extension of distributions



Given a suitable subspace inclusion XX^X \hookrightarrow \hat X and a distribution uu on XX, then an extension of uu to X^\hat X is a distribution u^\hat u on X^\hat X whose restriction of distributions to XX coincides with uu. If uu comes from an ordinary smooth function or else if uu is properly understood as a generalized function, then this corresponds to an ordinary extension

X u u^ X^ \array{ X &\overset{u}{\longrightarrow}& \mathbb{R} \\ \downarrow & \nearrow_{\mathrlap{\hat u}} \\ \hat X }

Regarding that we the distributions are not the maps from the underlying space we need to replace pre-composition XX˜X\to\tilde{X} by the appropriate pullback of distributions always defined say if XX˜X\to\tilde{X} is a submersion; in the case of an open embedding (see this example) this operation is the restriction of distributions, the operation dual to the operation of extension by zero of test functions.


For an inclusion of two open sets on a manifold UVU\subset V there is an operator of extension by zero E VU:C 0 (U)C 0 (V)E_{V U}: C^\infty_0(U)\to C^\infty_0(V) where E VU(f)(x)=f(x)E_{V U}(f)(x) = f(x) if xUx\in U and E UV(f)(x)=0E_{U V}(f)(x) = 0 otherwise. The restriction of distributions ρ UV:𝒟(V)𝒟(U)\rho_{U V} : \mathcal{D}'(V)\to \mathcal{D}'(U) is then defined by

ρ UV(ϕ),f=ϕ,E VUf,fC 0 (U),ϕ𝒟(V). \langle \rho_{U V}(\phi), f\rangle = \langle \phi, E_{V U} f\rangle, \,\,\,\,\,\,\,f\in C^\infty_0(U),\,\phi\in \mathcal{D}'(V).

(this example).

Now the diagram in the idea section makes sense in the following way: for an open embedding XX^X\hookrightarrow\hat{X}, u^𝒟(X^)\hat{u}\in\mathcal{D}'(\hat{X}) extends u𝒟(X)u\in \mathcal{D}'(X) if ρ XX^(u^)=u\rho_{X\hat{X}}(\hat{u}) = u.


(extension of distributions)

Let XιX^X \overset{\iota}{\subset} \hat X be an inclusion of open subsets of some Cartesian space. This induces the operation of restriction of distributions

𝒟(X^)ι *𝒟(X). \mathcal{D}'(\hat X) \overset{\iota^\ast}{\longrightarrow} \mathcal{D}'(X) \,.

Given a distribution u𝒟(X)u \in \mathcal{D}'(X), then an extension of uu to X^\hat X is a distribution u^𝒟(X^)\hat u \in \mathcal{D}'(\hat X) such that

ι *u^=u. \iota^\ast \hat u \;=\; u \,.



Of particular interest is the special case of extension of distributions to a single point pX^p \in \hat X, hence where X=X^{p}X = \hat X \setminus \{p\} is the complement of that point.

Contrary to ordinary smooth functions, distributions (generalized functions) in general have more than one extension to a point, where the freedom is in choosing a distribution with point support at that point, hence some derivative of the delta distribution supported at that point.

However, if one requires that the scaling degree of the extended distribution at the given point is compatible with that of the original distribution, then there is only a finite set of possible extensions, parameterized by the coefficients of a finite number of derivatives of the delta-distribution (prop. 2 below).

In the construction of perturbative quantum field theories via causal perturbation theory, where the (“operator-valued”) distributions in questions are time-ordered product coefficients in the scattering matrix and where the point being extended to corresponds to the point where an interaction happens, this finite set of choices is identified with the ("re"-)normalization freedom.

We discuss specifically the space of solutions of extending a distribution on the complement n{0}\mathbb{R}^n \setminus \{0\} of the origin inside a Cartesian space to the full space n\mathbb{R}^n.


(unique extension of distributions with negative degree of divergence)

For nn \in \mathbb{N}, let u𝒟( n{0})u \in \mathcal{D}'(\mathbb{R}^n \setminus \{0\}) be a distribution on the complement of the origin, with negative degree of divergence at the origin

deg(u)<0. deg(u) \lt 0 \,.

Then uu has a unique extension of distributions u^𝒟( n)\hat u \in \mathcal{D}'(\mathbb{R}^n) to the origin with the same degree of divergence

deg(u^)=deg(u). deg(\hat u) = deg(u) \,.

