Heat kernel

For $u_t=\kappa u_{xx}$ on the real line,has total mass one and converges to the delta distribution as $t$ decreases to zero. The solution is $u(t)=K_t\ast u_0$.

The displayed fundamental solution solves $u_t=u_{xx}-cu$ on the real line with a Dirac delta at the initial time. It is the ordinary heat kernel multiplied by $e^{-ct}$, because that factor removes the constant reaction term. Its integral is $e^{-ct}$ and it converges to the Dirac delta as $t\downarrow 0$. Convolution with initial data gives the corresponding solution. For $c>0$ the total mass decays by absorption; for $c<0$ it grows instead.

On ${\mathbb{R}}^d$, the heat equation with initial Gaussian function $u_0(x)=e^{-|x-x_0{|}^2/a}$, $a>0$, has the displayed solution. Convolving the initial profile with the heat kernel adds the Gaussian width parameters; conservation of its spatial integral fixes the prefactor. Direct differentiation provides another proof. The width increases while the peak decreases, with no loss of the total integral.

On a closed manifold, the positive Laplace-Beltrami operator has a smooth heat kernel and its heat semigroup is a trace-class operator for positive time. Expanding in a smooth orthonormal eigenbasis gives $K_t(x,y)={\sum }_j{e}^{-t{\lambda }_j}{\varphi }_j(x)\overline{{\varphi }_j(y)}$. Elliptic regularity bounds each derivative of ${\varphi }_j$ polynomially in ${\lambda }_j$, while the exponential dominates those powers, so the expansion converges with all derivatives. Integrating the diagonal and using the unit norm of each eigenfunction proves the formula. For a bundle operator, insert the fibre matrix trace inside the integral.

The real Gaussian function heat kernel $(4\pi \epsilon z{)}^{-1/2}\exp [-(\varphi -\theta {)}^2/(4\epsilon z)]$ is integrable only when $\epsilon z>0$. For forward time $z>0$, this requires positive diffusivity. A Cole-Hopf transformation is an algebraic substitution for any nonzero coefficient, but that does not make the real forward convolution valid for negative diffusivity. Constant initial heat data already produce a divergent integral with the wrong sign.

For the heat equation on ${\mathbb{R}}^n$ with diffusion coefficient one, the heat kernel is $K_t(x)=(4\pi t{)}^{-n/2}{e}^{-|x{|}^2/(4t)}$. Its convolution with bounded initial data is smooth for positive time, solves the equation by differentiation under the integral, and returns continuous initial data locally uniformly through the approximate identity property. It is Gaussian filtering with per-coordinate variance $2t$.

For the positive Laplace-Beltrami operator, the kernel solving $({\partial }_t+{\Delta }_x)H=0$ with initial Dirac delta distribution on the diagonal represents the heat semigroup. On a closed manifold with a Riemannian metric it is smooth for $t>0$, symmetric, nonnegative, preserves constants, and obeys the semigroup property.

For a finite free action by Riemannian isometries, the Riemannian heat kernel on the quotient is the image sum shown above. Integrating over a fundamental domain combines all images into the integral on the covering manifold, proving the initial condition and showing that there is no averaging factor in the kernel. The heat trace does have an averaging factor $1/|U|$, because the integral of a quotient function over the cover is $|U|$ times its quotient integral. Mere freeness of a general nondiscrete group action is not enough for this covering formula.

The normalized pullback is a unitary map from $L^2$ of the quotient onto the deck-invariant subspace of $L^2$ of the cover. Local isometry and the sheet-counting integration formula preserve the Dirichlet quadratic form. Therefore the Friedrichs extension and its heat semigroup intertwine with this map. A finite sum of lifted Riemannian heat kernels gives the quotient kernel; independence of lifts follows by reindexing the deck group. This formulation avoids a uniqueness claim for unrestricted heat-equation solutions on a noncompact or incomplete manifold.

