Convolution, Approximate Identities, and Poisson Summation

Convolution averages translated copies of one function against another. It composes translation-invariant kernels, moves derivatives onto a smooth factor, and turns into ordinary multiplication after Fourier transformation. A uniformly $L^1$-bounded, unit-mass family whose absolute mass concentrates at the origin is an approximate identity: it recovers the input in a declared topology as the scale is removed. When its kernels are smooth, it also regularizes at every positive scale. Periodization performs the discrete–continuous step. With the conventions fixed below, the three central statements are

$$\begin{aligned} (f*g)(x) &=\int_{\mathbb R^n}\mathrm d^n y\,f(x-y)g(y),\\ \widetilde{f*g}(p) &=\widetilde f(p)\widetilde g(p),\\ f*\rho_\varepsilon &\longrightarrow f \quad\text{in }L^p,\qquad f\in L^p(\mathbb R^n),\quad 1\leq p<\infty, \end{aligned}$$

and, for $f\in\mathcal S(\mathbb R^n)$ and a full-rank lattice $\Lambda$,

$$\sum_{\ell\in\Lambda}f(x+\ell) =\frac1{V_\Lambda} \sum_{k\in\Lambda^*} e^{-ik\cdot x}\widetilde f(k).$$

Each line has different authorizing hypotheses. A convolution may exist only almost everywhere, smoothing comes from regularity of the kernel, approximate identity convergence is not generally uniform or pointwise everywhere, and the displayed Poisson formula is not being asserted for arbitrary $L^1$ or $L^2$ functions.

Required background. Lebesgue Integration and Convergence Theorems supplies the dominated, monotone, and absolute-convergence arguments used throughout. This page will still identify the step at which a limit or integral exchange is used.

Helpful background. Fourier Series, Fourier Transforms, and Plancherel Theory develops the transform rules and norm extension in full. It is useful but not required: the formulas needed for periodization are repeated here.

For targeted review, Product Measures, Fubini–Tonelli, and Change of Variables and $L^p$ Spaces, Inequalities, and Weak Convergence are targeted review links for double integrals and norm estimates, not additional prerequisites.

Convolution domains and a reliable workflow

The noncompact calculations use $x,p\in\mathbb R^n$, Lebesgue measure, and the site’s positive-forward Fourier convention

$$\begin{aligned} \widetilde f(p) &=\int_{\mathbb R^n}\mathrm d^n x\, e^{+ip\cdot x}f(x),\\ f(x) &=\int_{\mathbb R^n} \frac{\mathrm d^n p}{(2\pi)^n}\, e^{-ip\cdot x}\widetilde f(p). \end{aligned}$$

The first formula is literal for $L^1$ inputs; both are pointwise safe on the Schwartz class. Periodic formulas will state the lattice, cell volume, reciprocal lattice, and phase separately.

A dependable convolution or lattice calculation has six steps:

1. Specify the domain, measure, and boundary conditions.
2. Name the input class and the theorem that makes the convolution or sum exist.
3. Before exchanging integrals or sums, establish nonnegativity or absolute integrability.
4. State the topology of every limit: pointwise, uniform, $L^p$, or test-function convergence.
5. Carry the Fourier phase, reciprocal-lattice definition, $(2\pi)^n$, and cell volume together.
6. Check a normalized Gaussian or the zero Fourier mode by a second route.

The stop rule is important. Delta functions, delta combs, singular kernels, and conditionally convergent mode sums require tempered distributions and Fourier calculus or additional summation and regularization arguments. They are not ordinary $L^1$ functions merely because they appear beneath an integral sign.

