6.3: Maximum Principle
( \newcommand{\kernel}{\mathrm{null}\,}\)
Let Ω⊂Rn be a bounded domain. Set
DT=Ω×(0,T), T>0,ST={(x,t): (x,t)∈Ω×{0} or (x,t)∈∂Ω×[0,T]},
see Figure 6.3.1.
Figure 6.3.1: Notations to the maximum principle
Theorem 6.2
Assume u∈C(¯DT), that ut, uxixk exist and are continuous in DT, and
$$u_t-\triangle u\le 0\ \ \mbox{in}\ D_T.\]
Then
$$\max_{\overline{D_T}} u(x,t)=\max_{S_T} u.\]
Proof
Assume initially ut−△u<0 in DT. Let ε>0 be small and 0<ε<T. Since u∈C(¯DT−ε), there is an (x0,t0)∈¯DT−ε such that
$$u(x_0,t_0)=\max_{\overline{D_{T-\varepsilon}}} u(x,t).\]
Case (i)
Let (x0,t0)∈DT−ε. Hence, since DT−ε is open,
ut(x0,t0)=0, uxl(x0,t0)=0, l=1,…,n and
$$\sum_{l,k=1}^n u_{x_lx_k}(x_0,t_0)\zeta_l\zeta_k\le0\ \ \mbox{for all}\ \zeta\in\mathbb{R}^n.\]
The previous inequality implies that uxkxk(x0,t0)≤0 for each k. Thus we arrived at a contradiction to ut−△u<0 in DT.
Case (ii)
Assume (x0,t0)∈Ω×{T−ε}. Then it follows as above △u≤0 in (x0,t0), and from u(x0,t0)≥u(x0,t), t≤t0, one concludes that ut(x0,t0)≥0. We arrived at a contradiction to ut−△u<0 in DT again.
Summarizing, we have shown that
$$\max_{\overline{D_{T-\varepsilon}}} u(x,t)=\max_{T-\varepsilon} u(x,t).\]
Thus there is an (xε,tε)∈ST−ε such that
$$u(x_\varepsilon,t_\varepsilon)=\max_{\overline{D_{T-\varepsilon}}} u(x,t).\]
Since u is continuous on ¯DT, we have
$$\lim_{\varepsilon\to 0}\max_{\overline{D_{T-\varepsilon}}} u(x,t)=\max_{\overline{D_T}} u(x,t).\]
It follows that there is (¯x,¯t)∈ST such that
$$u(\overline{x},\overline{t})=\max_{\overline{D_T}} u(x,t)\]
since ST−ε⊂ST and ST is compact. Thus, theorem is shown under the assumption ut−△u<0 in DT. Now assume ut−△u≤0 in DT. Set
$$v(x,t):=u(x,t)-kt,\]
where k is a positive constant. Then
$$v_t-\triangle v=u_t-\triangle u-k<0.\]
From above we have
max¯DTu(x,t)=max¯DT(v(x,t)+kt)≤max¯DTv(x,t)+kT=maxSTv(x,t)+kT≤maxSTu(x,t)+kT,
Letting k→0, we obtain
$$\max_{\overline{D_T}} u(x,t)\le\max_{S_T} u(x,t).\]
Since ST⊂¯DT, the theorem is shown.
◻
If we replace in the above theorem the bounded domain Ω by Rn, then the result remains true provided we assume an {\it additional} growth assumption for u. More precisely, we have the following result which is a corollary of the previous theorem. Set for a fixed T, 0<T<∞,
$$D_T=\{ (x,t):\ x\in\mathbb{R}^n,\ 0<t<T\}.\]
Proposition 6.2
Assume u∈C(¯DT), that ut, uxixk exist and are continuous in DT,
$$u_t-\triangle u\le 0\ \ \mbox{in}\ D_T,\]
and additionally that u satisfies the growth condition
$$u(x,t)\le Me^{a|x|^2},\]
where M and a are positive constants.
Then
$$\max_{\overline{D_T}} u(x,t)=\max_{S_T} u.\]
It follows immediately the
Corollary
The initial value problem ut−△u=0 in DT, u(x,0)=f(x), x∈Rn, has a unique solution in the class defined by u∈C(¯DT), ut, uxixk exist and are continuous in DT and |u(x,t)|≤Mea|x|2.
Proof of Proposition 6.2.