On a closed manifold with a Riemannian metric, an orthonormal eigenbasis for the positive Laplace-Beltrami operator gives this formula, with eigenvalues counted with multiplicity. Polynomial elliptic bounds and exponential time decay give convergence of every derivative away from time zero. The conjugate is necessary for a complex basis.

An approximate Riemannian heat kernel with the correct delta initial limit and a residual extending with enough regularity to time zero. A cutoff Gaussian times transport coefficients supplies a local construction. Increasing the expansion order, or summing the full asymptotic series smoothly, makes the residual regular enough for a Volterra parametrix correction.

In normal coordinates $q={\exp }_{p}x$, write the Riemannian volume form as $J(p,x)dx$ and let $r=|x|$. With the heat operator ${\partial }_t-{\Delta }_q$, the Gaussian times a power-series amplitude has its singular terms cancelled by the displayed radial equations for $i\ge 1$. The initial coefficient is $w_0=J^{-1/2}$, andMultiplying the equation by $r^{i-1}{J}^{1/2}$ gives the derivative of $r^i{J}^{1/2}{w}_i$. Smoothness forces the integration constant to vanish. The integral on a fixed compact interval proves smoothness across the diagonal by induction; it yields $w_i(p,p)={\Delta }_q{w}_{i-1}(p,p)/i$. A truncated heat parametrix therefore has residual $-g{t}^k{\Delta }_q{w}_k$.

If $R=({\partial }_t+{\Delta }_x)P$ is smooth and bounded up to time zero on a closed manifold, the Volterra convolution of kernels solves $Q+R+R\ast Q=0$ by $Q={\sum }_{k\ge 1}(-1{)}^k{R}^{\ast k}$. Bounds by $C^k\mathrm{vol}(M{)}^{k-1}{t}^{k-1}/(k-1)!$ prove convergence. Then $H=P+P\ast Q$ is the exact Riemannian heat kernel and has the same initial delta limit.

The normalized Gaussian mass of an interval isIt is the convolution of an interval indicator with the heat kernel, and has an error function representation. As the variance tends to zero it converges to the interval indicator away from its endpoints, with value one half at an endpoint. This is a concrete approximate identity.

For $x>0$ and the heat equation on a half-line, the zero-initial-data solution with prescribed Dirichlet boundary condition $g$ is ${\int }_0^{t}P(x,t-s)g(s)ds$, whereIts mass concentrates at $t=0$ as $x\downarrow 0$, recovering the boundary trace.

For the heat equation with zero Dirichlet boundary conditions on $(0,L)$,It is also obtained by the method of images, taking the difference between Gaussian sources at $y+2jL$ and $-y+2jL$. The eigenfunction expansion is useful at long times; the image expansion is useful at short times.

A transition density with absorbing boundaries counts the joint event of survival up to time $t$ and a specified final position. Its integrated mass is survival probability, not generally one. Dividing by that mass gives the endpoint density conditional on survival, a different object.

For diffusion starting at the centre of $(-L,L)$ with absorbing endpoints, only even-parity spatial eigenfunctions contribute. Integrating their cosine expansion gives the displayed series. The leading late-time term has prefactor $4/\pi$ and decay rate ${\pi }^{2}D/(4L^2)$; its first relative correction is $-\frac{1}{3}{e}^{-2{\pi }^{2}Dt/L^2}$.

For a particle with mass $m$ on $(0,L)$, put $K_0(x;\beta )=\sqrt{m/(2\pi {\hslash }^2\beta )}{e}^{-m{x}^2/(2{\hslash }^2\beta )}$. The method of images gives $K_D(q_f,q_i;\beta )={\sum }_r[K_0(q_f-q_i+2rL;\beta )-K_0(q_f+q_i+2rL;\beta )]$. The diagonal integral isThe reflected intervals tile the real line, giving the subtraction $1/2$. The Poisson summation formula converts this to ${\sum }_{n\ge 1}{e}^{-\beta {\hslash }^2{\pi }^2{n}^2/(2m{L}^2)}$, proving equality between the image and energy eigenstate calculations.