Convolution composes translation-invariant kernels

For measurable complex-valued functions, define

$$\begin{aligned} (f*g)(x) &=\int_{\mathbb R^n}\mathrm d^n y\,f(x-y)g(y)\\ &=\int_{\mathbb R^n}\mathrm d^n y\,f(y)g(x-y), \end{aligned}$$

whenever the integral exists. The second form follows from a change of variables. If $f,g\in L^1(\mathbb R^n)$, Tonelli’s theorem gives

$$\begin{aligned} \|f*g\|_1 &\leq \int\mathrm d^n x\int\mathrm d^n y\, |f(x-y)|\,|g(y)|\\ &=\|f\|_1\|g\|_1. \end{aligned}$$

Thus $f*g$ is finite almost everywhere and belongs to $L^1$. The qualifier “almost” cannot be dropped. On $\mathbb R$, the functions

$$f(x)=g(x)=|x|^{-2/3}\mathbf 1_{\{|x|<1\}}$$

are integrable, but $(f*g)(0)$ contains $\int_{-1}^{1}|y|^{-4/3}\,\mathrm dy$ and diverges.

Young’s inequality and the endpoint used below

The full Young convolution inequality says that if $f\in L^p(\mathbb R^n)$, $g\in L^q(\mathbb R^n)$, $1\leq p,q,r\leq\infty$, and

$$1+\frac1r=\frac1p+\frac1q,$$

then

$$\|f*g\|_r\leq\|f\|_p\|g\|_q.$$

The definition, theorem, and useful endpoints are stated in Hunter 2014, Definition 1.22, Theorem 1.23, and Examples 1.24–1.26, p. 9, official UC Davis lecture-note PDF.

The case used for approximate identities has $g\in L^1$ and $f\in L^p$:

$$\|f*g\|_p\leq\|f\|_p\|g\|_1, \qquad 1\leq p\leq\infty.$$

For $p<\infty$, Minkowski’s integral inequality and translation invariance of Lebesgue measure give the proof directly:

$$\begin{aligned} \|f*g\|_p &\leq\int_{\mathbb R^n}\mathrm d^n y\, |g(y)|\,\|f(\,\cdot-y)\|_p\\ &=\|f\|_p\|g\|_1. \end{aligned}$$

For $p=\infty$, take the essential supremum after applying the pointwise integral bound.

If $f,g,h\in L^1$, absolute integrability also licenses the changes of variables and regrouping that give, almost everywhere,

$$f*g=g*f, \qquad (f*g)*h=f*(g*h).$$

Consequently the convolution operators $C_fh=f*h$ compose by $C_fC_g=C_{f*g}$. This is the precise sense in which convolution composes translation-invariant kernels. Outside an integrable setting, formal associativity is not permission to rearrange divergent integrals.

Smoothing comes from the smooth factor

Let $\rho\in C_c^\infty(\mathbb R^n)$ and $f\in L^1_{\mathrm{loc}}(\mathbb R^n)$. Compact support makes

$$(f*\rho)(x) =\int\mathrm d^n y\,f(y)\rho(x-y)$$

finite for every $x$, and dominated convergence moves derivatives onto the kernel:

$$\partial^\alpha(f*\rho) =f*(\partial^\alpha\rho).$$

Hence $f*\rho$ is smooth even when $f$ is not. Convolution alone does not have this effect. For example, $\mathbf 1_{[0,1]}*\mathbf 1_{[0,1]}$ is a continuous triangular function but is not differentiable at all of its corners.

For Schwartz functions, Fubini also gives the convolution theorem in the positive-forward convention:

$$\begin{aligned} \widetilde{f*g}(p) &=\int\mathrm d^n x\,e^{ip\cdot x} \int\mathrm d^n y\,f(y)g(x-y)\\ &=\int\mathrm d^n y\,e^{ip\cdot y}f(y) \int\mathrm d^n z\,e^{ip\cdot z}g(z)\\ &=\widetilde f(p)\widetilde g(p). \end{aligned}$$

By contrast, pointwise multiplication obeys

$$\widetilde{fg}(p) =\int\frac{\mathrm d^n q}{(2\pi)^n}\, \widetilde f(q)\widetilde g(p-q).$$

The absence of a $2\pi$ factor in the first identity and its presence in the second are a normalization check, not a cosmetic choice. The two Fourier rules are treated in Dyatlov 2022, Propositions 11.10 and 11.18, pp. 122–125, official MIT lecture-note PDF; his negative-forward phase is translated here by $p\mapsto-p$.

