Wavelet characterization of Sobolev norms∗

Daewon Chung†

Here we quickly review the Sobolev space. And then we observe the use of wavelets in defining
equivalent norms on L2 -Sobolev spaces. Also we’ll discuss about some other spaces.

Sobolev space
Sobolev space is a vector space of functions equipped with a norm that is a combination of Lp norms
of the function itself as well as its derivatives up to a given order. The derivatives are understood
in a suitable weak sense to make the space complete, thus a Banach space. We begin with the
classical definition of Sobolev spaces.
Definition 1. Let k be a nonnegative integer and let 1 < p < ∞ . The Sobolev space W k,p (Rn )
is defined as the space of functions f in Lp (Rn ) all of whose distributional derivatives ∂ α f are also
in Lp (Rn ) for all multi-indices α that satisfy |α| ≤ k. This space is normed by the expression
X
kf kW k,p = k∂ α f kLp
|α|≤k

where ∂ (0,...,0) f = f .
For the simplicity and convenience of discuss, we will only deal in the case of one dimensional.
In the one-dimensional case it is enough to assume f (k−1) is differentiable almost everywhere and
is equal almost everywhere to the Lebesgue integral of its derivative. Also, one of the most elegant
and useful ways of measuring differentiability properties of functions is in terms of L2 norms. One
of the reason for this is L2 is a Hilbert space and the other is the Fourier transform is unitary
isomorphism on L2 . From now, we only deal with
W k (R) = W k,2 (R) = {f ∈ L2 (R) : f (n) ∈ L2 (R) for all 0 < n ≤ k} ,
where the derivatives are in the weak sense; that is, it is a function f (n) such that
Z Z
f (n) (x)ϕ(x)dx = (−1)n f (x)ϕ(n) (x)dx
R R

for every test function ϕ ∈ S . This definition of Wk

makes its meaning clear, but there is an
k
equivalent characterization of W in terms of the Fourier transform.
Theorem 2. f ∈ W k if and only if (1 + |ξ|2 )k/2 fb ∈ L2 , and the norms
·X k ¸1/2 ·Z ¸1/2
f→ (n) 2
kf kL2 and f → b 2 2 k
|f (ξ)| (1 + |ξ| ) dξ
n=0

are equivalent.
∗
This is a term project for the Math572. Fall 2007
†
Graduate student with department of Mathematics & Statistics, University of New Mexico.

1
Proof. Since (f (n) )b(ξ) = (2πiξ)n fb(ξ), the Plancherel’s theorem implies that
k
X k Z
X
kf (n) k2L2 = |fb(ξ)|2 |(2πξ)n |2 dξ ,
n=0 n=0
P
so the theorem amounts to proving that the quantities kn=0 | ξ n |2 and (1 + | ξ|2 )k are comparable,
i.e., that each is bounded by a constant multiple of the other. But it is clear that | ξ n | ≤ 1 for
| ξ| ≤ 1 , and | ξ n | ≤ | ξ|n for | ξ| > 1 and 0 ≤ n ≤ k, so
k
X
| ξ n |2 ≤ C1 max(1, | ξ|2k ) ≤ (1 + | ξ|2 )k .
n=0
Pk
On the other hand, since | ξ|2k ≤ n=1 | ξ
n |2 ,

µ k
X ¶ k
X
2 k k 2k k 2k n 2 kk
(1 + | ξ| ) ≤ 2 max(1, | ξ| ) ≤ 2 (1 + | ξ| ) ≤ 2 C2 1 + | ξ | = 2 C2 | ξ n |2 .
n=1 n=0

Thus, proof is completed.

This suggests a generalization of W k in which k is replaced by an arbitrary real number s.

Namely, if u(ξ) is a function on R such that (1 + | ξ|2 )s/2 u(ξ) 2 1
R ∈ L , then uφ ∈ L for any φ ∈ S(R),
so u is a tempered distribution (whose action on φ ∈ S is uφ). Since the Fourier transform maps
tempered distributions into tempered distributions, we can defined the Sobolev space of order s:
½ Z ¾
W s (R) = f ∈ S 0 (R) : fb is a function and kf k2W s ≡ |fb(ξ)|2 (1 + | ξ|2 )s dξ < ∞ .

