CHARACTERIZATION OF SOLUTIONS TO DISSIPATIVE SYSTEMS WITH

SHARP ALGEBRAIC DECAY

LORENZO BRANDOLESE
arXiv:1509.05928v2 [math.AP] 23 Mar 2016

Abstract. We characterize the set of functions u0 ∈ L2 (Rn ) such that the solution of the
problem ut = Lu in Rn × (0, ∞) starting from u0 satisfy upper and lower bounds of the form
c(1 + t)−γ ≤ ku(t)k2 ≤ c′ (1 + t)−γ . Here L is in a large class of linear pseudo-differential
operator with homogeneous symbol (including the Laplacian, the fractional Laplacian, etc.).
Applications to nonlinear PDEs will be discussed: in particular our characterization provides
necessary and sufficient conditions on u0 for a solution of the Navier–Stokes system to satisfy
sharp upper-lower decay estimates as above.
In doing so, we will revisit and improve the theory of decay characters by C. Bjorland,
C. Niche, and M.E. Schonbek, by getting advantage of the insight provided by the Littlewood–
Paley analysis and the use of Besov spaces.

1. Introduction
The theory of decay characters was first introduced by C. Bjorland and M.E. Schonbek in [2]
and further developped by C. Niche and M.E. Schonbek in [6], with the motivation of obtaining
sharp upper and lower bound estimates for the L2 -norm of solutions to a large class of lin-
ear or semilinear parabolic systems: so far, this theory has been successfully applied, e.g., to
the heat equation, the Navier–Stokes equations [2], the quasi-geostrophic equations [6], several
compressible approximations of Navier–Stokes [6,7] and to the Navier–Stokes–Voigt equation [5].
The first issue of the present paper is that a slight modification of the original definition of
decay character makes the theory more powerful and more widely applicable: our modification
allows in particular to get necessary and sufficient conditions on u0 for the validity of L2 -
estimates of the form
(1.1) c(1 + t)−γ ≤ ketL u0 kL2 (Rn ) ≤ c′ (1 + t)−γ ,
(where L is a suitable linear pseudo-differential operator, such as the Laplacian, a fractional
Laplacian, etc.), whereas the original approach [2, 6] only gave sufficient conditions on u0 for
the validity of (1.1). We also characterize the class of initial data u0 such that estimates (1.1)
hold in terms of suitable subsets of Besov spaces.
Applications of our analysis to nonlinear problems include the complete characterization of
divergence-free vector fields v0 ∈ L2 (R3 )3 such that the corresponding weak solutions of the
Navier–Stokes equations satisfy two-side bounds for the energy of the form
(1.2) c(1 + t)−γ ≤ kv(t)kL2 (Rn ) ≤ c′ (1 + t)−γ , 0 < γ < 5/4.

Date: May 23, 2022.

2000 Mathematics Subject Classification. 42B25, 35K05, 35Q30.
Key words and phrases. Heat equation, Navier–Stokes, Decay, Energy, Besov, Diffusion.
Supported by the ANR project DYFICOLTI ANR-13-BS01-0003-01.
1
2 LORENZO BRANDOLESE

One of our main results, Theorem 5.1, will provide three different equivalent conditions on v0 for
the vality of (1.2). Our characterization does not go through when γ = 5/4, because the lower
bound of the nonlinear problem in this borderline case is no longer driven by the corresponding
lower bound for heat kernel.
The other important results of this paper will be summarized by Eq. (7.2) in the last section.

Notation. We will often use the symbols . or & in chain of inequalities to avoid the proliferation
of different constants. For example, writing f (t) . g(t) we mean that f (t) ≤ Cg(t) for some
constant C > 0 independent on t. When we have both f (t) . g(t) and g(t) . f (t) we will often
write f (t) ≃ g(t).

2. Revisiting the theory of decay characters

2.1. Improvement of the basic definitions. The authors in [2] introduced the notion of De-
cay indicator. As the original definition looks somewhat too restrictive, we redefine it in the
following way:

Definition 2.1. Let u0 ∈ L2 (Rn ), Bρ = {ξ ∈ Rn : |ξ| ≤ ρ}. The lower and upper decay
indicators of u0 are the two lower and upper limits
Z Z
−2r−n 2 −2r−n
Pr (u0 )− = lim inf ρ |b
u0 (ξ)| dξ and Pr (u0 )+ = lim sup ρ u0 (ξ)|2 dξ.
|b
ρ→0+ Bρ ρ→0+ Bρ

When Pr (u0 )− = Pr (u0 )+ , then we can define the decay indicator of u0 as Pr (u0 ) = Pr (u0 )− =
Pr (u0 )+ .


This definition is interesting only for r ∈ − n2 , ∞ , as the decay indicator is always zero when
r ≤ −n/2.

Remark 2.2. Originally, the decay indicator was defined in [2] as the limit
Z
−2r−n
(2.1) Pr (u0 ) = lim ρ u0 (ξ)|2 dξ,
|b
ρ→0+ Bρ

implicitly assuming that the limit does exist. But this is not always the case for u0 ∈ L2 (Rn ).
When the above limit does not exist, the lower and upper decay indicators are convenient
substitutes for obtaining some relevant estimates, as we will see below. Of course, Pr (u0 )− and
Pr (u0 )+ are always well defined in [0, +∞]. An example of v0 ∈ L2 (Rn ) such that Pr (v0 ) is not
well defined by the limit (2.1) can be constructed putting fast oscillations for |b v0 (ξ)| near the
origin. See the following Example 6.1.
On the other hand, if u0 ∈ L2 (Rn ) is such that u b0 (ξ) ∼ |ξ|r as |ξ| → 0, with r ∈ (−n/2, ∞),
then Pr (u0 )− = Pr (u0 )+ = r and so Pr (u0 ) is well defined in this case and Pr (u0 ) = r as well.

Definition 2.3. The upper and lower decay characters of u0 ∈ L2 (Rn ) are respectively defined
by

(2.2) r(u0 )+ = sup{r ∈ R : Pr (u0 )+ < ∞},

(2.3) r(u0 )− = inf{r ∈ R : Pr (u0 )− > 0}.
 CHARACTERIZATION OF SHARP ALGEBRAIC DECAY 3

Remark 2.4. For any u0 ∈ L2 (Rn ), the upper and lower decay characters are always well defined
(adopting the usual convention that inf ∅ = +∞) and satisfy the inequality
(2.4) − n/2 ≤ r(u0 )+ ≤ r(u0 )− ≤ ∞.
Indeed, for all r ≤ −n/2, we have Pr (u0 )+ = 0, so the first inequality is immediate. Let us
now prove the inequality in the middle of (2.4). If, by contradiction, r(u0 )− < r(u0 )+ , then we
could choose two real numbers r < r ′ such that Pr (u0 )− > 0 and Pr′ (u0 )+ < ∞. But the last
′ ′ R
u0 (ξ)|2 dξ = 0. On the other hand,
inequality implies Pr (u0 )+ = lim supρ→0 ρ2(r −r) ρ−2r −n Bρ |b
0 ≤ Pr (u0 )− ≤ Pr (u0 )+ = 0, that contredics the inequality Pr (u0 )− > 0.