Approximate identities concentrate and recover

A family $(\rho_\varepsilon)_{\varepsilon>0}\subset L^1(\mathbb R^n)$ is an approximate identity if

$$\int\mathrm d^n x\,\rho_\varepsilon(x)=1, \qquad \sup_{\varepsilon>0}\|\rho_\varepsilon\|_1<\infty,$$

and, for every $\delta>0$,

$$\lim_{\varepsilon\downarrow0} \int_{|x|>\delta}\mathrm d^n x\, |\rho_\varepsilon(x)|=0.$$

The last condition says that the mass concentrates near the origin. Unit integral by itself is not enough.

A standard mollifier starts with a nonnegative $\rho\in C_c^\infty(\mathbb R^n)$ supported in the unit ball and normalized by $\int\rho=1$. Define

$$\rho_\varepsilon(x) =\varepsilon^{-n}\rho(x/\varepsilon).$$

Then $\rho_\varepsilon$ is supported in the ball of radius $\varepsilon$, has integral and $L^1$ norm equal to one, and forms an approximate identity.

The norm-convergence proof

Let $1\leq p<\infty$ and define the translation $\tau_yf(x)=f(x-y)$. Translations are continuous in $L^p$:

$$\|\tau_yf-f\|_p\longrightarrow0 \qquad\text{as }y\to0.$$

One proof first verifies this for compactly supported continuous functions by uniform continuity, then approximates a general $L^p$ function by such a function. The same density argument fails in $L^\infty$, where continuous functions are not norm-dense.

Normalization gives

$$(f*\rho_\varepsilon)(x)-f(x) =\int\mathrm d^n y\,\rho_\varepsilon(y) \bigl[f(x-y)-f(x)\bigr].$$

Minkowski therefore yields

$$\|f*\rho_\varepsilon-f\|_p \leq \int\mathrm d^n y\,|\rho_\varepsilon(y)| \|\tau_yf-f\|_p.$$

Choose $\delta$ so that the translation norm is small for $|y|<\delta$. On that ball, the uniform $L^1$ bound controls the integral. Outside it, use $\|\tau_yf-f\|_p\leq2\|f\|_p$ and the concentration condition. Both pieces vanish, proving

$$\lim_{\varepsilon\downarrow0} \|f*\rho_\varepsilon-f\|_p=0, \qquad 1\leq p<\infty.$$

The standard mollifier construction, smoothness, convergence at Lebesgue points, and local $L^p$ convergence are developed in Hunter 2014, § 1.9 and Theorem 1.28, pp. 11–12, official UC Davis lecture-note PDF.

Convergence supplied by common approximate-identity hypotheses

Input and kernel	Guaranteed convergence
f in Lᵖ, 1 ≤ p < ∞; bounded approximate identity	Lᵖ-norm convergence
Bounded uniformly continuous f; nonnegative normalized concentrating kernel	Uniform convergence
f locally integrable; standard nonnegative radial mollifier	Pointwise convergence at every Lebesgue point, hence almost everywhere
General f in L∞	No general L∞-norm convergence

For a domain $\Omega\subset\mathbb R^n$, ordinary convolution near the boundary also needs a choice: restrict to points farther than $\varepsilon$ from $\partial\Omega$, extend $f$ outside $\Omega$, or use a boundary-adapted kernel. Silent zero extension can create an artificial boundary layer.

Pointwise and uniform failure at a jump

Let $f=\mathbf 1_{[0,\infty)}$ on $\mathbb R$ and take an even mollifier. At the jump,

$$(f*\rho_\varepsilon)(0) =\int_{-\infty}^{0}\rho_\varepsilon(y)\,\mathrm dy =\frac12$$

for every $\varepsilon$. The value of the chosen representative $f(0)$ could instead be $0$ or $1$, and changing that one value changes no $L^p$ class. Moreover, continuous mollifications cannot converge uniformly to the discontinuous step. Thus $L^p$ convergence for finite $p$, almost-everywhere convergence, and uniform convergence are genuinely different statements.

