Asymptotic behaviour methods for the Heat

Equation. Convergence to the Gaussian

Juan Luis Vázquez
arXiv:1706.10034v3 [math.AP] 22 Nov 2018

Universidad Autónoma de Madrid

2018

Abstract
In this expository work we discuss the asymptotic behaviour of the solutions of the classical
heat equation posed in the whole Euclidean space. After an introductory review of the main
facts on the existence and properties of solutions, we proceed with the proofs of convergence
to the Gaussian fundamental solution, a result that holds for all integrable solutions, and
represents in the PDE setting the Central Limit Theorem of probability. We present several
methods of proof: first, the scaling method. Then several versions of the representation
method. This is followed by the functional analysis approach that leads to the famous related
equations, Fokker-Planck and Ornstein-Uhlenbeck. The analysis of this connection is also
given in rather complete form here. Finally, we present the Boltzmann entropy method,
coming from kinetic equations.
The different methods are interesting because of the possible extension to prove the asymptotic
behaviour or stabilization analysis for more general equations, linear or nonlinear. It all
depends a lot on the particular features, and only one or some of the methods work in each
case. Other settings of the Heat Equation are briefly discussed in Section 9 and a longer
mention of results for different equations is done in Section 10.

2010 Mathematics Subject Classification: 35K05, 35K08.

Keywords and phrases: Heat equation, asymptotic behaviour, convergence to the Gaussian solu-
tion.

1
Contents
1 The Cauchy Problem in RN for the Heat Equation 3

2 Asymptotic convergence to the Gaussian 8

3 Proof of convergence by scaling 10

4 Asymptotic convergence to the Gaussian via representation 11

5 Improved convergence for distributions with second moment 16

6 Functional Analysis approach for Heat Equations 19

7 Calculation of spectrum. Refined asymptotics 24

8 Convergence via the Boltzmann entropy approach 25

9 Brief review of other heat equation problems 26

10 Application to other diffusion equations 30

11 Historical comments 32

2
1 The Cauchy Problem in RN for the Heat Equation
The classical heat equation (HE) ∂t u = ∆u, is one of the most important objects in the theory of
partial differential equations, and it has great relevance in the applied sciences. It has been devel-
oped since the seminal work of J. Fourier, 1822, [32], to describe phenomena of heat transport and
diffusion in many contexts. The great progress of the mathematical theory in these two centuries
has had a strong influence not only on PDEs, but also on Probability, Functional Analysis, as
well as Numerics. It now has well-established connections with other subjects like viscous fluids,
differential geometry, image processing or finance. See more on motivation in standard textbooks,
like [31, 58] or in the survey paper [69].
The heat equation enjoys a well developed theory that has many distinctive features. It can be
solved in many settings, suitable for different interests, that lead to quite different results and use
different tools. In this paper we will study the most typical setting, the Cauchy Problem posed
in RN , N ≥ 1:
(
∂t u = ∆u in RN × (0, ∞)
(CPHE)
u(x, 0) = u0 (x) in RN ,

with initial data u0 ∈ L1 (RN ). The equation appears in the theory of Probability and Stochastic
Processes as the PDE description of Brownian motion, and in that application the function u(·, t)
R probability density of the process at time t, hence it must be nonnegative and the total
denotes the
integral u(x, t) dx, often called the mass, must be one. Such restrictions on sign and integral
are not needed in the analytical study, and the extra generality is convenient both for the theory
and for a number of other applications.
The first question to be addressed by mathematicians is to find appropriate functional spaces
to solve this problem and to prove that it is a well-posed problem in the sense of Hadamard’s
definition (existence, uniqueness and stability). In tackling such a problem we will be lucky
enough to find a very explicit representation of the solution by means of the so-called “fundamental
solution” procedure.

Definition. We call fundamental solution of the HE in RN the solution of the equation with
initial data a “Dirac delta”:
(
∂t U = ∆U en RN × (0, ∞)
(FS)
U (x, 0) = δ0 (x) en RN .
In probabilistic terms, we start from a unit point mass distribution concentrated at the origin
x0 = 0, and we look for the way it expands with time according to the heat equation.

Exercise 1. (i) Prove by the method of Fourier Transform that the fundamental solution is given
by the formula
N |x|2
(GF) U (x, t) = (4πt)− 2 e− 4t .
e = −|ξ|2 U
Sketch. The Fourier transform of U satisfies ∂t U e with U
e (ξ, 0) = 1, hence U
e (ξ, t) =
−|ξ| 2t
e . Then, invert the Fourier transform.
(ii) Prove that U (x, t) → δ0 (x) when t → 0 in the sense of distributions, i. e.,
Z
U (x, t)φ(x)dx → < δ0 , φ >:= φ(0), for all φ ∈ Cc∞ (RN ), as t → 0.
RN

3
(iii) Prove that for every sequence tn → 0 the functional sequence U (x, tn ) is a C ∞ approximation
of the identity.

C. F. Gauss and the Gaussian solution

Notations and remarks. This function is known as the Gaussian Function or Gaussian Kernel
and we use the notation G(x, t) or Gt (x) instead of U . We will also use the Gaussian measure
dµ = Gt dx; it is a Radon measure in RN with parameter t. The Gaussian kernel is named after
the famous German mathematician Carl Friedrich Gauss, the Prince of Mathematics.
In statistics and probability theory the Gaussian function appears as the density function of the
normal distribution. Then the standard deviation of the distribution is taken as the parameter in
the formula, and we have 2t = σ 2 . This relation : time proportional to square space deviation, is
called the Brownian scale and appears in many calculations of evolution processes related to heat
propagation or Brownian motion. Finally, the same function appears in Statistical Mechanics
with the name of Maxwellian distribution (after James Clerk Maxwell, 1860). See other historical
comments at the end of the paper.
A main result in those disciplines is the fact that G appears as the limiting probability distri-
bution of many discrete processes, according to the famous Central Limit Theorem. The limiting
behaviour of the solutions of the Heat Equation is our main concern in this paper, and the
Gaussian kernel will play a main role.

Exercise 2. Calculate the Gaussian function by the method of self-similar solutions, U (x, t) =
t−α F (x t−β ), plus the condition of mass conservation.
Hint. Substitute the self-similar form into the PDE and check that time disappears explicitly
from the resulting equation for F = F (ξ) (with ξ = x t−β ) if we put β = 1/2. Check that the
choice α = N/2 means conservation of mass. This selects both similarity exponents. Then write
the resulting elliptic equation for F (ξ) assuming that it is a function of r = |ξ| (it is a radial
function) to get the ODE
1
(rN −1 F 0 (r))0 + (rN F )0 = 0.
2
2 /4
Calculate that log F (r) = a − r2 /4, with a ∈ R, hence F = Ce−|x| .

Exercise 3. Write the formula of Gaussian with mass M as

M 2
U (x, t; M ) = N/2
e−x /4t
(4πt)

4
and check that it is dimensionally correct. Dimensional analysis is a powerful tester for correct
formulas (if applied well).

Exercise 4. Gaussian Representation Formula. Show that if u0 ∈ C(RN ) and is bounded,

then the convolution
Z
N |x−y|2
(GRF) u(x, t) = (U ∗ u0 )(x, t) := (4πt)− 2 u0 (y)e− 4t dy
RN

is a solution of the heat equation with initial data u0 . Check that it is a C ∞ function of x and t
for every x ∈ RN and t > 0. Prove that u(x, t) → u0 (x) pointwise when t → 0.
(ii) Prove that the initial data is taken uniformly if u0 ∈ BU C(RN ), the space of bounded and
uniformly continuous functions in RN .
(iii*) Consider whether the boundedness condition of point (i) can be relaxed or eliminated.

Exercise 5. Prove that formula (GRF) is still valid when the initial data belong to Lp (RN ),
1 ≤ p ≤ ∞.
(ii) What is the regularity of the solution?
(iii) In which sense are the initial data taken?

Exercise 6. (i) Prove that (GRF) generates a continuous contraction semigroup in the following
Banach spaces
X = BU C(RN ), L2 (RN ), Lp (RN ), 1 < p < ∞ ,
by means of the definition St : X → X, St u0 = u(·, t), where u is the solution of the heat equation
with initial data u0 .
Sketch. What you have to check is that, as a function of time t f (t) := St u0 belongs to
C([0, ∞ : X), that St ∗ Ss = St+s , and also St u0 → u0 as t → 0 in X. Moreover, contraction
means that kSt u0 kX ≤ ku0 kX for every t > 0 and every u0 ∈ X.
(ii) Study in which sense the semigroups agree (explain if they are the same or not). You may
define the core of the semigroup by restricting it to a nice and dense set of functions, and then
explain how the concept of density is used to recover the complete semigroup in each of the above
spaces.

Notation. We will often write u(t) instead of u(·, t) for brevity in the hope that no confusion
will arise. Hence, u(t) is a function of x for every fixed t.

Exercise 7. (i) Prove that for p = 1 we have the stronger property of mass conservation
Z Z
(MC) u(x, t) dx = u0 (x) dx ,
RN RN

for all data in L1 (RN ).

(ii) Note that this law also holds for signed solutions. Show that the version with absolute value
is not valid for signed solutions by means of an example (superposition of two Gaussians)
(iii) Prove on the fundamental solution that the Lp -integrals with p R> 1 are not conserved and
check the rate of decay they exhibit. Terminology: The p energy is |u|p dx. Those p-energies
are not conserved because they “undergo dissipation”.

5
 (iv*) The curious reader may want to learn more about that dissipation and how to measure it
accurately. Any ideas?

Exercise 8. Conservation of moments. (i) We define the first (signed) moment as the vector
quantity
Z
(1.1) N1 (u(t)) = x u(x, t) dx ,
RN

which in principle evolves with time. Prove that thisR quantity is actually conserved in time for
any solution of the HE with initial data such that (1 + |x|)| u0 (x)| dx is finite.
(ii) For data such that |x|2 |u0 (x)| dx is finite, we define the second moment as the scalar
R

time-dependent quantity
Z
(1.2) N2 (u(t)) = |x|2 u(x, t) dx .
RN

Prove that this quantity is not conserved in time for solutions of the HE, and in fact

(1.3) N2 (u(t)) = N2 (u0 ) + 2N t .

(Recall here that N is the space dimension). These results will be important to understand the
finer versions of the asymptotic convergence to the Gaussian. Deeper analysis proves that the
only integral quantities conserved by the solutions of the Heat Equation are the mass and the
first moment.

Exercise 9. (i) Prove the so-called ultra-contractive estimates, also called smoothing effects.
They allow to pass from data in Lp to solutions u(t) in L∞ for every t > 0. More precisely, prove
that there exists a constant C = C(p, N ) such that for every u0 ∈ Lp (RN )
ku0 kp
ku(t)k∞ ≤ C .
tN/2p
We may say that the constant is universal since it does not depend of the particular solution we
take.
(ii) Calculate the exponent N/2p using only dimensional calculus.