The norm k · kW s on W s thus defined is called the Sobolev norm of order s. Theorem 1 shows
that this definition agrees with the previous one when s is a nonnegative integer. In particular,
W 0 = L2 . Now we’ll state the very useful lemma.
Lemma 3. (Sobolev Embedding Theorem) If s > k + 12 , then W s ⊂ C k and there is a constant
C = Cs,k such that
sup sup |f (n) (x)| ≤ Ckf kW s .
n≤k x∈R

Proof. By the Fourier inversion theorem, if (f (n) )b ∈ L1 then f (n) is continuous and supx |f (n) (x)| ≤
k(f (n) )bkL1 . Hence, to complete the proof it is enough to prove that k(f (n) )bkL1 ≤ kf kW s when
n ≤ k . But (f (n) )b(ξ) = (2πiξ)n fb(ξ) , and by Schwarz inequality, for n ≤ k we have
Z Z
nb
|(2πiξ) f |dξ ≤ (2π) k
(1 + |ξ|2 )k/2 |fb(ξ)|dξ
Z
= (2π) k
(1 + |ξ|2 )s/2 |fb(ξ)|(1 + |ξ|2 )(k−s)/2 dξ
µZ Z ¶1/2
≤ (2π) k
(1 + |ξ| ) |fb(ξ)|2 dξ
2 s 2 k−s
(1 + |ξ| ) dξ

The first integral on the parenthesis is kf k2W s , and the second one is
Z Z ∞
(1 + |ξ|2 )k−s dξ = 2 (1 + r2 )k−s dr ,
0

which is finite if and only if k − s < − 12 since the integral is roughly r2(k−s) for large r .

2
Corollary 4. If f ∈ W s for all s ∈ R, then f ∈ C ∞ .
Remark A: The Lemma 3. can be sharpened a bit: if s = k + α + 21 where 0 < α < 1 , then
W s ⊂ C k+α .
Remark B: For the case p = ∞, the Sobolev space W k,∞ is defined to be the Holder space C n,α
where k = n + α and 0 < α ≤ 1 .

Wavelet characterization of Sobolev norms

Here we discuss the use of wavelets in defining equivalent norms on W s (R) where s ∈ R.
Theorem 5. Suppose that φ is the mother wavelet of a wavelet basis for L2 (R), that φ ∈ C α (R)
for some α > s ≥ 0 and that, for some P sufficiently large N , |φ(x)| ≤ (1 + |x|)−N . Then there are
constants C1 , C2 such that C1 kf kW s ≤ j,k (1 + 4js )|hf, φjk i|2 ≤ C2 kf k2W s .
2

First of all, we’ll prove Theorem 5 in the case of bandlimited wavelet which means the support
of the Fourier transform of this wavelet is contained in a finite interval. Then we’ll pass to the
case of orthonormal wavelets with a change of wavelet argument, showing that changing from one
suitably regular wavelet basis to another defines a bounded operator on W s (R). However, we’ll
skip the arguments for the biorthogonal case. The arguments for the this case are actually much
the same except the notion of change of wavelet is slightly more complicated.
Proof. (of Theorem 5 in the case of bandlimited wavelet) We begin with the specific form of the
wavelets. Let b(ξ) be a smooth,
P non-negative function supported inside [1/3, 4/3] and decreasing
away from ξ = 1 such that j |b(|ξ|/2j )|2 = 1 for all ξ 6= 0 . Set ω(ξ) = sign(ξ)eπiξ b(2|ξ|) . Then we
have the functions
j
ωnj (ξ) = 2−j/2 e−2πinξ/2 ω(ξ/2j ), (j, n ∈ Z)
which form an orthonormal basis basis for L2 (R) and its inverse Fourier transforms are wavelets
φjn (see [AWW]). By Plancherel,
hf, φjn i = hfb, ωnj i .
Due to Theorem 2. we can finish this proof by showing that
Z X X
|f (ξ)|2 (1 + | ξ|2 )s dξ < ∞ if and only if (1 + 4js ) |hf, ωnj i|2 < ∞ .
j∈Z

Let A j j j
Pj = {2 /32 ≤ | ξ| ≤ 42 · 2 /3} . Then 2 is a wavenumber associated with amplitude of |Aj |
and j kf χAj kL2 = 2kf kL2 . Thus with Parseval’s theorem and the fact that differentiation is
equivalent
P to multiplying in the frequency side. We have kf k2W s is equivalent in magnitude to
js 2 2 j
j (1 + (2 ) kf χAj kL2 . Since b(ξ) supported only inside [1/3, 4/3], b(| ξ|/2 ) survive only for
Pj+1
integer j − 1, j and j + 1 on Aj . Thus k=j−1 |b(| ξ|/2k )|2 = 1 on Aj , so we can complete this
proof by showing there is constants c, C so that
X X
c |hf, ωnj i|2 ≤ kf χAj k2L2 ≤ C |hf, ωnj i|2 .
n n

