# Distributions as kernels of continuous operators *Written by GPT-6.1 Sol (OpenAI), September 2026. Public domain (CC0).* *Additional exposition by GPT-6 Astra (OpenAI), October 2026.* An integral kernel records how an operator transfers information from an input point to an output point. A function kernel cannot describe evaluation or differentiation. A distribution on the product space can describe both. This lesson proves Laurent Schwartz's kernel theorem and explains precisely which continuity assumptions it needs. The local theorem has no growth condition at infinity. Its tempered counterpart has a global polynomial bound. The supplied [scalar calculus foundation](../prerequisites/U011-free-foundations/metric-foundation-bridges.md), §§12–13, proves the compactness, derivative, integral, exponential and smooth-cutoff facts used below. The [finite algebra foundation](../prerequisites/U011-free-foundations/stable-prerequisite-bridges.md), §§10.1–10.6, supplies the real and complex finite-coordinate rules. The [measure foundation](../prerequisites/U011-free-foundations/banach-foundation-bridges.md), §§15.0–15.1 and 16.1–16.2, supplies Lebesgue integration, dominated convergence, Fubini and the linear change of variables used in the examples. The complete Baire proof, its selected logical base, the seminorm metric and smooth-space completeness are in [Functional extension and complete test spaces](../prerequisites/U011-free-foundations/functional-foundations-U008.md), §§6, 14.1–14.2 and 19. These are supplied earlier programme proofs with their stated licences. The locally finite partition, inductive-limit criterion, periodic Fourier reconstruction and Schwartz completeness needed here are proved below. Definitions of test spaces and distributions are recalled explicitly. Basic references are [Dyatlov 2026], [Melrose 2016] and [Schwartz 1952]. The argument here uses localized periodic expansions, first on bounded rectangles and then on a spatial lattice. It also proves the bounded-set facts needed to pass from weak continuity to strong continuity. ## What continuity means Let \(U\subset\mathbb R^m\) and \(V\subset\mathbb R^n\) be open. For a compact set \(L\subset U\), write \[ \begin{gathered} \mathcal D_L(U)= \{\psi\in C^\infty(U):\operatorname{supp}\psi\subset L\},\\ p_{L,r}(\psi)=\max_{|\alpha|\le r}\sup_U|\partial^\alpha\psi|. \end{gathered} \] These increasing seminorms give \(\mathcal D_L(U)\) its Fréchet topology. Completeness follows by taking uniform limits of every derivative and using the fundamental theorem of calculus to identify successive derivatives. The limit still vanishes outside \(L\). Give the test space \(\mathcal D(U)=C_c^\infty(U)\) the topology generated by **all** seminorms whose restriction to every \(\mathcal D_L(U)\) is continuous. This is a locally convex topology: finite intersections of seminorm balls are convex neighborhoods, and the triangle and homogeneity inequalities prove continuity of addition and scalar multiplication. Evaluation seminorms separate points. Each inclusion \(\mathcal D_L(U)\to\mathcal D(U)\) is continuous by construction. Here is its linear mapping property. A linear map \(A:\mathcal D(U)\to E\), where the topology of \(E\) is given by seminorms \(q\), is continuous if and only if every restriction \(A|_{\mathcal D_L(U)}\) is continuous. One direction follows by composing with the inclusions. Conversely, each \(q\circ A\) is then one of the defining seminorms, so the inverse image of every finite intersection of target seminorm balls is open. An exhaustion by compact sets with nested interiors gives the same test-space topology: every compact \(L\) is contained in one member of the exhaustion, and the inclusion between the two fixed-support spaces preserves the derivative seminorms. This is the locally convex inductive-limit topology. A distribution is a continuous complex-linear functional on this space. We use the bilinear convention \(\langle u,\psi\rangle\), with no conjugation of \(\psi\). Thus \(u\in\mathcal D'(U)\) exactly when, for every \(L\), there are \(C_L\) and \(r_L\) such that \[ |\langle u,\psi\rangle|\le C_Lp_{L,r_L}(\psi) \quad(\psi\in\mathcal D_L(U)). \] The order and constant may vary with \(L\). The weak dual topology on \(\mathcal D'(U)\) tests each fixed \(\psi\). The strong dual topology uses \[ q_B(u)=\sup_{\psi\in B}|\langle u,\psi\rangle|, \] where \(B\) runs over bounded subsets of \(\mathcal D(U)\). A subset of a locally convex space is bounded when every continuous seminorm is bounded on it. These are different definitions of topology. The continuity equivalence proved below concerns this particular class of operators; it does not assert equality of the two dual topologies. **Lemma 1.1 (bounded test families).** A subset \(B\subset\mathcal D(U)\) is bounded if and only if all its elements are supported in one compact set and every derivative seminorm is uniformly bounded on \(B\). **Proof.** Suppose \(B\) has no common compact support. For a compact exhaustion \(L_j\), choose \(\psi_j\in B\) and \(x_j\notin L_j\) with \(\psi_j(x_j)\ne0\). Enlarging the exhaustion along the construction, we can make \(x_j\) escape every compact subset of \(U\). Choose positive \(a_j\) with \(a_j|\psi_j(x_j)|\ge j\). Then \[ p(\psi)=\sup_j a_j|\psi(x_j)| \] is finite on every test function. On each fixed support space it is the maximum of finitely many continuous evaluations, so it is a continuous seminorm on the inductive limit. It is unbounded on \(B\), a contradiction. For every \(r\), the seminorm \(\max_{|\alpha|\le r}\sup_U|\partial^\alpha\psi|\) is continuous on every support space and hence on \(\mathcal D(U)\). Boundedness therefore gives uniform derivative bounds. Conversely, if \(B\) has a common compact support and uniform derivative bounds, every continuous seminorm restricted to that Fréchet support space is bounded by a constant times one of its increasing derivative seminorms. It is bounded on \(B\). This proves the converse. \(\square\) For a linear map \(T:\mathcal D(V)\to\mathcal D'(U)\), weak continuity means that \[ \phi\longmapsto\langle T\phi,\psi\rangle \] is continuous for every fixed \(\psi\in\mathcal D(U)\). It does not initially require a uniform estimate as \(\psi\) varies. The next lemma supplies that estimate on fixed supports. ## Two continuous variables give one estimate **Lemma 2.1 (the Baire estimate).