When a kernel is smooth
Written by GPT-6.1 Sol (OpenAI), Ultra reasoning effort, September 2026. Checked once by GPT-6.1 Sol (OpenAI), Ultra reasoning effort. Public domain (CC0).
Source/proof self-check and prerequisite integration by GPT-6 Astra (OpenAI), Ultra, October 2026. Historical authorship and component terms are retained.
Smoothness of an operator's output on smooth inputs is a weak test. Differentiation passes that test. A more revealing question is what happens to point masses and their derivatives. If an operator sends every compactly supported distribution to a smooth function continuously, its entire kernel is smooth. This lesson proves both directions and explains the continuity behind the assertion.
The preceding kernel theorem lesson supplies the local kernel correspondence, test-space topology, support localization, partitions and periodic product expansions. The supplied scalar calculus foundation, §§12–13, gives compactness, uniform continuity, the fundamental theorem, derivative rules and smooth cutoffs. The functional foundation, §§6, 14.1–14.2 and 19, proves Baire's theorem and completeness for the exact smooth seminorms. The measure foundation, §§15.0–15.1, provides integration and Fubini. The compactness and Taylor estimates actually needed below are proved here. We use the bilinear distribution pairing and Euclidean Lebesgue measure. Basic source comparisons are [Dyatlov 2026], [Melrose 2016] and [Schwartz 1952].
Compact distributions test all smooth functions
For an open , put , with seminorms
An exhaustion by compact sets and the increasing derivative orders make this a Fréchet space. Its continuous dual is . We give this dual its strong topology, whose seminorms are
for bounded . Boundedness means uniform bounds for each smooth seminorm, on each compact set; it does not require a common support for the functions in .
Every has an estimate
for some compact and integer . Its restriction to test functions is therefore a distribution supported in . Conversely, a compactly supported distribution acts on arbitrary smooth functions by , where near its support and is compactly supported. This is independent of , since the difference vanishes near the support. A distribution estimate for gives (1.1), on a compact neighborhood of the support. Thus is precisely the space of compactly supported distributions.
The weak convergence used in this lesson is
Testing against all smooth functions, rather than only compactly supported ones, matters. A point mass escaping to infinity converges to zero against every compact test, but need not converge against a smooth function growing at infinity.
Taylor estimates on compact neighborhoods
The exact integral remainder follows from the fundamental theorem already supplied. For on a neighborhood of , , repeated integration by parts gives
Indeed is the fundamental theorem. Integrating the displayed remainder once by parts, with a primitive of equal to , expresses it as the order- remainder minus . This proves the identity by induction. The formulas apply componentwise to complex-valued functions.
Apply this identity to on a segment contained in the domain. The finite-coordinate chain rule, iterated times, gives
The coefficient counts the ordered coordinate choices with the given multiplicities; the multinomial count follows by choosing their positions successively. This proves the full multivariable Taylor identity and its finite derivative remainder on the actual segment. In particular,
The second remainder is bounded by times the supremum of that second derivative on the segment, including negative . Apply the same formulas to every to obtain the corresponding smooth-seminorm estimates. A compact neighborhood lying inside the domain contains all sufficiently short segments from a given compact set, so the derivative suprema are uniform there. These are the Taylor estimates used for regularization and parameter differentiation below.
Lemma 2.1 (weak convergence is strong convergence for sequences here). If as in (1.2), then for every bounded . Furthermore there exist , independent of , such that
The same uniform estimate holds for any pointwise bounded family of continuous functionals on .
Proof. Let . The sets
are closed and cover , because each scalar sequence converges. Baire's theorem gives one set with nonempty interior. Subtracting two elements in a small neighborhood of an interior point gives on a neighborhood of zero. That neighborhood contains a ball for one of the increasing smooth seminorms. Scaling, including the zero-seminorm case as in Lemma 2.1 of the preceding lesson, proves (2.1). The proof uses only pointwise boundedness, so it also proves the final assertion for an arbitrary family.
For bounded , here is a direct proof that it is totally bounded in . Cover by finitely many closed coordinate boxes whose slightly larger boxes lie in . Bounds through order on this finite compact union give a common constant such that every derivative through order changes by at most within any one box, using the segment formula and the finite-coordinate norm estimate. Their values are also uniformly bounded.
Given , put a finite grid of mesh at most in each box, meaning every point is within that distance of a grid point. Such a grid is obtained by subdividing each coordinate interval sufficiently finely. Divide the bounded real and imaginary parts of all derivative values at all these finitely many grid points into finitely many bins of diameter at most in . There are only finitely many combined bin patterns. Choose one representative from each nonempty pattern attained by a member of . If and its representative have the same pattern, their derivative values at a grid point differ by at most . At any point of a box, the two moves to that grid point add at most . Thus of their difference is less than . This is a finite net with centers in , including the empty-family case. No compactness theorem for function spaces is left as a prerequisite.
Choose with every within of some in . Then
The maximum tends to zero. Then let tend to zero. This proves strong convergence.
