = {Princeton University Press},
  series    = {Princeton Mathematical Series 30},
  year      = {1970}
}

@book{Federer1969,
  author    = {Federer, Herbert},
  title     = {Geometric Measure Theory},
  publisher = {Springer-Verlag},
  series    = {Grundlehren der mathematischen Wissenschaften 153},
  year      = {1969}
}

@book{Hormander1990,
  author    = {H{\"o}rmander, Lars},
  title     = {The Analysis of Linear Partial Differential Operators I},
  edition   = {2},
  publisher = {Springer-Verlag},
  year      = {1990}
}

@book{Malgrange1966,
  author    = {Malgrange, Bernard},
  title     = {Ideals of Differentiable Functions},
  publisher = {Oxford University Press},
  series    = {Tata Institute Studies in Mathematics 3},
  year      = {1966}
}

@article{Glaeser1958,
  author  = {Glaeser, Georges},
  title   = {{\'E}tude de quelques alg{\`e}bres tayloriennes},
  journal = {Journal d'Analyse Math{\'e}matique},
  volume  = {6},
  pages   = {1--124},
  year    = {1958}
}

@article{Fefferman2006,
  author  = {Fefferman, Charles},
  title   = {Whitney's extension problem for $C^m$},
  journal = {Annals of Mathematics},
  volume  = {164},
  number  = {1},
  pages   = {313--359},
  year    = {2006}
}