= Solution
Use the <positive Laplace-Beltrami operator> $\Delta=-\operatorname{div}\nabla$. For the global discrete spectral statements, take $M$ to be a <closed manifold> with a smooth <Riemannian metric>. The <Riemannian heat kernel> is the integral kernel $H(t,x,y)$ of the <heat semigroup> $e^{-t\Delta}$:
$$
(e^{-t\Delta}f)(x)=\int_MH(t,x,y)f(y)\,dV_y,\qquad
(\partial_t+\Delta_x)H=0,\qquad H(0,x,y)=\delta_y(x).
$$
The initial condition is distributional with respect to the <Riemannian volume form>. Equivalently, an orthonormal basis of <Laplacian eigenfunctions> gives the <spectral expansion of the Riemannian heat kernel>
$$
H(t,x,y)=\sum_j e^{-t\lambda_j}\phi_j(x)\overline{\phi_j(y)}.
$$
The usual compact elliptic spectral theorem supplies this basis; polynomial <elliptic regularity> bounds and exponential decay prove convergence with every space derivative on $t\geq\varepsilon>0$. These analytic results may be used as clearly stated external ingredients. We will also construct the kernel locally and prove its short-time expansion.
On a sufficiently small neighbourhood of the diagonal, let $r=d(x,y)$ and write $x=\exp_y\xi$ in <geodesic normal coordinates>, with $dV_x=J(y,\xi)\,d\xi$. The <heat kernel expansion> has the form
$$
H(t,x,y)\sim(4\pi t)^{-m/2}e^{-r^2/(4t)}\sum_{j\geq0}t^ju_j(x,y),\qquad m=\dim M.
$$
The <heat invariants> are the integrated diagonal coefficients $a_j=\int_Mu_j(x,x)\,dV_x$, so
$$
\operatorname{Tr}(e^{-t\Delta})\sim(4\pi t)^{-m/2}\sum_{j\geq0}a_jt^j.
$$
In particular
$$
a_0=\operatorname{vol}(M),\qquad a_1=\frac16\int_M R\,dV,\qquad
a_2=\frac1{360}\int_M(5R^2-2|\operatorname{Ric}|^2+2|\operatorname{Rm}|^2)\,dV.
$$
Here $R$ is <scalar curvature>, $\operatorname{Ric}$ is the <Ricci tensor> and $\operatorname{Rm}$ is the <Riemann curvature tensor>. The last expression is the <integrated second scalar heat coefficient>; a local divergence term integrates to zero on a <closed manifold>. The exponent of the leading heat-trace term determines dimension, and these coefficients determine volume, total <scalar curvature>, and the displayed integrated curvature combination. On a surface $R=2K$, where $K$ is <Gaussian curvature>, so $a_1=\frac13\int K\,dA=\frac{2\pi}{3}\chi(M)$ by the <Gauss-Bonnet theorem>, and $a_2=\frac1{15}\int K^2\,dA$. Thus the <Euler characteristic>, and the genus of a connected orientable closed surface, are determined. The first three coefficients also detect constant <Gaussian curvature>, since they determine $\int(K-\overline K)^2\,dA$. These conclusions do not claim that the pointwise curvature function or the metric is determined.
Here is the local construction. Put $g_t=(4\pi t)^{-m/2}e^{-r^2/(4t)}$. The <Gauss lemma> and the radial expression for the geometric Laplacian give, for a smooth amplitude $u$,
$$
(\partial_t+\Delta_x)(g_t t^ju)
=g_t\left[t^{j-1}\left(r\partial_ru+\frac r2\partial_r\log J\,u+ju\right)+t^j\Delta_xu\right].
$$
The $t^{-2}$ terms cancel because $|\nabla r|=1$. Define $D=r\partial_r+(r/2)\partial_r\log J$. The singular term disappears with $u_0=J^{-1/2}$, normalized by $u_0(y,y)=1$. For $j\geq1$, cancel the remaining powers of $t$ by the <heat-kernel transport equations>
$$
(D+j)u_j=-\Delta_xu_{j-1},
$$
which have the explicit solution
$$
u_j(\exp_y\xi,y)=-J(y,\xi)^{-1/2}\int_0^1s^{j-1}J(y,s\xi)^{1/2}(\Delta_xu_{j-1})(\exp_y(s\xi),y)\,ds.
$$
This formula is obtained by differentiating $r^jJ^{1/2}u_j$ along each radial geodesic. It proves smoothness jointly in $x,y$, including the diagonal, by induction: the integrand is smooth on a fixed compact integration interval. A homogeneous solution of $(D+j)u=0$ has $r^jJ^{1/2}u$ constant along a ray, and smoothness at $r=0$ forces this constant to vanish for $j>0$. Thus the smooth transport coefficients are unique. In particular $u_j(y,y)=-(\Delta_xu_{j-1})(y,y)/j$. Expanding $J=1-\operatorname{Ric}_{ab}(y)\xi^a\xi^b/6+O(|\xi|^3)$ gives $u_1(y,y)=R(y)/6$, checking the sign convention.