The upper decay character r(u0 )+ will play a role in obtaining decay estimates from above for
et∆ u
0 , and the lower decay character r(u0 )+ will be useful for obtaining estimates from below.
This is the reason of our terminology.
Depending on u0 ∈ L2 (Rn ), the sup and the inf in the definition of r(u0 )+ and r(u0 )− can
be achieved or not. This observation motivates next definition. Our definition below is an
improvement of that of [2].

Definition 2.5. If u0 ∈ L2 (Rn ) is such that there exists r ∗ ∈ (−n/2, ∞) such that
(2.5) r ∗ (u0 ) = max{r ∈ R : Pr (u0 )+ < ∞} = min{r ∈ R : Pr (u0 )− > 0}.
then we call this number r ∗ = r ∗ (u0 ) the decay character of u0 . We define also the decay
character of u0 in the two limit situations as follows:
r ∗ (u0 ) = +∞, if r(u0 )+ = r(u0 )− = +∞,
r ∗ (u0 ) = −n/2, if r(u0 )+ = r(u0 )− = −n/2.

Remark 2.6. For u0 ∈ L2 (Rn ) the decay character r ∗ (u0 ) does not always exist. One reason for
this is that it can happen that r(u0 )+ < r(u0 )− . See Example 6.3. Another reason is that one
could have r(u0 )+ = r(u0 )− , but the supremum or the infimum appearing in the definition of
r(u0 )+ and r(u0 )− are nor achieved. See Example 6.2.
In fact, r ∗ (u0 ) does exist and belongs to (−n/2, ∞) if and only if there exists r ∗ ∈ (−n/2, ∞)
such that 0 < Pr∗ (u0 )− ≤ Pr∗ (u0 )+ < ∞. The “only if” part of this claim is clear. The “if”
part holds because, if r ∗ < r ′ then Pr′ (u0 )+ = ∞ (otherwise we would get the contradiction
Pr∗ (u0 )+ = 0, as already observed right after (2.4)) and so r ∗ = max{r ∈ R : Pr (u0 )+ < ∞}. In
the same way one sees that r ∗ = min{r ∈ R : Pr (u0 )− > 0}.

Remark 2.7. Let us illustrate the difference between our improved definition of decay character
and the original definition in [2]. The main motivation of such an improvement is the validity
of the second assertion of Theorem 2.9 below. Our discussion concerns here only the case
r ∗ (u0 ) ∈ (−n/2, ∞), but it could be adapted also to the limit situations r ∗ (u0 ) = +∞ or −n/2.
Originally, the decay character r ∗ (u0 ) was defined in [2] implicitly assuming that, for all r ∈ R,
Pr (u0 )− = Pr (u0 )+ = Pr (u0 ) and also assuming that, for some r ∗ ∈ (−n/2, ∞), 0 < Pr∗ (u0 ) <
∞. Under such two conditions, we see that r ∗ = max{r ∈ R : 0 < Pr (u0 ) < ∞} = min{r ∈
R : 0 < Pr (u0 ) < ∞}. For this reason, if u0 admits a decay character in the sense of [2] then u0
admits a decay character in the sense of our definition and these are the same. The converse is
not true, so the original definition in [2] is indeed more restrictive than ours: for example, the
4 LORENZO BRANDOLESE

function v0 ∈ L2 (Rn ), constructed in Example 6.1 admits a decay character only in the sense of
our definition.

2.2. Applications to upper and lower decay estimates. We start with an application of
the notions of upper and lower decay indicators. As in [6], we consider the class of (matricial)
pseudo-differential operators L with symbol
M(ξ) = P (ξ)−1 D(ξ)P (ξ), for a.e. ξ ∈ Rn ,
where D(ξ) and P (ξ) are respectively diagonal and orthogonal matrices of order m, with D(ξ)ij =
−ci |ξ|2α δi,j , and ci ≥ c > 0, for all i = 1, . . . , m and α > 0. We also assume that P (ξ)ij are
homogeneous functions smooth outside ξ = 0.
We are interested in establishing L2 -estimates from above and below for solutions of the linear
problem
(
ut = Lu, t > 0, x ∈ Rn ,
(2.6)
u|t=0 = u0 ,
where u = (u1 , . . . , um ) and u0 = (u0,1 , . . . , u0,m ). Basic examples include the heat equation
vt = ∆v (in this case P (ξ) = Im and D(ξ) = −|ξ|2 Im ) or the evolution problem for the fractional
Laplacian (P (ξ) = Im and D(ξ) = −|ξ|2α Im ). Examples of physical interest with P 6= Im arise,
e.g., in fluid mechanics, see [6].
We will typically assume u0 ∈ (L2 (Rn ))m . However, from now on we will not distinguish in
our notations between scalar and vector-valued function spaces and write abusively u0 ∈ L2 (Rn )
also in the vector-valued case.

Proposition 2.8. Let u0 ∈ L2 (Rn ), let L be a pseudo-differential operator as above. Let u be

the solution of the linear problem (2.6).
- If Pr (u0 )− > 0 then there is a constant C1 > 0 such that, for all t > 0,
1
C1 (1 + t)− α (r+n/2) ≤ ku(t)k22 . .
- If Pr (u0 )+ < ∞, then there is a constant C2 > 0 such that, for all t > 0,
1
ku(t)k22 ≤ C2 (1 + t)− α (r+n/2) .

Proof. The proof below is a minor modification to that of [2, Theorem 5.7] (for the particular
case L = ∆) or that of [6, Theorem 2.10] (for the general case).
We have
2α
|eM(ξ)t u
b0 (ξ)| = |P (ξ)−1 eD(ξ)t P (ξ)b
u0 (ξ)| & e−c|ξ| t |b
u0 (ξ)|.
Hence, for any function ρ = ρ(t),
Z
2
ku(t)k2 ≥ |eM(ξ)t u
b0 (ξ)|2 dξ
|ξ|≤ρ(t)
Z Z
−2c|ξ|2α t 2 −2ctρ(t)2α
& e |b
u0 (ξ)| dξ ≥ e u0 (ξ)|2 dξ.
|b
|ξ|≤ρ(t) |ξ|≤ρ(t)
R
Let Φr (ρ) = ρ−2r−n |ξ|≤ρ |b u0 (ξ)|2 dξ. Under the assumption of the first assertion and by the
definition of lim inf, we have 0 < Pr (u0 )− = limǫ→0+ inf ρ∈(0,ǫ] Φr (ρ). Hence, for some c0 , ρ0 > 0
 CHARACTERIZATION OF SHARP ALGEBRAIC DECAY 5

and all 0 < ρ ≤ ρ0 , we have Φr (ρ) > c0 . The choice ρ(t) = ρ0 (1 + t)−1/(2α) then leads to
2α
ku(t)k22 & e−2ctρ(t) ρ(t)2r+n Φr (ρ(t)) & ρ(t)2r+n
and the first claim follows.
Let us prove the second assertion, following again the steps of [6]. First of all, the assumption
on the symbol M(ξ) allows us to write
1 d
ku(t)k22 = hb
u, P −1 DP u
bi = −h(−D)−1 P u
b, (−D)−1 P u
bi
2 dt Z
=− |(−D)−1/2 P u
b|2 dξ.

