Generalized Functions and Fourier Analysis: Michael Oberguggenberger Joachim Toft Jasson Vindas Patrik Wahlberg Editors

Operator Theory
Advances and Applications

260
Michael Oberguggenberger
Joachim Toft
Jasson Vindas
Patrik Wahlberg
Editors
Generalized
Functions and
Fourier Analysis
Dedicated to Stevan Pilipović on the
Occasion of his 65th Birthday
Operator Theory: Advances and Applications
Volume 260
Founded in 1979 by Israel Gohberg
Editors:
Joseph A. Ball (Blacksburg, VA, USA)
Harry Dym (Rehovot, Israel)
Marinus A. Kaashoek (Amsterdam, The Netherlands)
Heinz Langer (Wien, Austria)
Christiane Tretter (Bern, Switzerland)
Associate Editors: Honorary and Advisory Editorial Board:

Vadim Adamyan (Odessa, Ukraine) Lewis A. Coburn (Buffalo, NY, USA)
Wolfgang Arendt (Ulm, Germany) Ciprian Foias (College Station, TX, USA)
Albrecht Böttcher (Chemnitz, Germany) J.William Helton (San Diego, CA, USA)
B. Malcolm Brown (Cardiff, UK) Thomas Kailath (Stanford, CA, USA)
Raul Curto (Iowa, IA, USA) Peter Lancaster (Calgary, Canada)
Fritz Gesztesy (Columbia, MO, USA) Peter D. Lax (New York, NY, USA)
Pavel Kurasov (Stockholm, Sweden) Bernd Silbermann (Chemnitz, Germany)
Vern Paulsen (Houston, TX, USA) Harold Widom (Santa Cruz, CA, USA)
Mihai Putinar (Santa Barbara, CA, USA)
Ilya Spitkovsky (Abu Dhabi, UAE)
Subseries
Linear Operators and Linear Systems
Subseries editors:
Daniel Alpay (Orange, CA, USA)
Birgit Jacob (Wuppertal, Germany)
André C.M. Ran (Amsterdam, The Netherlands)
Subseries
Advances in Partial Differential Equations
Subseries editors:
Bert-Wolfgang Schulze (Potsdam, Germany)
Michael Demuth (Clausthal, Germany)
Jerome A. Goldstein (Memphis, TN, USA)
Nobuyuki Tose (Yokohama, Japan)
Ingo Witt (Göttingen, Germany)
More information about this series at http://www.springer.com/series/4850

Michael Oberguggenberger • Joachim Toft
Jasson Vindas • Patrik Wahlberg
Editors
Generalized Functions
and Fourier Analysis
Dedicated to Stevan Pilipoviü on the Occasion
of his 65th Birthday
Editors
Michael Oberguggenberger Joachim Toft
Arbeitsbereich für Technische Mathematik Department of Mathematics
Universität Innsbruck Linnaeus University
Innsbruck, Austria Växjö, Sweden
Jasson Vindas Patrik Wahlberg

Department of Mathematics Department of Mathematics
Ghent University Linnaeus University
Gent, Belgium Växjö, Sweden
ISSN 0255-0156 ISSN 2296-4878 (electronic)

Operator Theory: Advances and Applications
ISBN 978-3-319-51910-4 ISBN 978-3-319-51911-1 (eBook)
DOI 10.1007/978-3-319-51911-1
Library of Congress Control Number: 2017939710
© Springer International Publishing AG 2017

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, express or implied, with respect to the material contained herein or for any errors or
omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
Printed on acid-free paper
This book is published under the trade name Birkhäuser, www.birkhauser-science.com

The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
A. Ascanelli, S. Coriasco and A. Süß

On Temperate Distributions Decaying at Infinity . . . . . . . . . . . . . . . . . . . . 1
F. Baumgartner, M. Oberguggenberger and M. Schwarz

Transport in a Stochastic Goupillaud Medium . . . . . . . . . . . . . . . . . . . . . . . 19
Y. Chen, M. Signahl and J. Toft

Hilbert Space Embeddings for Gelfand–Shilov
and Pilipović Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
S.-Y. Chung
Blow-up Phenomena for Solutions of Discrete Nonlinear
p-Laplacian Parabolic Equations on Networks . . . . . . . . . . . . . . . . . . . . . . . 45
A. Debrouwere
Generalized Function Algebras Containing Spaces of Periodic
Ultradistributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
G. Debruyne and J. Vindas

On General Prime Number Theorems with Remainder . . . . . . . . . . . . . . 79
P. Giordano and M. Kunzinger

Inverse Function Theorems for Generalized Smooth Functions . . . . . . . 95
T. Levajković, H. Mena and A. Tuffaha

The Stochastic LQR Optimal Control with Fractional
Brownian Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
G. Omel’yanov
Multi-soliton Collision for Essentially Nonintegrable Equations . . . . . . 153
P.R. Popivanov
Microlocal Solvability and Subellipticity of Several Classes of
Pseudodifferential Operators with Involutive Characteristics . . . . . . . . . 171
vi Contents
Y. Sawano
An Observation of the Subspaces of S . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
N. Teofanov and F. Tomić