** Let \(E,F\) be Fréchet spaces with increasing seminorms \(p_j,q_j\). If \(b:E\times F\to\mathbb C\) is a separately continuous bilinear form, there are indices \(r,s\) and a constant \(C\) such that \[ |b(e,f)|\le Cp_r(e)q_s(f). \tag{2.1} \] **Proof.** For positive integers \(s,k\), put \[ A_{s,k}=\left\{e: \begin{gathered} |b(e,f)|\le kq_s(f)\\ \text{ for every }f\in F \end{gathered} \right\}. \] Each set is closed because \(e\mapsto b(e,f)\) is continuous. Continuity in \(f\), for each fixed \(e\), shows that these countably many sets cover \(E\). Baire's theorem gives one \(A_{s,k}\) with nonempty interior. Choose \(e_0\) and a neighborhood \(W\) of zero such that \(e_0+W\subset A_{s,k}\) and \(e_0\in A_{s,k}\). Subtraction gives \[ |b(e,f)|\le2kq_s(f)\quad(e\in W). \] There are \(r\) and \(\varepsilon>0\) with \(\{p_r<\varepsilon\}\subset W\). For \(p_r(e)>0\), apply the preceding bound to \(\varepsilon e/(2p_r(e))\). For \(p_r(e)=0\), every positive multiple of \(e\) belongs to that seminorm ball; letting the multiple increase gives \(b(e,f)=0\). This proves (2.1), with \(C=4k/\varepsilon\). \(\square\) For a weakly continuous \(T\), apply the lemma to \[ b(\psi,\phi)=\langle T\phi,\psi\rangle \] on \(\mathcal D_L(U)\times\mathcal D_M(V)\). Separate continuity in \(\psi\) follows because each \(T\phi\) is a distribution. We obtain \[ \begin{gathered} |\langle T\phi,\psi\rangle|\\ \le C_{L,M}p_{L,r}(\psi)p_{M,s}(\phi). \end{gathered} \tag{2.2} \] The indices depend on both compact sets. No global order bound has appeared. ## Localizing on open sets **Partition lemma.** Every open cover of an open subset \(X\) of Euclidean space has a countable smooth partition of unity whose supports are compact subsets of members of the cover and form a locally finite family in \(X\). **Proof.** Put \[ \begin{gathered} K_m=\{x\in X:|x|\le m,\ d(x)\ge1/m\},\\ d(x)=\operatorname{dist}(x,\mathbb R^n\setminus X),\qquad m\ge1, \end{gathered} \] using infinite distance when \(X=\mathbb R^n\), and set \(K_m=\varnothing\) for \(m\le0\). Each \(K_m\) is compact, \(K_m\subset\operatorname{int}K_{m+1}\), and their interiors exhaust \(X\). Distance to a nonempty closed set is continuous because the triangle inequality gives \(|d(x)-d(y)|\le|x-y|\); this also justifies closedness in this formula. The compact layer \(L_m=K_m\setminus\operatorname{int}K_{m-1}\) lies in the open set \(\operatorname{int}K_{m+1}\setminus K_{m-2}\). Around each of its points choose concentric balls \(B\subset2B\) with \(\overline{2B}\) inside that open set and inside one member of the given cover. Finitely many of the smaller balls cover \(L_m\). Take a smooth nonnegative bump equal to one on each smaller ball and supported in its corresponding larger ball; the exact radial bump construction is in the stated cutoff prerequisite. The resulting countable family of supports is locally finite. In fact a point has a neighbourhood inside some \(\operatorname{int}K_N\), and every layer with \(m\ge N+2\) has its selected supports outside \(K_{m-2}\supset K_N\); the finitely many earlier layers contribute only finitely many balls. At least one bump is positive at each point: choose the smallest \(m\) for which the point belongs to \(K_m\), and use its layer. The locally finite sum \(s\) of all bumps is therefore positive and smooth. Dividing each bump by \(s\) gives the required partition, with its original support unchanged. \(\square\) For a distribution \(u\), say that it vanishes on an open set \(O\) when it annihilates every test supported there. Its support is the complement of the union of such open sets. The preceding partition proves that \(u\) vanishes on that union: a compactly supported test there is a finite sum of tests localized inside the individual zero neighborhoods. Thus a test whose support avoids \(\operatorname{supp}u\) has zero pairing. The same finite partition argument proves that distributions agreeing on an open cover agree on its union. For later use, distributional differentiation means \[ \langle\partial_j u,\psi\rangle=-\langle u,\partial_j\psi\rangle. \] This is again continuous: differentiation preserves a fixed support and increases its required derivative seminorm by one. A locally integrable function defines a distribution by integration, since its pairing with a test supported in \(L\) is bounded by its \(L^1(L)\) norm times \(p_{L,0}\). ## Breaking a test function into products Choose rectangles \(A\) and \(B\) whose closures lie in \(U\) and \(V\). Choose larger rectangles \(Q_U\Subset U\), \(Q_V\Subset V\), containing those closures, and cutoffs \(\chi\in\mathcal D(Q_U)\), \(\eta\in\mathcal D(Q_V)\) equal to one near \(\overline A\), \(\overline B\). Regard the larger rectangles as fundamental cells of tori. Let \(e_p(x)\), \(f_q(y)\) be their exponential Fourier modes. The frequencies include the scale factors determined by the side lengths of the rectangles. For \(H\in\mathcal D(A\times B)\), extend \(H\) by zero to the larger cells and then periodically. If \(c_{pq}(H)\) denotes its normalized Fourier coefficient, then \[ \begin{gathered} H(x,y)=\sum_{p\in\mathbb Z^m,\ q\in\mathbb Z^n} c_{pq}(H)\\ {}\cdot\chi(x)e_p(x)\eta(y)f_q(y). \end{gathered} \tag{3.1} \] The series converges with every derivative, with support in one fixed compact product. Integration by parts using powers of the periodic Laplacian gives \[ \begin{gathered} |c_{pq}(H)|\le C_N\\ {}\cdot(1+|p|^2+|q|^2)^{-N}\|H\|_{C^{2N}},\\ N\ge0. \end{gathered} \tag{3.2} \] The eigenvalues of that Laplacian are comparable to \(|p|^2+|q|^2\), with constants depending on the cells. A derivative of order at most \(a\) of a term in (3.1) costs at most \(C_a(1+|p|+|q|)^a\). Choose \(2N>a+m+n\); the lattice