This insures the validity of the energy inequality

Z
d
ku(t)k2 ≤ −C |ξ|2α |b
2
u(ξ, t)|2 dξ.
dt
The classical Fourier splitting idea [8] is then used to deduce the estimate
Z
d 2 2α 2 2α
kuk2 + ρ(t) ku(t)k2 . ρ(t) u(ξ, t)|2 dξ.
|b
dt |ξ|≤ρ(t)

Our assumption now reads Pr (u0 )+ = limǫ→0+ supρ∈(0,ǫ] Φr (ρ) < ∞. It implies that for some
R
ρ0 , C > 0 and all 0 < ρ ≤ ρ0 , we have Φr (ρ) = ρ−2r−n |ξ|≤ρ(t) |bu0 (ξ)|2 dξ ≤ C. Hence,
Z
ρ(t)2α u(ξ, t)|2 dξ . ρ(t)2α+2r+n .
|b
|ξ|≤ρ(t)

Combining the two last estimates, choosing now ρ(t)2α = M (1 + t)−1 with M > r + n/2, and
multiplying by the integrating factor (1 + t)M we arrive at
d

(1 + t)M ku(t)k22 . (1 + t)M −1−(2r+n)/(2α) .
dt
The upper bound follows by integration. 
The following theorem clearly illustrates the importance of the notion of decay character.
Theorem 2.9.

(1) Let u0 ∈ L2 (Rn ) be such that the decay character r ∗ = r ∗ (u0 ) ∈ (−n/2, +∞) does exist.
Let L as in Proposition 2.8, and u be the solution of the problem (2.6). Then for some
C1 , C2 > 0, and all positive t,
1 ∗ +n/2) 1 ∗ +n/2)
(2.7) C1 (1 + t)− α (r ≤ ku(t)k22 ≤ C2 (1 + t)− α (r .
(2) Conversely, if u0 ∈ L2 (Rn ) is such that the solution u of the problem (2.6) satisfies
estimates (2.7) with r ∗ ∈ (−n/2, ∞) then u0 possess a decay character and r ∗ (u0 ) = r ∗ .
Assertion (1) appears also in [2, 6], but with their more restrictive definition of r ∗ (u0 ). This
first assertion is an immediate consequence of the estimates obtained in Proposition 2.8. On the
other hand, the validity of assertion (2) is made possible by the fact that, compared to [2, 6], in
our definition we relaxed a little bit the requirements for the existence of the decay character.
The proof of the converse part of Theorem 2.9 will be postponed in Remark 4.4, after the
characterization of decay characters in terms of subsets of Besov spaces.
6 LORENZO BRANDOLESE

3. Besov space approach to two-side bounds for linear dissipative systems

Let us recall the definition of the homogeneous Besov spaces via the Littlewood-Paley analysis.
Let ϕ be a smooth radial function with support contained in the annulus {ξ ∈ Rn : 3/4 ≤ |ξ| ≤
P
8/3}, such that j∈Z ϕ(ξ/2−j ) = 1 for all ξ ∈ Rn \{0}. Let ∆j be the usual Littlewood–Paley
localization operator around the frequency |ξ| ≃ 2j , j ∈ Z, namely, ∆ d j b
j f = ϕ(·/2 )f , see [1]. In
this paper we will only need to consider the case s < n/p, 1 ≤ p, q ≤ +∞. In this case, the
elements of Ḃps,q can be realized as tempered distributions:

n X
o
s
Ḃp,q = f ∈ S ′ (Rn ) : f = ∆j f, the series being convergent in S ′ (Rn ), 2js k∆j f kp j ∈ ℓq (Z) ,
j∈Z

normed by


kf kḂ s = 2js k∆j f kp .
p,q ℓq (Z)

The classical characterization of Besov spaces with negative regularity in terms of the heat
kernel reads (see, e.g., [1, Theorem 3.4]):
−2σ
(3.1) f ∈ Ḃp,q ⇐⇒ tσ ket∆ f kp ∈ Lq (R+ , dt/t), σ > 0, 1 ≤ p, q ≤ ∞,

−2σ
with equivalence of the corresponding norms. In particular, for u0 ∈ L2 (Rn ), we have u0 ∈ Ḃ2,∞
if and only if ket∆ u0 k2 . (1 + t)−σ .

−σ
3.1. Two useful subsets of Besov spaces. We now introduce two subsets of Ḃ2,∞ :
 

  

 
k∆j f k2 ≤ C2σj ∀j ∈ Z
Ȧ−σ
2,∞
−σ
= f ∈ Ḃ2,∞ : ∃ c, C, M > 0 s.t.

 k∆j f k2 ≥ c2σj for at least one j ∈ Z in 

 
any interval of lenght M

and

Ȧ−σ
2,∞
−σ
= f ∈ Ḃ2,∞ : ∃ c, C > 0, ∃(jk )k∈N ⊂ Z s.t. jk → −∞,
 
∞ k∆j f k2 ≤ C2σj ∀j ∈ Z
(jk − jk+1 ) ∈ ℓ (N), and .
k∆jk f k2 ≥ c2σjk ∀k ∈ N

These sets are not closed under summation and do not have a linear structure. We point out
the inclusions
Ȧ−σ −σ −σ
2,∞ ⊂ Ȧ2,∞ ⊂ Ḃ2,∞ .

It is worth observing that the intersection L2 (Rn ) ∩ Ȧ−σ

2,∞ is empty, since dyadic blocks of
L functions satisfy k∆j f k2 ≤ kf k2 for all j ∈ Z. On the other hand, Ȧ−σ
2
2,∞ has a nontrivial
2 n
intersection with L (R ), which we will characterize below.
 CHARACTERIZATION OF SHARP ALGEBRAIC DECAY 7

3.2. Two-side bounds for the heat kernel. Our first applications of the sets Ȧ−2σ −2σ
2,∞ and Ȧ2,∞
is the following theorem.

Theorem 3.1. Let σ > 0 and f be a tempered distribution.

(1) Then, f ∈ Ȧ−2σ
2,∞ if and only if

(3.2) ∃ c1 , c2 > 0 such that c1 t−σ ≤ ket∆ f k2 ≤ c2 t−σ , for all t > 0.
(2) Moreover, f ∈ Ȧ−2σ
2,∞ if and only if

(3.3) tσ ket∆ f k2 ∈ L∞ (R+ ), and lim inf tσ ket∆ f k2 > 0.

t→+∞

(3) In addition, f ∈ L2 (Rn ) ∩ Ȧ−2σ

2,∞ if and only if

(3.4) ∃ c, c′ > 0 such that c (1 + t)−σ ≤ ket∆ f k2 ≤ c′ (1 + t)−σ , for all t > 0.

