Heat Kernel and the Fundamental Solution of the Diffusion Equation

Release all the heat at a single point and a single instant, then step back and watch. What spreads out is the most important function in parabolic theory: the heat kernel, the Green's function of the diffusion equation $\partial_t u = D\nabla^2 u$ . It is the elementary response of the medium — the answer to "what does a unit point source do?" — and because the equation is linear, that one answer builds every other by superposition. Convolve the kernel with any initial temperature, concentration, or probability profile and the Cauchy problem is solved in a single stroke. This article takes the free-space Gaussian and develops it into a complete object: its three independent derivations, its behaviour at boundaries, its limiting cases, and the surprisingly deep places it reappears. Three threads run through everything below. The first is self-similarity: the kernel has exactly one length scale, the diffusion length $\ell=\sqrt{Dt}$ , and it simply rescales as time passes — the width grows as $\sqrt{t}$ and the peak falls as $1/\sqrt{t}$ , keeping the area fixed at one. The second is the semigroup structure: the blurring operators $\{e^{Dt\nabla^2}\}_{t\ge 0}$ compose by adding times and have no inverse — the precise statement that diffusion smooths, forgets, and runs only forward. The third is analytic continuation: rotate time onto the imaginary axis and the same Gaussian becomes the free quantum propagator, tying the kernel to path integrals, Euclidean field theory, and — through its short-time trace — the spectral geometry that lets one almost hear the shape of a drum.

Key takeaways

The heat kernel $K(x,t;x')=(4\pi Dt)^{-n/2}\exp(-|x-x'|^2/4Dt)$ is the fundamental solution of the diffusion equation: the field at later time from a unit point source released at $x'$ when $t=0$ .
It is a unit-area Gaussian whose width grows as $\sqrt{Dt}$ , so the mean-square displacement is $\langle r^2\rangle = 2nDt$ — the linear-in-time signature that distinguishes diffusion from ballistic $\sim t$ wave spreading.
Convolving the kernel with initial data solves the Cauchy problem, $u(x,t)=\int K\,u_0\,\mathrm{d}^n x'$ , and the kernel reproduces $u_0$ as $t\to 0^+$ (it is an approximate identity).
The kernels form a semigroup, $K_t * K_s = K_{t+s}$ , generated by $D\nabla^2$ ; Wick-rotating $t\to it$ turns it into the free Schrödinger propagator.
On a bounded domain it becomes the spectral sum $K=\sum_n e^{-\lambda_n D t}\phi_n(x)\phi_n^*(x')$ , whose short-time trace encodes the domain's volume, area, and curvature (Weyl asymptotics).
Key idea: the fundamental solution is a self-similar Gaussian whose only length scale is the diffusion length $\ell=\sqrt{Dt}$ ; convolving it with initial data solves any Cauchy problem, and the family $\{K_t\}_{t>0}$ is the smoothing semigroup generated by $D\nabla^2$ .

Where this lives today

The heat kernel is running inside working technology and open research right now. In machine vision and deep learning, Gaussian blurring is convolution with $K_t$ , and the "scale space" it generates underlies edge detection, scale-invariant feature matching, and the diffusion generative models that create images by reversing a noising process that is literally a discretized heat flow. In quantitative finance, the same Gaussian is the transition density behind the Black–Scholes equation, a heat equation in disguise. In quantum field theory, heat-kernel (Schwinger proper-time) regularization tames ultraviolet divergences and computes anomalies, while the spectral action builds Lagrangians directly from the short-time trace $\mathrm{Tr}\,e^{-tH}$ . And Kac's drum question has real teeth: Gordon, Webb, and Wolpert exhibited genuinely different domains with identical Laplacian spectra in 1992 (so the answer is "no, not always"), yet the heat-kernel coefficients still recover volume, boundary length, and topology — which is why the "heat kernel signature" is now a standard descriptor for matching and classifying 3D shapes.

The physical picture

Picture a droplet of dye released at one point of a perfectly still pool. Each dye molecule wanders by countless random collisions, and the probability of finding it a distance $x$ away after time $t$ is a bell curve that widens as $\sqrt{t}$ . The heat kernel is that transition probability density; the macroscopic "flow" of dye or heat is nothing more than an enormous ensemble of these random walkers relaxing toward uniformity. There is no wind carrying them and no front sweeping outward — just a Gaussian that flattens and spreads while conserving its total. The first figure shows this spreading quantitatively for three successive times, and makes the self-similar bookkeeping visible: as the width doubles the peak halves, keeping the area pinned at one.