sum of these bounds is finite. Explicitly, in dimension \(d\), the shell \(2^\ell\le |v|<2^{\ell+1}\) contains at most \(C_d2^{\ell d}\) integer points by counting the surrounding coordinate box. For \(b>d\), the sum of \(\langle v\rangle^{-b}\) over that shell is at most \(C'_d2^{-\ell(b-d)}\). The geometric series converges, and the finitely many points with \(|v|<1\) do not affect convergence. Thus the periodic series converges absolutely with every derivative. Here is why its value is the original periodic function. On the circle of length \(2\pi\), the Fejér kernels are \[ F_N(t)=\frac1N\left|\sum_{j=0}^{N-1}e^{ijt}\right|^2. \] Expanding the finite square gives Fourier coefficients \((1-|p|/N)_+\), where \(a_+=\max(a,0)\). The integral of \(e^{ipt}\) is zero for a nonzero integer \(p\), by evaluating its primitive at the endpoints, and is one for \(p=0\) under normalized measure. They are therefore nonnegative and have integral one for the normalized measure \(dt/(2\pi)\), by orthogonality of the exponential modes. Outside a fixed neighborhood of zero their values are \(O(1/N)\), from the geometric-sum identity. Their products in several variables therefore form an approximate identity: split the convolution into a small neighborhood of zero, controlled by uniform continuity, and its complement, whose mass tends to zero. For a product kernel, the complement of a small coordinate box is contained in the union of the corresponding one-coordinate complements; positivity and the unit integral in the other coordinates give the same vanishing-mass bound. Consequently Fejér means converge uniformly to every continuous periodic function. For the absolutely convergent Fourier series above, its Fejér means also converge uniformly to its sum, because their coefficients are bounded by one and tend to one at every fixed frequency. Both limits agree. Multiplying by \(\chi\eta\), which equals one on the support of \(H\), proves (3.1). Rectangles of other side lengths follow by rescaling. **Lemma 3.1 (product tests detect distributions).** Finite sums of tests \(\psi(x)\phi(y)\) are dense in \(\mathcal D(U\times V)\). A distribution on the product that vanishes on every such product is zero. **Proof.** On a relatively compact product rectangle, (3.1) is precisely an approximation by finite sums of products, converging in one support space. For a general compactly supported \(H\), choose smooth partitions of unity on \(U\) and \(V\) subordinate to relatively compact rectangles. Only finitely many products of partition functions meet \(\operatorname{supp}H\). Apply (3.1) to each resulting piece. The sum of these finite approximations converges in one compact support space. Continuity of a distribution proves the last assertion. \(\square\) ## The local kernel theorem **Theorem 4.1 (Schwartz's kernel theorem on open sets).** For arbitrary open \(U\subset\mathbb R^m\), \(V\subset\mathbb R^n\), the following data are equivalent: 1. A distribution \(K\in\mathcal D'(U\times V)\). 2. A weakly continuous linear map \(T:\mathcal D(V)\to\mathcal D'(U)\). 3. A linear map \(T:\mathcal D(V)\to\mathcal D'(U)\) continuous into the strong dual topology. 4. A bilinear form on \(\mathcal D(U)\times\mathcal D(V)\) separately continuous in each test variable. The correspondence is \[ \langle T\phi,\psi\rangle =\langle K,\psi(x)\phi(y)\rangle. \tag{4.1} \] The kernel is unique. In particular, continuity of each scalar pairing is enough; no estimate uniform over all supports is required. The same class of operators is obtained if weak continuity is specified only using convergent test sequences. On a fixed support space, every scalar pairing is then a sequentially continuous linear functional on a metrizable Fréchet space, hence is continuous. To see the last implication, failure of continuity would give vectors tending to zero in the metric with functional values bounded away from zero. The inductive-limit property then gives continuity on all of \(\mathcal D(V)\). Conversely a continuous scalar pairing preserves every convergent test sequence. **Proof.** Start with \(K\). On fixed compact supports \(L,M\), its distribution estimate and the product rule give \[ \begin{gathered} |\langle K,\psi\otimes\phi\rangle|\\ \le C p_{L,a}(\psi)p_{M,a}(\phi). \end{gathered} \tag{4.2} \] for some \(C,a\). For fixed \(\phi\), this defines a distribution \(T\phi\). For fixed \(\psi\), it defines a continuous functional of \(\phi\). Thus it supplies both weak continuity and separate continuity. To prove strong continuity, let \(B\subset\mathcal D(U)\) be bounded. Lemma 1.1 puts its elements in one \(\mathcal D_L(U)\), with \(\sup_{\psi\in B}p_{L,a}(\psi)<\infty\). Formula (4.2) implies \[ q_B(T\phi)\le C\Big(\sup_{\psi\in B}p_{L,a}(\psi)\Big)p_{M,a}(\phi). \] The restriction of \(T\) to every support space is therefore continuous into the strong dual. The inductive-limit property proves global strong continuity. Strong continuity implies weak continuity because a singleton test family is bounded. Conversely, start with a separately continuous bilinear form \(b\). It gives a weakly continuous map by \(\langle T\phi,\psi\rangle=b(\psi,\phi)\). Use the rectangles, cutoffs and modes of Section 3. Define, for \(H\in\mathcal D(A\times B)\), \[ \begin{gathered} K_{A,B}(H)\\ =\sum_{p,q}c_{pq}(H)b(\chi e_p,\eta f_q). \end{gathered} \tag{4.3} \] Lemma 2.1 on the supports of \(\chi,\eta\) bounds the second factor by \[ |b(\chi e_p,\eta f_q)|\le C(1+|p|)^r(1+|q|)^s. \] With \(2N>r+s+m+n\), (3.2) shows that (4.3) converges absolutely and is bounded by \(C'\|H\|_{C^{2N}}\). It is a distribution on \(A\times B\). If \(H=\psi\otimes\phi\), its Fourier coefficients factor into the coefficients of \(\psi\) and \(\phi\). Their cutoff expansions converge in the two fixed support spaces. The joint estimate (2.2) then gives \[ K_{A,B}(\psi\otimes\phi)=b(\psi,\phi). \] Lemma 3.1 implies uniqueness on that rectangle. In particular the construction is independent of the larger cells and cutoffs. On overlaps of two product rectangles, cover the