Proof. The property tσ ket∆ f k2 ∈ L∞ (R+ ) and the upper bound in (3.2), are equivalent to
−2σ
f ∈ Ḃ2,∞ , as noticed in (3.1). In the same way, the upper bound in (3.4) is equivalent to the
−2σ
condition f ∈ L2 (Rn ) ∩ Ḃ2,∞ . For this reason, in the proof of Theorem 3.1 we need to focus
only on the lower bound properties.
Let us start with a simple preliminary computation. For t > 0, let p = p(t) ∈ Z such that
4 ≤ t < 4p+1 . Letting dj = k∆j f k22 2−4jσ we have (the positive constant c is not the same from
p

line to line):
X 2j
ktσ et∆ f k22 ≃ t2σ e−ct2 k∆j f k22
j∈Z
X j

(3.5) ≃ t2σ e−ct4 24jσ dj

j∈Z
X j
≃ 42jσ e−c 4 dj−p ,
j∈Z

where in the last step was obtained by shifting the summation index.
Assume now f ∈ Ȧ−2σ −2σ
2,∞ . Then, for some constant M > 0 as in the definition of Ȧ2,∞ , and
some constant c > 0 independent on p we have (c is not the same from line to line),
X j
ktσ et∆ f k22 & 42jσ e−c 4 dj−p
|j|≤M

& max dj−p & 1,

|j|≤M

where in the last inequality we used max|j|≤M 2−4σ(j−p) k∆j−p f k22 ≥ c, that follows from the
definition of Ȧ−2σ
2,∞ . This proves the lower bound in (3.2).
If we assume instead the weaker condition f ∈ Ȧ−2σ 2,∞ , then take M = k(jk − jk+1 )kℓ∞ , where
−2σ
jk → −∞ is as in the definition of Ȧ2,∞ . Then, there exist an integer p0 and c > 0 such that for
all p ≥ p0 we have max|j|≤M dj−p = max|j|≤M 2−4σ(j−p) k∆j−p f k22 ≥ c > 0. As c is independent
on p, and hence on t, at least in some interval [t0 , +∞), we get now

ktσ et∆ f k22 & 1 on [t0 , ∞).

8 LORENZO BRANDOLESE

This in turn gives

lim inf tσ ket∆ f k2 > 0.
t→+∞

Conversely, if we assume (3.2), then the lower bound in (3.2) and the computation at the
beginning of the proof imply
X j
(3.6) 42jσ e−c 4 dj−p ≥ c′ > 0,
j∈Z

with 4p ≃ t and c′ independent on p. On the other hand, the upper bound (3.2) imply that
−2σ j
u0 ∈ Ḃ2,∞ , hence (dj ) ∈ ℓ∞ (Z). As (42jσ e−c 4 ) ∈ ℓ1 (Z), we can find M > 0 such that
X j
42jσ e−c 4 < c′ /(2kdj kℓ∞ ).
|j|≥M

Combining the two last estimates we get

X j
42jσ e−c4 dj−p ≥ c′ /2.
|j|≤M
P j
On the other hand, |j|≤M 42jσ e−c4 dj−p ≤ C max|j|≤M dj−p for some C > 0 independent on p.
It then follows that, for some c > 0 independent on p,

max dj−p ≥ c.
|j|≤M

This insures that f ∈ Ȧ−2σ

2,∞ .
If, instead of estimates (3.2) we have the weaker conditions (3.3), then (3.6) holds true only
for all p ≥ p0 , for some p0 ∈ Z large enough. Applying the above argument to p = p0 + kM ,
k ∈ N, we get, for some c > 0 independent on k,

max dj−(p0 +kM ) ≥ c.

|j|≤M

Therefore, we can then construct a sequence of integers (jk ) such that jk → −∞, |jk+1 − jk | ≤
3M , and djk ≥ c > 0 for all k ∈ N. This implies f ∈ Ȧ−2σ
2,∞ .
It only remains to prove the last claim of Theorem 3.1. If f ∈ L2 (Rn ) ∩ Ȧ−2σ t∆ 2
2,∞ then ke f k2 ≤
2
kf k2 and the upper bound estimate (3.4) immediately follows from the first of (3.3). The lower-
bound estimate in (3.4) follows combining the fact that ket∆ f k22 ≥ ct−σ , valid in [t0 , +∞) for
some large enough t0 > 0, by the second of (3.3), with the inequality ket∆ f k22 ≥ ket0 ∆ f k22 , valid
in the interval [0, t0 ].
Conversely, if estimates (3.4) hold then f ∈ Ȧ−2σ by the previous part of the Theorem.
2
R 2,∞
−2t|ξ| 2

Moreover, by Fatou’s lemma, kf k2 ≤ lim inf t→0+ e f (ξ) dξ ≤ (c′ )2 < ∞, so f ∈ L2 (Rn ).


3.3. A Besov space characterization through etL . In order establish the natural general-
ization of Theorem 3.1 to the solutions of the linear problem (2.6), we first need to extend the
classical heat kernel characterization of Besov space (3.1) to the setting of the operator etL .
Specifically, we establish the following theorem.
 CHARACTERIZATION OF SHARP ALGEBRAIC DECAY 9

Theorem 3.2. Let σ > 0, 1 ≤ p, q ≤ ∞. Then for any u ∈ S ′ (Rn ) such that the Littlewood–
P −2σ (Rn ) if and only if ktσ/α etL f k ∈
Paley series u = j∈Z ∆j u converges in S ′ (Rn ), we have f ∈ Ḃp,q p
q +
L (R , dt/t). Moreover, the two norms
kf kḂp,q
−2σ and ktσ/α etL f kp
Lq (R+ , dt/t)

are equivalent.
Proof. The proof is an adaptation of that in [1, Theorem 2.34] for the special case L = ∆. Recall
that
F(etL f )(ξ) = P (ξ)−1 etD(ξ) P (ξ)fb(ξ),
where D(ξ) = −|ξ|2α D, where D = diag(c1 , . . . , cn ), with α, cj > 0, and P (ξ) is an orthogonal
matrix. The entries P (ξ)j,k (j, k = 1, . . . , m) are assumed to be homogeneous functions smooth
for ξ 6= 0.
Lemma 3.3. Let p ∈ [1, ∞], and f be a tempered distribution with support in the annulus λC,
where C = {ξ ∈ Rn : 0 < r1 < |ξ| < r2 } and λ > 0. Then, for all t > 0.
2α
ketL f kp ≤ Ce−ctλ kf kp ,
for some c, C > 0 independent on λ and t.
Proof. The operator etL is a convolution operator of the form etL f = G(·, t) ∗ f . By the homo-
geneity assumptions on P and D, the symbol G(ξ, b t) = P (ξ)−1 etD(ξ) P (ξ) satisfies the scaling
b t) = G(ξt
relation G(ξ, b 1/(2α) , 1), and so G(λx, t) = λ−n G(x, tλ−2α ).
Replacing fb with its dilate fb(λ·) we can reduce to the case λ = 1. So assume now λ = 1
and consider a smooth cut-off function φ(ξ) identically equal to one on C with compact support
bounded away from ξ = 0. As f = φ ∗ f , we have
ketL f kp = kG(·, t) ∗ φk1 kf kp ,
and it only remains to show that kG(·, t) ∗ φk1 ≤ Ce−ct .
Indeed, computing the inverse Fourier transform and integrating by parts,
Z
−n
G(·, t) ∗ φ(x) = (2π) eiξ·x φ(ξ)P (ξ)−1 etD(ξ) P (ξ) dξ
Z
2 −n