Now, let us assume f is supported in [0, ∞) . Then

Z
hf, ωnj i = f (ξ)2−j/2 e−2πinξ/2j ω(ξ/2j )dξ
Z 2j+2 /3
j j
= e−πiξ/2 f (ξ)b(ξ/2j )2−j/2 e2πinξ/2 dξ
2j /3

3
 j
Since {2−j/2 e2πinξ/2 }n∈Z forms an orthonormal basis for L2 ([2j+1 /3, 2j−1 /3)) , we have
X Z
j
|hf, ωnj i| = |e−πinξ/2 |2 |f (ξ)|2 |b(ξ/2j )|2 dξ ≤ Ckf χAj k2L2 .
2

On the other hand,

j+1 Z
X j+1 X
X
kf χAj k2L2 ≤ k
|f (ξ)b(ξ/2 )| dξ =2
|hf, ωnj i|2 .
k=j−1 k=j−1 n

With the above two estimates, we complete the proof when f is supported in [0, ∞) . Again the
same methods apply to f supported in [−∞, 0], then we can get our desired result.

Actually, for this bandlimited case the restriction 0 < s < 1 is not necessary condition. Now,
we have to pass this result to other wavelets. However, we’ll discuss this very roughly. (You can
find the detail of this process in [HL].) It will be convenient to return to interval notation ψI = ψjk
where I = I(j, k) = [k/2j , (k + 1)/2j ) . The general case of Theorem 5 change to proving bounds
on a change of wavelet matrix. This is where we use the hypothesis that s < 1 .
P
By expanding the wavelet ψ 1 in terms of the basis ψI2 . that is ψ 1 = I∈D hψ
1, ψI2 iψI2 , any
f ∈ L2 (R) can be expressed as
X X X
f= hf, ψI1 iψI1 = hf, ψI1 i hψI1 , ψJ2 iψJ2 .
I I J

Now let ψ 1 = ψP b be the bandlimited wavelet considered above for which we have the norm equiva-
2
lence kf kW s ∼ I∈D (1+|I| −2s )|hf, ψ b i|2 . Let H2 be the Hilbert sequence space consisting of those
P I
sequence {cI } so that I∈D (1 + |I|−2s )|cI |2 < ∞ . Let the matrix AIJ = hψIb , ψJ i . To prove that
a more general wavelet ψ = ψ 2 under consideration also provides a norm equivalence between W s
and H2 , it is enough to show that the matrix AIJ is bounded and continuously P invertible on Hs or
equivalently, jointly on l2 (D) and on Ḣs which is the space of {cI } such that I∈D |I|−2s |cI |2 < ∞ .
Since {cI } ∈ Ḣs if and only of |I|−s cI ∈ l2 (D) , proving that AIJ is bounded on Ḣs is equivalent
to proving that BIJ = (|I|/|J|)2 AIJ is l2 (D)-bounded. Thus, one finds a sufficient condition for
l2 -boundedness of AIJ first, then verifies a corresponding condition for BIJ .
Remark C: One can find the characterizations of other function spaces using wavelets. Espe-
cially, one can see the characterization of more general case of Sobolev spaces W k,p (R), where
1 < p < ∞, k = 1, 2, 3, · · · (see [HW]).

Notes and more

Sobolev norms are often arise in matters of well-posedness for PDEs. The variational form of the
Navier-Stokes equations in Rn is:
Z t
v(t) = S(t)v0 − PS(t − s)∇ · (v ⊗ v)(s)ds .
0

One wished to solve for the velocity field v = v(x, t) . Here S(t) = exp(t4) is the heat semigroup and
P is the singular integral operator that projects a vector field onto its divergence-free component.
Cannone and Meyer [CM] developed a method for obtaining so called mild solution of the variational

4
form of the Navier-Stokes equations in Rn (VFNSE) with initial data in certain function spaces
including Sobolev spaces. A crucial step was to develop a notion under which a function space X is
adapted to the bilinear product estimates required for application of a Picard iteration method to
solve for v. Cannone and Meyer then used Littlewood-Paley theory to prove such estimates which
can be regarded, in a sense, as estimates for wavelets projections of pointwise products specific to
bandlimited wavelets. Extending such estimates to more general wavelets is one step required in
adapting the techniques in [CM] to the setting of domain in Rn with boundary such as half-spaces.
Here is an abstraction of the essential role of wavelets in obtaining useful estimates for operators such
as PS(t − s)∇ . As we deal in class, Pj and Qj denote the orthogonal projection of an orthonormal
MRA. One seeks estimates of the form
kAj (f g)kX ≤ ηj kf kX kgkX .
Then one can establish estimates of the above form for Sobolev spaces as like following.
Theorem 7. Let Qj denote the projection onto the jth wavelet space of multiresolution analysis of
L2 (R) where the wavelets are orthogonal and Lipschitz with compact support. Then for 0 ≤ α ≤ 1
and j ≥ 1,
kQk (f g)kW α ≤ C2j(1/2−α) kf kW α kgkW α .
One can find the proof of theorem 7 in [HL]. Several other issues relating the use of wavelets to
the study of Navier-Stokes equations are outlined by Katz and Pavlović in [KP].