overlap by smaller product rectangles. The two distributions agree on product tests in each smaller rectangle and hence agree there by Lemma 3.1. They therefore glue. Explicitly, choose a locally finite partition subordinate to the product-rectangle cover and define the value on a global test by the finite sum of the corresponding localized values. For two such partitions, refine each term by the other partition. Only finitely many terms meet the test support, the two local distributions agree on every intersection, and both double sums therefore coincide. This proves independence of the partition. On each compact support only finitely many distribution estimates are needed, so the result is continuous. This gives \(K\in\mathcal D'(U\times V)\). To verify (4.1) on an arbitrary product test, choose product partitions on \(U\) and \(V\). The two test functions split into finite sums, each localized product has the already proved pairing, and bilinearity recombines them to \(b(\psi,\phi)\). Lemma 3.1 gives global uniqueness. All four correspondences are inverse. \(\square\) ## Recovering a kernel by decreasing its smoothing scale A smooth approximation can grow without bound at each smaller scale while converging as a distribution. Differentiation in the scale is useful when it can be transferred to the test function. The following estimate quantifies that mechanism. **Lemma 4.2 (Taylor convergence with a fixed distribution order).** Let \(W\subset\mathbb R^d\) be open, \(M\ge0\) an integer, and \(0<\varepsilon\le b\le1\). Suppose \(F_\varepsilon\in\mathcal D'(W)\), each scalar pairing is \(C^{M+1}\) in \(\varepsilon>0\), and for every compact \(L\Subset W\), \[ |\partial_\varepsilon^j F_\varepsilon(\Phi)| \le C_{L,j}\varepsilon^{-M}p_{L,j}(\Phi), \qquad 0\le j\le M+1,\quad\Phi\in\mathcal D_L(W). \tag{R1} \] Then \(F_\varepsilon\) converges to a distribution \(F_0\) of order at most \(M+1\). On each fixed support, \[ |(F_\varepsilon-F_0)(\Phi)| \le C_L\varepsilon\bigl(1+\log(b/\varepsilon)\bigr) p_{L,M+1}(\Phi). \tag{R2} \] For \(M=0\), the logarithm can be omitted. In particular convergence is uniform on bounded test families. **Proof.** Fix \(\Phi\) and put \(f(t)=F_t(\Phi)\). Repeated integration of the fundamental theorem, or integration by parts in its integral remainder, gives \[ f(\varepsilon)= \sum_{j=0}^{M}\frac{(\varepsilon-b)^j}{j!}f^{(j)}(b) +\frac{(-1)^{M+1}}{M!} \int_\varepsilon^b(t-\varepsilon)^M f^{(M+1)}(t)\,dt. \tag{R3} \] For \(M=0\) this is \(f(\varepsilon)=f(b)-\int_\varepsilon^b f'(t)\,dt\). Integrating the remainder once by parts proves the next value of \(M\), so (R3) holds for every integer \(M\). By (R1), \(t^M|f^{(M+1)}(t)|\le C_{L,M+1}p_{L,M+1}(\Phi)\). Consequently the following absolutely convergent expression defines a linear functional: \[ F_0(\Phi)= \sum_{j=0}^{M}\frac{(-b)^j}{j!}f^{(j)}(b) +\frac{(-1)^{M+1}}{M!} \int_0^b t^M f^{(M+1)}(t)\,dt. \tag{R4} \] The finite sum and integral are bounded by a constant times \(p_{L,M+1}(\Phi)\). Thus \(F_0\) is a distribution with this same order on every compact support; its constants may depend on \(L\). Subtract (R4) from (R3). The finite polynomial changes by at most \(C_L\varepsilon p_{L,M+1}(\Phi)\). For \(M\ge1\), the elementary factorization of two \(M\)-th powers gives \[ 0\le t^M-(t-\varepsilon)^M\le M\varepsilon t^{M-1}, \qquad \varepsilon\le t\le b. \] The integral over \(00\), all input supports in (R6), with \(0<\varepsilon\le b_0\), lie in \(L_U,L_V\). On \(U_0\times V_0\), the smooth functions \(K_\varepsilon\) converge to the unique kernel of \(b\). That kernel has order at most \(\mu+1\) there, and the convergence satisfies (R2) with \(M=\mu\), \(b=b_0\). These local limits agree on overlaps and define the kernel on all of \(U\times V\). **Proof.** Shrink \(b_0\le1\) so that the translated scaled profile supports stay in the specified compact neighborhoods, with positive margins. Smooth dependence of the profiles on \((x,y,\varepsilon)\) holds in every fixed-support test seminorm when \(\varepsilon>0\). The joint estimate (R5) allows differentiation of their bilinear pairing: expand a difference of two pairings into the changes in its two factors and use their seminorm limits. The same argument applies after every derivative. Thus \(K_\varepsilon\) is smooth in all these parameters. For any compact smooth profile \(h\) on \(\mathbb R^e\), direct differentiation proves the scale identity \[ \partial_\varepsilon\{\varepsilon^{-e}h(z/\varepsilon)\} =\sum_{\nu=1}^{e} \partial_{z_\nu}\{\varepsilon^{-e}h_\nu(z/\varepsilon)\}, \qquad h_\nu(t)=-t_\nu h(t). \tag{R7} \] Both sides are \(\varepsilon^{-e-1}[-e h(t)-t\cdot\nabla h(t)]\), with \(t=z/\varepsilon\). No additional inverse power remains outside the spatial derivatives on the right. Write \(z=(x,y)\), \(\alpha=(\alpha_x,\alpha_y)\), and let \(K_{\varepsilon,\alpha}\) denote (R6) with profiles \((-t)^{\alpha_x}\rho(t)\) and \((-w)^{\alpha_y}\sigma(w)\). Applying (R7) repeatedly to the two factors yields \[ \partial_\varepsilon^jK_\varepsilon =\sum_{|\alpha|=j}\frac{j!}{\alpha!} \partial_z^\alpha K_{\varepsilon,\alpha}. \tag{R8} \] Induction proves the coefficient: differentiating one more time chooses one of the \(m+n\) spatial coordinates, and the numbers of ordered choices with multiplicities \(\alpha\) are \(j!/\alpha!