For standard Brownian motion in ${\mathbb{R}}^d$, $p_t(y-x)$ is the transition density from $x$ to $y$. It is the heat kernel for generator $\frac{1}{2}\Delta$. For every unit vector $e$, $\int |{\partial }_e{p}_t(x)|dx=\sqrt{2/\pi }/\sqrt{t}$, which bounds the change in this density under a spatial translation and proves the Gaussian heat-kernel proof of the harmonic Liouville theorem.

For $D,\gamma >0$, the displayed function solves the heat equation $u_t=D{u}_{xx}$ and has initial condition $u(x,0)=A{e}^{-\gamma (x-x_0{)}^2}$. Set $s=1+4\gamma Dt$. Direct differentiation givesThese identities prove the equation. The total integral is $A\sqrt{\pi /\gamma }$, independent of time. The real smooth formula is valid for $t>-1/(4\gamma D)$, including a finite interval of negative time for this particular initial profile. For nonzero $A$, its value at $x_0$ diverges as the lower time endpoint is approached. The superposition principle combines finitely many such profiles with different centers or widths.

For $D,\gamma ,a>0$, apply the superposition principle to two translated solutions for Gaussian initial data for the heat equation. The displayed solution of $u_t=D{u}_{xx}$ has odd initial profile $A(e^{-\gamma (x-a{)}^2}-e^{-\gamma (x+a{)}^2})$. Its total spatial integral is zero. For $A>0$, it is positive on $x>0$ and negative on $x<0$ at every $t\ge 0$.

Writing it as $2A{s}^{-1/2}{e}^{-\gamma (x^2+a^2)/s}\sinh (2\gamma ax/s)$ shows how cancellation affects its late-time shape. On the diffusive scale $x=O(\sqrt{s})$, its leading form is $4A\gamma ax{s}^{-3/2}{e}^{-\gamma x^2/s}$. Thus the two lobe maxima asymptotically lie at $x=\pm \sqrt{s/(2\gamma )}$, while their absolute heights decay as $s^{-1}$. Their centers are not fixed at the initial Gaussian centers. The first spatial moment remains ${\int }_{\mathbb{R}}xu(x,t)dx=2Aa\sqrt{\pi /\gamma }$, as follows directly from the Gaussian integral and translation.

For $d\ge 3$, $A_d=\Gamma (d/2-1)/(2{\pi }^{d/2})$. Substitute $u=|x{|}^2/(2s)$ to obtain the displayed identity for $x\ne 0$. The semigroup property and Tonelli theorem then give $P_t(|\cdot {|}^{2-d})(x)=A_d^{-1}{\int }_t^{\infty }{p}_s(x)ds$. The coefficient multiplying the time integral is the reciprocal of the coefficient in the full Green integral, a distinction important in Brownian motion hitting asymptotics.

For $L>0$ and $t>0$, the Sturm-Liouville eigenfunction expansion for $u_t=u_{xx}$ on $(0,L)$ with homogeneous Neumann boundary conditions givesThe constant eigenfunction mode preserves the spatial integral. This is the $\beta =0$ limit of the weighted Neumann heat kernel with constant drift, and its boundary inputs have the signs in the Neumann boundary-forcing formula.

The heat semigroup acts by convolution with the heat kernel: $P_{t}f=K_t\ast f$. The identity $K_s\ast K_t=K_{s+t}$ gives $P_s{P}_t=P_{s+t}$, and its infinitesimal generator is a constant multiple of the Laplacian.

For a time-independent strictly positive Dirichlet realization of an elliptic operator and $\psi \in L^2(U)$, the eigenfunction expansionsolves $u_t+Lu=0$ with homogeneous Dirichlet boundary conditions. For $t\ge \epsilon >0$, every power ${\lambda }_m^N$ times the exponential is bounded, so the series and all its time derivatives converge in all the Sobolev domains of powers of an elliptic Dirichlet operator. Elliptic regularity and the Sobolev embedding theorem give smoothness up to the spatial boundary for positive time. Parseval identity and the dominated convergence theorem give $u(t)\to \psi$ in $L^2$ as $t\downarrow 0$.