The Gaussian gives a semigroup and a second check

The heat kernel

$$H_t(x) =\frac1{(4\pi t)^{n/2}} e^{-|x|^2/(4t)}, \qquad t>0,$$

is a noncompactly supported approximate identity. Direct Gaussian integration and the transform convention above give

$$\int\mathrm d^n x\,H_t(x)=1, \qquad \widetilde H_t(p)=e^{-t|p|^2}.$$

The multiplier immediately checks the composition law:

$$\widetilde{H_s*H_t}(p) =e^{-s|p|^2}e^{-t|p|^2} =\widetilde H_{s+t}(p),$$

so Fourier inversion gives

$$H_s*H_t=H_{s+t}.$$

Completing the square in the convolution integral gives the same result without Fourier transformation. The normalization is dimensionally consistent: $[t]=\text{length}^2$, so $H_t$ has dimension $\text{length}^{-n}$.

For $f\in L^2(\mathbb R^n)$, the recommended Fourier page’s Plancherel theorem supplies an independent convergence proof:

$$\begin{aligned} \|H_t*f-f\|_2^2 &=\int\frac{\mathrm d^n p}{(2\pi)^n} \left|e^{-t|p|^2}-1\right|^2 |\widetilde f(p)|^2\\ &\longrightarrow0 \qquad (t\downarrow0), \end{aligned}$$

by dominated convergence. The heat-kernel formula, its unit mass, smoothing, and $L^p$ recovery are given in Hunter 2014, §§ 5.1.1–5.1.2, pp. 130–131, official UC Davis lecture-note PDF.

For every bounded continuous function $a$,

$$\int\mathrm d^n x\,H_t(x)a(x) \longrightarrow a(0).$$

This is the safe meaning of calling $(H_t)$ a delta sequence on this page. It does not say that $H_t$ converges pointwise or in $L^1$ to an ordinary function called $\delta$.

Periodization produces reciprocal-lattice coefficients

Let $A$ be an invertible real $n\times n$ matrix and define the full-rank lattice, a fundamental cell, its volume, and the reciprocal lattice by

$$\begin{aligned} \Lambda&=A\mathbb Z^n, &Q&=A[0,1)^n,\\ V_\Lambda&=|\det A|, &\Lambda^*&=2\pi A^{-T}\mathbb Z^n. \end{aligned}$$

The defining property is $k\cdot\ell\in2\pi\mathbb Z$ for every $k\in\Lambda^*$ and $\ell\in\Lambda$. For $f\in\mathcal S(\mathbb R^n)$, the periodization

$$(P_\Lambda f)(x) =\sum_{\ell\in\Lambda}f(x+\ell)$$

and all of its differentiated sums converge absolutely and locally uniformly. It is a smooth $\Lambda$-periodic function.

For $k\in\Lambda^*$, unfold the cell translates:

$$\begin{aligned} \int_Q\mathrm d^n x\, e^{+ik\cdot x}(P_\Lambda f)(x) &=\sum_{\ell\in\Lambda} \int_Q\mathrm d^n x\, e^{+ik\cdot x}f(x+\ell)\\ &=\sum_{\ell\in\Lambda} \int_{Q+\ell}\mathrm d^n y\, e^{+ik\cdot y}f(y)\\ &=\widetilde f(k). \end{aligned}$$

Absolute convergence authorizes both the sum–integral exchange and the unfolding. Fourier-series reconstruction on the cell now gives the shifted Poisson summation formula

$$\boxed{ \sum_{\ell\in\Lambda}f(x+\ell) =\frac1{V_\Lambda} \sum_{k\in\Lambda^*} e^{-ik\cdot x}\widetilde f(k) }$$

with pointwise and locally uniform convergence. At $x=0$,

$$\sum_{\ell\in\Lambda}f(\ell) =\frac1{V_\Lambda} \sum_{k\in\Lambda^*}\widetilde f(k).$$

For a rectangular box,

$$\begin{aligned} \Lambda &=L_1\mathbb Z\times\cdots\times L_n\mathbb Z,\\ V_\Lambda&=\prod_{j=1}^nL_j,\\ k_r &=\left( \frac{2\pi r_1}{L_1},\ldots, \frac{2\pi r_n}{L_n} \right), \qquad r\in\mathbb Z^n. \end{aligned}$$

The factor $2\pi$ in $\Lambda^*$ and the factor $1/V_\Lambda$ in the Fourier series are inseparable.