(Brunetti-Fredenhagen 00, theorem 5.2, Dütsch 18, theorem 3.35 a))


Regarding uniqueness:

Suppose u^\hat u and u^ {\hat u}^\prime are two extensions of uu with deg(u^)=deg(u^ )deg(\hat u) = deg({\hat u}^\prime). Both being extensions of a distribution defined on n{0}\mathbb{R}^n \setminus \{0\}, this difference has support at the origin {0} n\{0\} \subset \mathbb{R}^n. By this prop. this implies that it is a linear combination of derivatives of the delta distribution supported at the origin:

u^ u^=α nc α αδ 0 {\hat u}^\prime - \hat u \;=\; \underset{ {\alpha \in \mathbb{N}^n} }{\sum} c^\alpha \partial_\alpha \delta_0

for constants c αc^\alpha \in \mathbb{C}. But by this example the degree of divergence of these point-supported distributions is non-negative

deg( αδ 0)=|α|0. deg( \partial_\alpha \delta_0) = {\vert \alpha\vert} \geq 0 \,.

This implies that c α=0c^\alpha = 0 for all α\alpha, hence that the two extensions coincide.

Regarding existence:


bC cp ( n) b \in C^\infty_{cp}(\mathbb{R}^n)

be a bump function which is 1\leq 1 and constant on 1 over a neighbourhood of the origin. Write

χ1bC ( n) \chi \coloneqq 1 - b \;\in\; C^\infty(\mathbb{R}^n)

graphics grabbed from Dütsch 18, p. 108

and for λ(0,)\lambda \in (0,\infty) a positive real number, write

χ λ(x)χ(λx). \chi_\lambda(x) \coloneqq \chi(\lambda x) \,.

Since the product χ λu\chi_\lambda u has support of a distribution on a complement of a neighbourhood of the origin, we may extend it by zero to a distribution on all of n\mathbb{R}^n, which we will denote by the same symbols:

χ λu𝒟( n). \chi_\lambda u \in \mathcal{D}'(\mathbb{R}^n) \,.

By construction χ λu\chi_\lambda u coincides with uu away from a neighbourhood of the origin, which moreover becomes arbitrarily small as λ\lambda increases. This means that if the following limit exists

u^limλχ λu \hat u \;\coloneqq\; \underset{\lambda \to \infty}{\lim} \chi_\lambda u

then it is an extension of uu.

To see that the limit exists, it is sufficient to observe that we have a Cauchy sequence, hence that for all bC cp ( n)b\in C^\infty_{cp}(\mathbb{R}^n) the difference

(χ n+1uχ nu)(b)=u(b)(χ n+1+χ n) (\chi_{n+1} u - \chi_n u)(b) \;=\; u(b)( \chi_{n+1} + \chi_n )

becomes arbitrarily small.

It remains to see that the unique extension u^\hat u thus established has the same scaling degree as uu. This is shown in (Brunetti-Fredenhagen 00, p. 24).


(space of point-extensions of distributions)

For nn \in \mathbb{N}, let u𝒟( n{0})u \in \mathcal{D}'(\mathbb{R}^n \setminus \{0\}) be a distribution of degree of divergence deg(u)<deg(u) \lt \infty.

Then uu does admit at least one extension (def. 1) to a distribution u^𝒟( n)\hat u \in \mathcal{D}'(\mathbb{R}^n), and every choice of extension has the same degree of divergence as uu

deg(u^)=deg(u). deg(\hat u) = deg(u) \,.

Moreover, any two such extensions u^\hat u and u^ {\hat u}^\prime differ by a linear combination of partial derivatives of distributions of order deg(u)\leq deg(u) of the delta distribution δ 0\delta_0 supported at the origin:

u^ u^=α n|α|deg(u)q α αδ 0, {\hat u}^\prime - \hat u \;=\; \underset{ { \alpha \in \mathbb{N}^n } \atop { {\vert \alpha\vert} \leq deg(u) } }{\sum} q^\alpha \partial_\alpha \delta_0 \,,

for a finite number of constants q αq^\alpha \in \mathbb{C}.

This is essentially (Hörmander 90, thm. 3.2.4). We follow (Brunetti-Fredenhagen 00, theorem 5.3), which was inspired by (Epstein-Glaser 73, section 5). Review of this approach is in (Dütsch 18, theorem 3.35 (b)), see also remark 1 below.


For fC ( n)f \in C^\infty(\mathbb{R}^n) a smooth function, and ρ\rho \in \mathbb{N}, we say that ff vanishes to order ρ\rho at the origin if all partial derivatives with multi-index α n\alpha \in \mathbb{N}^n of total order |α|ρ{\vert \alpha\vert} \leq \rho vanish at the origin:

αf(0)=0AAA|α|ρ. \partial_\alpha f (0) = 0 \phantom{AAA} {\vert \alpha\vert} \leq \rho \,.

