# The Dirac Delta as a Tempered Distribution
**Context**: Companion notes to the "On the Function Spaces for Fourier Transform" slides. Fills in the derivations the slides state, and connects them back to where the delta first showed up in the courses.
**Topics**: Schwartz space, tempered distributions, the delta as a distributional limit, the distributional derivative, $H' = \delta$, Fourier transforms of distributions
**Related**: the Spring 2026 Day 12 notes (first appearance of $\delta$ in the transform of $\cos t$), the Spring 2026 Day 14 notes (the working definition, "a tool of the devil"), MATH235 Laplace unit ($\mathcal{L}\{\delta(t-a)\} = e^{-sa}$)
---
## Symbol Conventions
These are fixed for the whole document; the slides use the same ones.
| Object | Symbol |
|---|---|
| Heaviside (unit step) function | $H(t - t_0)$ |
| Schwartz space of test functions | $S(\mathbb{R})$, or $S(\mathbb{R},\mathbb{C})$ when the codomain matters |
| A test function | $\phi$; a sequence of them $\phi_n$ |
| A tempered distribution | $T \in S'(\mathbb{R})$ |
| Pairing, **distribution first, test function second** | $\langle T, \phi \rangle$ |
| The distribution induced by an ordinary function $f$ | $T_f$, with $\langle f, \phi \rangle := \int_{\mathbb{R}} f(t)\phi(t)\,dt$ |
| Distributional derivative | $\langle T', \phi \rangle := -\langle T, \phi' \rangle$ |
| Dirac delta at $t_0$ | $\delta_{t_0}$, with $\langle \delta_{t_0}, \phi \rangle = \phi(t_0)$; $\delta := \delta_0$ |
| Fourier transform (symmetric convention) | $\hat{\phi}(\omega) = \frac{1}{\sqrt{2\pi}} \int_{\mathbb{R}} \phi(t) e^{-i\omega t}\,dt$ |
Two rules keep the notation honest. Letters like $f$, $T$, $H$, $\delta$ always sit in the *first* slot of a pairing and name the (possibly generalized) object; letters from the $\phi$ family always sit in the *second* slot and name a test function. Never let $f$ name a test function.
---
## 0. Why We Are Here: Two Roads to the Same Object
The delta shows up in two places in the courses, and it arrives looking like an emergency each time.
**The Fourier road.** We would like to get all the way back to Fourier series from the Fourier transform: take the transform of something periodic. But $\cos t$ is not integrable on $\mathbb{R}$, so its classical transform $\frac{1}{\sqrt{2\pi}} \int \cos(t)\, e^{-i\omega t}\,dt$ does not converge. Day 12 wrote down the answer anyway,
$
\mathcal{F}\{\cos t\}(\omega) = \sqrt{\frac{\pi}{2}}\,\big[\delta(\omega - 1) + \delta(\omega + 1)\big],
$
and Day 14 called the object that makes this work "a tool of the devil." The Fall 2023 boards did the same thing from the other direction: send the period of a square pulse to infinity and watch $\hat{f}$ pile up into $\sqrt{2\pi}\,A\,\delta(\omega)$. The physical picture was a pop on a record or a CD: an impulse in time whose spectrum is a spike. Day 12's takeaway line, "delta functions bridge discrete and continuous frequency domains," is exactly right and is what Section 5 proves.
**The ODE road.** In differential equations we drive systems with switches. The Heaviside step
$
H(t - t_0) = \begin{cases} 0, & t < t_0 \\ 1, & t > t_0 \end{cases}
$
turns a forcing term on at $t_0$. A solution written in terms of $H$ has a corner, and it is natural to ask, "what is the derivative of the step?" Classically there is none at $t_0$. The Laplace unit of MATH235 answers with $\delta(t - t_0)$ and the transform pair $\mathcal{L}\{\delta(t-a)\} = e^{-sa}$, and uses it to model impulses: a hammer strike, a kick.
Both roads run into the same wall, and the wall is the word *function*.
---
## 1. Historical Sketch
The behavior had been in the air for a long time. Fourier's inversion formula (1822) already contains it in disguise, Cauchy used sharply peaked kernels, and Heaviside's operational calculus (1890s) manipulated "the derivative of the step" with a straight face and correct answers. Sobolev (1936) introduced weak derivatives for partial differential equations.
The notation and the name come from P. A. M. Dirac (1902–1984), who introduced $\delta(x)$ in 1927 to handle the normalization of continuous-spectrum states in quantum mechanics: he needed $\langle x | x' \rangle = \delta(x - x')$ to play the role that $\delta_{ij}$ plays for a discrete basis. He was open about it not being a function. The theory that makes it rigorous is Laurent Schwartz's theory of distributions, announced in a 1945 paper and laid out in full in the two-volume *Théorie des distributions* (1950–51), for which he received the Fields Medal in 1950. Schwartz's idea, which is the whole content of these notes: stop asking what $\delta$ *is* at each point and ask instead what it *does* to nice functions. A distribution is a linear map on a vector space of test functions.