Figure

The fundamental solution at three times t₁, 4t₁, 9t₁ for a solute with D = 1×10⁻⁹ m²/s: the Gaussian keeps unit area but its width grows as √(Dt) while its peak falls as 1/√(Dt), so the standard deviation runs 1.34 mm → 2.68 mm → 4.02 mm while the peaks drop in the ratio 3 : 1.5 : 1

There is a second, equally useful way to see the same physics: in Fourier space. Each plane wave $e^{ik\cdot x}$ is an eigenfunction of the Laplacian with eigenvalue $-k^2$ , so under diffusion it does nothing but decay, $\propto e^{-Dk^2 t}$ . Short-wavelength (large- $k$ ) features die first, which is exactly why the heat equation turns jagged data into something smooth and analytic — and why running it backward is catastrophic, since those same modes would then blow up. The next figure turns the spreading into its cleanest quantitative fingerprint: the diffusion length $\ell=\sqrt{2Dt}$ grows sublinearly, so waiting four times as long only doubles the spread. That concave, front-less growth is the visual opposite of a wave, whose sharp front advances linearly as $ct$ .

Figure

The diffusion length ℓ = √(2Dt) for the same solute D = 1×10⁻⁹ m²/s grows sublinearly: quadrupling the elapsed time from 900 s to 3600 s only doubles the spread, 1.34 mm → 2.68 mm — diffusion has no sharp front, unlike a wave whose front advances linearly as ct

The core equations

The starting point is the diffusion equation itself, the canonical parabolic PDE. It says that the field changes in time at a rate set by its own curvature in space: where a profile is concave down (a hot bump) it cools, where it is concave up (a cold dip) it warms, and the rate is set by the single material constant $D$ .

\frac{\partial u}{\partial t} = D\,\nabla^2 u \qquad (1)

where: $u(x,t)$ — the diffusing field (temperature, concentration, or probability density); $D$ — diffusion constant, units $\mathrm{length}^2/\mathrm{time}$ ; $\nabla^2$ — Laplacian in $n$ spatial dimensions. Reading it: the Laplacian is a local averaging operator, so the field is dragged toward the average of its neighbours at a rate set by $D$ . The heat kernel is defined as the causal Green's function of this operator — the response to a source that is a point in both space and time. Causality (the kernel vanishes before the source fires) picks out the physically forward-in-time solution and makes it unique.

\left(\frac{\partial}{\partial t} - D\nabla^2\right)K(x,t;x',t') = \delta^{(n)}(x-x')\,\delta(t-t'),\quad K=0\ \text{for }t<t' \qquad (2)

where: $K$ — causal fundamental solution of the diffusion operator; $\delta^{(n)}$ — $n$ -dimensional Dirac delta (a unit point source in space); $t'$ — the emission time. Reading it: constant coefficients make $K$ depend only on $x-x'$ and $t-t'$ , so we may set $x'=0,\,t'=0$ without loss. Solving that defining equation in free space (three independent ways, below) gives the Gaussian. This is the single formula the whole article orbits.

K(x,t;x',0) = \frac{1}{(4\pi D t)^{n/2}}\exp\!\left(-\frac{|x-x'|^2}{4 D t}\right) \qquad (3)

where: $n$ — number of spatial dimensions; $|x-x'|$ — Euclidean source-to-field distance; the standard deviation per axis is $\sqrt{2Dt}$ and the diffusion length is $\ell=\sqrt{Dt}$ . Reading it: the prefactor $(4\pi Dt)^{-n/2}$ conserves total mass as the exponential widens, and $t\to 0^+$ collapses it to $\delta^{(n)}(x-x')$ . The reason the kernel matters is that it solves every initial-value problem by a single integral: superpose the point-source responses weighted by the initial data and you have the field for all later times — the Poisson (convolution) formula.

u(x,t) = \int_{\mathbb{R}^n} K(x,t;x',0)\,u_0(x')\,\mathrm{d}^n x' \qquad (4)

where: $u_0(x')$ — prescribed initial data $u(x,0)$ ; $u(x,t)$ — the solution for $t>0$ , equal to the convolution $K_t * u_0$ . Reading it: linearity plus the kernel's two properties (it solves the PDE, and $K_t\to\delta$ ) guarantee this attains the data and evolves it correctly. Chaining two evolutions is the same as one longer evolution, because Gaussian variances add. This composition law — Chapman–Kolmogorov in probability, the semigroup property in analysis — is the algebraic heart of the object.