By Hadamard's lemma, such a function may be written in the form

(1)f(x)=α n|α|=ρ+1x αr α(x) f(x) \;=\; \underset{ {\alpha \in \mathbb{N}^n} \atop { {\vert \alpha \vert} = \rho + 1 } }{\sum} x^\alpha r_\alpha(x)

for smooth functions r αC cp ( n)r_\alpha \in C^\infty_{cp}(\mathbb{R}^n).


𝒟 ρ( n)𝒟( n)C cp ( n) \mathcal{D}_\rho(\mathbb{R}^n) \hookrightarrow \mathcal{D}(\mathbb{R}^n) \coloneqq C^\infty_{cp}(\mathbb{R}^n)

for the subspace of that of all bump functions on those that vanish to order ρ\rho at the origin.

By definition this is equivalently the joint kernel of the partial derivatives of distributions of order |α|{\vert \alpha\vert} of the delta distribution δ 0\delta_0 supported at the origin:

b𝒟 ρ( n)AAAAα n|α|ρ αδ 0,b=0. b \in \mathcal{D}_\rho(\mathbb{R}^n) \phantom{AA} \Leftrightarrow \phantom{AA} \underset{ {\alpha \in \mathbb{N}^n} \atop { {\vert \alpha\vert} \leq \rho } } {\forall} \left\langle \partial_\alpha \delta_0, b \right\rangle = 0 \,.

Therefore every continuous linear projection

p ρ:𝒟( n)𝒟 ρ( n) p_\rho \;\colon\; \mathcal{D}(\mathbb{R}^n) \longrightarrow \mathcal{D}_\rho(\mathbb{R}^n)

may be obtained from a choice of dual basis to the { αδ 0}\{\partial_\alpha \delta_0\}, hence a choice of smooth functions

{w βC cp ( n)} β n|β|ρ \left\{ w^\beta \in C^\infty_{cp}(\mathbb{R}^n) \right\}_{ { \beta \in \mathbb{N}^n } \atop { {\vert \beta\vert} \leq \rho } }

such that

αδ 0,w β=δ α βAAAAAA αw β(0)=δ α βAAAAfor|α|ρ, \left\langle \partial_\alpha \delta_0 \,,\, w^\beta \right\rangle \;=\; \delta_\alpha^\beta \phantom{AAA} \Leftrightarrow \phantom{AAA} \partial_\alpha w^\beta(0) \;=\; \delta_\alpha^\beta \phantom{AAAA} \text{for}\, {\vert \alpha\vert} \leq \rho \,,

by setting

(2)p ρidα n|α|ρw α αδ 0,(), p_\rho \;\coloneqq\; id \;-\; \left\langle \underset{ {\alpha \in \mathbb{N}^n} \atop { {\vert \alpha\vert} \leq \rho } }{\sum} w^\alpha \partial_\alpha \delta_0 \,,\, (-) \right\rangle \,,


p ρ:bbα n|α|ρ(1) |α|w α αb(0). p_\rho \;\colon\; b \mapsto b - \underset{ {\alpha \in \mathbb{N}^n} \atop { {\vert \alpha\vert} \leq \rho } }{\sum} (-1)^{{\vert \alpha\vert}} w^\alpha \partial_\alpha b(0) \,.

Together with Hadamard's lemma in the form (1) this means that every b𝒟( n)b \in \mathcal{D}(\mathbb{R}^n) is decomposed as

(3)b(x) =p ρ(b)(x)+(idp ρ)(b)(x) =α n|α|=ρ+1x αr α(x)+α n|α|ρ(1) |α|w α αb(0) \begin{aligned} b(x) & = p_\rho(b)(x) \;+\; (id - p_\rho)(b)(x) \\ & = \underset{ { \alpha \in \mathbb{N}^n } \atop { {\vert \alpha \vert} = \rho + 1 } }{\sum} x^\alpha r_\alpha(x) \;+\; \underset{ { \alpha \in \mathbb{N}^n } \atop { {\vert \alpha \vert} \leq \rho } }{\sum} (-1)^{{\vert \alpha \vert}} w^\alpha \partial_\alpha b(0) \end{aligned}

Now let

ρdeg(u). \rho \;\coloneqq\; deg(u) \,.