= (1 + |x| ) (I − ∆ξ )n eiξ·x φ(ξ)P (ξ)−1 etD(ξ) P (ξ) dξ
Z
2 −n


= (1 + |x| ) eiξ·x (I − ∆ξ )n φ(ξ)P (ξ)−1 etD(ξ) P (ξ) dξ


But (I − ∆ξ )n φ(ξ)P (ξ)−1 etD(ξ) P (ξ) can be bounded by a linear combination of functions of
the form |ψβ (ξ)| |∂ξβ etD(ξ) |, with ψβ (ξ) smooth with support contained in that of φ(ξ), β ∈ Nn
and |β| ≤ 2n. Hence, for all ξ,
2α
(I − ∆ξ )n φ(ξ)P (ξ)−1 e−tD(ξ) P (ξ) ≤ C(1 + t)2d e−ct|ξ| χ(ξ),
for some smooth χ with compact support contained in that of φ. It follows,
kG(·, t) ∗ φk1 ≤ Ce−ct
and the assertion of the Lemma follows. 
10 LORENZO BRANDOLESE

−2σ . Then, by Lemma 3.3,

Now assume that u ∈ Ḃp,q
X X jα
(3.7) ktσ/α etL f kp ≤ tσ/α k∆j etL f kp . tσ/α 4jσ e−ct 4 ηj ,
j∈Z j∈Z

with (ηj ) ∈ ℓq and k(ηj )kℓq = kf kḂp,q

−2σ . But, for all σ, c, α > 0,

X jα
(3.8) sup tσ/α 4jσ e−ct 4 < ∞,
t>0
j∈Z

as one can easily check splitting the summation in two terms corresponding to j such that
4j ≤ t−1/α , next 4j > t−1/α .
In the case q = ∞, we readily deduce

ktσ/α etL f kp . k(ηj )kℓ∞ ≤ kf kḂp,∞

−2σ ,

that agrees with (3.14).

In the case 1 ≤ q < ∞, we need to estimate the integral below first applying (3.7), next
jα
Hölder inequality to the summation with weight 4jσ e−ct 4 and (3.8), next Fubini’s theorem
and the definition of the Gamma function:
Z ∞ Z t X q
σq/α tL q σ/α jσ −ct 4jα
t ke f kp dt . t 4 e ηj dt/t
0 0 j∈Z
Z t X 
jα
. tσ/α 4jσ e−ct 4 ηjq dt/t
0 j∈Z
X
. Γ(σ/α) ηjq . kukqḂ −2σ .
p,q
j∈Z

To prove the reverse estimate, let us first establish the following identity:
Z ∞
tσ/α (−D(ξ))1+σ/α etD(ξ) vb(ξ) dt
0
Z ∞
1+σ/α −c1 t|ξ|2α 2α
= t1+σ/α |ξ|2α(1+σ/α) diag(c1 e , . . . , c1+σ/α
n e−cn t|ξ| )bv (ξ) dt/t
0
Z ∞
1+σ/α −c1 t
= t1+σ/α diag(c1 e , . . . , c1+σ/α
n e−cn t )b
v (ξ) dt/t
0
= Γ(1 + σ/α)b
v (ξ)

d
Applying this to vb(ξ) = P (ξ)∆ j f (ξ) we obtain
Z ∞
d 1 d
∆j f (ξ) = tσ/α P (ξ)−1 (−D(ξ))1+σ/α etD(ξ) P (ξ)∆ j u(ξ) dt
Γ(1 + σ/α) 0
Z ∞
1
= tσ/α P (ξ)−1 (−D(ξ))1+σ/α P (ξ)F(etL ∆j u)(ξ) dt
Γ(1 + σ/α) 0
We now make use of the homogeneity properties of P (ξ) and D(ξ) and apply a classical Fourier
multiplier theorem (see, e.g. [1, Lemma 2.2]), next the identity etL = etL/2 etL/2 with Lemma 3.3
 CHARACTERIZATION OF SHARP ALGEBRAIC DECAY 11

to get
Z ∞
k∆j f kp . tσ/α 4αj(1+σ/α) ketL ∆j f kp dt
0
Z ∞
jα
(3.9) . tσ/α 4αj(1+σ/α) e−ct 4 ketL/2 ∆j f kp dt
Z0 ∞
jα
. tσ/α 4αj(1+σ/α) e−2ct 4 ketL f kp dt
0

When q = ∞ we just have to observe that

(3.10) k∆j f kp . 22jσ sup tσ/α ketL f kp
t>0

to conclude that
σ/α tL


kf kḂp,∞
−2σ . sup t ke f kp .
t>0
When 1 ≤ q < ∞, we argue as follow:
X X Z ∞ q
q −2jqσ q −jqσ σ/α αj(1+σ/α) −ct 4jα tL
kf kḂ −2σ = 2 k∆j f kp . 4 t 4 e ke f kp dt
p,q
j∈Z j∈Z 0

X Z ∞ q−1 Z ∞ 
αjq −ct 4jα qσ/α tL jα
. 4 e dt t ke f kqp e−ct 4 dt
j∈Z 0 0

X Z ∞ 
jα qσ/α tL jα
. 4 t ke f kqp e−ct 4 dt
j∈Z 0
Z ∞ X 
αj −ct 4jα
. t4 e tqσ/α ketL f kqp dt/t
0 j∈Z
Z ∞
. tqσ/α ketL f kqp dt/t
0

Here we used first (3.9), next Hölder inequality, and in the last inequality a particular case of
Eq. (3.8). The proof of Theorem 3.2 is now complete. 

We are now in a position of generalizing Theorem 3.1 to the case of the operator L.

Theorem 3.4. Let σ > 0 and f be a tempered distribution. Let also L a pseudo-differential
operator as at the beginning of Section 2.2
(1) Then, f ∈ Ȧ−2σ
2,∞ if and only if

(3.11) ∃c1 , c2 > 0 such that c1 t−σ/α ≤ ketL (t)k2 ≤ c2 t−σ/α , for all t > 0.
(2) Moreover, f ∈ Ȧ−2σ
2,∞ if and only if

(3.12) tσ/α ketL (t)k2 ∈ L∞ (R+ ), and lim inf tσ/α ketL f k2 > 0,
t→+∞

(3) In addition, f ∈ L2 (Rn ) ∩ Ȧ−2σ

2,∞ if and only if

(3.13) ∃c, c′ > 0 such that c (1 + t)−σ/α ≤ ketL (t)k2 ≤ c′ (1 + t)−σ/α , for all t > 0.
12 LORENZO BRANDOLESE

Proof. Recalling that the multiplication by P (ξ) and P (ξ)−1 conserve the euclidean norm of a
vector and that D(ξ) and etD(ξ) are diagonal matrices, we can write
Z
ke f k2 = |P (ξ)−1 eD(ξ)t P (ξ)fb(ξ)|2 dξ
tL 2

Z
2α
≃ |e−ct|ξ| fb(ξ)|2 dξ,

By our previous theorem, we have the analogue of the characterization to (3.1) for the operator
L, namely,
−2σ
(3.14) f ∈ Ḃp,q ⇐⇒ tσ/α ketL f k2 ∈ Lq (R+ , dt/t),

with the equivalence of the respective norms.

We now argue as in (3.5): We set as before dj = k∆j f k22 2−4jσ . For t > 0, let p = p(t) ∈ Z
such that 4pα ≤ t < 4(p+1)α . Then we have
X 2jα
ktσ/α etL f k22 ≃ t2σ/α e−ct2 k∆j f k22
j∈Z
X jα

(3.15) ≃ t2σ/α e−ct4 24jσ dj

j∈Z
X jα
≃ 42jσ e−c 4 dj−p ,
j∈Z

The proof can now be carried on making obvious modification to that of Theorem 3.1. 

