The heat equation

We now consider the heat equation $$\label{e:H}\tag{H} \partial_t u - \lap u = 0 $$

Fundamental solution in $\R^d$

Remark 1. If $u$ is a solution to \eqref{e:H} that decays sufficiently rapidly in space, then $\partial_t \int_{\R^d} u \, dx = 0$.

Remark 2. If $u$ is a solution to \eqref{e:H} then so is $u_\lambda \defeq \lambda^{d} u(\lambda^2 t, \lambda x)$.

As a result it’s natural to look for scale invariant solutions of the form $$\label{e:R}\tag{R} u_t(x) = u(t, x) = \frac{1}{t^{d/2}} f\paren[\Big]{ \frac{\abs{x}}{\sqrt{t}} } \,. $$

Proposition 3. If a function $u$ of the form \eqref{e:R} solves \eqref{e:H}, then $$ f(r)=e^{-r^2/4}\left( c_1 + c_2 \int_1^r e^{s^2/4} \, \frac{ds}{s^{d - 1}} \right) \,. $$

Definition 4. The fundamental solution of the heat equation is $$ \phi_t(x) = \frac{1}{(4 \pi t)^{d/2}} e^{-\abs{x}^2 / (4t)} $$

Remark 5. The constants above are chosen so that $\int_{\R^d} \phi_t \, dx = 1$. Physically, the fundamental solution represents the temperature at $(x, t)$ if the initial temperature is a delta distribution at $0$.

Remark 6. The fundamental solution $\phi_t$ is the density of the normal distribution with mean $0$ and variance $2t$.

Proposition 7. Let $g \in C(\R^d) \cap L^\infty(\R^d)$ and define $$ u_t(x) = \phi_t * g(x) = \int_{\R^d} \phi_t(x - y) g(y) \, dy \,. $$ Then $u \in C^\infty(\R^d \times (0, \infty) )$, solves \eqref{e:H}, and $$\label{e:I}\tag{I} \lim_{t \to 0} u_t(x) = g(x) \,. $$

Remark 8 (Infinite speed of propagation). If $g \geq 0$ is not identically $0$ then $u_t(x) > 0$ for all $t > 0$, $x \in \R^d$. In other words, the heat from any localized initial temperature distribution is felt immediately everywhere.

Proposition 9 (Duhamel’s principle). Suppose $f \in C^{1, 2}_c(\R^d)$, $g \in C(\R^d) \cap L^\infty$, and define $$ u_t(x) = \phi_t * g + \int_0^t \phi_{t - s} * f_s \, ds = \int_{\R^d} \phi_t(x - y) g(y) \,dy + \int_0^t \int_{\R^d} \phi_{t - s}(x - y) f_s(y) \, dy \, ds \,. $$ Then $u \in C^{1,2}((0, \infty) \times \R^d)$, solves $\partial_t u - \lap u = f$ with initial condition \eqref{e:I}.

Mean value property

Definition 10. Given $x \in \R^d$, $t \geq 0$, $r > 0$ define the heat ball $$ E(t, x, r) = \set{ (s, y) \in \R^{d+1} \st s \leq t, \phi_{t -s}(x - y) \geq \frac{1}{r^d}} $$

Remark 11. Note $(t -h, x - z) \in E(t, x, r)$ if and only if $$ 4 \pi h \leq r^2 \quad\text{and}\quad \abs{z}^2 \leq 2d h \ln \paren[\Big]{\frac{r^2}{4 \pi h}} \,. $$

Theorem 12 (Mean value property). Suppose $u$ is a $C^{1, 2}$ solution to \eqref{e:H}. Then $$ u_t(x) = \frac{1}{4 r^d} \int_{E(t, x, r)} u_s(y) \frac{\abs{x - y}^2}{(t - s)^2} \, dy \, ds \,. $$

Remark 13. The intuition here is to set $\varphi = \log(r^d G(s, y))$, where $G(s, y) = \phi(t - s, x - y)$. Now $$ (\partial_s + \lap_y)(u \varphi) = 2 \grad u \cdot \grad \varphi - u \abs{\grad \varphi}^2 $$ and $\abs{\grad_y \varphi}^2$ is the weight in the mean value property. Now expanding and integrating formally gives the mean value property.

For harmonic functions, level sets of the Newton kernel were balls, and $\grad \Gamma$ is constant on the boundary of balls. For the heat equation, level sets of the “backward heat kernel”, $G$, are heat balls; but the gradient of $G$ isn’t constant on the boundary leading to the weights.

Making the above intuitive argument work requires a bit of care.

Proof sketch.