---
## 2. The Working Definition, and Why It Cannot Be a Function
Every course states the same three properties. For $t_0 \in \mathbb{R}$, the delta at $t_0$ is "the function" $\delta(t - t_0)$ with
1. **Localization**: $\delta(t - t_0) = 0$ for $t \neq t_0$.
2. **Normalization**: $\displaystyle\int_{\mathbb{R}} \delta(t - t_0)\,dt = 1$.
3. **Sifting**: $\displaystyle\int_{\mathbb{R}} \delta(t - t_0)\, f(t)\,dt = f(t_0)$ for every continuous $f$.
Properties (1) and (2) are already incompatible for a function. If a function vanishes everywhere except at one point, its integral (Riemann or Lebesgue) is zero: a single point has no length. So no function satisfies both, and "$\delta(t_0) = \infty
quot; does not repair this; $\infty \cdot 0$ is not $1$.
What is actually true, and what the three properties are shorthand for, is this: *there are sequences of honest functions $\phi_n$ for which the integrals $\int \phi_n(t) f(t)\,dt$ converge to $f(t_0)$.* The delta is the name we give to the limit of those integrals, viewed as a rule $f \mapsto f(t_0)$. In the language of the handwritten notes: the three properties "should be thought of as the value of a sequence of functionals." Making that sentence precise takes three steps.
---
## 3. Step 1: Schwartz Space, the "Really Nice" Functions
We need a supply of functions so well behaved that anything reasonable can be tested against them. The right class for Fourier analysis is the **Schwartz space**
$
S(\mathbb{R}, \mathbb{C}) = \left\{ \phi \in C^\infty(\mathbb{R}, \mathbb{C}) \;\middle|\; \forall \alpha, \beta \in \mathbb{N}_0,\; \sup_{x \in \mathbb{R}} \left| x^\alpha \frac{d^\beta \phi}{dx^\beta} \right| < \infty \right\}.
$
In words: smooth functions which, together with every derivative, decay faster than any power grows. The quantities
$
\|\phi\|_{\alpha,\beta} := \sup_{x \in \mathbb{R}} \left| x^\alpha \phi^{(\beta)}(x) \right|
$
are the **seminorms** of $S(\mathbb{R})$, and they define what convergence means there: $\phi_n \to \phi$ in $S(\mathbb{R})$ if and only if $\|\phi_n - \phi\|_{\alpha,\beta} \to 0$ for every pair $(\alpha, \beta)$. That is a strong requirement (uniform convergence of every derivative, with polynomial weights), which is why continuity with respect to it will be easy to satisfy.
**Examples.**
- $\phi(x) = x^a e^{-bx^2}$ with $b > 0$ and $a \in \mathbb{N}_0$. The integer restriction on $a$ matters: for fractional $a$ the function is not smooth at $0$, and for negative $x$ may not even be defined. These are the integrands in the moments of the normal distribution, $\mathbf{E}[X^a]$ for $X \sim \mathcal{N}(0, \sigma^2)$.
- Any $C^\infty$ function with compact support.
- $e^{-|x|}$ is **not** in $S(\mathbb{R})$ (not smooth at $0$); $(1 + x^2)^{-1}$ is not (decays only like $x^{-2}$); polynomials, $\cos x$, and $H$ are not (no decay at all).
**Why this space.** Schwartz space is stable under everything the Fourier transform does. Differentiation and multiplication by polynomials map $S(\mathbb{R})$ to itself, and the transform swaps them:
$
\widehat{\phi'}(\omega) = i\omega\, \hat{\phi}(\omega), \qquad \widehat{x\phi}(\omega) = i \frac{d}{d\omega} \hat{\phi}(\omega).
$
Decay in $x$ buys smoothness in $\omega$ and vice versa, so a function that has both stays a function that has both. Consequently $\mathcal{F} : S(\mathbb{R}) \to S(\mathbb{R})$ is a bijection, continuous in the seminorms, with inverse given by the usual inversion formula. Schwartz space is exactly the space on which the Fourier transform can do no harm. That is the fact Section 5 leans on.
---
## 4. Step 2: Distributions as Continuous Linear Functionals
Now generalize the notion of function. A **tempered distribution** is a map
$
T : S(\mathbb{R}) \to \mathbb{C}, \qquad \phi \mapsto \langle T, \phi \rangle,
$
that is
- **linear**: $\langle T, c_1 \phi_1 + c_2 \phi_2 \rangle = c_1 \langle T, \phi_1 \rangle + c_2 \langle T, \phi_2 \rangle$, and
- **continuous**: if $\phi_n \to \phi$ in $S(\mathbb{R})$ then $\langle T, \phi_n \rangle \to \langle T, \phi \rangle$.
A distribution takes a function in and returns a number. The set of all tempered distributions is the **dual space** of $S(\mathbb{R})$, written $S'(\mathbb{R})$.
**Ordinary functions are distributions.** Let $f$ be locally integrable with at most polynomial growth, $|f(t)| \le C(1 + |t|)^N$. Then
$
T_f(\phi) := \langle f, \phi \rangle := \int_{\mathbb{R}} f(t)\, \phi(t)\, dt
$
converges for every $\phi \in S(\mathbb{R})$ (the test function's decay beats $f