4. Characterization of L2 -functions admitting a decay character

For any u0 ∈ L2 (Rn ), we introduce the quantity σ(u0 ) ∈ [0, +∞] given by
 −σ
(4.1) σ(u0 ) = sup σ ≥ 0 : u0 ∈ Ḃ2,∞ .

Notice that, for u0 ∈ L2 (Rn ), σ(u0 ) is always well defined in [0, +∞]. Indeed, by the Plancherel
0 ⊂ Ḃ 0 .
theorem, L2 (Rn ) = Ḃ2,2 2,∞
The relation between the decay character r ∗ (u0 ) and σ(u0 ) is given by the following simple
result.

Proposition 4.1. Let u0 ∈ L2 (Rn ). When the decay character of u0 does exist, it is given by
the formula (valid also in the limit cases r ∗ (u0 ) = −n/2 or ∞):

(4.2) r ∗ (u0 ) + n/2 = σ(u0 ).

Proof. In the case r ∗ (u0 ) = −n/2, we have +∞ = Pr (u0 )− > 0 for all r > −n/2, and so
ket∆ u0 k22 ≥ cǫ (1 + t)−ǫ for all ǫ > 0 by the first part of Proposition 2.8. Applying the characteri-
zation (3.1) of Besov spaces with negative regularity in terms of the heat kernel we get σ(u0 ) = 0.
In the case r ∗ (u0 ) = +∞, we have Pr (u0 )+ < ∞ for all r ∈ R and so ket∆ u0 k22 ≤ Ck (1 + t)−k
for all k ≥ 0 by the second part of Proposition 2.8. We apply again (3.1) to conclude that
σ(u0 ) = +∞. In the case −n/2 < r ∗ (u0 ) < ∞ the conclusion follows immediately combining
both parts of Proposition 2.8 with (3.1). 
 CHARACTERIZATION OF SHARP ALGEBRAIC DECAY 13

Remark 4.2. C. Niche and M.E. Schonbek generalized the notion of decay character by intro-
ducing, for s ≥ 0 and u0 ∈ H s (Rn ), the notation
√
rs∗ (u0 ) = r ∗ (Λs u0 ), where Λ = −∆.
The condition u0 ∈ H s (Rn ) insures that both u0 and Λs u0 belong to L2 (Rn ). They proved
in [6, Theorem 2.11] that, for u0 ∈ H s (Rn ), rs∗ (u0 ) = r ∗ (u0 ) + s (this formula being valid
also in the limit cases). Proposition 4.1 provides a one-sentence proof for this formula to hold:
−σ −(σ+s)
indeed, Λs is known to be an isomorphism between Ḃ2,∞ and Ḃ2,∞ and therefore σ ∗ (Λs u0 ) =
∗
σ (u0 ) + s.

One can ask if it is possible to give a Besov space characterization of the existence of the
decay character of an L2 -function. Proposition 4.3 below will provide a positive answer.

Proposition 4.3. Let u0 ∈ L2 (Rn ). Then the decay character of u0 does exist and r ∗ = r ∗ (u0 ) ∈
∗
(−n/2, ∞) if and only if there is σ ∗ > 0 such that u0 ∈ Ȧ−σ ∗
2,∞ . In this case, σ = σ(u0 ) and
formula (4.2) holds.
∗
Proof. Let us start assuming u0 ∈ Ȧ−σ ∗
2,∞ , with σ ∈ (0, ∞). For any ρ > 0, take j ∈ Z such that
2j−1 j
< ρ ≤ 2 . Then,
Z Z
−2σ∗ 2 −2jσ∗
ρ |b
u0 (ξ)| dξ . 2 u0 (ξ)|2 dξ
|b
|ξ|≤ρ |ξ|≤2j
∗
X
. 2−2jσ k∆k u0 k22
−∞<k≤j
 X 
−2jσ∗ ∗
.2 22kσ ku0 k2Ḃ −σ∗
2,∞
−∞<k≤j

. ku0 k2Ḃ −σ∗ .

2,∞

Letting r ∗ + n/2 = σ ∗ , such estimates imply Pr∗ (u0 )+ < +∞.

∗
Moreover, from our assumption we can take (jk ) as in the definition of Ȧ−σ 2,∞ . Let ρ0 = 2
j0

and M = k(jk − jk+1 )kℓ∞ . For any 0 < ρ < ρ0 , let jk be the largest integer of the sequence
(jk ) such that 2jk ≤ ρ. Then we must have 2jk +M ≥ ρ, otherwise we would find another integer
jh ∈ [jk , jk + M ] such that 2jh ≤ ρ, thus contredicting the maximality of jk . Then we have
Z
−2σ∗ ∗
(4.3) ρ u0 (ξ)|2 dξ & 2−2(jk +M )σ k∆jk u0 k22 ≥ c > 0,
|b
|ξ|≤ρ

with c independent on ρ ∈ (0, ρ0 ). This implies, Pr∗ (u0 )− > 0. By Remark 2.6 we conclude that
the decay character does exist and r ∗ (u0 ) = r ∗ ∈ (−n/2, +∞).
Conversely, assume that the decay character of u0 does exist and r ∗ (u0 ) ∈ (−n/2, +∞). Then
we know by the first assertion of Theorem 2.9, that the solution of the heat equation et∆ u0
1 ∗
is bounded from below and above by C(1 + t)− 2 (r +n/2) . We conclude by the last claim of
∗
Theorem 3.1 that u0 ∈ Ȧ−σ ∗ ∗
2,∞ and that σ = r + n/2. 

Remark 4.4 (Proof of Theorem 2.9). This first assertion of Theorem 2.9 follows immediately from
the estimates obtained in Proposition 2.8. We can now prove the last assertion of Theorem 2.9.
∗ −n
If u0 ∈ L2 (Rn ) satisfies the estimates (2.7), then by Theorem 3.4-(3), we obtain u0 ∈ Ȧ−2r
2,∞
14 LORENZO BRANDOLESE

and the decay character of u0 does exist and equals r ∗ by Proposition 4.3. Theorem 2.9 is now
completely proved. 

5. Application to nonlinear dissipative systems: The Navier–Stokes equation

As an application of our analysis to nonlinear dissipative systems, let us consider the Navier–
Stokes equations in R3 :

 3
ut + u · ∇u + ∇p = ∆u, x ∈ R , t > 0

(NS) ∇·u =0


u(x, 0) = u (x),
0

where u = u(x, t) = (u1 , u2 , u3 )(x, t) is the velocity field of an incompressible viscous fluid flow,
p = p(x, t) is the pressure, and u0 = (u0,1 , u0,2 , u0,3 ) is the initial datum. For any u0 ∈ L2 (R3 ),
with ∇ · u0 = 0, we know since J. Leray’s classical paper [3] that there is at least one solution
u ∈ Cw ([0, ∞), L2 (Rn ))∩L2loc (R+ , H 1 (R3 )) solving (NS) in the distributional sense and satisfying
the strong energy inequality
Z t
2
(5.1) ku(t)k2 + 2 k∇u(r)k22 dr ≤ ku(s)k22 ,
s

for s = 0, almost s > 0 and all t ≥ s. The uniqueness and the regularity of such solutions are
well known open problems.
The following theorem completely characterize the solutions satisfying sharp two-side decay
estimates. It completes well known Wiegner’s Theorem by giving several necessary and sufficient
conditions for the validity of bounds from below for the decay of the energy. It improves [2,
Theorem 6.5], and completes an earlier study by Skalák [9] where only the upper bounds were
discussed.