  1. Without loss of generality $x = 0$, $t = 0$, and write $E_r = E(0, 0, r)$.
  2. Define $\displaystyle F(r) = \frac{1}{r^d} \int_{E_r} \frac{u_s(y) \abs{y}^2}{s^2} \, dy \, ds = \int_{E_1} \frac{u_{r^2 s}(r y) \abs{y}^2}{s^2} \, dy \, ds \,. $
  3. Compute $F’(r) = I_1 + I_2$ where $$ I_1 = \frac{1}{r^{d+1}} \int_{E_r} \partial_i u \frac{y_i \abs{y}^2}{s^2} \, dy \, ds \quad\text{and}\quad I_2 = \frac{1}{r^{d+1}} \int_{E_r} 2 \partial_s u \frac{\abs{y}^2}{s} \, dy \, ds \,. $$
  4. Let $\displaystyle \psi \defeq \frac{\abs{y}^2}{4s} + \frac{d}{2} \ln\paren[\Big]{\frac{-4\pi s}{r^2}} $ and note $\psi = 0$ on $\partial E_r$.
  5. So $I_2 = \frac{1}{r^{d+1}} \int_{E_r} 4 \partial_s u y \cdot \grad \psi$, integrate by parts (a few times) use \eqref{e:H} and check $I_1 + I_2 = 0$.
  6. Check $\lim_{r \to 0} F(r) = 4 u_0(0)$ to conclude.

Strong Maximum principle

Definition 14. Let $U \subseteq \R^d$ be open.

  1. The parabolic cylinder $U_T$ is $(0, T] \times U$ (including the top). (Sometimes $U_T$ is called parabolic interior of $[0, T] \times \bar U = \bar U_T$).
  2. The parabolic boundary of $U_T$ is $\Gamma_T \defeq \bar U_T - U_T = [0, T] \times \partial U \cup \set{0} \times U$

Theorem 15. Suppose $U$ is a bounded domain, $T > 0$, and $u \in C^{1, 2}(U_T) \cap C(\bar U)$ solves \eqref{e:H} in $U_T$.

  1. The maximum of $u$ in $\bar U_T$ is attained on the parabolic boundary $\Gamma_T$ (i.e. $\max_{\bar U_T} u = \max_{\Gamma_T} u$).
  2. If $u$ attains its maximum at some (parabolic) interior point $(t_0, x_0) \in U_T$, then $u$ is constant on $\bar U_{t_0}$.

Proof sketch. By the mean value property, $u$ must be constant on any heat ball based at $(t_0, x_0)$. Now join $(t_0, x_0)$ to $(s, y)$ by a piecewise linear path (that always has a strictly negative slope in the $t$ direction), and show that $u(s, y) = u(t_0, x_0)$.

Corollary 16. Let $f \in C(U_T)$, $g \in C(\Gamma_T)$. There is at most one $C^{1, 2}(U_T) \cap C(\bar U_T)$ solution of \begin{align} \partial_t u - \lap u &= f &&\text{in } U_T \,,\\ u &= g && \text{on } \Gamma_T \,. \end{align}

Maximum principles for parabolic equations.

Let $L$ be an uniformly elliptic operator, with bounded continuous coefficients (which are now also allowed to depend on time), and suppose $c \leq 0$.

Theorem 17 (Weak maximum principle). Suppose $U$ is bounded, $u \in C^{1,2}(U_T) \cap C(\bar U_T)$ and $\partial_t u - Lu \leq 0$ in $U_T$.

  1. A nonnegative maximum of $u$ is attained on the parabolic boundary. That is, $$ \max_{\bar U_T} u \leq \max_{\Gamma_T} u_+\,. $$
  2. If $c = 0$, then any (not necessarily nonnegative) maximum of $u$ is attained on the boundary. That is, $$ \max_{\bar U_T} u = \max_{\Gamma_T} u\,. $$

Proof sketch. First assume $(\partial_t - L)u < 0$ and show that $u$ cannot attain a nonnegative maximum in $U_T$. For the general case, use $u_\epsilon = u - \epsilon t$ and send $\epsilon \to 0$. (Your old barrier trick of $u_\epsilon = u + \epsilon e^{k x_1}$ will also work.)

Lemma 18 (Parabolic Hopf lemma). Suppose $u \in C^{1,2}(U_T) \cap C(\bar U_T)$ satisfies $(\partial_t - L)u \leq 0$. Let $(t_0, x_0) \in (0,T] \times \partial U$ be a point at which $u$ attains a nonnegative maximum $M$ on $\bar U_{t_0}$. Suppose there is an interior ball $B = B(y,R) \subseteq U$ tangent to $\partial U$ at $x_0$, and $u < M$ in $(0,t_0] \times B$. If $u$ is $C^1$ at $(t_0,x_0)$, then $\hat n \cdot \grad u(t_0,x_0) > 0$.