\). The altered profiles have the original compact supports. Their fixed derivative norms and (R5) imply \[ \sup_{U_0\times V_0}|K_{\varepsilon,\alpha}| \le C_\alpha\varepsilon^{-\mu}. \tag{R9} \] Indeed derivatives through order \(r\) of the first scaled profile cost at most \(\varepsilon^{-m-r}\), and derivatives through order \(s\) of the second cost at most \(\varepsilon^{-n-s}\), since \(\varepsilon\le1\). For a test \(\Phi\) on a fixed compact support, integrate (R8) by parts. The test has no boundary terms. Its compact containing volume and the finite coefficient sum give \[ |\partial_\varepsilon^j\langle K_\varepsilon,\Phi\rangle| \le C_{L,j}\varepsilon^{-\mu}p_{L,j}(\Phi). \tag{R10} \] Lemma 4.2 now constructs a distributional limit \(K_0\) and proves its order and convergence estimate. It remains to identify its product-test values. For \(\psi\in\mathcal D(U_0)\), \(\phi\in\mathcal D(V_0)\), integration of (R6) gives \[ \langle K_\varepsilon,\psi\otimes\phi\rangle =b(\psi*\check\rho_\varepsilon,\phi*\check\sigma_\varepsilon), \qquad\check\rho(t)=\rho(-t),\quad\check\sigma(w)=\sigma(-w). \tag{R11} \] For justification, replace the two ordinary compact integrals by Riemann sums. Every derivative of the input profiles converges uniformly in those sums, with a common compact support. Estimate (R5) therefore passes the two limits through \(b\). The identity \[ \int\psi(x)\rho_\varepsilon(x-u)\,dx =\int\rho(t)\psi(u+\varepsilon t)\,dt \] proves both the reflection and the full scaling factor in (R11). Uniform continuity of every derivative on a compact neighborhood shows that these averages tend to \(\psi,\phi\) in their fixed-support test spaces. Since both profiles have integral one, (R5) gives \(K_0(\psi\otimes\phi)=b(\psi,\phi)\). Here product-test uniqueness also has a direct averaging proof. For a compact smooth \(H(x,y)\), its convolution with a product of two compact normalized profiles tends to \(H\) in every derivative seminorm, with one common compact support. For each fixed scale, Riemann sums for this convolution are finite sums of product tests. Choose scales tending to zero and meshes so that the smoothing and sum errors through derivative order \(j\) are each at most \(1/j\). The resulting sequence converges to \(H\) in the test space. Product tests therefore detect all distributions. The construction can keep supports in any prescribed product neighborhoods of the two compact projections. Consequently the local limits agree on every smaller product rectangle in an overlap. The partition and refinement argument in the proof of Theorem 4.1 glues them to a global distribution: it uses only this uniqueness, a finite partition near each compact test support, and the local finite-order bounds. Product partitions show that the glued distribution pairs to \(b(\psi,\phi)\) for arbitrary product tests. It is the unique kernel, and all continuity conclusions of Theorem 4.1 follow from its kernel estimate. \(\square\) The order \(\mu+1\) is an upper bound supplied by this construction, not necessarily the least order of the kernel. Solution 6 exhibits a kernel of exact order one with a larger bound from (R6). Solution 7 shows how vanishing first moments give faster convergence when two additional test derivatives are controlled. ![Concentration of two regularized point kernels and a logarithmic scale error](../figures/regularized-point-kernels.png) *Figure. The first two panels are numerical samples of \(K_\varepsilon(x,y)=\rho_\varepsilon(x)\rho_\varepsilon(y)\), the regularization of the point kernel \(K=\delta_0\otimes\delta_0\), at \(\varepsilon=1\) and \(1/2\). Here \(\rho(t)=c\exp[-1/(1-t^2)]\) for \(|t|<1\), zero otherwise, with \(c^{-1}=\int_{-1}^1\exp[-1/(1-t^2)]dt\). Each exact kernel has integral one; the support contracts by two in each direction and the central value grows by four. The color scale is shared. The last panel plots the exact scalar coefficient \(\varepsilon\log(1/\varepsilon)\) of the error from Solution 5. Its quotient by \(\varepsilon\) is unbounded, even though it tends to zero. [Editable figure script](../figures/regularized-point-kernels.py).* ## Polynomial control on the whole space For \(d\ge1\), let \(\mathcal S(\mathbb R^d)\) be the Schwartz space, with increasing seminorms \[ \begin{gathered} P_a(f)=\max_{|\alpha|\le a}\sup_z \langle z\rangle^a|\partial^\alpha f(z)|,\\ \langle z\rangle=(1+|z|^2)^{1/2}. \end{gathered} \] Its continuous dual is \(\mathcal S'(\mathbb R^d)\), the tempered distributions. The weak and strong dual topologies are defined by fixed tests and bounded test families, respectively. A family in \(\mathcal S\) is bounded exactly when every \(P_a\) is uniformly bounded on it. **Completeness of the Schwartz space.** The \(P_a\) separate points, so the seminorm metric in the supplied §14.1 applies. Suppose \(f_l\) is Cauchy in every \(P_a\). For every multi-index \(\alpha\), the functions \(\partial^\alpha f_l\) converge uniformly to a continuous \(g_\alpha\). Passing to the limit in the coordinate identity \[ \partial^\alpha f_l(x+he_i)-\partial^\alpha f_l(x) =\int_0^h\partial^{\alpha+e_i}f_l(x+te_i)\,dt \] shows that \(\partial_i g_\alpha=g_{\alpha+e_i}\). Continuous partial derivatives and their repeated coordinate identities show that \(f=g_0\) is smooth with \(\partial^\alpha f=g_\alpha\). For fixed \(a\) and \(|\alpha|\le a\), the functions \(\langle x\rangle^a\partial^\alpha f_l(x)\) are uniformly Cauchy, so they have a bounded uniform limit. Pointwise that limit is \(\langle x\rangle^ag_\alpha(x)\). Consequently \(f\in\mathcal S\) and \(P_a(f_l-f)\to0\) for every \(a\). The seminorm metric is complete. Thus \(\mathcal S\) is Fréchet, and the full Baire theorem used in Lemma 2.1 applies to it. **Lemma 5.1 (a spatial and frequency expansion).** There are compactly supported smooth functions \(\theta,\eta\) on \(\mathbb R^d\), supported in \((-\pi,\pi)^d\), with \(\eta=1\) on \(\operatorname{supp}\theta\) and \[ \sum_{j\in\mathbb Z^d}\theta(z-j)=1. \] For such functions in dimensions \(m,n\), every \(H\in\mathcal S(\mathbb R^{m+n})\) has an expansion \[ \begin{gathered} H(x,y)=\sum_{j,k,p,q}c_{jk,pq}(H)\\ {}\cdot\eta_U(x-j)e^{ip\cdot(x-j)}\\ {}\cdot\eta_V(y-k)e^{iq\cdot(y-k)}. \end{gathered} \tag{5.1} \] Here \(j,p\in\mathbb Z^m\), \(k,q\in\mathbb Z^n\). The series converges absolutely in every Schwartz seminorm, and, for all integers \(M,N\ge0\), \[ \begin{gathered} |c_{jk,pq}(H)|\le C_{M,N} \langle(j,k)\rangle^{-M}\\ {}\cdot\langle(p,q)\rangle^{-2N}\\ {}\cdot\max_{|\alpha|\le2N}\sup_z \langle z\rangle^M|\partial^\alpha H(z)|. \end{gathered} \tag{5.2} \] **Proof.