\int_{\mathbb{R}^n}K(x,t;x'')\,K(x'',s;x')\,\mathrm{d}^n x'' = K(x,t+s;x') \qquad (5)

where: $t,s>0$ — successive evolution times whose variances $2Dt$ and $2Ds$ add to $2D(t+s)$ . Reading it: the operators $e^{Dt\nabla^2}$ form a one-parameter semigroup — closed under composition, no inverse; irreversibility written as algebra. Weight the kernel by $|x-x'|^2$ and integrate, and you recover the defining kinematic signature of diffusion: the mean-square displacement grows linearly in time, with slope $2D$ per axis.

\langle |x-x'|^2\rangle = \int |x-x'|^2\, K\,\mathrm{d}^n x = 2 n D t \qquad (6)

where: $n$ — spatial dimension; per axis $\langle x^2\rangle = 2Dt$ . Reading it: this is Einstein's Brownian-motion relation; linear growth ( $\propto t$ , not $\propto t^2$ ) separates a random walk from ballistic motion. Finally, on a bounded or curved domain the free-space Gaussian is replaced by a spectral sum over the eigenfunctions of $-\nabla^2$ , each decaying at its own rate. The short-time behaviour of the trace of this sum then reads off the geometry of the domain.

K(x,t;x') = \sum_{n} e^{-\lambda_n D t}\,\phi_n(x)\,\phi_n^*(x'),\qquad \mathrm{Tr}\,e^{Dt\nabla^2}\;\sim\;\frac{\mathrm{Vol}+a_1 t + a_2 t^2 + \cdots}{(4\pi D t)^{n/2}} \qquad (7)

where: $\phi_n,\lambda_n$ — orthonormal eigenfunctions and eigenvalues of $-\nabla^2$ under self-adjoint boundary conditions ( $\lambda_n\ge 0$ ); $\mathrm{Vol}$ — domain volume (Weyl's leading term); $a_1,a_2$ — geometric invariants (boundary area, integrated curvature). Reading it: the spectrum determines, and is partly determined by, the shape — the mathematics behind "hearing the shape of a drum".

Derivation ABuilding the free-space kernel from Fourier modes

The fastest honest route to the Gaussian is to diagonalize the Laplacian. Every plane wave $e^{ik\cdot x}$ is an eigenfunction of $\nabla^2$ with eigenvalue $-k^2$ , so the spatial Fourier transform turns the diffusion PDE into an uncoupled family of first-order ordinary differential equations, one per wavevector, each solved by a decaying exponential. Inverting the transform then reassembles the decayed modes into position space, and a single completed-square Gaussian integral delivers both the $(4\pi Dt)^{-n/2}$ prefactor and the $-|x|^2/4Dt$ exponent with no guessing of the answer in advance. We do it in one dimension; the $n$ -dimensional kernel is the product of $n$ identical factors.

Step 1. Transform the defining equation (with source at the origin) over space, $\tilde K(k,t)=\int K(x,t)\,e^{-ikx}\,\mathrm{d}x$ . The Laplacian becomes multiplication, $\partial_x^2\to -k^2$ , so the PDE collapses to the ODE $\partial_t\tilde K = -Dk^2\,\tilde K$ . Why: the Fourier transform is precisely the change of basis in which $\nabla^2$ is diagonal, converting a coupled spatial problem into independent evolution of each mode.

Step 2. Solve each mode with the point-source initial condition $\tilde K(k,0)=1$ (the transform of $\delta(x)$ is unity), giving $\tilde K(k,t)=e^{-Dk^2 t}$ . Why: a point source excites every wavenumber equally at $t=0$ , and thereafter each simply decays at its own rate $Dk^2$ — fast for short wavelengths, slow for long ones, which is the smoothing property in disguise.

Step 3. Invert the transform, $K(x,t)=\dfrac{1}{2\pi}\displaystyle\int_{-\infty}^{\infty} e^{ikx}\,e^{-Dk^2 t}\,\mathrm{d}k$ . Why: superposing the decayed modes with their original phases $e^{ikx}$ rebuilds the position-space field; the integral is Gaussian in $k$ , which is what makes it doable in closed form.