Ultradifferentiable Functions of Class Mpτ,σ and Microlocal
Regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
J. Toft
Matrix Parameterized Pseudo-differential Calculi on
Modulation Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
H. Vernaeve
An Application of Internal Objects to Microlocal Analysis in
Generalized Function Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
-D. Vučković and J. Vindas
Rotation Invariant Ultradistributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
K. Yoshino
Eigenvalue Problems of Toeplitz Operators in
Bargmann–Fock Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
Operator Theory:
Advances and Applications, Vol. 260, vii–viii

c 2017 Springer International Publishing
Preface
In the past decade, the fields of Generalized Functions and Fourier Analysis have
converged and made joint progress, notably in the core disciplines of pseudodiffer-
ential operators, microlocal analysis and time-frequency analysis. The volume is
a compilation of chapters highlighting new developments in the two subjects and
their interaction. The focus is on concepts as well as applications in adjacent fields,
such as partial differential equations and stochastics. The collection is based on sci-
entific activities related to the International Association for Generalized Functions
(IAGF) and the ISAAC interest groups on Pseudo-Differential Operators (IGPDO)
and on Generalized Functions (IGGF), notably on the longstanding collaboration
of these groups within the International Society for Analysis, Applications and
Computation (ISAAC). Indeed, many researchers work in both fields.
This volume is dedicated to Stevan Pilipović at the occasion of his sixty-
fifth birthday. Stevan Pilipović is a leading personality in these fields. His research
extends through the full range of topics addressed here: from generalized func-
tions, pseudodifferential operators, microlocal analysis, time-frequency analysis,
linear, nonlinear and stochastic partial differential equations, stochastic processes,
integral equations, to fractional calculus and reaches out as far as mechanics and
analytic number theory. With a large number of publications (and an impressive
number of co-authors) as well as an equally impressive number of students he has
shaped and influenced the fields over a number of generations of mathematicians.
Stevan Pilipović is highly appreciated as a scientist, an impressive personality and
a dedicated and warm-hearted colleague and supervisor. He has led the analysis
branch of the University of Novi Sad through several decades and is also a member
of the Serbian Academy of Sciences.
The joint volume is an independent sequel to the volumes “Pseudo-Differential
Operators, Generalized Functions and Asymptotics” and “Pseudo-Differential Op-
erators and Generalized Functions”, which collected contributions to the two fields
from the ISAAC congresses 2011 in Moscow and 2013 in Krakow, respectively. It
contains 16 original research contributions, part of which were presented at the
Macau ISAAC Congress 2015, part of which were solicited by the Editors especially
for this volume. All papers were peer-reviewed.
The contents of the volume can be roughly categorized in the groups: distri-
bution spaces, operator theory (papers by Ascanelli/Coriasco/Süß, Chen/Signahl/
Toft, Sawano, and Yoshino); ultradistributions and microlocal analysis (papers by
viii Preface
Debrouwere, Teofanov/Tomić, and Vučković/Vindas); Colombeau theory (papers

by Giordano/Kunzinger, and Vernaeve); pseudodifferential operators and microlo-
cal analysis (papers by Popivanov, and Toft); qualitative analysis of partial differ-
ential equations (papers by Chung, and Omel’yanov); stochastic differential equa-
tions and generalized stochastic processes (papers by Levajković/Mena/Tuffaha,
and Baumgartner/Oberguggenberger/Schwarz); and finally, generalized functions
and analytic number theory (paper by Debruyne/Vindas).
The editors Michael Oberguggenberger

Joachim Toft
Jasson Vindas
Patrik Wahlberg
Operator Theory:
Advances and Applications, Vol. 260, 1–18

On Temperate Distributions
Decaying at Infinity
Alessia Ascanelli, Sandro Coriasco and André Süß
Abstract. We describe classes of temperate distributions with prescribed de-

cay properties at infinity. The definition of the elements of such classes is
given in terms of the Schwartz’ bounded distributions, and we discuss their
characterization in terms of convolution and of decomposition as a finite sum
of derivatives of suitable functions. We also prove mapping properties under
the action of a class of Fourier integral operators, with inhomogeneous phase
function and polynomially bounded symbol globally defined on Rd .
1. Introduction
In the analysis of partial differential equations, the well-known class of tem-
pered distributions S (Rd ) and related techniques based on Fourier transforms
are widely used. However, in some kind of problems global in the space variable,
where there is the need of knowing the behavior of the solution to a Cauchy prob-
lem as |x| → ∞, one would have to consider tempered distributions with a pre-
scribed decay at infinity to infer, from this decay and from the Cauchy data, what
is the precise decay of the solution. This actually happened to the authors, while
studying the existence and uniqueness of solutions to certain stochastic partial dif-
ferential equations of hyperbolic type, with (t, x)-depending coefficients admitting
a polynomial behavior with respect to x as |x| → ∞.
The present paper is so devoted to the study of subclasses of the space of tem-
pered distributions having a prescribed decay at infinity, to the characterization
of the elements of these classes, and to the statement and proof of the mapping
properties of these distributions under the action of Fourier integral operators
with symbols satisfying polynomially bounds at infinity in both the variable and
covariable.
More precisely, in Section 3 we introduce, for every s ∈ R, the space S (Rd )s
of s-decreasing distributions, that is, the tempered distributions u such that ·s u is
a bounded distribution in the sense of [13], see Section 2 and Remark 3.12 below
2 A. Ascanelli, S. Coriasco and A. Süß
for the precise definition of the concept of bounded distribution. The notation
x = (1 + |x|2 )1/2 when x ∈ Rd is currently used in this paper. We also recall the
class of tempered distributions with rapid decrease S (Rd )∞ .
In Theorems 3.3 and 3.15 we give useful characterizations of S (Rd )∞ and