Proof sketch. On the annular cylinder $$ Q = (B(y,R) - \bar B(y,R/2)) \times (t_0-\delta,t_0] $$ construct a barrier of the form $$ v(t,x) = e^{-\alpha\abs{x-y}^2 - \beta(t_0-t)} - e^{-\alpha R^2}\,. $$ Choose $\alpha,\beta$ sufficiently large (and $\delta$ sufficiently small) so that $(\partial_t - L)v < 0$ in $Q$, and $v \leq 0$ on the bottom and outer spatial boundary. Apply the weak maximum principle to compare $u$ with $M-\epsilon v$.

Theorem 19 (Strong maximum principle). Suppose $U$ is a bounded connected domain, $u \in C^{1,2}(U_T) \cap C(\bar U_T)$ and $(\partial_t - L) u \leq 0$ in $U_T$.

If $u$ attains a nonnegative maximum on $\bar U_T$ at some $(t_0,x_0) \in U_T$, then $u$ is constant on $\bar U_{t_0}$.

Proof sketch. Use the parabolic Hopf lemma on an interior ball that touches the set where $u$ attains its maximum. Show that the maximum propagates spatially and backward in time.

The Cauchy problem

Proposition 20. Suppose $u \in C^{1, 2}((0, T] \times \R^d) \cap C([0, T] \times \R^d)$ solves \eqref{e:H}, and there exists $a, A > 0$ such that for all $t \in [0, T]$, $x \in \R^d$ we have $$\label{e:G}\tag{G} u_t(x) \leq A e^{a \abs{x}^2} \,. $$ Then $\sup \set{ u_t(x) \st t \in [0, T], x \in \R^d} \leq \sup_{\R^d} g $.

Proof sketch.

  1. For some constant $T_1 > 0$ to be chosen shortly, define $$ \psi = \frac{1}{(2 T_1 - t)^{d/2}} e^{\textcolor{red}{+} \abs{x}^2 / (4(2T_1 - t))} $$
  2. Let $y \in \R^d$ be arbitrary, $v_t(x) = u_t(x) - \epsilon \psi_t(x - y)$, and observe $v$ satisfies \eqref{e:H}.
  3. Let $r > 0$, $U = B(y, r)$, and check for $T_1 = 1/8a \varmin T$ and large enough $r = r(\epsilon)$ we have $v \leq \sup g$ on $\Gamma_{T_1}$.
  4. Use the maximum principle on $U_{T_1}$ to show $v \leq \sup g$. Send $\epsilon \to 0$ to conclude the same for $u_t(y)$ for all $t \in [0, T_1]$.
  5. This will prove the proposition on $[0, T_1]$, and iterate if needed to prove it on $[0, T]$.

Corollary 21 (Uniqueness). If $f \in C([0, T] \times \R^d)$, $g \in C(\R^d)$, then there is at most one solution to \eqref{e:H} with initial data $u_0 = g$ that satisfies the growth estimate \eqref{e:G}.

Remark 22. When solving the Laplace equation $- \lap u = 0$ in half space $H$, you have uniqueness to the boundary value problem if the solution is bounded. If the solution grows even polynomially, uniqueness fails (e.g. $u = x_d$ is a Harmonic function which is $0$ on $\partial H$). In contrast, for the heat equation, uniqueness holds under the exponential growth assumption \eqref{e:G}. (The reason for this is connected to the fact that the heat kernel has a Gaussian tail; where as the Poisson kernel has a polynomial tail.)

Example 23 (Tychonoff). Define $$ u(t,x)=\sum_{n=0}^\infty \frac{f^{(n)}(t)}{(2n)!}x_1^{2n}\,, \quad\text{where}\quad f(t)= \begin{cases} e^{-1/t^4},&t>0,\\ 0,&t = 0. \end{cases} $$ Then $u$ is a non-zero solution to \eqref{e:H} with $u_0 = 0$.

Interior regularity

Proposition 24. If $u \in C^{1, 2}(U_T)$ solves \eqref{e:H} in $U_T$ then $u \in C^\infty(U_T)$.

Proof sketch.