See [10], pp. 217. We can assume that 4aT<1, since the finite interval can be divided into finite many intervals of equal length τ with
4aτ<1. Then we conclude successively for k that
$$u(x,t)\le\sup_{y\in\mathbb{R}^n}u(y,k\tau)\le\sup_{y\in\mathbb{R}^n}u(y,0)\]
for kτ≤t≤(k+1)τ, k=0,…,N−1, where N=T/τ.
There is an ϵ>0 such that 4a(T+ϵ)<1. Consider the comparison function
vμ(x,t):=u(x,t)−μ(4π(T+ϵ−t))−n/2e|x−y|2/(4(T+ϵ−t))=u(x,t)−μK(ix,iy,T+ϵ−t)
for fixed y∈Rn and for a constant μ>0. Since the heat kernel K(ix,iy,t) satisfies Kt=△Kx, we obtain
$$\frac{\partial}{\partial t}v_\mu-\triangle v_\mu=u_t-\triangle u\le0.\]
Set for a constant ρ>0
$$D_{T,\rho}=\{(x,t):\ |x-y|<\rho,\ 0<t<T\}.\]
Then we obtain from Theorem 6.2 that
$$v_\mu(y,t)\le\max_{S_{T,\rho}}v_\mu,\]
where ST,ρ≡ST of Theorem 6.2 with Ω=Bρ(y), see Figure 6.3.1.
On the bottom of ST,ρ we have, since μK>0,
$$v_\mu(x,0)\le u(x,0)\le\sup_{z\in\mathbb{R}^n}f(z).\]
On the cylinder part |x−y|=ρ, 0≤t≤T, of ST,ρ it is
vμ(x,t)≤Mea|x|2−μ(4π(T+ϵ−t))−n/2eρ2/(4(T+ϵ−t))≤Mea(|y|+ρ)2−μ(4π(T+ϵ))−n/2eρ2/(4(T+ϵ))≤supz∈Rnf(z)
for all ρ>ρ0(μ), ρ0 sufficiently large. We recall that 4a(T+ϵ)<1.
Summarizing, we have
$$\max_{S_{T,\rho}}v_\mu(x,t)\le\sup_{z\in\mathbb{R}^n}f(z)\]
if ρ>ρ0(μ). Thus
$$v_\mu(y,t)\le\max_{S_{T,\rho}}v_\mu(x,t)\le\sup_{z\in\mathbb{R}^n}f(z)\]
if ρ>ρ0(μ).
Since
$$v_\mu(y,t)=u(y,t)-\mu\left(4\pi(T+\epsilon-t)\right)^{-n/2}\]
it follows
$$u(y,t)-\mu\left(4\pi(T+\epsilon-t)\right)^{-n/2}\le\sup_{z\in\mathbb{R}^n}f(z).\]
Letting μ→0, we obtain the assertion of the proposition.
◻
The above maximum principle of Theorem 6.2 holds for a large class of parabolic differential operators, even for degenerate equations.
Set
$$Lu=\sum_{i,j=1}^na^{ij}(x,t)u_{x_ix_j},\]
where aij∈C(DT) are real, aij=aji, and the matrix (aij) is non-negative, that is,
$$\sum_{i,j=1}^na^{ij}(x,t)\zeta_i\zeta_j\ge 0\ \ \mbox{for all}\ \zeta\in\mathbb{R}^n,\]
and (x,t)∈DT.
Theorem 6.3
Assume u∈C(¯DT), that ut, uxixk exist and are continuous in DT, and
$$u_t-L u\le 0\ \ \mbox{in}\ D_T.\]
Then
$$\max_{\overline{D_T}} u(x,t)=\max_{S_T} u.\]
Proof
(i) One proof is a consequence of the following lemma: let A, B real, symmetric and non-negative matrices. Non-negative means that all eigenvalues are non-negative. Then trace~(AB)≡∑ni,j=1aijbij≥0, see an exercise.
(ii) Another proof is more directly: let U=(z1,…,zn), where zl is an orthonormal system of eigenvectors to the eigenvalues λl of the matrix A=(ai,j(x0,t0)). Set ζ=Uη, x=UT(x−x0)y and v(y)=u(x0+Uy,t0), then
0≤n∑i,j=1aij(x0,t0)ζiζj=n∑i=1λiη2i0≥n∑i,j=1uxixjζiζj=n∑i=1vyiyiη2i.
It follows λi≥0 and vyiyi≤0 for all i.
Consequently
$$\sum_{i,j=1}^na^{ij}(x_0,t_0)u_{x_ix_j}(x_0,t_0)=\sum_{i=1}^n\lambda_iv_{y_iy_i}\le0.\]
◻
Contributors and Attributions
Integrated by Justin Marshall.