Dyatlov states the lattice-comb identity and explains its relation to Fourier-series periodization in Dyatlov 2022, Theorem 11.32, pp. 133–134, official MIT lecture-note PDF. His transform has negative forward phase, so $\widetilde f(p)=\widehat f_D(-p)$. Because $\Lambda^*=-\Lambda^*$, the unshifted sum is unchanged; the sign in the shifted exponential is fixed by the reconstruction formula displayed here.

The Schwartz hypothesis is a safe theorem, not a claim of optimality. Weaker versions exist when the periodized sum and the Fourier series have sufficient convergence or when both sides are interpreted distributionally. Membership in $L^1$ alone does not guarantee pointwise sampling, absolute summability of the samples, or pointwise Fourier-series reconstruction.

Mode sums equal a continuum term plus images

Let $h\in\mathcal S(\mathbb R^n)$ be a momentum-space function and define its inverse transform

$$\check h(x) =\int\frac{\mathrm d^n p}{(2\pi)^n}\, e^{-ip\cdot x}h(p).$$

Apply Poisson summation to $\check h$. At $x=0$ this gives the exact identity

$$\frac1{V_\Lambda} \sum_{k\in\Lambda^*}h(k) =\sum_{\ell\in\Lambda}\check h(\ell).$$

Separating the zero image produces

$$\begin{aligned} \frac1{V_\Lambda} \sum_{k\in\Lambda^*}h(k) &=\int\frac{\mathrm d^n p}{(2\pi)^n}\,h(p)\\ &\quad+ \sum_{\substack{\ell\in\Lambda\\\ell\ne0}} \check h(\ell). \end{aligned}$$

The first term is the continuum momentum integral. The nonzero images are the exact finite-volume correction under these hypotheses. This is stronger than the heuristic replacement $V_\Lambda^{-1}\sum_k\to \int\mathrm d^n p/(2\pi)^n$: it displays the remainder rather than discarding it.

The shifted form is

$$\frac1{V_\Lambda} \sum_{k\in\Lambda^*} h(k)e^{-ik\cdot x} =\sum_{\ell\in\Lambda}\check h(x+\ell).$$

Generic loop integrands are not Schwartz, and their sum and integral may both diverge. In that situation the displayed identity is a guide to a regulated calculation, not permission to subtract infinities term by term.

A Gaussian round trip

Set $h_t(p)=e^{-t|p|^2}$. Its inverse transform is $H_t$, so the image formula becomes

$$\frac1{V_\Lambda} \sum_{k\in\Lambda^*}e^{-t|k|^2} =\frac1{(4\pi t)^{n/2}} \sum_{\ell\in\Lambda} e^{-|\ell|^2/(4t)}.$$

For the one-dimensional lattice $L\mathbb Z$, the shifted identity is

$$\begin{aligned} &\sum_{m\in\mathbb Z} \frac1{\sqrt{4\pi t}} \exp\!\left[-\frac{(x+mL)^2}{4t}\right]\\ &\qquad= \frac1L\sum_{r\in\mathbb Z} \exp\!\left[-t\left(\frac{2\pi r}{L}\right)^2\right] e^{-2\pi i r x/L}. \end{aligned}$$

Integrating either side over $0\leq x<L$ gives $1$. On the image side, the cells tile $\mathbb R$ and the Gaussian has unit mass. On the mode side, every nonzero Fourier mode integrates to zero. This is an independent check of the phase, $1/L$, Gaussian exponent, and normalization.