Exercise 10. Regularity. (i) Prove that there exists a universal constant C = C(p, N ) > 0
such that for every p ≥ 1 and every t > 0
du(t) ku0 kp
k kp ≤ C .
dt t
(ii) Prove that for nonnegative solutions we have a stronger pointwise inequality
∂u u
≥ −C1 with C1 = N/2 .
∂t t
Prove that this constant is optimal. Hint. Check the properties first for the fundamental solution,
then use convolution.
(iii) Prove that for every t > 0 the x-derivatives of u satisfy
 
 ≤ C ||u0 ||1 L∞ .
 ∂ 

 ∂xi u(x, t)
t2


6
Obtain a similar estimate for data in Lp RN .

(iv) Check that the dimensions are all correct in these formulas.
Exercise 11. Gaussian Representation Formula with measure data. (i) Prove that the
representation formula can be used with any bounded Radon measure as initial data, in the form
Z
|x−y|2
−N
(GRFm) u(x, t) = (Gt ∗ µ)(x, t) := (4πt) 2 e− 4t dµ(y) .
RN

(ii) Prove that it produces a classical solution of the HE for all t > 0 and that the aRpriori
estimates hold with the L1 norm of u0 replaced by the total mass of the measure, |µ(RN )| = |dµ|.
(iii) Prove that the initial data are taken in the weak sense of measures.
(iv) Observe that the fundamental solution U = Gt falls into this class.
Exercise 12. (i) Find explicit solutions of polynomial type of heat equation. Suggestion: try the
solution types
u = A0 (x), u = A0 (x) + A1 (x)t, u = A0 (x) + A1 (x)t + A2 (x)t2 , . . .
Find all the solutions of the first two types. Maybe the best known solutions in this class are
u = 1 and u = |x|2 + 2N t.
(ii) Prove that they also satisfy the representation formula, even though they are not integrable
or bounded.
The general theory says that the (GRF) holds for all initial data which are locally bounded
measures µ with a weighted integrability condition
Z
2
e−c|x| |dµ(x)| < ∞ for some c > 0 ,
RN
see [70], but we will not use that sharp result in these notes. We will only point out the following
example of blow-up solution with very large growth as |x| → ∞
2 /4(T −t)
(1.4) Ub (x, t) = C (T − t)−N/2 e|x| .

Exercise 13. Check that Ub is a classical solution of the heat equation for −∞ < t < T and
blows up everywhere in x as t → T .

Exercise 14. Construction of new solutions.

(i) Show that if u(x, t) is a solution of the heat equation, so are u1 = ∂t u, and u2 = ∂xi u.
Rx
(ii) Show that if u(x, t) is a solution of the heat equation in 1D then so is also u3 = −∞ u(s, t) ds.
(iii) Prove that if u(x, t) is a solution of the heat equation so is also
v(x, t) = Au(Bx, B 2 t)
for any choice of the parameters A and B. This is called the scaling property. A and B need not
be positive.
(v*) Show that if u(x, t) is a solution of the heat equation, then so is also
v(x, t) = xux + 2tut .
This is a 1D notation. But the result is valid in any dimension if the notation is correctly
interpreted.

7
2 Asymptotic convergence to the Gaussian
The main result on the asymptotic behaviour of general integrable solutions of the heat equation
consists in proving that they look increasingly like the fundamental solution. Since this solution
goes to zero uniformly with time, the estimate of the convergence has to take into account that
fact and compensate for it. This happens by considering a renormalized error that divides the
standard error in some norm by the size of the Gaussian solution U (t) = Gt in the same norm.
For instance, in the case of the sup norm we know that

kGt kL∞ (RN ) = Ct−N/2 , kGt kL1 (RN ) = 1.

This is the basic result we want to prove

Theorem 2.1 Let u0 ∈ L1 (RN ) and let u0 (x)dx = M be its mass. Then the solution u(t) =
R

u(·, t) of the HE in the whole space ends up by looking like M times the fundamental solution
U (t) = Gt in the sense that

(2.1) lim ku(t) − M Gt k1 → 0

t→∞

and also that

(2.2) lim tN/2 ku(t) − M Gt k∞ → 0 .

t→∞

By interpolation we get the convergence result for all Lp norms

(2.3) lim tN (p−1)/2p ku(t) − M Gt kLp (RN ) → 0 .

t→∞

for all 1 ≤ p ≤ ∞.

We add important information to this result in a series of remarks.

• First, one comment about the spatial domain. The fact that we are working in the whole space
is crucial for the result of Theorem 2.1. The behaviour of the solutions of the heat equation
posed in a bounded domain with different kinds of boundary conditions is also known, and the
asymptotic behaviour does not follow the Gaussian pattern.
• As we will see below, convergence to the Gaussian happens on the condition that the data
belong to the class of integrable functions, that can be extended without problem to bounded
Radon measures. This is actually no news since the theory says that the solution corresponding
to an initial measure µ ∈ M(RN ) is integrable and bounded for any positive time, so we may
change the origin of time and make the assumption of integrable and bounded data. But we point
out the some of the proofs work directly for measures without any problem.
• We recall that it is usually assumed that u ≥ 0 on physical grounds but such assumption is not
at all needed for the analytical study of this paper. Thus, the basic result holds also for signed
solutions even if the total integral is negative, M ≤ 0. There is no change in the proofs. We may
also put M = 1 by linearity as long as M 6= 0.
• M = 0 is a special case that deserves attention: even if Theorem (2.1) is true, the statement
does not imply that the solution looks asymptotically like a Gaussian; to be more precise, it only
says that the previous first-order approximation disappears. If we want more precise details about

8
what the solution looks like, we have to search further to identify the terms that may give us the
size and shape of such a solution. This question will be addressed below. For the moment let us
point out that differentiation in xi of the Heat Kernel produces a new solution with zero integral
2 /4t
ui (x, t) = ∂xi Gt (x) = C t−(N +2)/2 xi e−|x| ,

to which we can apply the above comments. In particular,

tN/2 ui (x, t) = O(t−1/2 ) ,

where O(·) is the Landau O-notation for orders of magnitude.

• It must be stressed that convergence to the Gaussian does not hold for other data. Maybe
the simplest example of solution that does not approach the Gaussian is given by any non zero
constant solution, but it could be objected that L∞ (RN ) is very far from L1 (RN ). Actually, the
same happens for all Lp (RN ) spaces, p > 1. Indeed,
R a simple argument based on approximation
and comparison shows that for any u0 ≥ 0 with u0 (x) dx = +∞ we have

lim tN/2 u(x, t) = +∞

t→∞

everywhere in x ∈ RN (and the divergence is locally uniform).

• The way different classes of non-integrable solutions actually behave for large time is an inter-
esting question that we will not address here. Thus, the reader may prove using the convolution
formula that for locally integrable data that converge to a constant C as |x| → ∞ the solution
u(x, t) stabilizes to that constant as t → ∞. Taking growing data may produce solutions that
tend to infinity with time, like the 1D family of travelling waves
2 t+cx
UT W (x, t) = C ec

defined for real constants C, c 6= 0. A more extreme case is the blow-up solution Ub (x, t) of formula
(1.4) that not only does not stay bounded with time, it even blows up in finite time.
• About the three convergence results of the Theorem, it is clear that (2.3) follows from (2.1)
and (2.2). Now, what is interesting is that (2.2) follows from (2.1) and the smoothing effect
(Exercise 9). We ask the reader to prove this fact. Hint: use (2.1) between 0 and t/2 and then
the smoothing effect between t/2 and t.
• It is interesting to note that the convergence in L1 norm of formula (2.1) can be formulated
without mention to the existence of a fundamental solution in the following form:
Alternative Theorem. Let u(t) and v(t) be any two solutions of the RCauchy problem
R for the
HE in the whole space, and let us assume that their initial data satisfy u0 dx = v0 dx. Then
we have

(2.4) lim ku(t) − v(t)k1 → 0 .

t→∞

We could use a similar approach for the L∞ estimate but some Gaussian information appears
in the weight tN/2 . The alternative approach has been first remarked by specialists in stochastic
processes of Brownian type, and it is known as a mixing property. It has been generalized to
many variants of the heat equation.

9
3 Proof of convergence by scaling
There are many approaches to the proof of the main result, and we will show some of the best
known below. They are interesting for their possible extension to similar asymptotic results for
other equations, both linear and nonlinear, see Sections 9 and 10. The first proof we give is based
on scaling arguments. In the proposed method the proof is divided into 5 steps.

Step 1. Scaling transformation. It is easy to check that the HE is invariant under the following
one-parameter family of transformations Tk defined on space-time functions by the formula

Tk u(x, t) = k N u(kx, k 2 t), k > 0,

see also Exercise 14 (iii). This transformation maps solutions u into new solutions uk = Tk u,
called rescalings of the original solution. It also conserves the mass of the solutions
Z Z
Tk u(x, t) dx = u(y, k 2 t) dy = constant.
RN RN

Step 2. Uniform estimates. The estimates proved on the general solutions in Section 1 imply
that the whole family uk is uniformly bounded in space and time if time is not so small: x ∈ RN
and t ≥ t1 > 0 (cf. Exercise 9). They also have uniform estimates on all derivatives under the
same restriction on time (cf. Exercise 10).

Step 3. Limit problem. We can now use functional compactness and pass to the limit k →
∞ along suitable subsequences to obtain a function Û (x, t) that satisfies the same estimates
mentioned above and is a weak solution of the HE in Q = RN × (0, ∞). By the estimates on
derivatives the solution is classical. The convergence takes place locally in Q in the sup norm for
the functions and their space and time derivatives.

Step 4. Identifying the limit. We now have to check that the limit solution Û (x, t) has the same
mass M as the sequence uk and that is takes the Dirac delta as initial data. By uniqueness for
solutions with measure data (which we accept as part of the theory) we will then conclude that
Û (x, t) is just a fundamental solution, Û (x, t) = M Gt .
(i) The proof of this step is best done when u0 is nonnegative, compactly supported and bounded,
since in that case the solution is bounded above by constant times the fundamental solution,
|u(x, t)| ≤ C G(x, t + 1). Applying the transformation we get

|uk (x, t)| ≤ C Tk (G(x, t + 1)) = C G(x, t + k −2 ).

This means a uniform control from above of the mass of all the tail mass of all the solutions uk ,
by which we mean the mass lying in exterior regions of space. Such control allows to avoid the
loss of mass that could occur in the limit by Fatou’s Theorem. Hence, the mass of Û is the same,
i.e., M .
The convergence of Û to M δ(x) as t → 0 happens because uk (x, 0) → M δ(x), and the previous
tail analysis shows that Û takes zero initial values for x 6= 0.
(ii) To recover the same result for a general u0 ≥ we use approximation by data as above and
then the L1 contraction of the heat semigroup. For signed solutions, separate the positive and
negative parts of the data and solve separately.

10
Step 5. Recovering the result. (i) We now use the convergence of the rescaled solutions at a
fixed time, say t = 1,
lim |uk (x, 1) − M G1 (x)| → 0.
k→∞

This convergence takes place locally in RN in all Lp norms. By the tail analysis, it also happens
in L1 (RN ). But since the derivatives are also uniformly bounded we have also an L∞ estimate.
1/2
(ii) We now use the meaning to the transformation Tk and write k = t1 to get
N/2 1/2
lim |t1 u(xt1 , t1 ) − M G1 (x)| → 0.
t1 →∞

−1/2
But this is just the result we wanted to prove after writing x = y t1 and observing that
−N/2
Gt1 (y) = t1 G1 (x). Similarly for the L1 norm.

This 5-step proof is taken from paper [48], where it was applied to the p-Laplacian equation. See
whole details for the porous medium equation in the book [66]. It has had further applicability.

Exercise 15. Fill in the details of the above proof.

4 Asymptotic convergence to the Gaussian via representation

The second proof we give of the main result, Theorem 2.1, is based on the examination of the error
in terms of the representation formula. This is a very direct approach, but it needs a previous step,
whereby the proof is done under the further restriction that the data have a finite first moment.
Then the convergence result is more precise and quantitative. This particular case has an interest
in itself since it shows the importance of controlling the first moment of a mass distribution. We
already know that the first moment is associated to a conserved quantity (see Exercise 8).

