TMA4305 PDEs 2017

Weak maximum principle for the heat equation

Harald Hanche-Olsen

In this note, we consider the standard heat equation

𝑢𝑡 − △ 𝑢 = 0 in Ω𝑇

in which Ω𝑇 = (0, 𝑇) × Ω where Ω ⊂ ℝ𝑛 is a bounded region, 𝑇 > 0, and

𝑢 ∈ 𝐶(Ω𝑇 ) ∩ 𝐶 2 (Ω𝑇 ).

We think of Ω𝑇 as an open cylinder with base Ω and height 𝑇. Its closure is a

closed cylinder: Ω𝑇 = [0, 𝑇] × Ω.

Definition. The parabolic boundary of Ω𝑇 is the set

Γ = ({ 0 } × Ω) ∪ ([0, 𝑡] × 𝜕Ω).

Clearly, Γ is contained in the normal boundary 𝜕Ω𝑇 ; the difference is

𝜕Ω𝑇 ⧵ Γ = { 𝑇 } × Ω.

We call { 𝑇 } × Ω the final boundary of Ω𝑇 (nonstandard nomenclature).

Observation. If a 𝐶 2 function 𝑣 has a maximum at some point in Ω𝑇 , then 𝑣𝑡 = 0

and △ 𝑣 ≤ 0 at that point, so we get 𝑣𝑡 − △ 𝑣 ≥ 0 there. Moreover, this holds
at the final boundary as well, the only difference being that there, we can only
conclude 𝑣𝑡 ≥ 0 and △ 𝑣 ≤ 0. In other words,

𝑣𝑡 − △ 𝑣 ≥ 0 at any maximum in Ω𝑇 ⧵ Γ.

We must face a minor technical glitch: The above statement requires that 𝑣 is 𝐶 2
up to and including the final boundary of Ω𝑇 . This complicates the proof of the
following theorem, but only a little.

Theorem 1 (The weak maximum principle). Assume that 𝑢 ∈ 𝐶(Ω𝑇 ) ∩ 𝐶 2 (Ω𝑇 )

satisfies
𝑢𝑡 − △ 𝑢 ≤ 0.

Then 𝑢(𝑡, 𝒙) ≤ maxΓ 𝑢 for all (𝑡, 𝒙) ∈ Ω𝑇 . In other words, 𝑢 achieves its maximum
on the parabolic boundary.

Version 2017-09-25
Weak maximum principle for the heat equation 2

Proof. First, to deal with the “minor technical glitch” mentioned above, we shall
strengthen the assumptions somewhat, and assume that 𝑢 ∈ 𝐶 2 ((0, 𝑇] × Ω). We
will remove this extra assumption at the end.
Now let 𝜀 > 0, and put 𝑣(𝑡, 𝒙) = 𝑢(𝑡, 𝒙) − 𝜀𝑡. Then 𝑣𝑡 − △ 𝑣 ≤ −𝜀 < 0, and
so it follows immediately from the Observation above that 𝑣 cannot achieve its
maximum anywhere outside Γ. On the other hand, since 𝑣 is continuous and Ω𝑇
is compact, 𝑣 does have a maximum in Ω𝑇 , and so we must conclude that 𝑣(𝑡, 𝒙) ≤
maxΓ 𝑣 for any (𝑡, 𝒙) ∈ Ω𝑇 . But then 𝑢(𝑡, 𝒙) = 𝑣(𝑡, 𝒙) + 𝜀𝑡 ≤ maxΓ 𝑣 + 𝜀𝑇 ≤
maxΓ 𝑢+𝜀𝑇. Since this holds for any 𝜀 > 0, it finally follows that 𝑢(𝑡, 𝒙) ≤ maxΓ 𝑢,
and the proof is complete, with the strengthened assumptions.
We now drop the requirement that 𝑢 ∈ 𝐶 2 ((0, 𝑇]×Ω). However, it is still true that
𝑢 ∈ 𝐶 2 ((0, 𝑇 ′ ] × Ω), for any 𝑇 ′ < 𝑇, so the first part shows that 𝑢(𝑡, 𝒙) ≤ maxΓ𝑇′ 𝑢
for all (𝑡, 𝒙) ∈ Ω𝑇 ′ . Here Γ𝑇 ′ is tbe parabolic boundary of Ω𝑇 ′ . But Γ𝑇 ′ ⊂ Γ, so we
also have 𝑢(𝑡, 𝒙) ≤ maxΓ 𝑢. For any 𝑡 < 𝑇, we can pick 𝑇 ′ with 𝑡 < 𝑇 ′ < 𝑇, so
the inequality holds. Finally, it also holds for 𝑡 = 𝑇, since 𝑢 is continuous on Ω𝑇 .
This, at last, completes the proof.
It should come as no surprise that there is also a minimum principle. It is
proved by replacing 𝑢 by −𝑢 in Theorem 1.
Corollary 2 (The weak minimum principle). Assume that 𝑢 ∈ 𝐶(Ω𝑇 ) ∩ 𝐶 2 (Ω𝑇 )
satisfies
𝑢𝑡 − △ 𝑢 ≥ 0.
Then 𝑢(𝑡, 𝒙) ≥ minΓ 𝑢 for all (𝑡, 𝒙) ∈ Ω𝑇 . In other words, 𝑢 achieves its minimum
on the parabolic boundary.
We will mostly be concerned with solutions of the heat equation 𝑢𝑡 −△ 𝑢 = 0,
and for these, both the maximum principle and the minimum principle can be
used. But we may also wish to study inhomogeneous equations 𝑢𝑡 − △ 𝑢 = 𝑓,
and if 𝑓 has a definite sign, one or the other principle will apply.
Corollary 3 (Uniqueness for the heat equation). There exists at most one solution
𝑢 ∈ 𝐶(Ω𝑇 ) ∩ 𝐶 2 (Ω𝑇 ) to the problem
𝑢𝑡 − △ 𝑢 = 𝑓 in Ω𝑇 ,
𝑢=𝑔 on Γ.
Here, 𝑓 and 𝑔 are given functions on Ω𝑇 and Γ, respectively. (Thus 𝑔 combines initial
values and boundary values in one function.)
Proof. Let 𝑢 be the difference between two solutions to this problem: Then 𝑢
solves the same problem, but with 𝑓 = 0 and 𝑔 = 0. Thus 𝑢 achieves both its
minimum and maximum on Γ, but 𝑢 = 0 there, so 𝑢 = 0 everywhere.