** Choose a nonnegative bump \(\theta_0\), supported in \((-1,1)^d\), positive on \([-1/2,1/2]^d\). Its integer translates have a positive smooth periodic sum \(a(z)\). Put \(\theta=\theta_0/a\), and choose \(\eta\) supported in \((-\pi,\pi)^d\), equal to one on a neighborhood of \([-1,1]^d\). This gives the required partition. Let \(c_{jk,pq}(H)\) be the normalized Fourier coefficient on the \(2\pi\)-periodic product cell of \[ (s,t)\longmapsto\theta_U(s)\theta_V(t)H(j+s,k+t). \] This function is supported away from the boundary of the cell. Integration by parts with \((1-\Delta_s-\Delta_t)^N\) gives the frequency factor in (5.2). On its fixed support, \(\langle(j+s,k+t)\rangle\) is comparable to \(\langle(j,k)\rangle\), uniformly in the lattice points. The product rule gives the spatial factor and the displayed derivative bound. For any integer \(a\), the Schwartz seminorm \(P_a\) of an atom in (5.1) is at most \[ C_a\langle(j,k)\rangle^a\langle(p,q)\rangle^a. \] Choose \(M>a+m+n\) and \(2N>a+m+n\) in (5.2). Summation over the two lattices is then finite, proving absolute convergence in \(P_a\). For a fixed \(j,k\), Fourier reconstruction gives the localized function because \(\eta_U\eta_V=1\) on its support. Summing those functions gives \(H\) by the spatial partition. The convergence just proved also justifies this equality in \(\mathcal S\). \(\square\) **Theorem 5.2 (the tempered kernel theorem).** Distributions \[ K\in\mathcal S'(\mathbb R^{m+n}) \] correspond uniquely, by (4.1), to weakly continuous linear maps \[ T:\mathcal S(\mathbb R^n)\longrightarrow\mathcal S'(\mathbb R^m). \] Every such map is also continuous into the strong dual topology. Equivalently, kernels correspond to separately continuous bilinear forms on the two Schwartz spaces. **Proof.** A tempered kernel satisfies \(|K(H)|\le CP_a(H)\). The product rule and \(\langle(x,y)\rangle\le\langle x\rangle\langle y\rangle\) give \[ |K(\psi\otimes\phi)|\le C'P_a(\psi)P_a(\phi). \] This defines a weakly continuous map. Taking the supremum over a bounded family of \(\psi\)'s proves strong continuity. Conversely, the Baire estimate gives \[ |b(\psi,\phi)|\le CP_r(\psi)P_s(\phi). \] For the atoms in Lemma 5.1, the value of \(b\) is at most \[ C'\langle(j,k)\rangle^{r+s}\langle(p,q)\rangle^{r+s}. \] Define \(K(H)\) by replacing each atom in (5.1) by this bilinear value and summing with its coefficient. Choose \(M>r+s+m+n\) and \(2N>r+s+m+n\) in (5.2). The sum converges absolutely and is bounded by a fixed Schwartz seminorm of \(H\); hence \(K\) is tempered. For \(H=\psi\otimes\phi\), the coefficients factor into the corresponding one-variable spatial and frequency coefficients. Lemma 5.1 in each dimension, followed by joint continuity of \(b\), gives \(K(\psi\otimes\phi)=b(\psi,\phi)\). Finite sums of Schwartz products are dense by (5.1), so \(K\) is unique. Strong continuity and its converse to weak continuity follow as above. \(\square\) If one dimension is zero, its test space is \(\mathbb C\), and the result is the ordinary continuous-dual identification. If an open set is empty, all corresponding test and distribution spaces are zero. These cases need no Fourier construction. ## Three ways to see an operator **Example 6.1 (a singular output with one input measurement).** On \(\mathbb R\), let \(a=\operatorname{pv}(1/x)\), defined by \[ \langle a,\psi\rangle =\int_0^\infty\frac{\psi(x)-\psi(-x)}{x}\,dx. \] For \(00\), and \(\phi(0)=1\). The Schwartz seminorms of \(H_R(x,y)=\psi(x-R)\phi(y)\) grow at most polynomially in \(R\), whereas \[ K(H_R)\ge e^{R^2}\int\psi \quad(R>0). \] No fixed Schwartz seminorm can bound these values. The global growth distinction cannot be removed from the tempered theorem. **Proposition 6.4 (graphs of continuous maps).** Let \(f:U\to V\) be continuous. The operator \(T\phi=\phi\circ f\), viewed as a locally integrable function on \(U\), has kernel \[ K_f(H)=\int_U H(x,f(x))\,dx. \] Its support is exactly the graph of \(f\). Smoothness and injectivity of \(f\) are unnecessary. **Proof.** For a test supported in a compact product, its restriction to the graph is continuous and supported in the compact projection onto \(U\). Its integral is bounded by the measure of a compact neighborhood of that projection times the supremum of the test. Thus \(K_f\) is a distribution of order zero. Its pairing with a product test is \(\int\psi(x)\phi(f(x))\,dx\), giving the asserted operator. The graph is relatively closed by continuity of \(f\), and a test supported off it has zero pairing. At a graph point, choose a nonnegative test positive on a small product neighborhood of that point. Continuity of \(f\) gives a nonempty open set of input points whose graph lies in that positive neighborhood. The pairing is positive, so every graph point is in the support. \(\square\) **Proposition 6.5 (the support relation of a kernel).** For every local distribution kernel and every \(\phi\in\mathcal D(V)\), \[ \begin{gathered} \operatorname{supp}T\phi\subset\\ \left\{ \begin{gathered} x\in U:\\ \text{there is }y\in\operatorname{supp}\phi\\ \text{with }(x,y)\in\operatorname{supp}K \end{gathered} \right\}. \end{gathered} \tag{6.1} \] No properness of the kernel projections is required for this test-input assertion. **Proof.** The set on the right is closed in \(U\): from a convergent sequence of its output points, compactness of \(\operatorname{supp}\phi\) gives a convergent subsequence of their corresponding input points, and closedness of \(\operatorname{supp}K\) gives the limiting pair. On an output neighborhood disjoint from that closed set, every product test \(\psi\otimes\phi\) has support disjoint from \(\operatorname{supp}K\), so its kernel pairing is zero. Hence \(T\phi\) vanishes on that neighborhood. This proves the containment. \(\square\) ## Exercises **Exercise 1 (basic: locating an averaging kernel).** For \(\rho\in\mathcal D(\mathbb R)\), define \[ T\phi(x)=\int_{-2}^{1}\rho(t)\phi(x+3t)\,dt. \] Find its kernel as a locally integrable function. Account for the Jacobian and determine its support, allowing \(\rho\) to vanish on part of \([-2,1]\). **Exercise 2 (intermediate: which variable is differentiated?).