We’ve discussed the characterization of Sobolev norms in terms of magnitudes of wavelet coef-
ficients. The change of basis matrix that maps coefficients in one wavelet basis to coefficients in
another plays an important role here. Actually the magnitudes of the entries of this matrix depend
on the regularity of the the wavelets. Consequently, the wavelet characterization of Sobolev spaces
depends only on the regularity of the mother wavelet. This observation also can be extended to
Besov spaces which are extensions of Sobolev spaces. Membership in Besov and Sobolev spaces is
determined by the ”smoothness” of the functions concerned; the Besov norms involve differences,
the Sobolev norms use derivatives.

Let us define the Besov spaces in a classical way. As a first step, we define some operators. For
each h ∈ R, the translation operator Th is defined on functions f on R by Th f (x) = f (x + h) . The
first-difference operator 4h ≡ 41h is defined by
4h f (x) = (Th − I)f (x) = f (x + h) − f (x) ,
and higher-order differences are defined inductively by
4r+1 r
h f (x) = 4h (4h f )(x), (r = 1, 2, 3, . . .).
Since Thk f (x) = Tkh f (x) = f (x + kh) , it is clear that
X r µ ¶
r r r
4h f (x) = (Th − I) f (x) = (−1)r−k f (x + kh) .
k
k=0

Let the space of bounded continuous functions on R be denoted by C = C(R). It is a Banach space
under the supremum norm f → kf k∞ = sup{|f (x)| : x ∈ R} . The r-th order modulus of continuity
of a function f in Lp (R), 1 ≤ p < ∞ , is defined by
wr (f, t)p = sup k4rh f kp (f ∈ Lp , t > 0) ,
|h|≤t

5
when p = ∞ , the Lp -norm is replaced by the norm in C. Each modulus wr (f, t)p , (1 ≤ p ≤ ∞) ,
is a nonnegative increasing function of t > 0; furthermore, for each fixed t, wr (·, t)p is a seminorm
on Lp or C. It is following from the definition of higher order differences that

wr (f, t)p ≤ 2r kf kp .

Since 42h = Th2 − I = (Th + I)4h , one sees also that

wr (f, 2t)p ≤ 2r wr (f, t)p .

Definition 8. Suppose α > 0 and 1 ≤ p, q ≤ ∞ . Let r be a positive integer with r > α . The Besov
space Bpα,q consist of those f in Lp (if p < ∞) or C (if p = ∞) for which the norm
 ½Z ∞ ¾

 dt 1/q
 kf kp + [t−α wr (f, t)p ]q , (q < ∞) ,
kf kBpα,q = 0 t
 © ª

 kf kp + sup t−α wr (f, t)p , (q = ∞) ,
0<t<∞

is finite.

Besov spaces are family of function spaces characterized in terms of moduli of smoothness and
sharing many of the fundamental properties of Sobolev spaces, including embedding, restriction
and extension, and interpolation properties. Peetre [P] contains an excellent historical account of
these developments and a definition of Besov norms in terms of more general Littlewood-Paley
decomposition. Working in R, one fixes φ and ψj , j = 1, 2, . . . such that φb is supported in (−2, 2)
P
and ψbj is supported in (−2j+1 , 2j+1 )\(−2j−1 , 2j−1 ) , while φb + ψbj ≡ 1 and 2−jβ |dβ ψj /dxβ | ≤ cβ
for β ∈ N . For p ∈ [1, ∞], q > 0 and α ∈ R one define the space
½ µX
∞ ¶1/q ¾
Bpα,q (R) = f ∈ S 0 (R) : kφ ∗ f kLp + (2αj kφj ∗ f kLp )q <∞ .
j=1

In 1985, Frazier and Jawerth [FJ] provided a discrete description, essentially in terms of wavelet
frames, followed shortly thereafter by Lemarié and Meyer’s [LM] characterization of Bpα,q by wavelet
coefficient norms
µX
∞ ¶1/q
n jα q
kf kBpα,q (R ) = k{hf, φk i}klp + (2 k{hf, ψjk,p0 i}k klp ) .
j=1

Here, φ is the scaling function of an MRA and ψ is the corresponding mother wavelet, while
ψjk,p = 2j/p ψ(2j x − k) , also written ψI,p = |I|1/2−1/p ψI . To define Bpα,∞ one replaces the sum over
j by a corresponding supremum. Originally the typical bandlimited wavelets were used but, as in
Theorem 5, any sufficiently regular orthogonal or biorthogonal wavelet will do.