Finite-volume mode sums and regulated delta sequences

The periodized heat kernel

$$\begin{aligned} H_t^\Lambda(x) &=\sum_{\ell\in\Lambda}H_t(x+\ell)\\ &=\frac1{V_\Lambda} \sum_{k\in\Lambda^*} e^{-t|k|^2}e^{-ik\cdot x} \end{aligned}$$

is an ordinary smooth function for every $t>0$. The image representation shows that it is nonnegative and $\Lambda$-periodic; the zero Fourier mode shows that

$$\int_Q\mathrm d^n x\,H_t^\Lambda(x)=1.$$

Define periodic convolution by

$$(F*_\Lambda G)(x) =\int_Q\mathrm d^n y\,F(x-y)G(y).$$

The Fourier multiplier of $H_t^\Lambda$ is $e^{-t|k|^2}$. Hence

$$H_s^\Lambda *_\Lambda H_t^\Lambda =H_{s+t}^\Lambda,$$

and for every continuous periodic $a$,

$$H_t^\Lambda *_\Lambda a \longrightarrow a \qquad\text{uniformly as }t\downarrow0.$$

For periodic $L^p$ data, $1\leq p<\infty$, convergence instead holds in $L^p(Q)$. Thus $(H_t^\Lambda)$ is a regulated periodic delta sequence. The formal limit

$$\delta_\Lambda(x) =\frac1{V_\Lambda} \sum_{k\in\Lambda^*}e^{-ik\cdot x}$$

is meaningful through its action on smooth periodic test functions, not as a pointwise-convergent series or an ordinary value at $x=0$. In the opposite limit, $t\to\infty$, every nonzero mode is suppressed and $H_t^\Lambda(x)\to1/V_\Lambda$, the convolution kernel for projection onto the constant mode.

In canonical field notation, the continuum equal-time relation is written formally as

$$[\phi(t,\mathbf x),\pi(t,\mathbf y)] =i\delta^{(n)}(\mathbf x-\mathbf y).$$

Tong gives the three-dimensional canonical relation in Tong 2006–2007, § 2.1, p. 21, equation (2.2), official lecture-note PDF. In a periodic-box calculation, Tong also identifies $(2\pi)^3\delta^{(3)}(0)$ with the spatial volume in Tong 2006–2007, § 2.3, p. 25, equation (2.25), official lecture-note PDF. Discrete modes then replace the continuum delta by $\delta_\Lambda$; inserting $e^{-t|k|^2}$ replaces it by the smooth kernel $H_t^\Lambda$. At finite $t$ this is a regulated approximation that modifies the exact canonical commutator; the canonical algebra is recovered only in the $t\downarrow0$ test-function limit. Fourier Series, Fourier Transforms, and Plancherel Theory uses $e^{+ik\cdot x}$ for fixed-time spatial reconstruction. Replacing $k$ by $-k$ changes the displayed phase, but the heat sum is unchanged because both the reciprocal lattice and the multiplier are even.

This example is deliberately limited. The heat factor is a spatial smoothing choice, not automatically a Lorentz-invariant regulator, and it leaves the $k=0$ coefficient equal to one. It therefore does not cure an infrared zero-mode problem; Massless Scalars, Zero Modes, and Infrared Limits develops that separate issue. The meaning and domains of smeared quantum fields are treated on Quantum Fields as Operator-Valued Distributions, within the broader Foundations treatment. Tong motivates the same distinction between point notation and smeared fields in Tong 2006–2007, § 2.4, p. 31, equations (2.52)–(2.53), official lecture-note PDF.

Diagnostics and common pitfalls

Using associativity before proving integrability. Algebraic parentheses do not license a change in the order of divergent integrals. Establish a Tonelli, Fubini, or Young bound first.

Assuming convolution always smooths. Regularity is inherited from a factor whose derivatives can be moved through the integral. Two rough $L^1$ functions need not have a smooth convolution.

Checking only unit mass. An approximate identity must also concentrate at the origin and have uniformly bounded $L^1$ norm. Mass can otherwise escape to infinity or grow by cancellation.