Version 2017-09-25
3 Weak maximum principle for the heat equation

The following corollary is proved in essentially the same way, by applying the
minimum and maximum principles to 𝑢1 − 𝑢2 . Note that it immediately implies
the preceding corollary by taking 𝑔1 = 𝑔2 .
Corollary 4 (Continuous dependence on data). Let 𝑢1 and 𝑢2 satisfy
𝑢𝑖𝑡 − △ 𝑢𝑖 = 𝑓 in Ω𝑇 ,
} for 𝑖 = 1, 2.
𝑢𝑖 = 𝑔 𝑖 on Γ,
Then |𝑢1 − 𝑢2 | ≤ maxΓ |𝑔1 − 𝑔2 |.
Exercise 1 (Continuous dependence on data, improved). Assume that 𝑢1 and 𝑢2
satisfy
𝑢𝑖𝑡 − △ 𝑢𝑖 = 𝑓𝑖 in Ω𝑇 ,
} for 𝑖 = 1, 2.
𝑢𝑖 = 𝑔𝑖 on Γ,
Let 𝜑 = supΩ |𝑓1 − 𝑓2 | and 𝛾 = maxΓ |𝑔1 − 𝑔2 |, and show that |𝑢1 − 𝑢2 | ≤ 𝛾 + 𝜑𝑇.
𝑇
Note that for any 𝑡, we can pick 𝑇 = 𝑡, so we really get |𝑢1 − 𝑢2 | ≤ 𝛾 + 𝜑𝑡.
Hint: Apply the maximum principle to 𝑢1 − 𝑢2 − 𝜑𝑡 and 𝑢2 − 𝑢1 − 𝜑𝑡.
Exercise 2. Show that the maximum (and minimum) principle continues to hold
if 𝑢𝑡 − △ 𝑢 is replaced by the more general
𝑢𝑡 − ∇ ⋅ (𝐴∇𝑢)
where the (constant) real 𝑛 × 𝑛 matrix 𝐴 is symmetric and positive definite.
Here are some ingredients for a proof:
• The Hessian of 𝑢 is defined to be the (symmetric!) 𝑛 × 𝑛 matrix H𝑢 with
entries 𝑢𝑥𝑖𝑥𝑗 . At an interior maximum point, H𝑢 is negative semidefinite,
i.e., 𝒚𝑇 H𝒚 ≤ 0 for all 𝒚 ∈ ℝ𝑛 . (Short proof: Take the second derivative of
𝑢(𝒙 + 𝑠𝒚) with respect to 𝑠 where 𝑥 is a maximum point, and put 𝑠 = 0.)
• The Frobenius inner product of two real matrices 𝐴 and 𝐵 is
𝑛 𝑛
⟨𝐴, 𝐵⟩F = ∑ ∑ 𝑎𝑖𝑗 𝑏𝑖𝑗 = tr(𝐴𝑇 𝐵).
𝑖=1 𝑗=1

It turns out that

∇ ⋅ (𝐴∇𝑢) = ⟨𝐴, H𝑢⟩F .
• It is known that if 𝐴 and 𝐵 are positive semidefinite, then ⟨𝐴, 𝐵⟩F ≥ 0.
(Short proof: Since 𝐴 is symmetric, we can write ⟨𝐴, 𝐵⟩F = tr(𝐴𝐵). 𝐴 will
have a positive semidefinite square root 𝐴1/2 . A standard result on the trace
gives tr(𝐴𝐵) = tr(𝐴1/2 𝐴1/2 𝐵) = tr(𝐴1/2 𝐵𝐴1/2 ), but 𝐴1/2 𝐵𝐴1/2 is positive
semidefinite, and such matrices have nonnegative trace.)

Version 2017-09-25
Weak maximum principle for the heat equation 4

Exercise 3. Show that the maximum (and minimum) principle continues to hold
if 𝑢𝑡 − △ 𝑢 is replaced by the even more general

𝑢𝑡 − ∇ ⋅ (𝐴∇𝑢) + 𝑏(∇𝑢),

where the real matrix 𝐴 is symmetric and positive definite, provided the function
𝑏 satisfies 𝑏(𝟎) = 0. (For a simple and common example, let 𝑏(∇𝑢) = 𝒃 ⋅ ∇𝑢.)
Remark. In many PDE texts, the term ∇ ⋅ (𝐴∇𝑢) is written out in detail as
𝑛 𝑛
∇ ⋅ (𝐴∇𝑢) = ∑ ∑ 𝑎𝑖𝑗 𝑢𝑥𝑖𝑥𝑗 .
𝑖=1 𝑗=1

Pedantically speaking, considering the order in which derivatives are taken, that
should be 𝑛 𝑛
∇ ⋅ (𝐴∇𝑢) = ∑ ∑ 𝑎𝑖𝑗 𝑢𝑥𝑗𝑥𝑖 ,
𝑖=1 𝑗=1

but this makes no difference, due to symmetry.

Version 2017-09-25