Theorem 4.1 Under the assumptions that u0 ∈ L1 (RN ) and that the first absolute moment is
finite
Z
(4.1) N1 = |u0 (y)y| dy < ∞ ,

we get the convergence

(4.2) tN/2 |u(x, t) − M Gt (x)| ≤ CN1 t−1/2 ,

as well as

(4.3) ku(x, t) − M Gt (x)kL1 (RN ) ≤ CN1 t−1/2 .

The rate O(t−1/2 ) is optimal under such assumptions.

Proof. (i) We may perform the proof under the further restriction that u0 ≥ 0. For a signed
solution we must only separate the positive and negative parts of the data and apply the results
to both partial solutions.

11
 (ii) Let us do first the sup convergence. We have
Z Z
u(x, t) − M Gt (x) = u0 (y)Gt (x − y) dy − Gt (x) u0 (y) dy
Z
= u0 (y)(Gt (x − y) − Gt (x)) dy
Z Z 1 
= u0 (y) ∂s (Gt (x − sy)) ds dy
Z 10
x − sy −|x−sy|2 /4t
Z
= Ct−N/2 dy ds u0 (y)hy, ie
0 2t

with C = (4π)−N/2 . Consider the piece of the integrand of the form

x − sy −|x−sy|2 /4t 2
f= e = ξe−|ξ| /4 ,
t1/2
where we have put ξ = (x − sy)/t1/2 . We observe that the vector function f is bounded by a
numerical constant, hence
Z
|u(x, t) − M Gt (x)| ≤ C1 t−(N +1)/2 |u0 (y)y| dy .

Taking into account that Gt is of order t−N/2 in sup norm, we write the result as (4.2), and
C = C(N ) is a universal constant.

(iii) For the L1 convergence we start in the same way and arrive at
Z 1
x − sy −(x−sy)2 /4t
Z
−N/2
u(x, t) − M Gt (x) = C t dy ds u0 (y)hy, ie .
0 2t
Now we integrate to get
1
|x − sy| −|x−sy|2 /4t
Z Z Z
−N/2
ku(x, t) − M Gt (x)k1 ≤ C t dx dy ds |u0 (y)y| e
Z 1 RN ZRN 0 2t
|x − sy|
Z
2
=C dy ds |u0 (y)y|t−1/2 ( t−N/2 1/2 e−|x−sy| /4t dx)
R N 0 RN t

With the change of variables x − sy = t1/2 ξ we already know that the last integral is a constant
independent of u, hence the formula for the L1 error. Note that now we are speaking of masses
and we do not need any renormalization time factor.
(iv) We ask the reader to prove the optimality as an exercise.

Exercise 16. Take as solution the Gaussian after a space displacement, u(x, t) = G(x + h, t), and
find the convergence rate to be exactly O(t−1/2 ). This is just a calculus exercise but attention to
details is needed. Hint: Write
h2 2
G(x + h, t) = Gt (x) + h∂x Gt (x) + D Gt (ξ)
2 x
(where ξ = x + sh, 0 < s < 1), and check that
2 /4
∂x Gt (x) = Ct−(N +1)/2 ξe−ξ = O(t−(N +1)/2 ) ,

12
uniformly in x, and Dx2 Gt (ξ) = O(t−(N +2)/2 ) uniformly in x. This exercise shows that the term
h∂x Gt (x) is the precise corrector with relative error O(t−1/2 ). We could continue the analysis by
expanding in Taylor series with further terms, see below. Exact correctors and longer expansions
for general solutions will be done later in Section 7 by the methods of Functional Analysis.

Remark. The factor tN/2 is the appropriate weight to consider relative error. We point out that
the more precise relative error formula

|u(x, t) − Gt (x)|
(4.4) rel (t) =
Gt (x)

does not admit a sup bound, as can be observed by choosing u(x, t) = Gt (x−h) for some constant
h since then
2
rel (t) = |exh/2t e−h /4t − 1| ,
which is not even bounded. It is then quite good that our weaker form does admit a good
estimate. Same happens for the L1 norm. This comment wants to show that error calculations
with Gaussians are delicate because of the tail (i.e., the behaviour for large |x|.

• Proof of Theorem 2.1 in this approach. Given an initial function u0 ∈ L1 (RN ) without any
assumption on the first moment, we argue by approximation plus the triangular inequality. In
the end we get a convergence result, but less precise. To quote, if u0 is integrable with integral
M and let us fix an error δ > 0. First, we find an approximation u01 with compact support and
such that
ku0 − u01 k1 < δ.
Due to the already mentioned effect L1 → L∞ , and applied to the solution u − u1 , we know that
for all t > 0
ku(t) − u1 (t)k∞ < C δ t−N/2 .
On the other hand, we have just proved that for data with finite moment:

ku1 (t) − M1 Gt k∞ ≤ CN1δ t−(N +1)/2 .

In this way, for sufficiently large t (depending on δ) we have

tN/2 ku1 (t) − M1 Gt k∞ ≤ Cδ

with C a universal constant. Next se recall that |M − M1 | ≤ δ as well as Gt ≤ Ct−N/2 . Using

the triangular inequality we arrive at

(4.5) lim tN/2 ku(t) − M Gt k∞ = 0 ,

t→∞

which ends the proof.

Remark. In the general conditions of Theorem 2.1 we still obtain convergence, but we no longer
obtain a rate. In fact, we show next that no convergence speed can be found without further
information on the data other than integrability of u0 .

13
4.1 No explicit rates for general data. Counterexample

Let us explain how the lack of a rate for the whole class of L1 functions is shown. Given any
decreasing and positive rate function φ such that φ(t) → 0 as t → ∞ we construct a modification
of the Gaussian kernel that produces a solution with the same mass M = 1 and such that it
satisfies a lower bound for the error of the form

tN/2 ku(x, t) − Gt (x)k∞ ≥ nφ(tn )

at a sequence of times tn → ∞ to be chosen.

P
Construcion. The idea is to find a choice of small masses m1 , m2 , . . . with n mn = δ < 1, and
locations xn with |xn | = rn → ∞ and consider the solution
∞
X
u(x, t) = (1 − δ)Gt (x) + mn Gt (x − xn ) .
n=1

Let us be precise. The error u(x, t) − Gt (x) is calculated at x = 0 as

∞ ∞
2 2
X X
(4πt)N/2 |u(0, t) − Gt (0)| = |δ − mn e−xn /4t | = mn (1 − e−xn /4t ) .
n=1 n=1

Put mn = 2−n (any other summable series will do). Choose iteratively tn and xn as follows.
Given choices for the steps 1, 2, . . . , n − 1, pick tn to be much larger than tn−1 and such that
φ(tn ) ≤ mn /2n. This is where we use the fact that φ(t) tends to zero, even if it may decrease in
2
a very slow way. Choose now |xn | = rn so large that e−rn /4tn < 1/2. Essentially, the mass has
1/2
to be displaced at distance equal or larger than O(tn ). Then,
2
mn (1 − e−xn /4tn ) ≥ nφ(tn ) .

Let us make some further practical calculations (with no precise scope in mind): Let for example
φ(t) = t− . Choose mn = 2−n . Then tn ≥ 2n ,

tn ∼ 2n/ , rn ∼ 2n/2

and the mass in the outer region {|x| ≥ rn } is approximately 2−n ; hence such a mass is M (r) ∼
r−2 , which is not so small if  → 0.

4.2 Infinite propagation in space

The representation formula immediately shows that a solution u corresponding to nonnegative

initial data will be strictly positive at all points x ∈ RN for any time t > 0. The infinite speed
of propagation of the heat equation with the instantaneous formation of a thin tail at infinity
is considered an un-physical property by many authors, one of the not many drawbacks of this
wonderful equation. However, it is essential to the equation and creates some curious effects.

• Spatial tails for positive solutions. Let us examine the precise form of the tail in the
simplest case where u0 has compact support. For simplicity and w.l.o.g. we assume that u0 is
supported in the ball BR of radius R > 0 centered at x = 0. Take x1 ∈ RN such that |x1 | = r1 > 1,

14
say x1 = r1 e1 . Then the representation formula implies that a bound from above is obtained by
displacing all the mass to the nearest point to x1 inside BR , which is x00 = Re1 , and we get the
upper bound
M −(|x1 |−R)2 /4t
u(x1 , t) ≤ M Gt (x1 − x00 ) = e .
(4πt)N/2
In the opposite direction, moving all the mass to x000 = −Re1 we get the lower bound

M 2
u(x1 , t) ≥ M Gt (x1 − x000 ) = e−(|x1 |+R) /4t .
(4πt)N/2

Both estimates are clearly optimal in this context. Since the equation is invariant under rotations
the result holds for all x such that |x| > R instead of our restricted choice of x1 . Moreover, we see
that the ratio of both estimates tends to infinity as |x| → ∞. It is therefore convenient to take
logarithms, and then we easily get a general formula where BR is centered at xc with |xc | = Rc .
R
Proposition 4.2 Let u0 be supported in the ball BR (xc ) and let M = u0 (x) dx. Then for every
t > 0 and every x 6∈ BR (xc ) we have

N 1 N 1
(4.6) log(4πt) + (|x − xc | − R)2 ≤ − log(u(x, t)/M ) ≤ log(4πt) + (|x − xc | + R)2 .
2 4t 2 4t
It follows that

log(u(x, t)/M ) 1
(4.7) lim =− .
|x|→∞ |x|2 4t

We conclude that in first approximation the tail at infinity of all solutions with compactly
supported initial data is universal in shape and depends only on the mass of the data and time.
Of course, the second term in the expansion depends also on the radii Rc and R.
Another observation is that the asymptotic space behaviour allows to calculate the time elapsed
since the solution had compact support (if we already know the mass M ).
Let us also remark that solutions with more general nonnegative data can have other type of
tails and we invite the reader to calculate some of them, both for integrable and non-integrable
data. Here is an example that decays like a simple exponential.
Exercise 17. Consider the heat equation in 1D. Show that when u0 is integrable, positive and
bounded and u0 (x) = e−x for x ≥ 0, then for all times

lim ex u(x, t) = et .
x→∞

Sketch: Putting u0 (x) = 0 for x < 0 for simplicity, use the representation formula to write
Z x
1 2
x−t
e u(x, t) = 1/2
e−(y−2t) /4t dy
(4πt) −∞

Let then x → ∞. You may also use the explicit solution U (x, t) = et−x+c .
In any case the behaviour described in Proposition 4.2 is the minimal one for nonnegative
solutions.

15
• Signed solutions do not become everywhere positive.
The square exponential tail behaviour of the fundamental solution has more curious consequences.
R + if a signed solution u(x, t) has initial data such that theRmass
Thus, of the positive part M+ =
u0 (x)dx is larger than the mass of the negative part M− = u− 0 (x)dx, then we know that it
converges as t → ∞ to the positive Gaussian M Gt with M = M+ − M− > 0 so that

(4.8) lim tN/2 u(x, t) = (4π)−N/2 M > 0.

t→∞

for every x ∈ RN . It would be natural to expect that when M+ is much larger than M− then the
solution u is indeed positive everywhere for large enough, finite times. Now, this is true in most
of the space because of the previous convergence, but it is not true in all the space.