The conclusion is about sequences in this particular dual. It does not identify the weak and strong topologies on all subsets or all nets.
Lemma 2.2 (smooth tests approximate compact distributions strongly). For each , there exist , all supported in one compact subset of , such that in the strong topology of .
Proof. Extend the compact distribution to using a cutoff supported in , and let have integral one. Put , and regularize by
For fixed positive , differentiation in of this test family is valid in every finite derivative seminorm in : the segment formulas bound each difference-quotient remainder on the controlling compact. Pairing that remainder with proves smoothness of , with derivatives obtained by differentiating . Its support lies in , inside one compact subset of for small .
Let be bounded and use (1.1) for . Testing the regularization against gives
The displayed function is needed only near . Choose the compact controlling set in (1.1) inside a neighborhood on which the cutoff defining the extension is one, and a slightly larger compact containing all relevant short segments. The first segment formula bounds its derivatives through order by
The last supremum is finite. The distribution estimate proves , so convergence is strong. The pairing identity follows by approximating the integral by Riemann sums in the finite derivative seminorm used by . Thus no unproved interchange with an arbitrary functional is needed.
Point masses vary smoothly
For , let . Spatial differentiation of a distribution and differentiation of its parameter have opposite signs:
where denotes the distribution's test variable. Indeed the left side applied to is , while .
Lemma 3.1. The map has continuous derivatives of every order into the strong dual . They are
Proof. On a compact parameter neighborhood , Taylor's formula gives, for each bounded ,
whenever the segment stays in . This proves the first derivative in the strong topology. The same estimate applied to derivatives of proves (3.2) inductively. Continuity of each derivative follows from the mean-value bound with one further derivative, uniformly on .
The smoothing equivalence
Let be weakly continuous, and let be its kernel from the preceding lesson.
Theorem 4.1 (smooth kernels and smoothing operators). The following statements are equivalent:
- is represented by a function .
- extends to a continuous linear map , where the input has its strong dual topology.
- extends to a linear map carrying every sequence converging as in (1.2) to a sequence converging in every smooth seminorm on compact subsets of .
The extension is unique in either class, and is
Neither nor its derivatives need be bounded on the whole product. There is no proper-support assumption in this theorem.
Proof that a smooth kernel gives a continuous extension. For a compact distribution , (1.1) and Taylor's formula for the smooth function show that (4.1) is smooth and
For example, the first difference quotient in converges in the seminorm in on the compact set controlling ; its pairing therefore converges. The same reasoning works at every order.
Fix and an integer . The family
is bounded in . For every and every , all mixed derivatives in question are bounded on , by compactness. Formula (4.2) gives
This proves strong-dual continuity. For a smooth compactly supported input, (4.1) is the integral against , so it agrees with by kernel uniqueness. Lemma 2.1 shows that a weakly convergent input sequence is strongly convergent, proving statement 3 too.
Proof that a continuous extension has a smooth kernel. First assume statement 2 and set
Lemma 3.1 and continuity of imply that has continuous derivatives of every order with values in . Hence
These derivatives are jointly continuous. To check this, let in one compact product neighborhood. The corresponding functions of on the right of (4.5) converge uniformly on that neighborhood as ; evaluation at , followed by continuity of the limiting function, gives convergence of their values. Thus .
If only statement 3 is assumed, the same argument still works. Every parameter difference quotient in Lemma 3.1 converges weakly as well as strongly. Applying gives convergence in . The parameter derivatives are continuous because a convergent sequence of parameters gives weak convergence of the corresponding derivatives of point masses. Sequential continuity suffices to establish continuity for this finite-dimensional parameter space. Induction gives (4.5) and the same joint smoothness.
It remains to identify this smooth function with the kernel of . For , approximate the integral by Riemann sums in a finite collection of rectangles covering its support. These sums converge to the regular distribution strongly in : on a bounded family of smooth functions the integrands have a uniform derivative bound on the fixed compact integration region, so the Riemann-sum errors tend to zero uniformly. They also converge weakly. Applying either type of therefore gives
The last Riemann sums converge in each smooth output seminorm by smoothness on compact products. Thus the function kernel gives the same test-input operator as ; kernel uniqueness shows they agree as distributions.
Finally Lemma 2.2 approximates every compact distribution by smooth compactly supported inputs. Continuity, or the sequential property in statement 3, proves that the extension must be (4.1) and is unique. Since the smooth-kernel construction already satisfies (4.3), statement 3 also implies statement 2.
This proof explains why output smoothness on test inputs alone cannot establish the theorem: (4.4) needs the operator to act on point masses, and (4.5) needs controlled limits of their difference quotients.
A smooth kernel can have no global size bound
Consider
This function and its mixed derivatives are bounded on every compact product but have no useful uniform bound on the whole plane. The theorem still supplies a continuous map . For instance,
and
The minus sign is the sign of spatial distributional differentiation. Its input can have any fixed compact support; the output estimate uses a bounded family in , not a bound uniform over the whole plane.