Step 4. Complete the square in the exponent, $ikx - Dk^2 t = -Dt\,(k - ix/2Dt)^2 - x^2/4Dt$ , and use $\int e^{-Dt\,q^2}\mathrm{d}q=\sqrt{\pi/Dt}$ to obtain $K(x,t)=\dfrac{1}{\sqrt{4\pi Dt}}\,e^{-x^2/4Dt}$ . Why: the shifted contour integral is a standard real Gaussian; taking the product of $n$ such one-dimensional factors gives the full kernel of equation (3). *Verified: substituting the Gaussian back into $\partial_t K = D\partial_x^2 K$ satisfies it identically; $\int_{-\infty}^{\infty}K\,\mathrm{d}x=1$ ; and $K\to\delta(x)$ as $t\to 0^+$ . Matches Arfken–Weber–Harris and Stone & Goldbart.*

Go deeper · Derivation B · 1 minThe similarity solution and why the exponent is universal

Derivation BThe similarity solution and why the exponent is universal

The heat equation hides a scaling symmetry: it is unchanged by $x\to\lambda x$ together with $t\to\lambda^2 t$ , which forces its fundamental solution to depend on space and time only through the single dimensionless group $\eta = x/\sqrt{4Dt}$ . Feeding this self-similar ansatz into the PDE collapses a partial differential equation in two variables to an ordinary differential equation in $\eta$ alone, and the fact that every explicit factor of $t$ cancels is the tell-tale sign that the similarity variable was chosen correctly. This derivation needs no Fourier transform and shows why the profile must be a Gaussian rather than merely confirming that it is.

Step 1. Write the scaling ansatz $K(x,t)=t^{-1/2} f(\eta)$ with $\eta = x/\sqrt{4Dt}$ . Why: the $t^{-1/2}$ prefactor is dictated by conservation — since a natural width scales as $\mathrm{d}x\sim t^{1/2}$ , the amplitude must fall as $t^{-1/2}$ to keep $\int K\,\mathrm{d}x$ constant, exactly the self-similar form the scaling symmetry permits.

Step 2. Substitute the derivatives $\partial_t\eta = -\eta/2t$ and $\partial_x\eta = 1/\sqrt{4Dt}$ to get $\partial_t K = -\tfrac12 t^{-3/2}(f+\eta f')$ and $D\,\partial_x^2 K = \tfrac14 t^{-3/2}f''$ . Why: expressing both sides of the PDE through the chain rule in $\eta$ is what exposes the common $t^{-3/2}$ factor and lets time drop out entirely.

Step 3. Equate the two sides; the $t^{-3/2}$ cancels and leaves the reduced ODE $f'' + 2\eta f' + 2f = 0$ , which is exactly $(f' + 2\eta f)' = 0$ . Why: the total-derivative form is the gift of the correct similarity variable — it means the equation integrates once immediately, turning a second-order ODE into a first-order one.

Step 4. Integrate once to $f' + 2\eta f = C$ ; demanding decay $f,f'\to 0$ as $\eta\to\pm\infty$ forces the constant $C=0$ , so $f' = -2\eta f$ and $f(\eta)=A\,e^{-\eta^2}$ . Why: physical solutions must vanish far from the source, which kills the constant and selects the pure Gaussian branch out of the general solution.

Step 5. Fix $A$ by normalization: $\int K\,\mathrm{d}x = A\!\int e^{-\eta^2}\sqrt{4Dt}\,\mathrm{d}\eta = A\sqrt{4\pi Dt}\cdot t^{-1/2}\cdot t^{1/2}=1$ gives $A=1/\sqrt{4\pi D}$ , hence $K=(4\pi Dt)^{-1/2}e^{-x^2/4Dt}$ . Why: conservation of total heat is the last physical input, and it pins the one free constant, reproducing the Fourier result from a completely independent argument. *Verified: the boxed 1D kernel $K=(4\pi Dt)^{-1/2}e^{-x^2/4Dt}$ solves $\partial_t K=D\partial_x^2 K$ by direct substitution and integrates to one; the ODE factorization $f''+2\eta f'+2f=(f'+2\eta f)'$ is exact.*

Go deeper · Derivation C · 2 minHeat kernels on a boundary by the method of images

Derivation CHeat kernels on a boundary by the method of images

Real diffusion happens in domains with walls, and the free-space Gaussian must be corrected to respect them. When the boundary is a flat plane the fix is purely geometric: place a fictitious image source on the far side of the wall, choose its sign to enforce the boundary condition, and the superposition automatically solves the heat equation in the physical region while meeting the condition at the wall exactly, for all time. A minus sign builds the Dirichlet (perfectly absorbing) kernel; a plus sign builds the Neumann (perfectly insulating) one. We work the half-line $x>0$ with a source at $x'>0$ .

Step 1. For the absorbing wall $u(0,t)=0$ , add a negative image at $-x'$ : $K_D(x,t;x') = K(x-x',t) - K(x+x',t)$ . Why: both terms individually satisfy the heat equation for $x>0$ because their singularities sit at $\pm x'$ , and a linear combination of solutions is again a solution — the image contributes no spurious source inside the domain.