Exercise 18. (i) Take N = 1. Show that for the choice u0 (x) = δ(x) − ε δ(x − 1) with 0 < ε < 1
(a combination of delta functions) the solution u is positive on the left of a line x = r(t) and
negative for x > r(t). Show that for large times r(t) ∼ 2 log(1/ε)t.
Remark. We see that the problem arises at the far away tail. Of course, u(x, t) becomes positive
for large times at all points located at or less than the typical distance, |x| ≤ Ct1/2 .
(ii) Show that a similar result is true for integrable data if u0 is positive for x < 0 and negative
for x > 1, and zero in the middle. Show in particular that u(x, t) < 0 if x > 2t log(1/ε) + 1/2.
(iii) State similar results in several dimensions.

5 Improved convergence for distributions with second moment

Better convergence rates can be obtained by asking a better decay at infinity of u0 . The tech-
nical condition we use is having a finite second moment, a condition that is very popular in the
literature. In probability this is known as having a finite variation. The motivating example is
described next.

Exercise 19. Consider a time displacement of the fundamental solution and show that u(x, t) =
G(x, t + t0 ) has the precise convergence rate O(1/t) towards G(x, t).
The result we prove is as follows.

Theorem 5.1 Under the assumptions that u0 ∈ L1 (RN ) and that the signed first moment
Z
(5.1) N1,i = yi u0 (y) dy
RN

is finite (for all coordinates), as well as the second moment:

Z
(5.2) N2 = y 2 |u0 |(y) dy < ∞,
RN

we get the convergence

X
(5.3) tN/2 |u(x, t) − M Gt (x) + N1,i ∂xi Gt (x)| ≤ CN2 t−1 ,
i

16
and
X
(5.4) ku(x, t) − M Gt (x) + N1,i ∂xi Gt (x)kL1 (RN ) ≤ C N2 t−1 .
i

The rate O(t−1 ) is optimal under such assumptions.

Remark. The signed first moment is called in Mechanics the center of mass, for M 6= 1, M 6= 0,
we use the formula
Z
1
(5.5) xc = x u0 (x)dx .
M
In probability it is the average location of the sample and M = 1. When xc is finite and M 6= 0,
the center of mass can be reduced to zero by just a displacement of the spatial axis. This very
much simplifies formulas (5.3), (5.4), see below.

Proof. (i) Starting as in Theorem 4.1 we arrive at the formula

Z  
2 2
D := u(x, t) − M Gt (x) = (4πt)−N/2 e−|x−y| /4t − e−|x| /4t u0 (y) dy.
RN
2 /4t
Let us consider the 1D function f (s) = e−|x−sy| , s ∈ R, and let us use the Taylor formula
Z 1
f (1) = f (0) + f 0 (0) + f 00 (s)(1 − s) ds .
0

We have
1 2 1 2 1 2
f 0 (s) = hy, x − syie−|x−sy| /4t , f 00 (s) = (− |y| + 2 |hy, x − syi|2 )e−|x−sy| /4t .
2t 2t 4t
Using these results, we get D = D1 + D2 , where
Z
x X Z 
−(N +1)/2 −|x2 |2 /4t
D1 = (4πt) u0 (y)hy, 1/2 ie dy = − yi u0 (y) dy ∂i Gt (x) .
2t i
√
On the other hand, putting ξ = (x − sy)/2 t and yb = y/|y|, we also have
Z Z 1  
−(N +2)/2 1 2
D2 = (4πt) 2
|y| u0 (y) − + (hb y , ξi) e−ξ dyds.
2
RN 0 2
Since the factor dependent on ξ is uniformly bounded for all ξ we have
Z 
|D2 | ≤ C(N ) u0 (y) |y| dy) t−(N +2)/2 .
2
RN

This proves the result.

(ii) We leave to the reader to prove the corresponding statement in L1 norm.
(iii) Optimality of the rate follows from Exercise 19.

Exercise 20. Give examples of well-known probability distributions for which the first moment
is finite or infinite. Same for the second moment. Try with examples of the form

f (x) = C(1 + |x|2 )−a .

17
For a = 1 we get the well-known Cauchy distribution that is integrable only in 1D.
Answers. Integrable 2a > N , 1st moment 2a > N + 1, 2nd moment 2a > N + 2.

Reformulation of the result. It is well-known that the first moment can be eliminated to
moving the origin of coordinates to the center of mass xc defined in formula (5.5) when M 6= 0.
If we do that and then apply Theorem 5.1 we get the following result

Corollary 5.1.1 Under the assumptions of Theorem 5.1 we have

|u(x, t) − M Gt (x − xc )| ≤ C N2∗ (u0 ) t−(N +2)/2 ),

(
(5.6)
ku(x, t) − M Gt (x − xc )k1 ≤ C N2∗ (u0 ) t−1 ) ,

where N2∗ (u0 ) = RN |x − xc |2 u0 (x) dx is the centered second moment.

R

Higher development result. A continuation of this method into stricter convergence rates
using higher moments can be done by using further terms in the Taylor series development. We
will not do it but only quote the statement that can be found in [30].
Theorem 4 [30] Let G(x, t) be the heat kernel. For any 1 ≤ p ≤ N/(N − 1) and k ≥ 0 an integer
the solution of initial value problem for the heat equation satisfies:
X (−1)|α| Z 
ku(x, t) − f (x)x dx Dα G(x, t)kp ≤ Ck t−(k+1)/2 k|x|k+1 f (x)kp ,
α
α!
|α|≤k−j

for any initial data f ∈ L1 (RN , 1 + |x|k ) such that |x|k+1 f ∈ Lp (RN ).
Let us note that their proof is different from the previous ones and interesting. As a conclusion,
we also have a result about convergence in the first moment norm.

Theorem 5.2 Under the assumptions that u0 ∈ L1 (RN ) and that the first, second and third
moments are finite, we get the convergence
X
(5.7) ku(x, t) − M Gt (x) + N1,i ∂xi Gt (x)kL1 (|x|dx) ≤ C t−1/2 .
i

The constant C > 0 depends on u0 through the second and third moments.

Proof. (i) We start with formulas from Theorem (5.1) where it is proved that
X
|u(x, t) − M Gt (x) + N1,i ∂xi Gt (x)| = D2
i

with Z Z 1  
1 2
D2 = (4πt)−(N +2)/2 y , ξi)2 e−ξ dyds,
|y|2 u0 (y) − + (hb
RN 0 2
√
and ξ = (x − sy)/2 t. Multiplying by |x| and integrating we have
Z ZZZ  
−(N +2)/2 1 2
D2 |x|dx = (4πt) 2
y , ξi) e−ξ |x|dx dyds.
|y| u0 (y) − + (hb 2
RN 2

18
 √
Writing now |x| ≤ 2|ξ| t + s|y|, we split the upper estimate of this integral into I1 + I2 , where
ZZZ  
−(N +1)/2 1 2
I1 = Ct 2
|y| u0 (y) − + (hb y , ξi) e−ξ 2|ξ|dx dyds ,
2
2

and ZZZ  
−(N +2)/2 1 2
I2 = Ct 2
y , ξi) e−ξ s|y|dx dyds .
|y| u0 (y) − + (hb 2
2
As for the first integral, the separate integrals are bounded and only the ξ integral gets a time
factor tN/2 from dx = tN/2 dξ, so that

I1 ≤ C t−1/2 .

The second integral easily gives I2 ≤ C t−1 by the assumption on the third moment of u0 . The
proof is complete.

General conclusion. All integrable solutions of HE in the whole space converge to Gaussian (in
the renormalized forms we have written) if the initial mass is finite, u0 ∈ L1 (RN ). But the speed
with which they do depends on how much initial mass is located far away, in other colloquial words
for probabilists, on “how populated the tails are”. The quantitative versions we have established
use mainly the moments of order 1 and 2. The moment of order 2 is called in Probability the
(square of) the standard deviation. We remind the reader that not all probability distributions
have a finite standard deviation (see Exercise 20).

6 Functional Analysis approach for Heat Equations

We are going to use energy functions of different types to study the evolution of dissipation
equations. The basic equation is the classical heat equation, but the scope is quite general. Our
aim is not to establish the convergence of general solutions to the fundamental solution (which
is well done by other methods, as we have shown), but a bit more, namely, to find the speed
of convergence. After change of variables (renormalization) this reads as rate of convergence
to equilibrium and relies on functional inequalities. These functional inequalities also play an
important in other areas.
The methods we will introduce next will apply to more general linear parabolic equations that
generate semigroups. The method also works for equations evolving on manifolds as a base space.
Since around the year 2000 we have been studying these questions for nonlinear diffusion equations.
The main nonlinear models are: the porous medium equation, the fast diffusion equation, the p-
Laplacian evolution equation, the chemotaxis system, some thin film equations, ... Recently, the
fractional heat equation and various fractional porous medium equations have been much studied.

6.1 Heat Equation Transformations

Take the classical Heat Equation posed in the whole space RN for τ > 0:
1
uτ = ∆y u
2

19
with notation u = u(y, τ ) that is useful since we want to save the standard notation (x, t) for later
use. We know the (self-similar) fundamental solution, also called the evolution Gaussian profile
2 /2τ
U (y, τ ) = C τ −N/2 e−y .

It was proved in previous sections that this Gaussian is an attractor for all solutions in its basin
of attraction, consisting on all solutions with initial data that belong to L1 (RN ) with integral
M = 1. See Sections 3, 4, and 5.
Remark. We have inserted a harmless factor 1/2 in front of the left-hand side following the prob-
abilistic convention in order to get a Gaussian with clean exponent −|y|2 /2t which has standard
deviation t1/2 with no extra factors. Eliminating the prefactor leads to the exponential expres-
sion with exponent −|y|2 /4t, usual in PDE books. Our convention leads to some other simpler
constants.

• Fokker-Planck equation. It is the first step in this approach to the asymptotic study. The
scaling on the variables u and y to factor out the expected size of both which must mimic the
Gaussian sizes, and then take logarithmic scale for the new time

u(y, τ ) = v(x, t) (1 + τ )−N/2 , 2t = log(1 + τ ).

After some simple computations this leads to the well-known Fokker-Plank equation for v(x, t):

(6.1) vt = ∆x v + ∇x · (x v) = ∇x (∇x v + xv) .

We can write it as vt = L1 (v), where the Fokker-Planck operator L1 = LF P can be written in

more explicit form as
L1 (v) = ∆v + x · ∇v + N v .
We check now that when we look for stationary solutions by putting vt = 0 we get as easiest case
the equation ∇v+x·v = 0 (after cancelling a divergence). Integrating it under the radial symmetry
2
assumption is the simplest way to get the Gaussian distribution G = c e−x /2 , and indirectly, the
−N/2 to
R solution of the original heat equation. We choose the constant c = (2π)
fundamental
normalize G dx = 1.
The asymptotic result we are aiming at consists precisely of proving that when v0 (x) is integrable
with mass 1 then v(x, t) will tend to G as t → ∞. For a general presentation of the FP equation see
[57]. We will keep the notation v(x, t) for the solutions of the Fokker-Planck equation throughout
this section.

• The Ornstein-Uhlenbeck semigroup. (i) In order to study relative error convergence it

seems reasonable to pass to the quotient w = v/G, where G is the stationary state, to get the
Ornstein-Uhlenbeck version

wt = L2 (w) = G−1 ∇ · G ∇w = ∆w − x · ∇w ,


(6.2)

a symmetrically weighted heat equation with Gaussian weight. Note that the corresponding
stationary solution is now W = 1, much easier.