A heat kernel expansion is the short-time asymptotic expansion of the trace or diagonal kernel of an elliptic operator. Its local coefficients are polynomials in background curvatures and field strengths; in quantum field theory they extract ultraviolet divergences and anomalies.

For the positive scalar Laplace-Beltrami operator on a compact manifold without boundary, the heat invariants are the integrated coefficients $a_j={\int }_M{u}_j(x,x)d{V}_g$ in the short-time heat kernel expansion. In particular $a_0=\mathrm{Vol}(M)$ and $a_1=\frac{1}{6}{\int }_M{R}_{g}d{V}_g$, with $R_g$ the scalar curvature. They are determined by the heat trace, hence by the spectrum with multiplicities. The leading positive coefficient recovers the dimension from ${\lim }_{t\downarrow 0}\log \mathrm{Tr}(e^{-t\Delta })/\log (1/t)=d/2$.

For the positive scalar Laplace-Beltrami operator on a closed manifold with a Riemannian metric, the coefficient of $t^2$ in the normalized heat trace is the displayed integral. The local coefficient additionally has a scalar-curvature divergence term whose integral vanishes. In dimension two, $R=2K$, $|\mathrm{Ric}{|}^2=2K^2$, $|\mathrm{Rm}{|}^2=4K^2$, giving $a_2=\frac{1}{15}{\int }_M{K}^{2}dA$. Together with volume and total Gaussian curvature, it determines the variance of curvature and detects constant curvature.

On a closed two-dimensional Riemannian manifold, the first three integrated scalar heat invariants are $A$, $\frac{1}{3}\int KdA$, and $\frac{1}{15}\int K^{2}dA$, where $K$ is Gaussian curvature. If its scalar spectrum agrees with that of a hyperbolic surface of area $A$, these integrals are $A,-A/3,A/15$. Expanding the displayed square then forces $K=-1$ everywhere. Thus constant negative curvature is spectrally determined in this setting, although the global hyperbolic metric need not be determined up to isometry.

For $L=-{\nabla }^a{\nabla }_a$ on scalar functions, the leading coefficient of the heat parametrix in normal coordinates is the inverse square root of the metric volume density: $u_0=1+R_{ab}{\xi }^a{\xi }^b/12+O(|\xi {|}^3)$. The heat-kernel transport equations give $u_1(x,x)={\nabla }^a{\nabla }_a{u}_0(x,x)=R(x)/6$. This gives the displayed diagonal expansion. The coefficient is specific to the minimally coupled scalar operator; a bundle connection or a curvature-dependent potential changes the general Laplace-type coefficients.

For $L=-{\partial }_j(g^{ij}{\partial }_i)$ in geodesic normal coordinates, the Gauss lemma gives $g^{ij}{x}_i=x_j$. Substituting $F=(4\pi t{)}^{-n/2}{e}^{-|x{|}^2/(4t)}{\sum }_{k\ge 0}{t}^k{B}_k$ into $({\partial }_t+L)F=0$ therefore yields $(x\cdot \nabla +k)B_k=-L{B}_{k-1}$, with $B_0=1$. Integration along radial rays gives the displayed formula. Every coefficient is smooth at the centre. Any smooth solution of $(x\cdot \nabla +k)h=0$ is zero for $k>0$, since $s^{k}h(sx)$ is constant and tends to zero at $s=0$. This proves uniqueness. In this special divergence-form case $L1=0$, so every $B_k$ for $k>0$ is zero: the Gaussian itself solves the equation exactly on the coordinate ball. The positive Laplace-Beltrami operator instead has a metric volume-density term and a nonconstant leading amplitude.