Theorem 5.1. Let u0 ∈ L2 (R3 ) and σ > 0. The three following properties are equivalent:
Z Z
−2σ 2 −2σ
(i) lim inf ρ |b
u0 (ξ)| dξ > 0 and lim sup ρ u0 (ξ)|2 dξ < ∞,
|b
ρ→0 |ξ|≤ρ ρ→0 |ξ|≤ρ
(ii) u0 ∈ Ȧ−2σ
2,∞ ,
(iii) −σ
(1 + t) . ket∆ u0 k2 . (1 + t)−σ .
If, moreover, u0 is a divergence-free vector-field, if u is a weak solution of (NS) as above and
0 < σ < 5/4, then the three previous properties are equivalent to
(iv) (1 + t)−σ . ku(t)k2 . (1 + t)−σ .

Proof. The equivalence (i) ⇐⇒ (ii) is just a reformulation of Proposition 4.3. Indeed, as ex-
plained in Remark 2.6, the decay character r ∗ (u0 ) does exist and belongs to (−n/2, ∞) if and
only if there exists r ∗ ∈ (−n/2, ∞) such that 0 < Pr∗ (u0 )− ≤ Pr∗ (u0 )+ < ∞.
The equivalence (ii) ⇐⇒ (iii) was established in Theorem 3.1.
The implication (iii) ⇒ (iv) relies on classical Wiegner’s theorem [10]: in the 3D case and in
the absence of external forces, the Theorem in [10] can restated as follows: if u0 ∈ L2 (Rn ) is a
divergence-free vector field such that ket∆ u0 k2 . (1 + t)−σ , and if u is a weak solution of (NS)
 CHARACTERIZATION OF SHARP ALGEBRAIC DECAY 15

as above, then the difference w(t) = u(t) − et∆ u0 satisfies the decay estimates

 −1/4−σ , if 0 ≤ σ < 1,
(1 + t)

kw(t)k2 . ln(e + t)(1 + t)−5/4 , if σ = 1,


(1 + t)−5/4 , if σ > 1.
Notice in particular that, for 0 < σ < 5/4, one gets tσ kw(t)k2 → 0 as t → ∞. Therefore, under
the same restrictions on σ and the conditions of item (iii), it easily follows that u = et∆ u0 + w
satisfies the estimates as in (5.1).
The proof of the implication (iv) ⇒ (iii) relies on the so-called inverse Wiegner’s theorem
recently established by Z. Skalák [9]: his results asserts (among other things) that if u is a
weak solution of the Navier–Stokes equation as above, satisfying ku(t)k2 . (1 + t)−σ (with
0 ≤ σ ≤ 5/4) then ket∆ u0 k2 . (1 + t)−σ . Hence, Wiegner’s theorem applies and decomposing
et∆ u0 = u − w property (iii) follows in the range 0 < σ < 5/4. 
In the borderline case σ = 5/4, the implication (iii)⇒(iv) still holds. But the converse
implication is no longer true. Indeed, it is possible to construct examples of Navier–Stokes
flows such that limt→+∞ t5/4 ket∆ u0 k2 = 0 and (1 + t)−5/4 . ku(t)k2 . (1 + t)−5/4 . This can
be achieved as follows: one starts with a divergence-free initial datum u0 ∈ L2 (R3 ) such that
b0 vanishes in a neighborhood of the origin. This insures a fast (exponential) L2 -decay of the
u
solution of the t∆
R ∞Rheat equation e u0 . Generically (i.e. in the absence of special symmetries),
the matrix 0 (u ⊗ u)(y, s) dy dt will not be a scalar multiple of the identity matrix. By a
theorem of T. Miyakawa and M.E. Schonbek [4], one obtains a solution u of the Navier–Stokes
equations that satisfies the lower bound estimate ku(t)k2 & (1 + t)−5/4 , and, by Wiegner’s
theorem, ku(t)k2 . (1 + t)−5/4 .

6. Examples
Example 6.1 (Construction of v0 ∈ L2 (Rn ) such that the limit Pr (v0 ) is not well defined).
Let r ∈ (−n/2, ∞). Decompose the unit ball B1 = {ξ : |ξ| ≤ 1} in concentric dyadic annuli:
B1 = Γ0 ∪ Γ−1 ∪ Γ−2 ∪ . . ., with Γj = {ξ : 2j−1 ≤ |ξ| ≤ 2j }. Then set vb0 (ξ) = |ξ|r on Γ0 ∪ Γ−2 ∪ . . .
and vb0 (ξ) = 0 elsewhere. A direct computation shows that 0 < Pr (v0 )− < Pr (v0 )+ < ∞. Indeed,
passing to spherical coordinates in the computation of Pr (v0 )+ and Pr (v0 )− ,
Z ∞
Pr (v0 )+ = ωn lim 2−2j(2r+n) λ2r+n−1 1Γ2j ∪Γ2(j−1) ∪... dλ
j→−∞ 0
and Z ∞
−(2j+1)(2r+n)
Pr (v0 )− = ωn lim 2 λ2r+n−1 1Γ2j ∪Γ2(j−1) ∪... dλ,
j→−∞ 0
where ωn is the surface of the unit sphere of Rn
and 1A denotes the indicator function of
the set A. Both limits could be easily computed, but in fact it is simpler to observe that
Pr (v0 )− −(2r+n) < 1 to conclude that P (v ) is not well defined by formula (2.1). In fact,
Pr (v0 )+ = 2 r 0
0 < Pr (u0 )− < Pr (u0 )+ < ∞ and by Remark 2.6 v0 admits a decay character in the sense of
Definition 2.5 and r ∗ (v0 ) = r.

Example 6.2. It can happen that r(u0 )+ = r(u0 )− , yet the decay character r ∗ (u0 ) does not
exist. The example is elementary. Let u0 ∈ L2 (Rn ) be such that u b0 (ξ) = |ξ|r0 log |ξ| in a
16 LORENZO BRANDOLESE