(ii) The two-term form of the r.h.s. looks easier, with a diffusion and a convection term. Indeed,
the weighted form of the Ornstein-Uhlenbeck operator L2 = LOU is very convenient for our

20
calculations. To begin with, it allows to prove the symmetry of the operator in the weighted
space X = L2 (Gdx): for every two convenient functions w1 and w2 we have
Z Z Z
(6.3) (L2 w1 ) w2 Gdx = w1 (L2 w2 ) Gdx = − h∇w1 , ∇w2 iGdx.
RN RN RN

It seems natural to introduce the Gaussian measure dµ = Gdx as a reference measure in the
calculations. The important consequence of this computation is that A = −L2 is a positive and
self-adjoint operator in the Hilbert space X = L2 (dµ). This is a rather large space that includes
all functions with polynomial growth. We will keep the notation w(x, t) for the solutions of the
Ornstein-Uhlenbeck equation throughout this section.

(iii) We may also observe that L1 (v) = G L2 (v/G) and that L2 is the adjoint to L1 in the sense
that for conveniently smooth and decaying functions
Z Z
(L1 v)w dx = v (L2 w)dx.
RN RN

More formally, we can consider the duality between the spaces X = L2 (Gdx) and X 0 = L2 (G−1 dx)
given precisely by the integral of the product and the operators are adjoint.

(iv) Finally, to complete the comparison we can write the Fokker-Planck equation as

∂t v = ∇ · (G∇(v/G)) .

The analogy says that this operator is negative and self-adjoint in the stranger space X1 =
L2 (G−1 dx), that is much smaller than L2 (RN ). For a detailed mathematical presentation of the
Ornstein-Uhlenbeck semigroup we refer to [2].

• The Hamiltonian connection. Start from the Fokker-Planck equation and use now the
change of variables v = zG1/2 . Then,

G1/2 zt = ∇ · (G∇(zG−1/2 )).

We have

∇ · (G∇(zG−1/2 )) = G1/2 ∆z + ∇G1/2 · ∇z + G∇G−1/2 · ∇z + z∇(G∇G−1/2 ),

so that the equation for z becomes:

zt = ∆z − V (x)z, V (x) = G−1/2 ∆G1/2 ,

that we may write as a real Schrödinger Equation zt = L3 (z) = −H(z) with Hamiltonian operator
1 N
H(z) = −∆z + V (x)z, V = |x|2 − ,
4 2
2
In calculating the Schrödinger potential we have used G1/2 = e−|x| /4 . Operator H is directly
symmetric in L2 with no weight. The fact that it is positive is not clear from the formulas but it
will follow from the equivalence with the Ornstein-Uhlenbeck operator.

• The equivalence of this equation with the former ones comes from the transformation formulas

L1 (v) = G1/2 L3 (v/G1/2 ) = G L2 (v/G),

21
and also that if vi = G1/2 zi = Gwi , i = 1, 2, we get
Z Z Z
v1 v2 G−1 dx = z1 z2 dx = w1 w2 Gdx,
RN RN RN

which allows to show that all three operators Li are self-adjoint and dissipative since we have
already proved it for L2 . The interesting remark for the Schrödinger representation is that it does
not need any weighted space, X3 = L2 (RN ).

The rich equivalence among the three equations and also with the heat equation is a favorite
topic in Linear Diffusion and Semigroup Theory.

• General Fokker-Planck Equation. It is based on generalising the coefficient x of the con-

vection term into a more general term that is the gradient of a potential that we call S(x).1 The
General Fokker-Planck equation (GFP) reads

(6.4) vt = ∆v + ∇ · (∇S v) .

Standard assumption is that S must be a positive and convex function in RN , called the potential.
The stationary state is now G(x)
e = C e−S(x) , and the equation reads then vt = L1 (v) with

L1 (v) = ∇ · (G
e ∇(v/G)).
e

The other equations, General Ornstein-Uhlenbeck and General Hamiltonian Equation, follow in
the same way as before using only the expressions in terms of G that is replaced by G. e The
weighted scalar products have no difference and the relation of norms still holds. In the Hamilto-
nian representation we get a potential (put G = e−S )
1 1
V (x) = −G−1/2 ∆(G1/2 ) = |∇S|2 − ∆S .
4 2

On the other hand, the analysis of the complete spectrum is not possible unless we have very
particular cases of potentials S. Moreover, the connection with a renormalization of the heat
equation is completely lost.

6.2 Asymptotic Energy Method via the Ornstein-Uhlenbeck Equation

The Ornstein-Uhlenbeck formulation allows for a very clear and simple treatment of the problem
of convergence with rate to the Gaussian profile. We may assume without lack of generality that
Z Z Z
w dµ = v dx = u dy = 1 ,

with the above notations for u, v, and w. We now make a simple but crucial calculation on the
time decay of the energy for the OUE:
Z Z
dF(w(t))
(6.5) F(w(t)) = |w − 1|2 G dx, = −2 |∇w|2 G dx = −D(w(t)).
RN dt R N

1
The standard notation is U (x) but we will change the notation here to S(x) to avoid confusion with other uses
of U .

22
We can now use a result from abstract functional analysis: the Gaussian Poincaré inequality
with measure dµ = G(x) dx:
Z Z 2 Z
2
w dµ − w dµ ≤ CGP |∇w|2 dµ
RN RN RN

The sharp constant in this inequality is precisely CGP = 1, with no dependence on dimension.
Moreover the functions that realize the optimal constant are w(x) = xi for any i = 1, 2, . . . , N .
2 We will give below a proof based on the analysis of the spectrum of the Ornstein-Uhlenbeck

operator.
Then, the left-hand side is just F and the inequality implies
d
− F(w(t)) ≥ 2F(w(t)),
dt
which after integration gives F(w(t)) ≤ F(w0 )e−2t , i.e.:
Z Z
|w − 1|2 dµ ≤ e−2t |w0 − 1|2 dµ ∀ t ≥ 0
RN RN

We have proved the convergence to equilibrium in the following form.

Theorem 6.1 Under the assumptions on the initial data

Z Z
w0 dµ = 1, w02 dµ < ∞ ,
RN RN

the solutions of the OUE satisfy the following stabilization estimate

kw(t) − 1kL2 (dµ) ≤ kw0 − 1kL2 (dµ) e−t .

This is theR first of the well known Functional Estimates for the solutions to the HE. The renor-
malization w0 dµ = 1 is no restriction.
R
In terms of v the hypotheses are v0 dx = 1 and kv0 − GkL2 (G−1 dx) < ∞. The weight is now
K = G−1 and the measure dν = K dx, which behaves inversely in infinity. The result is
Z Z
2 −2t
|v − G| dν ≤ e |v0 − G|2 dν ∀ t ≥ 0.
RN Rd

Recall that we have a new logarithmic time t = log(1 + τ ). The rate of convergence in real time
in both variables is then O((1 + τ )−1/2 ) = O(τ −1/2 ) as τ → ∞.

• Then by regularity theory (regularizing effect from L1 to L∞ ) for the heat equation and the
other equations, we get estimates in the sup norm with similar relative rates at least locally in
space. But note that since w = xe−t is a solution of the OUE, we cannot get a uniform estimate
in L∞ of the Ornstein-Uhlenbeck variable.
2
This is an old inequality in the folklore of Hermite polynomials, and probably was known in one dimension to
both mathematicians and physicists in the 1930’s in relation to eigenvalue problems, as mentioned in [9].

23
7 Calculation of spectrum. Refined asymptotics
We will proceed with a further step in the analysis to get more precise asymptotics. Indeed, the
knowledge of the spectrum of the equivalent operators allows to obtain a complete description
of the long-time behaviour in weighted spaces. This is a well-known fact in the study of the HE
posed in bounded domains, that has a parallel here.

7.1 Spectrum

We will make all computations on the Ornstein-Uhlenbeck operator. The Gaussian Poincaré
inequality is a simple consequence of the following analysis of the spectrum of the Ornstein-
Uhlenbeck Operator in L2 (dµ):
(a) Since the FP and the OU operators have a compact inverse, we conclude that they have a
discrete spectrum, [50]. The ground state of the OUE is formed by the constant function w = 1
(which comes from the Gaussian function v = G for the FPE) with eigenvalue λ0 = 0. The next
eigenfunctions are the coordinate functions φi = xi corresponding to the eigenvalue λ1 = 1 with
multiplicity N .
(b) In 1D we find the rest of the eigenfunctions and eigenvectors as the family of Hermite
polynomials, given by the compact formula

(7.1) Hk (x) = (−1)k G(x)−1 (d/dx)k G(x) ,

that tells much about how we will see them. Indeed, the formula can be derived from the fact
that the derivatives in y of the Gaussian evolution solution U (y, τ ) are still solutions of the heat
equation with different decay rate. Passing to the FPE we conclude that the derivatives Dk G
are eigenfunctions of L1 (here D = d/dx). It is easy to see that these solutions have the form
Hk (x)G(x) where Hk is a polynomial of degree k (proof by induction). Passing to the OUE we
get the formula above. More precisely, we have the recursion formula:
d
GHk+1 = −d/dx(Hk G) = −(Hk0 − xHk )G, Hk+1 = (x − )Hk .
dx
The first members of the family Hk are 1, x, x2 − 1, . . .
The corresponding eigenvalue to Hk for L2 = LOU is λk = k. This can be seen from the heat
equation formula since differentiating in x adds a factor τ −1/2 to the decay, which goes over as
e−t for every derivative we take. Induction proof at the FP level: if we assume that

L1 (Dk G) = (G(Dk G/G)0 )0 = (Dk G)00 + x(Dk G)0 = −λk Dk G ,

then

L1 (Dk+1 G) = (Dk G)000 + x(Dk G)00 = −λk Dk G0 − (x(Dk G)0 )0 + x(Dk G)00 = −(λk + 1)Dk+1 G.

Therefore, λk+1 = λk + 1. Since λ0 = 0 the proof is done.

It is then proved that the Hk (x) form a basis in L2 (dµ) in 1D.
(c) For several dimensions we have the functorial property: if x = (x1 , x2 ) and w(x) = w1 (x1 )w2 (x2 )
we get
L2 (w) = w1 L2 (w2 ) + L(w1 ) w2 .

24
This produces new eigenfunctions in higher dimensions, and it gives for the product function the
sum of the eigenvalues: λ(w) = λ(w1 ) + λ(w2 ). The set of combinations generates a base of
eigenfunctions, this is essentially due to Fubini’s theorem, cf. [10]. Our account is very short but
we consider this part an extension, and it is well documented in the corresponding literature.

7.2 Refined asymptotics

From the spectrum we can get a very precise description of the convergence in the weighted
spaces by using the equivalent of the Fourier analysis on bounded domains. The meaning of the
coefficients for the original equation has to be understood.
We get X X
w(x, t) = cα Hα (x)e−kt , v(x, t) = cα ∂α G(x)e−kt
α α

where α is an N -multi-index, k = |α|, and Hα is the corresponding multidimensional Hermite

polynomial, after renormalization in L2 (dµ). We have

hw0 , Hα iL2µ
R
v0 (x)Hα (x) dµ(x)
cα = = R ,
hHα Hα iL2µ Hα2 (x) dµ(x)
R
so that c0 is the mass and for k = 1, and ci are the first coordinate moments v0 (x)xi dx after
normalization, and so on. The convergence of the w series holds in L2 (dµ), with errors of the
order of the first term that is left out. The convergence of the v series holds in L2 (G−1 dx), with
errors of the order of the first term that is left out.