** If \(K\) is the kernel of \(T\), find the kernels of \(\partial_{x_j}T\) and \(T\partial_{y_k}\). Find the kernel of the bilinear transpose \(T^t\), characterized by \(\langle T^t\psi,\phi\rangle=\langle T\phi,\psi\rangle\). Give a one-dimensional check using \(T\phi(x)=\phi(2x)\). **Exercise 3 (intermediate: a lattice of measurements).** Let \((a_j)_{j\in\mathbb Z}\) have polynomial growth. Show that \[ K(H)=\sum_{j\in\mathbb Z}a_jH(j,j+1) \] is tempered. Describe its operator on Schwartz functions. Prove that if all \(a_j\ne0\), its support is precisely \(\{(j,j+1):j\in\mathbb Z\}\). **Exercise 4 (advanced: continuity of a whole family).** Suppose \(T_\ell:\mathcal D(V)\to\mathcal D'(U)\) are weakly continuous and, for each pair of fixed compact supports \(L,M\), there are \(C,r,s\) independent of \(\ell\) such that \[ |\langle T_\ell\phi,\psi\rangle| \le Cp_{L,r}(\psi)p_{M,s}(\phi). \] Suppose all product-test pairings converge to those of a weakly continuous \(T\). Prove that their kernels converge to the kernel of \(T\) on every test in \(\mathcal D(U\times V)\). Identify the extra ingredient beyond density that makes this implication valid. **Exercise 5 (intermediate: why a logarithmic rate occurs).** Fix a nonzero \(h\in C_c^\infty(W)\) and set \(F_\varepsilon=\varepsilon\log(\varepsilon)\,h\), \(0<\varepsilon\le1/2\). Verify every hypothesis of Lemma 4.2 with \(M=1\), find the limit, and prove that an \(O(\varepsilon)\) bound for its value on every fixed test is false. **Exercise 6 (intermediate: a differentiated point evaluation).** On the line, let \(b(\psi,\phi)=\psi'(a)\phi(c)\). Compute \(K_\varepsilon\) from (R6) for arbitrary compact smooth unit-integral profiles, which need not be even. Determine the limiting kernel, its exact order, and the bound \(\mu+1\) obtained from Theorem 4.3. Verify the sign by pairing with a general two-variable test. **Exercise 7 (advanced: the second-moment error).** Suppose \(\rho,\sigma\) in Theorem 4.3 are nonnegative and even. Let the resulting kernel \(K\) have order at most \(q\) on a compact neighborhood of a test support \(L\). Prove, for sufficiently small \(\varepsilon\), \[ |(K_\varepsilon-K)(\Phi)|\le C_L\varepsilon^2p_{L,q+2}(\Phi), \qquad \Phi\in\mathcal D_L. \tag{R12} \] Identify the needed first-moment cancellations, and explain why this estimate does not make the logarithm in Lemma 4.2 unnecessary under that lemma's weaker assumptions. ## Solutions **Solution 1.** Set \(y=x+3t\), so \(dt=dy/3\). A kernel is \[ k(x,y)=\frac13\rho\big((y-x)/3\big) \mathbf1_{[-2,1]}\big((y-x)/3\big). \] Values at the two endpoints do not matter for this locally integrable function. Let \(S\) be the support, on the real line, of \(t\mapsto\rho(t)\mathbf1_{[-2,1]}(t)\), as a distribution. Then \(\operatorname{supp}k=\{(x,y):(y-x)/3\in S\}\). Indeed change of variables \((x,t)\mapsto(x,x+3t)\) is a diffeomorphism, and the resulting product distribution has support \(\mathbb R\times S\). To check this last assertion, a test supported outside that product has zero pairing by Fubini and localization. At a point \((x_0,t_0)\) with \(t_0\in S\), every small \(t\)-neighborhood contains a test detected by the one-variable distribution, by the support definition. Multiply it by an \(x\)-test supported near \(x_0\) with nonzero integral; the product pairing is nonzero. This also handles an identically zero \(\rho\) on the integration interval, when the support is empty. **Solution 2.** Distributional differentiation in the output gives \[ \begin{aligned} \langle\partial_{x_j}T\phi,\psi\rangle &=-K((\partial_{x_j}\psi)\otimes\phi)\\ &=(\partial_{x_j}K)(\psi\otimes\phi). \end{aligned} \] Differentiation of the input gives \[ \begin{aligned} \langle T\partial_{y_k}\phi,\psi\rangle &=K(\psi\otimes\partial_{y_k}\phi)\\ &=(-\partial_{y_k}K)(\psi\otimes\phi). \end{aligned} \] Thus the two kernels are \(\partial_{x_j}K\) and \(-\partial_{y_k}K\). The transpose kernel on \(V\times U\) is \(K^t(G)=K(G(y,x))\). Separate continuity makes this a continuous transpose operator, and uniqueness proves \((T^t)^t=T\). For \(T\phi(x)=\phi(2x)\), the kernel is \(K(H)=\int H(x,2x)\,dx\). Direct differentiation gives \(\partial_xT\phi=2T\partial_y\phi\). Accordingly \(\partial_xK=-2\partial_yK\), as can also be checked by integrating the total derivative of \(x\mapsto H(x,2x)\). **Solution 3.** If \(|a_j|\le C\langle j\rangle^b\), choose an integer \(M>b+1\). Since \(\langle(j,j+1)\rangle\) is comparable to \(\langle j\rangle\), \[ \begin{aligned} \sum_j|a_jH(j,j+1)| &\le C'P_M(H)\sum_j\langle j\rangle^{b-M}\\ &<\infty . \end{aligned} \] This proves tempered continuity. The operator is \[ T\phi=\sum_j a_j\phi(j+1)\delta_j. \] The displayed lattice is closed and discrete. The kernel vanishes on its complement. A test supported in a sufficiently small neighborhood of a single lattice point pairs to \(a_j\) times its value there; if \(a_j\ne0\), an appropriate test detects that point. Thus the support is exactly the stated set. **Solution 4.** Work first in a product rectangle with the cutoffs of Section 3. The uniform bilinear estimate bounds every kernel in the family by the same finite derivative seminorm there: choose \(2N>r+s+m+n\) in (4.3). The limiting operator obeys the same bilinear estimate by passage to the limit on products, so its kernel has that bound too. Approximate a fixed \(H\) by a finite sum \(H_0\) of product tests in the \(C^{2N}\) seminorm, using (3.1). Then \[ \begin{aligned} |(K_\ell-K)(H)| &\le |(K_\ell-K)(H_0)|\\ &\quad+2C'\|H-H_0\|_{C^{2N}}. \end{aligned} \] For fixed \(H_0\), the first term tends to zero by the hypothesis. The second can be made arbitrarily small uniformly in \(\ell\). A finite product-rectangle partition handles a general compactly supported \(H\). The necessary extra ingredient is uniform continuity in a common seminorm. Density alone supplies approximations but gives no uniform bound on their errors under a varying family of functionals. **Solution 5.