neighborhood of ξ = 0, for some r0 > −n/2. Then by a simple computation r(u0 )+ = r(u0 )− =
r0 . But the decay character r ∗ (u0 ) is not well defined, because the sup and the inf in the
definition of r(u0 )+ and r(u0 )− (see Eq: (2.2)-(2.3)) are not achieved. For this specific initial
datum, the corresponding solution of the linear problem (2.6) satisfies estimates similar to those
in (2.7), but with a corrective time-dependent logarithmic factor. Notice that for this example
there is no r such that 0 < Pr (u0 ) < ∞ (the existence of such an r was needed in the original
definition of decay character [2]).
Example 6.3. Let us construct an example of w0 ∈ L2 (Rn ) verifying r(w0 )+ < r(w0 )− . Fix
r ∈ (−n/2, +∞). Let (ak ) and (bk ) two real decreasing sequences such that 0 < bk+1 ≤ ak < bk
and bk → 0 for all k ∈ N, and consider two more positive sequences (ηk ), (hk ) to be chosen later.
We define w0 through its Fourier transform:
∞
X
wb0 (ξ) = hk 1ak ≤|ξ|≤bk .
k=0
P∞
The condition insuring w0 ∈ L2 (Rn ) is k=0 h2k (bnk − ank ) < ∞. It is then convenient to set ηk =
R
h2k (bnk −ank ), and the first condition is that (ηk ) is summable. Let Φr (ρ) = ρ−2r−n |ξ|≤ρ |w b0 (ξ)|2 dξ.
Then,
X∞
−(2r+n)
Pr (w0 )+ = lim sup Φr (ρ) = lim Φr (bk ) = lim ωn bk ηj ,
ρ→0 k→∞ k→∞
j=k
and
∞
X
−(2r+n)
Pr (w0 )− = lim inf Φr (ρ) = lim Φr (ak ) = lim ωn ak ηj ,
ρ→0 k→∞ k→∞
j=k+1
where ωn is the measure of the unit sphere. Now, we choose,
k
bk = 2−2 , ank = bnk − b2n
k , ηk = bk2r+n − bk+1
2r+n
.
P
Then, ∞ 2r+n
j=k ηj = bk . These choices insure that w0 ∈ L2 (Rn ) as (ηk ) is indeed summable and
determine (hk ). Moreover, we get 0 < Pr (w0 )+ = ωn < ∞ and we conclude that
r(w0 )+ = r.
P
Let us now study r(w0 )− . First observe that ∞ 2r+n 2r+n 2
j=k+1 ηj = bk+1 = (bk ) . Then,
∞
X
−(2r+n)
ak ηj = [bk (1 − bnk )−n ]−2r−n (bk2r+n )2 → 0 as k → ∞,
j=k+1

so that
Pr (w0 )− = lim inf Φr (ak ) = 0.
k→∞
For this reason, r(w0 )− 6= r and r ∗ (w0 ) does not exist. In fact r(w0 )− is given by r ′ ∈ (−n/2, ∞),
such that the limit
∞
X
−(2r ′ +n)
lim inf Φ (ρ) = lim Φ (ak ) =
r′ r′ ω n ak ηj ,
ρ→0 k→∞
j=k+1

is a strictly positive real: we need 2r ′ + n = 2(2r + n). In conclusion,

r(w0 )− = 2r + n/2, r(w0 )+ = r.
 CHARACTERIZATION OF SHARP ALGEBRAIC DECAY 17

7. Conclusions
Let σ > 0 and f ∈ L2 (Rn ). In [2] C. Bjorland and M.E. Schonbek, proved that:
Z
(7.1a) 0 < lim ρ −2σ
|fb(ξ)|2 dξ < ∞ =⇒ (1 + t)−σ . ket∆ f k2 . (1 + t)−σ ,
ρ→0+ |ξ|≤ρ

where et∆denotes the heat kernel in Rn and the symbol . means that the inequality ≤ holds up
to a multiplicative constant independent on time. In [6] C. Niche and M.E. Schonbek extended
this to a large class of pseudo-differential operator L, with homogeneous symbol of degree 2α,
proving that
Z
(7.1b) 0 < lim ρ−2σ |fb(ξ)|2 dξ < ∞ =⇒ (1 + t)−σ/α . ketL f k2 . (1 + t)−σ/α .
ρ→0+ |ξ|≤ρ

In the present paper, we showed that the reverse implications in (7.1a)-(7.1b) do not hold, as
the limit on the LHS might not exist. Next, suitably relaxing the condition on the LHS, we
got the following characterization of the class of L2 -functions satisfying sharp two-side decay
estimates. Namely, we established that:
(7.2)
 Z

 lim inf ρ−2σ
|fb(ξ)|2 dξ > 0

ρ→0+ Z|ξ|≤ρ −2σ
⇐⇒ f ∈ Ȧ2,∞ , ⇐⇒ (1+t)−σ/α . ketL f k2 . (1+t)−σ/α ,

 −2σ b 2
|f (ξ)| dξ < ∞
lim sup ρ
ρ→0+ |ξ|≤ρ

where Ȧ−2σ −2σ n

2,∞ is the subset of the Besov space Ḃ2,∞ (R ) introduced in Section 3.1.
Moreover, if f is a divergence-free vector field, then in the special case n = 3, L = ∆ and
0 < σ < 5/4, we proved in Theorem 5.1 that the three previous conditions are in turn equivalent
to the two-side energy estimate of weak solutions of the Navier–Stokes equation starting from f :
(7.3) (1 + t)−σ . ku(t)k2 . (1 + t)−σ .
When an L2 function satisfy any of the above equivalent conditions (7.2), the exponent σ ∈
(0, ∞) is uniquely determined. We coin this exponent σ = σ(f ) the Besov character of f . It
is related to (a suitable improvement of) Bjorland–Schonbek’s [2] notion of “decay character”,
denoted r ∗ (f ), by the relation σ(f ) = r ∗ (f ) + n/2.

References
[1] H. Bahouri, J.-Y. Chemin, and R. Danchin, Fourier analysis and nonlinear partial differential equa-
tions, Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences],
vol. 343, Springer, Heidelberg, 2011.
[2] C. Bjorland and M.E. Schonbek, Poincaré’s inequality and diffusive evolution equations, Adv. Differential
Equations 14 (2009), no. 3-4, 241–260.
[3] J. Leray, Sur le mouvement d’un liquide visqueux emplissant l’espace, Acta Math. 63 (1934), no. 1, 193–248.
[4] T. Miyakawa and M. E. Schonbek, On optimal decay rates for weak solutions to the Navier-Stokes equations
in Rn , Proceedings of Partial Differential Equations and Applications (Olomouc, 1999), 2001, pp. 443–455.
[5] C. Niche, Decay characterization of solutions to the Navier-Stokes-Voigt equations in terms of the initial
datum, preprint, arXiv:1507.06964 (2015).
[6] C. Niche and M.E. Schonbek, Decay characterization to dissipative equations, J. London Math. Soc. (2) 91
(2015), 573–595.
18 LORENZO BRANDOLESE

[7] , Comparison of Decay of Solutions to Two Compressible Approximations to Navier–Stokes Equation,

Bull. Brazilian Math. Soc. (to appear), arXiv:1501.0205v1.
[8] M.E. Schonbek, The Fourier splitting method, Advances in geometric analysis and continuum mechanics
(Stanford, CA, 1993), Int. Press, Cambridge, MA, 1995, pp. 269–274.
[9] Z. Skalák, On the characterization of the Navier-Stokes flows with the power-like energy decay, J. Math. Fluid
Mech. 16 (2014), no. 3, 431–446.
[10] M. Wiegner, Decay results for weak solutions of the Navier-Stokes equations on Rn , J. London Math. Soc.
(2) 35 (1987), no. 2, 303–313.

L. Brandolese: Université de Lyon, Université Lyon 1, CNRS UMR 5208 Institut Camille Jordan,
43 bd. du 11 novembre, Villeurbanne Cedex F-69622, France.
E-mail address: [email protected]
URL: http://math.univ-lyon1.fr/∼brandolese