8 Convergence via the Boltzmann entropy approach

There is another approach for the convergence to the Gaussian that starts the analysis from
Boltzmann’s ideas on entropy dissipation. We start now from the Fokker-Planck equation vt =
∆v + ∇ · (xv) and consider the functional called entropy
Z Z Z
1
E(v) := v log(v/G) dx = v log(v)dx + x2 v dx + C .
RN RN 2 RN

and we assume that the data are such that the initial entropy is finite. We recall that no decay
is possible without some restriction on the data.
Differentiating along the flow (i.e., for a solution) leads to
 2
 ∇v
Z Z
dE(v)
v |∇ log(v/G)|2 dx .

= −I(v), I(u) = v  + x dx =
dt RN v RN

For some reasons the dissipation I(v) is called Fisher information. Let us continue the proof.
Putting now v = Gf 2 we find that
Z Z
2
E(v) = 2 f log(f ) dµ, I(v) = 4 |∇f |2 dµ.
RN RN

25
The famous logarithmic Sobolev inequality proved by Gross in 1975, [42], says than that (for all
suitable functions, not only solutions)
1
E ≤ I,
2
−2t
and we obtain the decay E(t) ≤ E(0) e . This means a precise decay for the entropy functional.
The calculations are justified for smooth solutions, and then we can pass to the limit for general
solutions with finite entropy.
In order to obtain decay in standard norms, there are formulas connecting the entropy with
other norms, like the Cziszar-Kullback inequality that implies that that

kf − Gk2L1 (RN ) ≤ 2E(f, G),

for any f, G ∈ L1 (RN ) positive with equal mass, see [8]. This is paper is a very good early
reference to the subject of entropies and the central limit theorem.
There are many works dealing with the use of functionals and functional inequalities to arrive at
asymptotic behaviour results plus a rate of convergence for this kind of equations. Let us mention
here [3, 4, 25, 53, 63] and the references to be mentioned in Section 10.

8.1 About entropy in Physics

Entropy has been introduced as a state function in thermodynamics by R. Clausius in 1865, in

the framework of the second law of thermodynamics, in order to interpret the results of S. Carnot.
A statistical physics approach: Boltzmann’s formula (1877) defines the entropy of a physical
system in terms of a counting of its micro-states. Boltzmann’s equation:

∂t f + v · ∇x f = Q(f, f ) .

It describes the evolution of a gas of particles having binary collisions at the kinetic level; f (t, x, v)
is a time dependent distribution function (probability density) defined on the phase space (x, v) ∈
N N
RR
R ×R . The Boltzmann entropy: H[f ] := f log(f )dxdv measures irreversibility. The famous
H-Theorem (1872) says that
ZZ
d
H[f ] = Q(f, f )log(f )dxdv ≤ 0 .
dt
Other approaches to thermodynamic entropy are due to Carathéodory (1908), Lieb-Yngvason
(1997),... see [52].
An important version of entropy appears in Information Theory. In 1948, while working at
Bell Telephone Laboratories Claude Shannon, an electrical engineer, set out to mathematically
quantify the statistical nature of “lost information” in phone-line signals (cf. Wikipedia article).
He arrived at an analog to thermodynamic entropy for use in information entropy.
There is also a concept of entropy in probability theory (with reference to an arbitrary measure).

9 Brief review of other heat equation problems

The methods presented above have been applied to prove convergence to a distinguished solution
(that plays the role of the Gaussian fundamental solution) in different contexts. We have not

26
mentioned some other methods like the transport method of Jordan-Kinderlehrer-Otto [44], 1998,
where the Fokker-Planck equation is interpreted as the steepest descent for a free energy related
to Boltzmann-Gibbs entropy, taken with respect to the Wasserstein metric. This novel technique,
based on mass transportation [71], has played an increasing role since then

9.1 Equation with forcing

A modification that still keeps the flavor of this presentation consists of considering a forcing term

(9.1) ∂t u = ∆u + f ,

where f is an integrable function of (x, t) ∈ QT = RN × (0, T ), T > 0. It can be easily proved

that a representation formula holds, [31]. From it we can derive asymptotic results that we leave
as exercises.

Exercise 21. (i) Prove that the L1 estimate of Theorem 2.1, formula (2.1), holds if f ∈ L1 (Q∞ ),
and we take as M the accumulated mass defined as
Z ZZ
(9.2) M= u0 (x) dx + f (x, t) dxdt.
RN Q∞

(ii) The L∞ statement (2.2) needs some further decay condition on f as t → ∞. Prove that it
holds e.g. if kf (·, t)k∞ ≤ Ct−γ with γ ≥ 1 + N/2.

The results of Section 4 can be repeated under conditions that we invite the reader to provide.

• Let us continue with a variation of the heat equation with forcing. In fact, there are many studies
where the forcing term takes the form f = f (u), and they fall into what is called reaction-diffusion,
[60]. The simplest case correspond to linear forcing, f (u) = κu with κ 6= 0. The modifications in
the analysis are minimal since the change of variables v(x, t) = u(x, t) e−κt transforms equation

(9.3) ∂t u = ∆u + κu ,

into the classical form vt = ∆v, to which our results apply. We thus conclude that for very large
t we have the asymptotic behaviour

(9.4) u(x, t) ∼ M G(x, t) eκt ,

which preserves the Gaussian profile as asymptotic shape but not the decay rates in time. The
reader is asked to write the precise theorems using the results of Sections 3, 4, and 5.

9.2 Dipoles and related issues

1
RWe have already seen that the heat equation with signed L initial data and zero mass, i. e.,
u0 (x) dx = 0 in the sense that the term representing the Gaussian approximation vanishes, so
that the rate of decay as t → ∞ is faster and the first approximation is given by a first term that
combines partial derivatives of the Gaussian :
X
(9.5) u(x, t) ∼ Dv Gt (x) , v= N1,i (u0 ) ei ,
i

27
of course under the condition that this vector v does not vanish. The ei are the canonical basis
and Dv denotes directional derivative. Therefore, u(x, t) = O(t−(n+1)/2 ).

• There is a very interesting application of this result in N = 1. Indeed, we may solve the problem
of asymptotic behaviour of the solutions of the heat equation posed in a half line I = (0, ∞) ⊂ R
for t > 0 with lateral Dirichlet data u(0, t) = 0, let us call it (DP-HE-HL). We assume that
u0 ∈ L1 (I) and u0 ≥ 0 (the last assumption is made for simplicity). The idea is tho extend the
initial data to the whole line by putting

u
b0 (x) = u0 (x) if x > 0, b0 (x) = −u0 (−x) if x < 0
u

(called anti-symmetric reflection). Note that u b ∈ L1 (R), its total mass is zero, and the first
moment 2N1 is not zero. Solving the heat equation with the usual representation formula we
obtain a solution ub(x, t) defined for x ∈ R and t > 0. This solution must be antisymmetric,
b(x, t) = −b
u u(−x, t) by the form of the data and an elementary symmetry property of the heat
equation. Restricting u
b to x > 0 we find a unique solution of Problem (DP-HE-HL). We can now
copy our asymptotic results to u b and translate them to u. Let us write
x 2 /4t
(9.6) D(x, t) = −∂x Gt (x) = e−|x| .
t3/2
Theorems 5.1 and 5.2 imply that D(x, t) is the asymptotic attractor of the evolution in the half
line.

Theorem 9.1 Let us assume that u0 ∈ L1 (I) and the first moment in I
Z ∞
(9.7) N1 = y u0 (y) dy
0
R∞
is finite, as well as the second moment : N2 = 0 |y 2 u0 |(y) dy < ∞. Then, we get the following
convergence formulas with rate for the solutions of (DP-HE-HL)

ku(x, t) − 2N1 D(x, t)kL1 (I) ≤ C N2 t−1 ,

t1/2 |u(x, t) − 2N1 D(x, t)| ≤ CN2 t−1 .

The rate O(t−1 ) is optimal under such assumptions. If the third moment is finite we also have

(9.8) ku(x, t) − 2N1 D(x, t)kL1 (I;|x|dx) ≤ C (N2 + N3 ) t−1/2 .

Function D(x, t) given by (9.6) is called the dipole solution because it takes the derivative of
the unit Dirac delta as initial data. It hasR a constant-in-time first moment that characterizes its
∞
strength. Not that the mass in I, M+ = 0 u(x, t) dx, is not conserved in time, but decays like
O(t−1/2 ). 3
We can derive a more general convergence result

Theorem 9.2 Let us assume that u0 ∈ L1 (I) and the first moment in I, N1 , is finite. Then,

(9.9) ku(x, t) − 2N1 D(x, t)kL1 (I;|x|dx) → 0

as t → ∞.
3
Dipole solutions appear often in Physics, specially in Electromagnetism.

28
Proof. (i) Assume that u ≥ 0. Approximate u0 by a compactly supported function u e0 ≤ u0 , so
that the error in the first moment is ε. Then,
Z ∞
ku(x, t) − u
e(x, t)kL1 (I;|x|dx) = (u(x, t) − u
e(x, t))xdx ≤ ε
0

because this expression is conserved with time for solutions of Problem (DP-HE-HL) in I (check
this!). The previous theorem for the solution u
e gives

ke e1 D(x, t)kL1 (I;|x|dx) ≤ K t−1/2 .

u(x, t) − 2N

that can be made less than ε if t is large enough. Finally, |N1 − N

e1 | ≤ ε and D has a finite
moment. Combining all this, the result holds.
(ii) For signed data, split into positive and negative part, and then combine the results.

• Similar formulas hold for the heat equation posed in a half space Ω in N dimensions. After
rotation and translation we may take Ω = {x ∈ RN : x1 > 0}. We consider zero lateral Dirichlet
data: u(x, t) = 0 for x = (0, x2 , . . . , xN ). n this case a multi-dimensional dipole solution appears,
D = −∂x1 Gt . Problems in half spaces can be solved too. The equivalent of Theorem 9.1 holds.
We leave the easy details to the reader.
Finally, solutions of the type D12 = ∂x1 ∂x2 Gt are attractors for problems posed in quadrant
domains {x ∈ RN : x1 > 0, x2 > 0}. And so on.

• By using the symmetric extension instead of the anti-symmetric one we can solve the heat
equation posed in a half line I = (0, ∞) ⊂ R for t > 0 with lateral Neumann data ux (0, t) = 0,
let us call it (NP-HE-HL). Let the reader fill in the details if needed.
• A related problem in several dimensions occurs when the domain is an exterior domain with
one or several holes and appropriate boundary conditions. Thus, with zero Dirichlet boundary
conditions and integrable data convergence to the Gaussian holds, while in 1D we fall back into
the dipole problem.

9.3 Other problems in subsets of RN

There are a number of problems involving the heat equation that have been studied in great
detail, like the heat equation posed in a bounded domain with boundary conditions of different
types (Dirichlet, Neumann, mixed, or other), but these settings lead to quite different results that
depart too much from the picture presented here, so we will not comment on them.
There are also equations with coefficients or weights; they form a large topic that leads also very
far from the present presentation.

9.4 Heat equation on manifolds

The construction of the heat equation has been carefully studied when the equation is posed on
a Riemannian manifold (M N , g). The equation then takes the form
N
X  
(9.10) ∂t u(x, t) = ∆g u(x, t) = |g(x)|−N/2 ∂i g(x)ij |g(x)|N/2 ∂j u(x, t) ,
i,j=1

29
where gij is the metric tensor, g ij its inverse, |g| its determinant, so that ∆g is the Laplace-
Beltrami operator, [40]. Two particular manifolds are specially relevant because their internal
symmetries and homogeneity make the theory specially strong and mathematically appealing:
the N -dimensional sphere SN and the hyperbolic space HN . The heat flow on the former is easily
shown to stabilize to a constant (much like a Neumann problem in RN ). The flow on HN is more
interesting, and typical solutions with finite mass converge to a modified Gaussian function, the
hyperbolic fundamental solution, that is described in detail in [41], see also [39].