  1. Let $\zeta$ be a smooth bump function which is $0$ near $\Gamma_T$, and let $u^\epsilon = u * \varphi_\epsilon$ for some $C^\infty_c$ bump function $\varphi$.
  2. Compute $\partial_t (u^\epsilon \zeta) - \lap (u^\epsilon \zeta) = u^\epsilon (\partial_t - \lap) \zeta - 2 \grad u^\epsilon \cdot \grad \zeta$, use Corollary 21 and Duhamel to show $$ u^\epsilon_t \zeta = \int_0^t \phi_{t - s} * \paren[\big]{ u^\epsilon (\partial_s \zeta + \lap \zeta) } \, ds +2 \sum_j \int_0^t \partial_j \phi_{t - s} * (\partial_j \zeta_s u^\epsilon_s) \, ds \,, $$ and send $\epsilon \to 0$.
  3. Fix $(t, x)$ near which you want to show smoothness. Make $\zeta$ a bump near $(t, x)$ so that $\partial_t \zeta = \partial_j \zeta = 0$, and rewrite \begin{align} u_t (x) &= \int_0^t \int_{B(x, r)^c} \phi_{t - s}(x - y) \paren[\big]{ u_s(y) (\partial_s \zeta_s(y) + \lap \zeta_s(y)) } \, ds \\ \label{e:Z}\tag{Z} &\qquad + 2 \sum_j \int_0^t \int_{B(x, r)^c} \partial_j \phi_{t - s}(x - y) (\partial_j \zeta_s(y) u_s(y)) \, ds \,, \end{align} and show this is smooth at $x$.

Proposition 25. Let $C(x, t, r) = \set{ (y, s) \st \abs{x - y} \leq r\,,~ t-r^2 \leq s \leq t }$. For every $k, \ell \in \N$, there exists $C_{k, l}$ such that for all multi-indices $\alpha$ of order $k$, for all solutions of the heat equation $u$ in $U_T$, and for all cylinders $C(x, t, r) \subseteq U_T$ we have $$ \norm{D^\alpha_x \partial_t^\ell u}_{L^\infty(C(x, t, r/2))} \leq \frac{C_{k, \ell}}{r^{k + 2\ell + d + 2}} \norm{D^\alpha_x \partial_t^\ell u}_{L^1(C(x, t, r))} \,. $$

Proof sketch. Use \eqref{e:Z} and prove this for the largest $r_0$ for which $C(x, t, r_0) \subseteq U_T$. Then rescale, to obtain it for all $r \leq r_0$.

Energy methods

Proposition 26. Let $f \in C(U_T)$, $g \in C(\Gamma_T)$. There is at most one $C^{1, 2}(\bar U_T)$ solution of \begin{align} \partial_t u - \lap u &= f &&\text{in } U_T \,,\\ u &= g && \text{on } \Gamma_T \,. \end{align}

Remark 27. If $U$ is $C^1$, then one can get a similar uniqueness result for the heat equation with Neumann boundary conditions.

Proposition 28 (Backward uniqueness). Suppose $u^1, u^2$ both solve \begin{align} \partial_t u - \lap u &= 0 && \text{in } U_T \,,\\ u &= g &&\text{on } \textcolor{red}{[0, T] \times \partial U} \,. \end{align} If $u^1_T = u^2_T$ in $U$, then $u^1 = u^2$ in $U_T$.

Remark 29. Note we didn’t assume $u^1 = u^2$ when $t = 0$ and $x \in U$.

Proof sketch.

  1. Set $w = u_1 - u_2$, $E(t) = \norm{\grad w}_{L^2}^2$ and show $\dot E_t^2 \leq E \ddot E$.
  2. If there is a interval where $E > 0$ in some interval, use the above to show $\ln E$ is convex and approaches $-\infty$ at the right end point to produce a contradiction.

Proposition 30 (Li Yau inequality). If $u > 0$ is a $C^{1, 2}((0, \infty) \times \R^d) \cap C([0, \infty) \times \R^d)$ solution of \eqref{e:H} with initial data $u_0 = g \in C(\R^d) \cap L^\infty(\R^d)$. Suppose further $u$ satisfies the growth estimate \eqref{e:G}. Then $$ \frac{|\grad u|^2}{u^2} \leq \frac{\lap u}{u} + \frac{d}{2t} $$

Proof sketch. Write $u = g * \phi_t$, and show $$ \abs{\grad u_t(x)}^2 = \abs[\Bigg]{ \int_{\R^d} \frac{(x - y)}{2t} \phi_t(x - y) g(y) \, dy }^2 \leq u_t(x) \int_{\R^d} \frac{\abs{x-y}^2}{4 t^2} \phi_{t}(x - y) g(y) \, dy \,. $$

Proposition 31 (Parabolic Harnack inequality). Suppose $u$ satisfies the same assumptions as Proposition 30. Then for every $0 < s < t$, $x, y \in \R^d$ $$ u_s(x) \leq u_t(y) \paren[\Big]{ \frac{t}{s}}^{d/2} \exp\paren[\Big]{ \frac{|x-y|^2}{4(t-s)} } \,. $$

Proof sketch. Let $\gamma$ be the straight line joining $(s, x)$ and $(t, x)$. Set $f(r) = \log u \circ \gamma(r)$ and bound $\dot f$ using the Li–Yau inequality.