Choose a smooth cutoff $\eta(x,y)$ that is one near the diagonal and supported where these <normal coordinates> exist, and form the <heat parametrix>
$$
P_N(t,x,y)=\eta(x,y)g_t(x,y)\sum_{j=0}^Nt^ju_j(x,y).
$$
It tends to the identity kernel as $t\downarrow0$: in <normal coordinates> the Gaussian is an <approximate identity>, and $u_0J=J^{1/2}\to1$ at the centre. Its residual $R_N=(\partial_t+\Delta_x)P_N$ is $\eta g_tt^N\Delta_xu_N$ plus cutoff-derivative terms supported a positive distance from the diagonal. Those latter terms are smaller than every power of $t$. Hence $R_N=O(t^{N-m/2})$ uniformly. Every prescribed finite number of derivatives has the same kind of estimate after increasing $N$; differentiating a Gaussian costs only finitely many powers of $t$.
For $N$ large enough that the residual is continuous and bounded up to time zero, use the <Volterra parametrix correction>. For kernels define
$$
(F*G)(t,x,y)=\int_0^t\int_MF(t-s,x,z)G(s,z,y)\,dV_z\,ds.
$$
Set $Q=-R_N+R_N*R_N-R_N*R_N*R_N+\cdots$. A bounded residual with bound $C$ gives
$$
|R_N^{*k}(t,x,y)|\leq C^k\operatorname{vol}(M)^{k-1}\frac{t^{k-1}}{(k-1)!},
$$
so this series converges and $Q+R_N+R_N*Q=0$. Because $P_N$ has identity initial kernel, differentiating the convolution gives $(\partial_t+\Delta_x)(P_N*Q)=Q+R_N*Q$. Thus
$$
\boxed{H=P_N+P_N*Q}
$$
is an exact kernel with the correct initial condition. Taking $N$ sufficiently large gives any required finite regularity; uniqueness of the heat equation makes the kernels obtained from different $N$ equal. For uniqueness, the difference applied to smooth initial data has zero initial value, and the energy identity $\frac d{dt}\|v\|_2^2=-2\|\nabla v\|_2^2$ forces $v=0$. Parabolic regularity then gives smoothness for positive time.
The Gaussian part of $P_N$ has uniformly bounded spatial $L^1$ norm, so the correction is $O(t^{N+1-m/2})$. This proves, uniformly near the diagonal,
$$
H(t,x,y)-g_t(x,y)\sum_{j=0}^Nt^ju_j(x,y)=O(t^{N+1-m/2}),
$$
with corresponding derivative estimates on increasing the truncation order. In particular on every parabolic neighbourhood $r\leq B\sqrt t$ it is the stated Gaussian expansion with relative remainder $O(t^{N+1})$, and on the diagonal it yields the complete <heat trace> expansion. The standard off-diagonal Gaussian version on compact subsets of a sufficiently small convex normal neighbourhood follows from the same convolution estimates with the Gaussian factor retained: the inequality
$$
\frac{d(x,z)^2}{t-s}+\frac{d(z,y)^2}{s}\geq\frac{d(x,y)^2}{t}
$$
localizes the convolution to the unique joining geodesic; its transverse Hessian is positive, and Gaussian integration gives the successive transport orders. The additive estimate above already specifies an appropriate uniform near-diagonal remainder without asserting a relative estimate far from the diagonal on the basis of a crude bound.
If $M$ is noncompact, use the minimal <Friedrichs extension> heat semigroup; if there is a boundary, specify boundary conditions. The same construction is local in the interior. One may extend a relatively compact coordinate neighbourhood to a closed auxiliary manifold; localization errors are supported a positive distance away and are exponentially small near the smaller diagonal neighbourhood. Thus the local coefficients remain the same. Finite integrated <heat invariants> and a discrete <spectrum> require the global hypotheses used above.
Back to article page