Demanding the wrong convergence. Finite-$p$ norm convergence does not imply uniform convergence or determine values at jumps. State the topology before taking the limit.

Treating a sharp mode cutoff as a positive mollifier. A Dirichlet kernel has unit integral but oscillates and its $L^1$ norm is not uniformly bounded. Mode coefficients tending individually to one do not by themselves define a well-behaved approximate identity.

Applying Poisson summation from $L^1$ alone. Integrability controls an integral, not point samples or absolute convergence of two lattice sums. Use a safe class such as $\mathcal S$, or state the stronger theorem or distributional interpretation being used.

Dropping the reciprocal-lattice normalization. If $\Lambda=A\mathbb Z^n$, then $\Lambda^*=2\pi A^{-T}\mathbb Z^n$ and $V_\Lambda=|\det A|$. Omitting either factor breaks the Gaussian check.

Setting the regulator to zero inside the series. For $t>0$ the periodic heat kernel is smooth. At $t=0$ its Fourier series represents a delta comb only in a test-function sense; $\delta_\Lambda(0)$ is not an ordinary finite value.

Confusing ultraviolet smoothing with infrared control. The factor $e^{-t|k|^2}$ damps large momentum but leaves $k=0$ untouched. A regulator must be matched to the limit or singularity it is intended to control.

What the method now controls

Convolution composes translation-invariant kernels and obeys norm bounds that make its existence checkable. A smooth approximate identity regularizes rough data and then recovers it in the topology justified by the input class. Periodization converts a noncompact Schwartz function into a smooth periodic function whose Fourier coefficients sample the continuous transform. Poisson summation makes that relation exact, while the image expansion separates a finite-volume mode sum into its continuum contribution and controlled image corrections.

Useful continuations are:

• Fourier Series, Fourier Transforms, and Plancherel Theory supplies the full transform and norm framework used in the optional momentum-space checks.
• Tempered Distributions and Fourier Calculus defines delta distributions and supplies the machinery for interpreting combs, convolution, and Fourier operations beyond ordinary functions.
• Heat Kernels, Zeta Functions, and Spectral Determinants develops heat kernels as spectral objects rather than only mollifiers.
• Quantum Fields as Operator-Valued Distributions gives the physical treatment of smearing and point-field notation.

Exercises

A convolution that is not smooth. Compute $\mathbf 1_{[0,1]}*\mathbf 1_{[0,1]}$ and check its support, integral, and maximum.

Solution

The convolution is the length of the overlap between $[0,1]$ and $[x-1,x]$:

$$(\mathbf 1_{[0,1]}*\mathbf 1_{[0,1]})(x) =\begin{cases} x,&0\leq x\leq1,\\ 2-x,&1\leq x\leq2,\\ 0,&\text{otherwise}. \end{cases}$$

Its support is $[0,2]$, its maximum is $1$, and its integral is the area of the triangle, equal to $1=\|\mathbf 1_{[0,1]}\|_1^2$. The corners show that convolution of two $L^1$ functions need not be smooth.

Scaling a mollifier. Let $\rho_\varepsilon(x)=\varepsilon^{-n}\rho(x/\varepsilon)$ with $\rho\in L^1$ and $\int\rho=1$. Verify its mass, $L^1$ norm, and Fourier transform.

Solution

With $x=\varepsilon z$,

$$\int\mathrm d^n x\,\rho_\varepsilon(x) =\int\mathrm d^n z\,\rho(z)=1,$$

and the same substitution with absolute values gives $\|\rho_\varepsilon\|_1=\|\rho\|_1$. For the positive-forward transform,

$$\begin{aligned} \widetilde\rho_\varepsilon(p) &=\int\mathrm d^n x\, e^{ip\cdot x}\varepsilon^{-n}\rho(x/\varepsilon)\\ &=\int\mathrm d^n z\, e^{i(\varepsilon p)\cdot z}\rho(z) =\widetilde\rho(\varepsilon p). \end{aligned}$$

Thus shrinking in position space broadens the momentum-space multiplier.