The centered normal densitiesform an approximate identity: for every $f\in L^1(\mathbb{R})$, $f\ast g_t\to f$ at almost every point where the Lebesgue differentiation theorem applies. Their Fourier transforms are $e^{-t{\xi }^2/2}$ in the angular-frequency convention.

For initial data $u_0$ on the real line, Fourier transformation or convolution with the fundamental solution gives

For a bounded function $u_0$, the convolution is well defined at every positive time because the heat kernel has integral one. The error-function diffusion front uses a Heaviside step function initial condition and gives one smoothed interface. The heat equation with interval-indicator initial data instead has two interfaces; subtracting their two error functions produces the finite-interval profile.

Let $\phi$ be continuous and suppose that the displayed improper integral converges separately at both infinite endpoints for some $T>0$. Absolute convergence is not required. Then its heat-kernel solution exists, is smooth and solves the heat equation for $0<t<T$, and tends to $\phi (x)$ as $t\downarrow 0$ at each fixed $x$.

To prove the smoothing claim, take a primitive $H$ of $\phi (\xi )e^{-{\xi }^2/(4T)}$. The assumed convergence makes $H$ bounded. For $q(x,\xi ,t)=K_t(x-\xi )e^{{\xi }^2/(4T)}$, integration by parts givesThe endpoint term vanishes. The coefficient of ${\xi }^2$ in the exponent of $q$ is $-(T-t)/(4tT)<0$. Its derivatives therefore have Gaussian decay on every compact subset of $\mathbb{R}\times (0,T)$, justifying differentiation under the integral sign in the absolutely convergent right-hand expression. Since $q_t=q_{xx}$, it satisfies the heat equation.

For the initial limit, apply the same construction to $\phi (\xi )-\phi (x)$ and split at $|\xi -x|=\delta$. Continuity controls the local part by its small oscillation times the unit kernel mass. On each remaining tail, $q$ is monotone for sufficiently small $t$: its maximum lies at $\xi =xT/(T-t)$, approaching $x$. If the relevant primitive has absolute bound $M$, integration by parts bounds that tail by $2M$ times the value of $q$ at its finite endpoint. For either endpoint $\xi =x\pm \delta$, that value tends to zero like $t^{-1/2}{e}^{-{\delta }^2/(4t)}$ times a fixed constant. Both tails vanish, proving the limit. Separate endpoint convergence matters here; merely assuming a symmetric Cauchy principal value does not provide the bounded primitive used in this proof.

Suppose $\phi$ is continuous at $x$ and, for some $T>0$,For $0<t<T$, its convolution with the heat kernel $K_t(s)=(4\pi t{)}^{-1/2}{e}^{-s^2/(4t)}$ is smooth and solves the heat equation. On every compact set in $\mathbb{R}\times (0,T)$, spatial or time derivatives of the heat kernel are bounded by the displayed integrable weight times a constant: their polynomial factors are absorbed by the strictly narrower Gaussian. This justifies differentiation under the integral sign.

Moreover, $\int K_t=1$, soInside $|\xi -x|<\delta$, continuity makes the difference uniformly small. Outside that interval, for $t\le T/4$, the inequality ${\xi }^2\le 2(\xi -x{)}^2+2x^2$ givesThe right side tends to zero, proving that the weighted tail contribution vanishes; the constant term $\phi (x)$ is controlled by the vanishing Gaussian tail mass. Thus $(K_t\ast \phi )(x)\to \phi (x)$ as $t\downarrow 0$. In particular, continuous data of polynomial growth satisfy this hypothesis, even when they are unbounded. This extends the Gaussian approximate identity argument beyond bounded data.

For $u_t=D{u}_{xx}$ and $u(x,0)=1_{[-a,a]}(x)$, convolution with the heat kernel gives

For the heat equation $u_t=\frac{1}{2}{u}_{xx}$ on $x>0$ with the Neumann boundary condition $u_x(t,0)=0$, the method of images givesThus $u(t,x)={\int }_0^{\infty }{K}_t^N(x,y)f(y)dy$. This is also the transition kernel of Reflected Brownian motion.