Observe that (by this prop.) the degree of divergence of the product of distributions x αux^\alpha u with |α|=ρ+1{\vert \alpha\vert} = \rho + 1 is negative

deg(x αu) =ρ|α|1 \begin{aligned} deg\left( x^\alpha u \right) & = \rho - {\vert \alpha \vert} \leq -1 \end{aligned}

Therefore prop. 1 says that each x αux^\alpha u for |α|=ρ+1{\vert \alpha\vert} = \rho + 1 has a unique extension x αu^\widehat{ x^\alpha u} to the origin. Accordingly the composition up ρu \circ p_\rho has a unique extension, by (3):

u^,b =u^,p ρ(b)+u^,(idp ρ)(b) =α n|α|=ρ+1x αu^,r αunique+α n|α|ρu^,w αchoice αδ 0,b \begin{aligned} \left\langle \hat u \,,\, b \right\rangle & = \left\langle \hat u , p_\rho(b) \right\rangle + \left\langle \hat u , (id - p_\rho)(b) \right\rangle \\ & = \underset{ { \alpha \in \mathbb{N}^n } \atop { {\vert \alpha \vert} = \rho + 1 } }{\sum} \underset{ \text{unique} }{ \underbrace{ \left\langle \widehat{x^\alpha u} \,,\, r_\alpha \right\rangle } } \;+\; \underset{ { \alpha \in \mathbb{N}^n } \atop { {\vert \alpha\vert} \leq \rho } }{\sum} \underset{ \text{choice} }{ \underbrace{ \langle \hat u \,,\, w^\alpha \rangle } } \left\langle \partial_\alpha \delta_0 \,,\, b \right\rangle \end{aligned}

That says that u^\hat u is of the form

u^=up ρ^unique+α n|α|ρc α αδ 0 \hat u \;=\; \underset{ \text{unique} }{ \underbrace{ \widehat{ u \circ p_\rho } } } + \underset{ { \alpha \in \mathbb{N}^n } \atop { {\vert \alpha\vert} \leq \rho } }{\sum} c^\alpha \, \partial_\alpha \delta_0

for a finite number of constants c αc^\alpha \in \mathbb{C}.

Notice that for any extension u^\hat u the exact value of the c αc^\alpha here depends on the arbitrary choice of dual basis {w α}\{w^\alpha\} used for this construction. But the uniqueness of the first summand means that for any two choices of extensions u^\hat u and u^ {\hat u}^\prime, their difference is of the form

u^ u^=α n|α|ρ((c) αc α) αδ 0, {\hat u}^\prime - \hat u \;=\; \underset{ { \alpha \in \mathbb{N}^n } \atop { {\vert \alpha\vert} \leq \rho } }{\sum} ( (c')^\alpha - c^\alpha ) \, \partial_\alpha \delta_0 \,,

where the constants q α((c) αc α)q^\alpha \coloneqq ( (c')^\alpha - c^\alpha ) \in \mathbb{C} are independent of any choices.

It remains to see that all these u^\hat u in fact have the same degree of divergence as uu.

By this example the degree of divergence of the point-supported distributions on the right is deg( αδ 0)=|α|ρdeg(\partial_\alpha \delta_0) = {\vert \alpha\vert} \leq \rho.

Therefore to conclude it is now sufficient to show that

deg(up ρ^)=ρ. deg\left( \widehat{ u \circ p_\rho } \right) \;=\; \rho \,.

This is shown in (Brunetti-Fredenhagen 00, p. 25).



Since in Brunetti-Fredenhagen 00, (38) the projectors (2) are denoted “WW”, the construction of extensions of distributions via the proof of prop. 2 has come to be called “W-extensions” (e.g Dütsch 18).


Powers of Feynman propagators

Extension of powers of Feynman propagators on globally hyperbolic spacetimes to the diagonal are worked out explicitly in (Hollands 07, section 3.4).


The argument for the characterization of the point extension of distributions goes back to

thereby laying the foundation for causal perturbation theory. A textbook account in functional analysis is in

  • Lars Hörmander, theorem 3.2.4 of The Analysis of Linear Partial Differential Operators I (Springer, 1990, 2nd ed.)

A more concise formulation and proof is due to

reviewed in

Exposition of the application to renormalization of Feynman diagrams is in

The refinement to the point-extension problem for distributions in the solution space of a given system of differential equations is discussed in

  • Dorothea Bahns, Michał Wrochna, On-shell extension of distributions, arXiv:1210.5448

Examples and applications to renormalization in perturbative quantum field theory are discussed for instance in

  • Stefan Hollands, sections 3.3 and 3.4 of Renormalized Quantum Yang-Mills Fields in Curved Spacetime, Rev. Math. Phys. 20:1033-1172, 2008 (arXiv:0705.3340)

For more on extension of distributions in renormalization see the references at causal perturbation theory and locally covariant perturbative quantum field theory.

Revised on February 8, 2018 16:06:28 by Urs Schreiber (