10 Application to other diffusion equations

We now examine some nonlinear diffusion equations where similar methods and results have been
successfully proved in the last half century, where the Gaussian profile is replaced by some other
attractive object.

• A prominent example that has been much studied is the Porous Medium Equation, ∂t u = ∆um ,
or better ∂t u = ∆(|u|m−1 u), m > 1. The asymptotic study depends crucially on the existence and
properties of a distinguished family of solutions, the Barenblatt solutions, [6], that are compactly
supported and self-similar, one for every mass M . Explicit formulas exist for them (1950, 1952):

1/(m−1)
(10.1) B(x, t; M ) = t−α F(x/tβ ), F(ξ) = C − kξ 2 +

where C is a free constant (to be determined by the mass M > 0) and k > 0 is a function of
m, N (k = (m − 1)/2m(N (m − 1) + 2)). They replace the Gaussian fundamental solutions in the
statement of the asymptotic theorems. Convergence of finite mass solutions to the Barenblatt
solution with the same mass is proved by the scaling method in [65]. A much earlier proof used
a method of optimal upper bounds, [33], 1980.
The faster convergence of solutions of the porous medium equation with changing sign is studied
in [49]. Here dipole solutions appear when the total mass is zero. Dipole solutions for nonlinear
parabolic problems were also studied in [7, 34, 43]. Convergence to the Barenblatt kernel for the
PME posed in RN minus one or several holes was studied in [19, 37].
These methods did not produce rates of convergence, but such rates were established using the
entropy method by Carrillo-Toscani in [22], 2000, via Bakry-Emery inequalities [4, 5], and then in
Del Pino-Dolbeault [27], 2002, using Gagliardo-Nirenberg inequalities. Convergence in the sense
of Wasserstein distances was introduced by F. Otto in 2001, [54]. Entropy methods suitable for
weighted porous media equations are used in [28].

• The Fast diffusion equation (i. e., the PME for m < 1) behaves much like the porous medium
equation for m close to 1, even if the shape of the selfsimilar profile is different,

−1/(1−m)
(10.2) F(ξ) = C + k(m, N )ξ 2

with fat tails as |x| → ∞. This is not the case for m < (N − 2)/N due to the phenomenon of
extinction in finite time. A detailed analysis of convergence to so-called pseudo-Barenblatt profiles
is done in [14, 17] by an entropy method which relies on some Hardy-Poincaré inequalities. This
is anyway a quite different scenario.

• The asymptotic convergence for the p-Laplacian equation ∂t u = ∆p u has been treated in [48] by
the scaling method, after settling the uniqueness of the fundamental solution of Barenblatt type.

30
This was done for p > 2, and it extends to some p < 2, to be precise to 2N/(N + 1) < p < 1.
The entropy method is used in [26, 29] using the analogy with the porous medium equation. The
doubly nonlinear equation ∂t u = ∆p (um ) was studied in [1] and subsequent works.

• There many variants of heat equation with lower-order terms, either first order or zero order. If
these terms are strong enough they will destroy the convergence towards some attractive solution
with a Gaussian shape. This is a huge research field and we will give only some ideas. Maybe
the best known models correspond to the case that can be written as ∂t u = ∆u + f (u) with
f (u) = −up , u ≥ 0; we then have so-called diffusion-absorption equations. Let us take and
p > 1, since the case p = 1 was explained before. Even if conservation of mass does not hold,
convergence of finite-mass solutions to a Gaussian profile with some positive mass M∞ is proved
when p > pc = (2 + N )/N . The limit case pc is very interesting and was studied in [38]. For
p < pc we enter into completely new asymptotic profiles. See also [45, 64].
There are many extensions of these ideas. The reaction cases, f (u) = +up , lead to the existence
of blow-up in finite time, a huge topic that falls complete out of the scope of these notes, see
[35, 61, 59]. Another reaction case that has attracted the attention of researchers is f (u) =
u(1 − u), called the Fisher-KPP model, of interest in biology and chemistry; here the long term
behaviour takes the form of expanding travelling waves and no trace of a Gaussian is seen, [51].

• Let us now examine some of the recent work on diffusion with fractional Laplace operators. The
linear fractional heat equation ∂t u + (−∆)s u = 0 has a rather complete theory in the paper [18].
It is well known that the self-similar fundamental solution exists for every N ≥ 1, 0 < s < 1 and
has the form P (x, t; s) = t−N/2s Fs (|x|t−1/2s ) where Fs is a smooth and positive profile function
with a fat tail as |x| → ∞

(10.3) Fs (ξ) ∼ c(N, s)|ξ|−(N +2s) ,

see [16]. Convergence to the self-similar fundamental solution can be proved for finite-mass
solutions by the scaling method (no rates), or by the representation analysis (with rates). See
separate notes by the author. For the entropy method see [12, 36].

• We continue with nonlinear fractional heat equations of porous medium type. The model studied
by Caffarelli and Vázquez [20] admits self-similar solutions that we may call fractional Barenblatt
solutions [13, 21, 11]. The entropy method is used in [21] to establish asymptotic convergence
without rates. Rates in 1D were obtained in [24]. Convergence with rates in several dimensions
is not known.
The alternative model of fractional porous equation, ∂t u+(−∆)s (|u|m−1 u) = 0 was studied in [55,
56]. Unique fundamental solutions of Barenblatt type were described in [67], where convergence
(without rates) was proved by the scaling method.

• There are a number of other equations that have been studied, like thin film equations [23],
the Barenblatt equation of elastoplastic filtration KPV91, inhomogeneous heat or porous medium
equations with weights, [47], or chemotaxis models, like [15]. A very important topic is the study
of the heat equation and the nonlinear diffusion models on manifolds, like the hyperbolic space,
see [41, 68].

• See [69] for a general presentation of linear and nonlinear diffusion equations including a detailed
survey of recent research work.

31
11 Historical comments
We add some historical notes on the origins and development of the Gaussian function, borrowed
from Wikipedia and other widely available sources, with no claim to be a rigorous historical
presentation. It seems that the so-called Gaussian function has its origin in Statistics. The 18th
century statistician Abraham de Moivre, a Frenchman exiled in England, seems to have been
the first person who noticed the existence of a bell-shaped curve as the limit of the probability
distributions of repeated random trials done independently. He was led by the practical problems
of calculating odds in gambling, not a every elevated motivation indeed. But he was a very fine
mathematician, appreciated by Newton. The curve he discovered is now called the “normal curve”.
One of the first applications of the normal distribution was to the analysis of errors of measure-
ments made in astronomical observations. A century later than De Moivre, the mathematicians
Adrain in 1808 and Gauss in 1809 developed independently the formula for the normal distribu-
tion and showed that errors were fit well by this distribution. The brilliant Gauss received much
credit, Adrain’s work was not known for many years.
This same distribution had been discovered by Laplace in 1778 when he derived the extremely
important central limit theorem, a main topic of this paper. Laplace showed that even if a
distribution is not normally distributed, the means of repeated samples from the distribution
would be very nearly normally distributed, and that the larger the sample size, the closer the
distribution of means would be to a normal distribution.
The distribution appeared later in another disguise in Statistical Mechanics as the Maxwell-
Boltzmann distribution, shortly called the Maxwellian. The original derivation in 1860 by James
Clerk Maxwell was an argument based on molecular collisions of the kinetic theory of gases as well
as certain symmetries in the speed distribution function; Maxwell also gave an early argument
that these molecular collisions entail a tendency towards equilibrium. After Maxwell, Ludwig
Boltzmann in 1872 also derived the distribution on mechanical grounds and argued that gases
should over time tend toward this distribution, due to collisions (see H-theorem).
The normal distribution has a wide implication to social issues. Thus, Quételet seems to have
been the first to apply the normal distribution to human characteristics. He noted that charac-
teristics such as height, weight, and strength were normally distributed.
Evidence for the Gaussian function as the fundamental solution of the heat equation came
after the work of probabilists in the 20th century to establish the link between heat equation and
Brownian diffusion, which is turn is the limit of discrete processes based on iterated random trials.
The close connection between stochastic differential equations and parabolic partial differential
equations is very much influenced by the role of the Gaussian function in both theories.

Acknowledgment. Work partially supported by Spanish Project MTM2014-52240-P. These

notes developed from Ph. D. courses and lectures given by the author at different events in recent
years, the last one was the Annual Meeting of the Red de Análisis Funcional y Aplicaciones, held
in Cáceres, Spain, in March 2017.

32
References
[1] M. Agueh. Asymptotic behavior for doubly degenerate parabolic equations. C. R. Math.
Acad. Sci. Paris 337 (2003), no. 5, 33–336.

[2] A. Andersson, P. Sjögren. Ornstein-Uhlenbeck theory in finite dimension, Preprint 2012:12,

Matematiska vetenskaper, Göteborg 2012, 40 pp.

[3] A. Arnold, P. Markowich, G. Toscani, A. Unterreiter. On convex Sobolev inequalities and

the rate of convergence to equilibrium for Fokker-Planck type equations, Comm. Partial
Differential Equations 26, 1–2 (2001), 43–100.

[4] D. Bakry, M. Émery. Hypercontractivité de semi-groupes de diffusion, C. R. Acad. Sci. Paris

Sér. I Math. 299 (1984), no. 15, 775–778. MR 772092.

[5] D. Bakry, I. Gentil, M. Ledoux. “Analysis and geometry of Markov diffusion operators”,
Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical
Sciences], 348, Springer, Cham, 2014.

[6] G. I. Barenblatt. On self-similar motions of compressible fluids in porous media, Prikl. Math.,
16 (1952), pp. 679–698 (in Russian).

[7] G. I. Barenblatt. On dipole-solutions in problems of nonstationary filtration of gas under

polytropic regime, Prikl. Mat. Mekh. 21 (1957), 718–720.

[8] A. R. Barron. Entropy and the central limit theorem, Ann. Probab., 14, (1986), 336–342.

[9] W. Beckner. A generalized Poincaré inequality for Gaussian measures, Proc. Amer. Math.
Soc. 105 (1989), 397–400.

[10] J. Bell. Gaussian measures, Hermite polynomials, and the Ornstein-Uhlenbeck semigroup,
2015; online.

[11] P. Biler, C. Imbert, G. Karch. Nonlocal porous medium equation: Barenblatt profiles and
other weak solutions. Arch. Ration. Mech. Anal. 215 (2015), no. 2, 497–529.

[12] P. Biler, G. Karch. Generalized Fokker-Planck equations and convergence to their equilibria.
In Evolution equations (Warsaw, 2001), vol. 60 of Banach Center Publ. Polish Acad. Sci.,
Warsaw, 2003, pp. 307–318.

[13] P. Biler, G. Karch, R. Monneau. Nonlinear diffusion of dislocation density and self-similar
solutions. Comm. Math. Phys. 294 (2010), no. 1, 145–168.

[14] A. Blanchet, M. Bonforte, J. Dolbeault, G. Grillo, J. L. Vázquez. Asymptotics of the fast

diffusion equation via entropy estimates, Arch. Rat. Mech. Anal. 191 (2009), 347–385.

[15] A. Blanchet, J. Dolbeault, M. Escobedo, J. Fernández. Asymptotic behaviour for small mass
in the two-dimensional parabolic-elliptic Keller-Segel model, J. Math. Anal. Appl. 361 (2010),
no. 2, 533–542.