In this note, we discussed two families of function spaces. One is the family of Sobolev spaces
and the other is the family of Besov spaces. Why we are interested in these spaces? Sobolev
spaces are named after the Russian mathematician Sergei L. Sobolev. There are many criteria
for smoothness of mathematical functions. The most basic criterion may be that of continuity. A
considerably stronger notion of smoothness is that of differentiability (because functions that are
differentiable are also continuous) and a yet stronger notion of smoothness is that the derivative

6
also be continuous so called C 1 . Differentiable functions are important in many areas, and in par-
ticular for differential equations. However, in the twentieth century, it was observed that the space
C 1 (or C 2 , etc.) was not exactly the right space to study solutions of differential equations. The
Sobolev spaces are the modern replacement for these spaces in which to look for solutions of partial
differential equations. Thus their importance lies in the fact that solutions of partial differential
equations are naturally in Sobolev spaces rather than in the classical spaces of continuous functions.

In image processing, one of most important class of functions is a Bounded variation which is
defined by
½X
M ¾
f ∈ BV ([a, b]) ⇔ Vba (f ) = sup |f (xj ) − f (xj−1 )| : a = x0 < x1 < · · · < xM = b < ∞ ,
j=1

Most images can be thought to belong the space of functions with bounded variation. Although,
images are 2-dimensional functions, we just deal with one dimensional case. Let Bpα,q (C) = {f :
kf kBpα,q ≤ C} and BV (C) = {f : Vba ≤ C}. Then, for the periodic Besov spaces Bpα,q ([a, b]) with
corresponding periodic wavelet bases, there are constants C1 ≤ C2 such that (see [D])
∞
1 X j/2 1
2 k{hf, ψjk i}k kl1 ≥ Vba (f ) ≥ sup 2j/2 k{hf, ψjk i}k kl1 .
C1 C2 j
j=1

That is,
B11,1 (C1 ) ⊂ BV (1) ⊂ B11,∞ (C2 ) .
We can see Bounded Variation actually sit in between two Besov spaces, thus understanding Besov
spaces suffices to understand Bounded Variation.

References
[AWW] P. Auscher, G. Weiss, and M.V. Wickerhauser, Local sine and cosine bases of Coifman
and Meyer and the construction of smooth wavelets, In C.K. Chui, editor, Wavelets: A
tutorial in Theory and Applications, p.237-256, Academic Press (1992)

[BS] C. Bennet and R. Sharpley Interpolation of Operators, Academic Press (1988)

[CM] M. Cannone and Y.Meyer. Littlewood-Paley decomposition and Navier-Stokes equations,

Meth. and Math. of Analysis, 2:307-319 (1995)

[D] D.Donoho, Unconditional bases are optimal bases for data compression and statistical es-
timation, Appl. Comput. harmon. Anal., 1:100-115 (1993)

[FJ] M. Frazier amd B. Jawerth. Decomposition of Besov Space, Indiana Univ. Math. J., 34:777-
799 (1985)

[HW] E. Hernández and G. Weiss, A first course on Wavelets, CRC Press (1996)

[HL] Jeffrey A. Hogan and Joseph D. Lakey, Time-Frequency and Time-Scale Methods,
Birkhäuser (2005)

7
[KP] N. Katz and N. Pavlović Finite time blow-up for a dyadic model of the Euler equations,
Trans. Amer. Math. Soc., 357:695-708 (2005)

[LM] P.G. Lemarié and Y. Meyer. Ondelettes et bases hilbertiennes, Rev. Mat. Iberoamericana,
2:1-18 (1986)

[P] J. Peetre. New thoughts on Besov Spaces, Duke University Mathematics Serise. Duke
University, Raleigh-Durham (1975)

[S] E. Stein, Singular integrals and differentiability properties of Functions, Princeton Univ.
Press (1970)

Daewon Chung
Department of Mathematics and Statistics
University of New Mexico
Albuquerque, NM 87131-001

E-Mail Address : dwchung@unm.edu

Webpage : http://math.unm.edu/∼midiking