Step 2. Check the wall: at $x=0$ , $K_D(0,t;x') = K(-x',t) - K(x',t) = 0$ by the even symmetry $K(-a)=K(a)$ of the Gaussian. Why: the two Gaussians are mirror images that coincide exactly on the boundary, so their difference cancels there identically at every instant — the Dirichlet condition holds not approximately but exactly.

Step 3. Check the start: as $t\to 0^+$ , $K_D\to\delta(x-x')$ for $x,x'>0$ , since the image term $K(x+x',t)$ concentrates at $x=-x'<0$ , outside the physical region. Why: only the true source lies in the domain, so the image leaves the initial condition untouched where it matters — the kernel still represents a unit source at $x'$ .

Step 4. For the insulating wall $\partial_x u|_{0}=0$ , flip the sign: $K_N(x,t;x') = K(x-x',t) + K(x+x',t)$ , whose $x$ -derivative vanishes at $x=0$ by odd symmetry of $\partial_x K$ . Why: a no-flux boundary is the reflecting counterpart of an absorbing one, and adding rather than subtracting the image builds an even profile whose slope is automatically zero at the wall, conserving all the heat. *Verified: both $K_D$ and $K_N$ solve the PDE, satisfy their boundary conditions identically for all $t>0$ , and reduce to $\delta(x-x')$ as $t\to 0^+$ ; standard in Arfken and Barton.*

Limiting cases

Reading the kernel $K=(4\pi Dt)^{-n/2}e^{-r^2/4Dt}$ in extreme regimes exposes what each factor is doing:

Case	Result	Meaning
Short time $t\to 0^+$	$K(x,t;x')\longrightarrow \delta^{(n)}(x-x')$	The Gaussian collapses to a delta: the kernel reproduces the initial data exactly and is the identity element of the semigroup at $t=0$ .
Long time on $\mathbb{R}^n$	$\int K\,\mathrm{d}^n x = 1$ for all $t$ , yet $K\to 0$ pointwise	Total heat is conserved forever; the profile flattens toward zero everywhere without ever losing area — mass spreads to infinity, it does not vanish.
Imaginary time (Wick rotation)	$t\to \dfrac{i\hbar t}{2mD}\Rightarrow K\to\Big(\dfrac{m}{2\pi i\hbar t}\Big)^{n/2}e^{\,i m\|x-x'\|^2/2\hbar t}$	With $D=\hbar/2m$ the kernel becomes the free Schrödinger propagator — same Gaussian shape, now oscillatory and unitary rather than decaying.
Bounded domain, large $t$	$K\sim e^{-\lambda_1 D t}\,\phi_1(x)\phi_1^*(x')$	The lowest non-trivial eigenvalue $\lambda_1$ sets the slowest relaxation; the system forgets its initial detail and settles to equilibrium at rate $\lambda_1 D$ .

Common trap"Diffusion spreads out at a definite speed with a front" — the infinite-speed trap

The heat kernel is strictly positive everywhere for any $t>0$ : a source switched on at the origin is felt, however faintly, arbitrarily far away the very next instant. Diffusion therefore has infinite signal speed and no sharp front. Do not confuse the spreading width $x\sim\sqrt{Dt}$ — a statistical scale, the place where the bulk of the heat sits — with a wave's light cone, beyond which the field is exactly zero. There is no such cone here; the exponentially small tail is genuinely nonzero everywhere. The flip side is equally important and equally abused: never integrate the heat equation backward in time. For $t<0$ every Fourier mode grows as $e^{+Dk^2|t|}$ , so the tiniest high-frequency noise in the data is amplified without bound — the backward problem is ill-posed in the Hadamard sense (it violates continuous dependence on data), which is why deblurring and thermal back-extrapolation are so treacherous.

Worked examples

1 · How far and how fast does a solute actually diffuse (fully worked)

A dye of diffusion constant $D = 1.0\times10^{-9}\ \mathrm{m^2/s}$ (a typical small molecule in water) is released as a sharp point in one dimension. Find its spatial spread after 1 hour, the peak of its concentration profile, and how long it must wait to spread over 1 cm. Then use the answer to state the scaling law in one sentence.

Step 1. The standard deviation of the 1D kernel is $\sigma=\sqrt{2Dt}$ . At $t=3600\ \mathrm{s}$ , $\sigma=\sqrt{2(1.0\times10^{-9})(3600)}=\sqrt{7.2\times10^{-6}}=2.68\times10^{-3}\ \mathrm{m}$ . So after a full hour the dye has spread only about $\boldsymbol{\sigma\approx 2.7\ \mathrm{mm}}$ — diffusion over macroscopic distances is astonishingly slow, which is exactly why life relies on stirring and blood flow rather than diffusion for bulk transport.