[16] R. M. Blumenthal, R. K. Getoor. Some Theorems on Stable Processes, Trans. Amer. Math.
Soc. 95 (1960), 263–273.

33
[17] M. Bonforte, J. Dolbeault, G. Grillo, J. L. Vázquez. Sharp rates of decay of solutions to the
nonlinear fast diffusion equation via functional inequalities, Proc. Natl. Acad. Sci. USA 107
(2010), no. 38, 16459–16464.

[18] M. Bonforte, Y. Sire, J. L. Vázquez. Optimal Existence and Uniqueness Theory for the
Fractional Heat Equation, arXiv:1606.00873v1, to appear in Nonlin. Anal.

[19] C. Brändle, F. Quirós, J. L. Vázquez. Asymptotic behaviour of the porous media equation
in domains with holes, Interfaces Free Bound. 9 (2007), no. 2, 211–232.

[20] L. A. Caffarelli, J. L. Vázquez. Nonlinear porous medium flow with fractional potential
pressure. Arch. Rational Mech. Anal. 202 (2011), 537–565.

[21] L. A. Caffarelli, J. L. Vázquez. Asymptotic behaviour of a porous medium equation with

fractional diffusion, Discrete Cont. Dyn. Systems-A 29, no. 4 (2011), 1393–1404.

[22] J. A. Carrillo, G. Toscani. Asymptotic L1 -decay of solutions of the porous medium equation
to self-similarity. Indiana Univ. Math. J. 49 (2000), 113–141.

[23] J. A. Carrillo, G. Toscani. Long-time asymptotics for strong solutions of the thin film equa-
tion, Comm. Math. Phys. 225 (2002), no. 3, 551–571.

[24] J. A. Carrillo, Y. Huang, M. C. Santos, J. L. Vázquez. Exponential convergence towards

stationary states for the 1D porous medium equation with fractional pressure J. Differential
Equations 258 (2015), no. 3, 736–763.

[25] D. Chafaï. Entropies, convexity, and functional inequalities: on Φ-entropies and Φ-Sobolev
inequalities, J. Math. Kyoto Univ. 44 (2004), no. 2, 325–363.

[26] M. Del Pino, J. Dolbeault. Nonlinear diffusions and optimal constants in Sobolev type in-
equalities: asymptotic behaviour of equations involving the p-Laplacian. C. R. Math. Acad.
Sci. Paris 334 (2002), no. 5, 365–370.

[27] M. Del Pino, J. Dolbeault. Best constants for Gagliardo-Nirenberg inequalities and applica-
tion to nonlinear diffusions, J. Math. Pures Appl. (9) 81 (2002), 847–875.

[28] J. Dolbeault, I. Gentil, A. Guillin, F.Y. Wang. Lq -functional inequalities and weighted porous
media equations, Potential Anal. 28 (2008), no. 1, 35–59.

[29] M. Del Pino, J. Dolbeault. Asymptotic behavior of nonlinear diffusions. Math. Res. Lett. 10
(2003), no. 4, 551–557.

[30] J. Duoandikoetxea, E. Zuazua. Moments, masses de Dirac et décomposition de fonctions, C.

R. Acad. Sci. Paris Sér. I Math. 315 (6) (1992), 693–698.

[31] L. C. Evans. “Partial differential equations”. Graduate Studies in Mathematics, 19. American
Mathematical Society, Providence, RI, 1998.

[32] J. Fourier. “Théorie analytique de la Chaleur”; reprint of the 1822 original: Éditions Jacques
Gabay, Paris, 1988. English version: “The Analytical Theory of Heat”, Dover, New York,
1955.

34
[33] A. Friedman, S. Kamin. The asymptotic behavior of gas in an N-dimensional porous medium,
Trans. Amer. Math. Soc. 262 (1980), 551–563.

[34] V. A. Galaktionov, S. Posashkov, J. L. Vázquez. Asymptotic convergence to dipole solutions

in nonlinear parabolic equations, Proc. Royal Soc. Edinburgh 125A (1995), 877–900.

[35] V. A. Galaktionov, J. L. Vázquez. “A stability technique for evolution partial differential

equations. A dynamical systems approach”. Progress in Nonlinear Differential Equations and
their Applications, 56. Birkhäuser Boston, Inc., Boston, MA, 2004.

[36] I. Gentil, C. Imbert. The Lévy-Fokker-Planck equation: Φ-entropies and convergence to

equilibrium, Asymptot. Anal. 59, 3–4 (2008), 125–138.

[37] B. H. Gilding, J. Goncerzewicz. Large-time behaviour of solutions of the exterior-domain

Cauchy-Dirichlet problem for the porous media equation with homogeneous boundary data,
Monatsh. Math. 150 (2007), no. 1, 11–39.

[38] A. Gmira, L. Véron. Large time behaviour of the solutions of a semilinear parabolic equation
in RN . J. Differential Equations 53 (1984), no. 2, 258–276.

[39] A. A. Grigor’yan. On the fundamental solution of the heat equation on an arbitrary Rieman-
nian manifold. (Russian) Mat. Zametki 41 (1987), no. 5, 687–692, 765. English translation:
Math. Notes 41 (1987), no. 5–6, 386–389.

[40] A. Grigor’yan. “Heat Kernel and Analysis on Manifolds”, American Mathematical Society,
Providence, 2009.

[41] A. A. Grigor’yan, M. Noguchi. The heat kernel on hyperbolic space, Bull. Lond. Math. Soc.
30 (6) (1998) 643–650.

[42] L. Gross. Logarithmic Sobolev inequalities, Amer. J. Math. 97 (1975), 1061-1083.

[43] J. Hulshof, J. L. Vázquez. The Dipole Solution for the Porous Medium Equation, Annali
Scuola Normale Sup. Pisa, Serie IV, 20 (1993), 193–217.

[44] R. Jordan, D. Kinderlehrer, F. Otto. The variational formulation of the Fokker-Planck equa-
tion, SIAM J. Math. Anal., 29 (1998), 1–17.

[45] S. Kamin, L. A. Peletier, J. L. Vázquez. Classification of singular solutions of a nonlinear

heat equation, Duke Math. J. 58 (1989), no. 3, 601–615.

[46] S. Kamin, L. A. Peletier, J. L. Vázquez. On the Barenblatt equation of elastoplastic filtration,

Indiana Univ. Math. J. 40 (1991), no. 4, 1333–1362.

[47] S. Kamin, G. Reyes, J. L. Vázquez. Long time behavior for the inhomogeneous PME in a
medium with rapidly decaying density, Discrete Contin. Dyn. Syst. 26 (2010), 521–549.

[48] S. Kamin, J. L. Vázquez. Fundamental solutions and asymptotic behaviour for the p-
Laplacian equation, Rev. Mat. Iberoamericana 4:2 (1988), 339–354.

[49] S. Kamin, J. L. Vázquez. Asymptotic behaviour of solutions of the porous medium equation
with changing sign, SIAM J. Math. Anal. 22 (1991), no. 1, 34–45.

35
[50] O. Kavian. Remarks on the large time behavior of a nonlinear diflusion equation, Ann. Inst.
H. Poincare Anal. Non Linéaire 4. No. 5 (1987), 423–452.

[51] A. N. Kolmogorov, I. Petrovsky, N. Piscounoff. Étude de l’équation de la diffusion avec

croissance de la quantité de matière et son application a un problème biologique, Bull. Univ.
Moskou, Ser. Internat., Sec. A, 1 (1937), 1–25.

[52] E. H. Lieb, J. Yngvason. A Guide to Entropy and the Second Law of Thermodynamics,
Notices Amer. Math. Soc. 45 (1998), no. 5, 571–581.

[53] P. A. Markowich, C. Villani. On the trend to equilibrium for the Fokker-Planck equation: an
interplay between physics and functional analysis. Mat. Contemp. 19 (2000), 1–29.

[54] F. Otto. The geometry of dissipative evolution equations: the porous medium equations,
Comm. Partial Diff. Equations, 26 (1–2) (2001), pp. 101–174.

[55] A. De Pablo, F. Quirós, A. Rodríguez, J. L. Vázquez. A fractional porous medium equation,

Advances in Mathematics 226 (2011), no. 2, 1378–1409.

[56] A. De Pablo, F. Quirós, A. Rodríguez, J. L. Vázquez. A general fractional porous medium

equation. Comm. Pure Appl. Math. 65 (2012), no. 9, 1242–1284.

[57] H. Risken. “The Fokker-Planck equation, methods of solutions and applications”, 2nd ed.,
Springer-Verlag, berlin, Heidelberg, (1989).

[58] S. Salsa. “Partial differential equations in action. From modelling to theory.” Third edition,
Unitext, 99. Springer, 2016.

[59] A. A. Samarskii, V. A. Galaktionov, S. P. Kurdyumov, A. P. Mikhailov. “Blow-up in quasi-

linear parabolic equations”. Translated from the 1987 Russian original by Michael Grinfeld
and revised by the authors. De Gruyter Expositions in Mathematics, 19. Walter de Gruyter
& Co., Berlin, 1995

[60] J. A. Smoller. "Shock Waves and Reaction-Diffusion Equations," Springer-Verlag, New

York/Heidelberg/Berlin, 1982.

[61] P. Quittner, P. Souplet. “Superlinear parabolic problems. Blow-up, global existence and
steady states”. Birkhäuser Advanced Texts, Birkhäuser Verlag, Basel, 2007.

[62] G. E. Uhlenbeck, L. S. Ornstein. On the theory of Brownian motion, Physical Review 36,
(1930), 823–84.

[63] G. Toscani. Entropy production and the rate of convergence to equilibrium for the Fokker-
Planck equation, Quart. Appl. Math. 57 (1999), 521–541.

[64] J. L. Vázquez. Asymptotic behaviour of nonlinear parabolic equations. Anomalous exponents.

In “Degenerate Diffusions", IMA Volumes in Mathematics 47, Springer Verlag (1993), pp.
215–228.

[65] J. L. Vázquez. Asymptotic behaviour for the Porous Medium Equation posed in the whole
space. Journal of Evolution Equations 3 (2003), 67–118.

36
[66] J. L. Vázquez. “The Porous Medium Equation. Mathematical Theory”, vol. Oxford Math-
ematical Monographs, Oxford University Press, Oxford, 2007.

[67] J. L. Vázquez. Barenblatt solutions and asymptotic behaviour for a nonlinear fractional heat
equation of porous medium type, J. Eur. Math. Soc. (JEMS) 16 (2014), no. 4, 769–803.

[68] J. L. Vázquez. Fundamental solution and long time behaviour of the Porous Medium Equation
in hyperbolic space, J. Math. Pures Appl. (9) 104 (2015), no. 3, 454–484

[69] J. L. Vázquez. The mathematical theories of diffusion. Nonlinear and fractional diffusion, in
“Nonlocal and Nonlinear Diffusions and Interactions: New Methods and Directions”. Springer
Lecture Notes in Mathematics, 2186, C.I.M.E. Foundation Subseries, yr 2017, pg. 205-278.

[70] D. V. Widder. “The Heat Equation”, Academic Press, New York, 1975.

[71] C. Villani. “Topics in Optimal Transportation”, Graduate Studies in Math, 58, AMS, 2003.

Address:
Juan Luis Vázquez. Departamento de Matemáticas, Universidad Autónoma de Madrid,
Campus de Cantoblanco, 28049 Madrid, Spain. e-mail address: [email protected]

37