A jump and the $L^\infty$ endpoint. For an even nonnegative mollifier, show that mollifying $\mathbf 1_{[0,\infty)}$ cannot converge in the $L^\infty$ norm.

Solution

At the origin the mollified value is $1/2$. More strongly, continuity forces a transition region in which the mollification takes values between $0$ and $1$, whereas the step takes only those endpoint values almost everywhere. For every $\varepsilon$, the essential supremum of the difference is at least $1/2$. Hence the difference cannot tend to zero in $L^\infty$, even though it tends to zero in $L^p$ for every finite $p$ on bounded intervals.

The Gaussian semigroup by completing the square. Verify directly that $H_s*H_t=H_{s+t}$.

Solution

The exponent in the convolution satisfies

$$\frac{|x-y|^2}{4s}+\frac{|y|^2}{4t} =\frac{s+t}{4st} \left|y-\frac{t}{s+t}x\right|^2 +\frac{|x|^2}{4(s+t)}.$$

The centered Gaussian integral is

$$\int_{\mathbb R^n}\mathrm d^n y\, \exp\!\left[ -\frac{s+t}{4st} \left|y-\frac{t}{s+t}x\right|^2 \right] =\left(\frac{4\pi st}{s+t}\right)^{n/2}.$$

Combining it with the two kernel prefactors leaves

$$(H_s*H_t)(x) =\frac1{(4\pi(s+t))^{n/2}} e^{-|x|^2/[4(s+t)]} =H_{s+t}(x).$$

This agrees with the multiplier calculation $e^{-s|p|^2}e^{-t|p|^2}=e^{-(s+t)|p|^2}$.

The theta transformation. In the one-dimensional Gaussian Poisson formula, set $x=0$, $L=1$, and $t=a/(4\pi)$ with $a>0$. Derive the resulting identity.

Solution

The image side becomes

$$\frac1{\sqrt a} \sum_{m\in\mathbb Z}e^{-\pi m^2/a},$$

while the mode side becomes

$$\sum_{r\in\mathbb Z}e^{-\pi a r^2}.$$

Therefore

$$\sum_{r\in\mathbb Z}e^{-\pi a r^2} =\frac1{\sqrt a} \sum_{m\in\mathbb Z}e^{-\pi m^2/a}.$$

Applying the transformation twice returns the original series, which checks the reciprocal scaling.

The regulated periodic delta. Show from its mode expansion that $H_t^\Lambda$ has unit mass, composes as a semigroup, and leaves the zero mode undamped.

Solution

Cell integration kills every term with $k\ne0$ and multiplies the $k=0$ term by $V_\Lambda$, so the mass is one. Periodic convolution multiplies Fourier coefficients, and

$$e^{-s|k|^2}e^{-t|k|^2}=e^{-(s+t)|k|^2},$$

which proves the semigroup law. At $k=0$ the multiplier is $e^{-t|0|^2}=1$ for every $t$. The regulator therefore suppresses high modes but does not alter the constant mode.

References

• Semyon Dyatlov, Lecture Notes for 18.155: Distributions, Elliptic Regularity, and Applications to PDEs, MIT (2022), Propositions 11.10 and 11.18 and Theorem 11.32, pp. 122–125 and 133–134, official lecture-note PDF. Schwartz convolution, Fourier multiplication, and Poisson summation.
• John K. Hunter, Notes on Partial Differential Equations, revised 2014, Hunter 2014, §§ 1.7, 1.9, and 5.1, pp. 9–12 and 130–131, official UC Davis lecture-note PDF. Young’s inequality, mollification, approximate-identity convergence, and the Gaussian heat kernel.
• David Tong, Quantum Field Theory, Cambridge Part III lecture notes (2006–2007), Tong 2006–2007, §§ 2.1, 2.3, and 2.4, pp. 21, 25, and 31, official lecture-note PDF. Canonical equal-time field commutators and their delta normalization.