Step 2. The peak value (at $x=0$ ) is the prefactor $1/\sqrt{4\pi Dt}=1/\sqrt{4\pi(1.0\times10^{-9})(3600)}=1/\sqrt{4.52\times10^{-5}}=\boldsymbol{149\ \mathrm{m^{-1}}}$ (units of inverse length because the 1D kernel is a probability density per unit length). About $68\%$ of the total sits within $\pm\sigma$ , since $\mathrm{erf}(1/\sqrt2)=0.683$ — the ordinary Gaussian one-sigma rule.

Step 3. To spread over a half-width of $1\ \mathrm{cm}$ we set $\sigma=10^{-2}\ \mathrm{m}$ and invert: $t=\sigma^2/(2D)=(10^{-2})^2/(2\times10^{-9})=\boldsymbol{5.0\times10^{4}\ \mathrm{s}\approx 13.9\ \mathrm{hours}}$ . Compare with Step 1: the distance grew by a factor $10^{-2}/2.68\times10^{-3}\approx 3.7$ , but the time grew by $5.0\times10^4/3600\approx 13.9\approx 3.7^2$ .

Step 4. That is the whole story of diffusion in one line: because $\sigma\propto\sqrt{Dt}$ , covering $k$ times the distance costs $k^2$ times the wait — reaching 10 cm instead of 1 cm would take a hundredfold longer, roughly 58 days. Ballistic transport (a wave, a current) would cover ten times the distance in ten times the time; diffusion demands a hundred.

2 · The relaxation time of a heated rod (fully worked)

A copper rod of length $L=0.10\ \mathrm{m}$ (thermal diffusivity $D=1.11\times10^{-4}\ \mathrm{m^2/s}$ ) is held at fixed temperature at both ends (Dirichlet). After the ends are clamped, how long does it take an arbitrary initial temperature bump to relax toward the steady state?

Step 1. On $[0,L]$ with Dirichlet ends the Laplacian eigenfunctions are $\phi_m(x)=\sqrt{2/L}\,\sin(m\pi x/L)$ with eigenvalues $\lambda_m=(m\pi/L)^2$ . The spectral kernel (equation 7) evolves mode $m$ as $e^{-\lambda_m D t}$ , so every mode decays exponentially and the slowest one dominates at late times.

Step 2. The lowest eigenvalue is $\lambda_1=(\pi/L)^2=(\pi/0.10)^2=986.96\ \mathrm{m^{-2}}$ . Its decay rate is $\lambda_1 D=(986.96)(1.11\times10^{-4})=0.1096\ \mathrm{s^{-1}}$ , so the relaxation (e-folding) time is

\tau_1 = \frac{1}{\lambda_1 D} = \frac{L^2}{\pi^2 D} = \frac{(0.10)^2}{\pi^2(1.11\times10^{-4})} = \boldsymbol{9.1\ \mathrm{s}}.

Step 3. Higher modes vanish far faster: the second mode has $\lambda_2=4\lambda_1$ , so it decays four times as quickly, $\tau_2=\tau_1/4\approx 2.3\ \mathrm{s}$ , and the $m$ -th mode as $\tau_1/m^2$ . After a few $\tau_1\approx 9\ \mathrm{s}$ the rod's profile is essentially the single lowest sine mode, and it settles to equilibrium on that timescale. The $L^2$ scaling is the practical face of $\sqrt{Dt}$ : double the rod and the relaxation time quadruples.

3 · From heat kernel to quantum propagator (Wick rotation)

Show that the free-particle Schrödinger propagator is the heat kernel continued to imaginary time, and check that the width scaling makes physical sense.

Step 1. Set $D=\hbar/2m$ so $4Dt=2\hbar t/m$ ; the 1D kernel is $K=(2\pi\hbar t/m)^{-1/2}\exp(-m x^2/2\hbar t)$ . Rotate $t\to it$ : since $-1/i=i$ , the exponent turns oscillatory, giving $K_{\text{Sch}}=\big(m/2\pi i\hbar t\big)^{1/2}\exp\!\big(i m x^2/2\hbar t\big)$ — exactly the standard free propagator.

Step 2. The real, decaying Gaussian has become complex and oscillatory. Diffusion dissipated probability irreversibly; the propagator redistributes it unitarily, and the smoothing semigroup has become a reversible unitary group — one sign in the exponent flips the arrow of time.