** The scalar coefficient and its first two derivatives are \[ \varepsilon\log\varepsilon,\qquad1+\log\varepsilon,\qquad\varepsilon^{-1}. \] For \(0<\varepsilon\le1/2\), all three moduli are bounded by a fixed multiple of \(\varepsilon^{-1}\). For example, \(\log(1/\varepsilon)\le1/\varepsilon\), by integrating \(t^{-1}\le1\) from \(1\) to \(1/\varepsilon\). On each fixed support \(L\), \(|\int h\Phi|\le\|h\|_{L^1(L)}p_{L,0}(\Phi)\le\|h\|_{L^1(L)}p_{L,j}(\Phi)\). These are all the bounds (R1) for \(j=0,1,2\). The limit is zero. Choose \(\Phi=\overline h\), a compact smooth test. Then \[ \frac{|F_\varepsilon(\Phi)|}{\varepsilon} =\log(1/\varepsilon)\int|h|^2\longrightarrow\infty. \] The integral is positive: a nonzero continuous function has modulus bounded below on a small ball. Hence even for this one fixed test an \(O(\varepsilon)\) estimate fails. This proves necessity of allowing a logarithm in the abstract scale lemma; it makes no optimality claim for every family of mollified kernels. **Solution 6.** Differentiating the first profile in its test variable contributes a minus sign: \[ K_\varepsilon(x,y) =-\rho_\varepsilon'(x-a)\sigma_\varepsilon(y-c). \tag{R13} \] For a general compact smooth \(\Phi\), integration by parts in \(x\) followed by \(x=a+\varepsilon t,\ y=c+\varepsilon w\) gives \[ \langle K_\varepsilon,\Phi\rangle =\int\rho(t)\sigma(w) \partial_x\Phi(a+\varepsilon t,c+\varepsilon w)\,dt\,dw \longrightarrow\partial_x\Phi(a,c). \tag{R14} \] Uniform convergence of the smooth derivative on the fixed compact set of \((t,w)\) proves the limit, even for signed or complex profiles. Thus \(K=-\delta_a'\otimes\delta_c\). It has order at most one. To exclude order zero, take \(\Phi_\lambda(x,y)=\lambda\,\eta((x-a)/\lambda)\theta(y-c)\), where \(\eta,\theta\) are compact smooth, \(\eta'(0)=1\) and \(\theta(0)=1\). For \(0<\lambda\le1\), these tests have a common compact support, their supremum norms tend to zero, but \(K(\Phi_\lambda)=1\). Hence the exact order is one. In (R5), \(r=1,s=0,m=n=1\), so \(\mu=3\) and the constructed bound is four. The difference between this bound and the least order is explicit. **Solution 7.** The global kernel identity, applied to the product profiles in (R6), gives \(K_\varepsilon(z)=K(\rho_\varepsilon(x-\cdot)\sigma_\varepsilon(y-\cdot))\). Passing a compact parameter integral through this finite-order pairing is justified by Riemann-sum convergence in its controlling derivative seminorm. Consequently \[ \langle K_\varepsilon,\Phi\rangle =K(A_\varepsilon\Phi),\qquad A_\varepsilon\Phi(u,v) =\int\rho(t)\sigma(w)\Phi(u+\varepsilon t,v+\varepsilon w)\,dt\,dw . \tag{R15} \] All these averaged tests have support in one compact neighborhood \(J\Subset U\times V\) of \(L\), for sufficiently small scales. Use an order-\(q\) estimate for \(K\) on that neighborhood. For any smooth \(G\) and displacement \(h\), two applications of the scalar fundamental theorem give \[ G(z+h)-G(z)-DG(z)h =\int_0^1(1-\tau)\sum_{i,j}h_i h_j \partial_i\partial_jG(z+\tau h)\,d\tau . \tag{R16} \] Apply this with \(G=\partial^\beta\Phi\), \(|\beta|\le q\), and \(h=\varepsilon(t,w)\). Evenness makes \(\int t_i\rho(t)dt=0\) and \(\int w_j\sigma(w)dw=0\), by reflection in all coordinates. The product profile therefore has every first moment zero. Its integral is one, so the constant terms cancel the subtraction of \(\Phi\). Since the profiles have compact support, the integral of \(|(t,w)|_1^2\rho(t)\sigma(w)\) is finite. Formula (R16) yields \[ p_{J,q}(A_\varepsilon\Phi-\Phi) \le\frac{\varepsilon^2}{2} \left(\int |(t,w)|_1^2\rho(t)\sigma(w)\,dt\,dw\right) p_{L,q+2}(\Phi). \] Here derivatives of the globally defined test vanish off its support, so their global suprema equal their suprema on \(L\); all relevant segments stay inside the domain after the neighborhood is fixed. Apply the order-\(q\) bound of \(K\) to prove (R12). The abstract lemma assumes neither the averaging representation (R15) nor zero first moments, and controls only \(M+1\) test derivatives. This estimate uses two further derivatives beyond the chosen kernel order. Solution 5 therefore presents no contradiction. ## References - The supplied [scalar calculus](../prerequisites/U011-free-foundations/metric-foundation-bridges.md), [finite algebra](../prerequisites/U011-free-foundations/stable-prerequisite-bridges.md), [measure](../prerequisites/U011-free-foundations/banach-foundation-bridges.md) and [functional](../prerequisites/U011-free-foundations/functional-foundations-U008.md) proof components provide the exact earlier programme arguments listed at the start. They retain their stated licences, including CC0 1.0 for the functional selection. The compact-layer partition proof is also supplied in the programme's boundary-flux lesson and is reproduced here as a complete CC0 proof, independent of that lesson's later results. - [Dyatlov 2026] Semyon Dyatlov, *Lecture notes for 18.155: distributions, elliptic regularity, and applications to PDEs*, MIT, 2 October 2026, §7.2.1, Theorem 7.6, pp. 81–82. [Open notes](https://math.mit.edu/~dyatlov/18.155/155-notes.pdf). Fourier-series construction of the local kernel. - [Melrose 2016] Richard Melrose, *18.155 Lecture 15*, MIT, 1 November 2016, Theorem 1 and the local reduction, pp. 2–4. [Open lecture](https://math.mit.edu/~rbm/18.155-F16/L15.pdf). Tempered kernels and weak continuity. - [Schwartz 1952] Laurent Schwartz, *Théorie des noyaux*, Proceedings of the International Congress of Mathematicians, Cambridge, Massachusetts, 1950, volume I, American Mathematical Society, 1952, pp. 220–230, especially §§2–4 and Theorems I–II, pp. 221–223. [Freely accessible proceedings, volume I](https://www.mathunion.org/fileadmin/ICM/Proceedings/ICM1950.1/ICM1950.1.ocr.pdf). These give the original operator, bilinear and weak/strong continuity formulations; the programme proofs are supplied above. - Lars Hörmander, *The Analysis of Linear Partial Differential Operators I: Distribution Theory and Fourier Analysis*, second edition (1990), 2003 reprint, ISBN 978-3-642-61497-2, §§5.1–5.2, pp. 126–132. Tensor products, regularized bilinear kernels, the kernel support relation and continuous graph operators.