The theorem's domain is compactly supported distributions. Extensions to arbitrary distributions require additional input-support control. Such extensions are outside the theorem proved here. Smoothness and support control answer different questions.
Exercises
Exercise 1 (basic: a dipole under Gaussian smoothing). For , compute . State a formula for , with the sign convention made explicit.
Exercise 2 (intermediate: smoothing only to finite order). Fix an integer . Let
Show that is continuous, but its kernel is not smooth. Explain the role of all derivative orders in Theorem 4.1.
Exercise 3 (intermediate: smooth outputs on test inputs). Show that the identity is continuous, but has no extension of either type in Theorem 4.1. Prove this directly using an approximate identity converging to .
Exercise 4 (advanced: extracting a kernel from moving point masses). Suppose is continuous as in statement 2 of Theorem 4.1. Derive a compact-set estimate for
that tends to zero uniformly in on a fixed compact set, in the one-dimensional input case. Explain why differentiating is legitimate, while applying an arbitrary linear map to a difference quotient would not be.
Exercise 5 (advanced: finite-rank approximation of a smoothing operator). Let converge to with all derivatives on compact products. Show that the corresponding maps converge to , uniformly on each bounded subset of , in every smooth output seminorm. Explain how locally cut-off Fourier expansions give a sequence of finite sums with this convergence.
Solutions
Solution 1. Since ,
Generally the spatial derivative of the input point mass satisfies
Here denotes distributional differentiation in its test variable; parameter differentiation in would have the opposite sign at odd orders.
Solution 2. On either side of zero, a derivative of order is a constant times , with a possible sign factor. It tends to zero at zero. For , the quotient defining the -th derivative at zero is bounded by a constant times , once the preceding derivative is assigned value zero there. It tends to zero. Induction therefore both proves existence of these derivatives at zero and identifies their continuous extensions, so . The derivative of order on a half-line is a nonzero constant times , which cannot extend continuously. For a compact ,
proving continuity. The kernel is the function , which is not near . The smooth conclusion of Theorem 4.1 requires output estimates and parameter derivatives at every order, rather than at one prescribed finite order.
Solution 3. On every fixed test-support space, the inclusion into is continuous, so the inductive-limit property gives continuity of . Choose a nonnegative bump of integral one with . Its rescalings converge strongly, and therefore weakly, to in , by Lemma 2.2. An extension would make converge in . But their values at zero are , which diverge. Even uniform convergence near zero is impossible. The identity kernel is the distribution on the diagonal, rather than a smooth function.
Solution 4. Continuity of for the output seminorm gives finitely many bounded test families and constants controlling it. Combine those families, with the constants absorbed by scaling, into one bounded ; then . For in a fixed compact interval , Lemma 3.1 gives
The right side is finite and tends to zero. Replacing the output seminorm by proves the same result with every output derivative. The operation is legitimate because continuity of transfers a proved strong-dual remainder estimate to a smooth-output remainder estimate. Algebraic linearity alone provides no such transfer of limits.
Solution 5. A bounded subset is pointwise bounded, since each singleton is bounded in . The last assertion of Lemma 2.1 gives one such that
Consequently
which tends to zero.
For the finite-sum construction, exhaust and by compact sets with interiors, and choose compact cutoffs equal to one on the -th sets. Split their product times into finitely many pieces supported in product rectangles, using product partitions. Expand each piece periodically on a larger product rectangle and multiply the modes by separate cutoffs, as in Section 3 of the preceding lesson. Finite Fourier sums are finite sums of functions of times functions of . Choose their truncation sufficiently large to approximate derivatives through order on the -th compact product within . Summing the finitely many pieces gives . On every fixed compact product, the outer cutoffs eventually equal one and the derivative errors tend to zero. Thus in on compact products. Each corresponding operator has finite-dimensional range, spanned by its finitely many 's. The estimate above proves the claimed operator convergence.
References
- Distributions as kernels of continuous operators, especially the topology, support and partition arguments, Lemma 2.1, the local kernel theorem and Section 3's product expansions. The supplied scalar, functional and measure proof components retain their stated licences; the functional selection retains CC0 1.0.
- [Dyatlov 2026] Semyon Dyatlov, Lecture notes for 18.155: distributions, elliptic regularity, and applications to PDEs, MIT, 2 October 2026, §7.2.2, Proposition 7.10, p. 83. Open notes. The forward smoothing implication is compared with the full argument above.
- [Melrose 2016] Richard Melrose, 18.155 Lecture 15, MIT, 1 November 2016, p. 1. Open lecture. This gives a related formulation in terms of an operator and its transpose.
- [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; §7, Theorem VIII, p. 228. Freely accessible proceedings, volume I. The smoothing-kernel characterization is stated there; this lesson supplies the complete proof.
- 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, §5.2, Theorem 5.2.6, p. 132. The exact copy supplies the comparison for the moving-point-mass argument. The lesson retains its full strong-dual and sequential formulations, explicit compactness and approximation proofs, quantitative estimates and five original worked problems.