Step 3. Check the scale: the real kernel spreads as $\sigma\sim\sqrt{Dt}$ , so under $D\to\hbar/2m$ a wavepacket delocalizes over $\sim\sqrt{\hbar t/m}$ — heavier particles spread more slowly, exactly the free Gaussian wavepacket of quantum mechanics. The same machinery governs a cooling rod and a spreading electron; only the reality of time differs.

Physics of the portrait

The heat kernel is one of those ideas that took two centuries and four disciplines to fully understand: born in Fourier's theory of heat, made physical by Einstein's account of Brownian motion, made rigorous by Wiener's measure on paths, and made geometric by Kac's playful question about drums.

Origins · 1 minheat kernel and how this was discovered

heat kernel

the Gaussian fundamental solution K = (4πDt)^(−n/2) e^(−r²/4Dt) · Green's function of the diffusion equation

The heat kernel began as a computational trick and grew into a bridge between analysis, probability, geometry, and quantum physics. Joseph Fourier introduced the heat equation and the idea of building solutions by superposing elementary modes in his 1822 Théorie analytique de la chaleur; the Gaussian point-source solution is implicit there, though Fourier himself worked mostly with series on bounded intervals rather than the free-space kernel. The probabilistic meaning arrived with Einstein's 1905 analysis of Brownian motion, which identified the spreading Gaussian as the transition probability of a diffusing particle and gave $\langle x^2\rangle = 2Dt$ — the relation Jean Perrin soon used to weigh atoms and pin down Avogadro's number. Norbert Wiener made it rigorous in 1923 by constructing Brownian motion as a genuine measure on continuous paths (the Wiener process), so the kernel became the density of an honest random object, not just a PDE solution. From there the story turned geometric: Minakshisundaram and Pleijel derived the short-time expansion of the kernel's trace on a curved manifold in 1949, reading volume, area, and curvature off its leading coefficients, and Mark Kac crystallized the circle of ideas in 1966 with "Can one hear the shape of a drum?" The same Gaussian, continued to imaginary time, is the free quantum propagator behind Feynman's path integral — one formula at the crossroads of Fourier analysis, statistical physics, spectral geometry, and quantum field theory.

History

1822

— Fourier's Théorie analytique de la chaleur founds the heat equation and modal superposition

1905

— Einstein identifies the spreading Gaussian as the Brownian transition density, with mean-square displacement 2Dt

1923

— Wiener constructs Brownian motion rigorously as a measure on continuous paths

1949

— Minakshisundaram and Pleijel derive the short-time heat-kernel trace expansion on a manifold

1966

— Kac asks "Can one hear the shape of a drum?", framing modern spectral geometry

Going further

Texts. Stone & Goldbart, Mathematics for Physics, chapters on Green's functions and PDEs · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., Green's-function chapter · Hassani, Mathematical Physics, Green's-functions and PDE chapters · Barton, Elements of Green's Functions and Propagation · Vladimirov, Equations of Mathematical Physics; for the geometric and probabilistic sides, Kac's 1966 article "Can one hear the shape of a drum?" and Rosenberg, The Laplacian on a Riemannian Manifold.

Provenance. The free-space Gaussian, the similarity ODE $f''+2\eta f'+2f=0$ , the normalization $\int K\,\mathrm{d}x=1$ , and $\langle x^2\rangle=2Dt$ are verifiable in Stone & Goldbart, Arfken–Weber–Harris (7th ed.), and Hassani. A verifier should: (i) substitute the Gaussian into $\partial_t u = D\nabla^2 u$ ; (ii) confirm the Gaussian integral $\int e^{-Dk^2 t}e^{ikx}\,\mathrm{d}k = \sqrt{\pi/Dt}\,e^{-x^2/4Dt}$ ; (iii) check the image-kernel boundary cancellations $K_D(0)=0$ and $\partial_x K_N|_0=0$ ; (iv) verify the Wick rotation reproduces the standard free propagator $(m/2\pi i\hbar t)^{n/2}$ with $D=\hbar/2m$ . Worked numbers: solute $D=10^{-9}\ \mathrm{m^2/s}$ gives $\sigma=2.68\ \mathrm{mm}$ at $t=3600\ \mathrm{s}$ and $t=5.0\times10^4\ \mathrm{s}$ to reach $\sigma=1\ \mathrm{cm}$ ; copper rod $L=0.10\ \mathrm{m}$ , $D=1.11\times10^{-4}\ \mathrm{m^2/s}$ gives $\tau_1=L^2/\pi^2 D=9.1\ \mathrm{s}$ . No non-standard constants are asserted.

Prerequisites. The Heat (Diffusion) Equation · Classification of Second-Order Linear PDEs · Boundary Conditions and Well-Posedness · Sturm–Liouville Theory and Eigenfunction Expansions · Separation of Variables in Cartesian Coordinates. Leads to. The Helmholtz Equation · Laplace's and Poisson's Equations · The Wave Equation · Method of Characteristics.

Ask the tutor: Ask me to spread a specific initial temperature or dye profile with the kernel, build the image kernel for your boundary, or Wick-rotate it into the quantum propagator step by step.

What exactly is a "fundamental solution", and how is it different from an ordinary solution?

A fundamental solution is the response of a linear differential operator to a point source — a Dirac delta in space (and, for a time-dependent problem, in time). For the diffusion operator that response is the heat kernel. Its power: once you know it you never solve the equation from scratch again, because any initial distribution is a superposition of point sources, so the solution is the same superposition of kernels — a convolution. An ordinary solution answers one initial-value problem; the fundamental solution answers all of them at once. This is why Green's functions are the organizing idea of linear PDE theory.

Why does the heat kernel have a √t width instead of growing linearly like a wave?

Because diffusion is a random walk, not a march. A walker taking $N$ independent steps has a displacement whose variance grows as $N$ (variances add), so the typical distance grows as $\sqrt{N}$ — and since the step count is proportional to time, the spread grows as $\sqrt{t}$ . A wave propagates coherently at fixed speed, so its front advances linearly as $ct$ . The fingerprint is in the equations: the heat equation's one time derivative against two space derivatives forces $x\sim\sqrt{Dt}$ by dimensional balance, whereas the wave equation's two time derivatives give $x\sim ct$ .

Does the heat kernel really imply signals travel infinitely fast?

Within the pure diffusion equation, yes — and that is a genuine limitation of the model, not a paradox to explain away. The kernel is strictly positive everywhere for any $t>0$ , so a source is felt (exponentially faintly) arbitrarily far away immediately. This does not conflict with relativity: the heat equation is a coarse-grained approximation to an underlying microscopic dynamics (molecular collisions) that is causal. When infinite speed is unacceptable one upgrades to a hyperbolic heat equation (the Cattaneo equation), which adds a relaxation time and restores a finite propagation speed. For almost all laboratory and engineering purposes the parabolic model is superbly accurate, because the far tail carries negligible energy.

What is the semigroup property, and why can't I run diffusion backward?

Evolving for time $t$ then for time $s$ is identical to evolving once for $t+s$ : $K_t * K_s = K_{t+s}$ . The operators $e^{Dt\nabla^2}$ compose by adding time labels, exactly as Gaussian variances add. The word semigroup (not group) is the point: there is no inverse. Undoing a forward evolution would need a "negative-time" kernel, and the backward heat equation blows up its high-frequency modes as $e^{+Dk^2|t|}$ . Physically this is irreversibility — diffusion destroys information about fine initial structure, and no bounded operation recovers it.

How does the heat kernel encode the "shape" of a domain?

On a bounded domain the kernel is a sum over Laplacian eigenmodes, $K=\sum_n e^{-\lambda_n D t}\phi_n(x)\phi_n^*(x')$ . Set $x'=x$ , integrate to form the trace $\sum_n e^{-\lambda_n D t}$ , and expand for small $t$ : the leading term is the volume over $(4\pi Dt)^{n/2}$ (Weyl's law), the next is proportional to boundary area, and further terms carry curvature and topology. So how fast heat initially escapes a point reads off the geometry — the meaning of "hearing the shape of a drum": the eigenvalue spectrum determines much, though as it turns out not quite all, of the shape.

When does the free-space Gaussian stop being the right kernel?

The $(4\pi Dt)^{-n/2}$ Gaussian is exact only in infinite, homogeneous, isotropic space with constant $D$ . Three things break it. Boundaries: near a wall you add image sources (Dirichlet/Neumann) or, on a bounded domain, switch to the eigenfunction sum. Inhomogeneity or anisotropy: if $D$ varies in space or is a tensor, the kernel stays Gaussian-like locally but loses the simple isotropic form, and on a curved manifold it acquires Minakshisundaram–Pleijel curvature corrections. Nonlinearity or drift: an advection term or a state-dependent $D$ destroys the exact Gaussian entirely. In every case the free-space kernel remains the leading short-time, short-distance approximation, which is why it stays the right starting point.