S (Rd )s , by means of convolutions and by means of a structure result. Theorems
3.7 and 3.10 are the first main results of this paper, and deal with the action of
suitable Fourier integral operators on distributions of class S (Rd )∞ . In short,
given a symbol a ∈ S m,μ (Rd ) and an SG-phase function ϕ (see Section 2 and (2.5)
for the precise definitions), we show that the Fourier integral operator which for
f ∈ S (Rd ) has the form

(Opϕ (a)f )(x) = (2π)−d eiϕ(x,ξ) a(x, ξ)f(ξ) dξ, (1.1)
Rd

maps u ∈ S (R )∞ to Opϕ (a)u ∈ S (Rd )∞ . A similar mapping property holds
d
for Fourier integral operators in (1.1) with regular SG-phase function (see Section
2 and (2.5), (2.6) for the precise definition), when acting on the space S (Rd )s ,
namely, Opϕ (a) : S (Rd )s → S (Rd )s−m− d −ε , under the additional hypotheses
2
that s > d2 > 0, ε ∈ (0, δ], δ = s− d2 . This is the content of our third main Theorem
3.18.
Notice that in this paper we will not address either the problem of the conti-
nuity of the mappings studied in our main theorems, or the sharpness of Theorem
3.18. These questions are, in fact, at the moment still open, and are left for future
research projects.
For an application of the results obtained here to the study of stochastic
partial differential equations, see the forthcoming paper [2]. See also [9], for an
introduction to a different approach to the employment of microlocal analysis
tools to stochastic differential equations.
2. Preliminaries
This section is devoted to recalling the basic definitions of the subspaces DL p (Rd ),
p ∈ (1, +∞], of the space of temperate distributions, and of the so-called SG-
symbol classes, together with their corresponding calculi of pseudo-differential and
Fourier integral operators.
2.1. Subspaces of temperate distributions

The following remarkable subspaces of S (Rd ) are introduced in [13, Chapter VI,
§8], where more details and the proofs of the results recalled below, which we need
in Section 3, can be found. Of course, it is possible to give a more general definition
of DLq (Ω) and DL p (Ω), 1 < p ≤ ∞, 1/p + 1/q = 1, on an arbitrary open subset
Ω ⊆ Rd , but, since in the sequel we will only deal with the globally defined ones,
and since we will often rely on their characterization in terms of convolution, we
restrict ourselves to this setting.
On Temperate Distributions Decaying at Infinity 3
Definition 2.1. Let DLq (Rd ) = {ψ ∈ E (Rd ) : ∀α ∈ Zd+ ∂ α ψ ∈ Lq (Rd )}, q ∈ [1, ∞].
1. We say that {ψk }k∈N ⊂ DLq (Rd ) satisfies ψk → 0 in DLq (Rd ) as k → +∞ if,
for any α ∈ Zd+ , ∂ α ψk Lq (Rd ) → 0 as k → +∞.
2. We define
1 1
DL p (Rd ) = (DLq (Rd )) , p ∈ (1, +∞], + = 1.
p q
Remark 2.2.
1. The space DL 1 (Rd ) can be defined as well, as the dual of an appropriate
subspace of DL∞ (Rd ), see [13, Chapter VI, §8].
2. In [13, Chapter VI, §8], the distributions belonging to the space DL ∞ (Rd )
are called bounded distributions on Rd . This space is denoted there by B .
3. Notice that D(Rd ) ⊂ DLq (Rd ) ⊂ L∞ (Rd ) ∩ C ∞ (Rd ), q ∈ [1, +∞], and all
elements f ∈ DLq (Rd ), q ∈ [1, +∞) satisfy lim f (x) = 0. Moreover,
|x|→+∞
DLp (Rd ) ⊆ DLq (Rd ) for q ≥ p, and D(Rd ) is densely embedded into DLq (Rd ),
q ∈ [1, +∞), but not in DL∞ (Rd ).
Proposition 2.3. For all p ∈ (1, +∞], DLp (Rd ) ⊂ Lp (Rd ) ⊂ DL p (Rd ). Moreover,
for all p ∈ (1, +∞], u ∈ DL p (Rd ), α ∈ Zd+ , we have ∂ α u ∈ DL p (Rd ).
Theorem 2.4. Let p ∈ (1, +∞]. Then, the following three properties are equivalent:
1. u ∈ DL p (Rd );
2. for all χ ∈ D(Rd ), u ∗ χ ∈ Lp (Rd );
3. there exist M ∈ N, βj ∈ Zd+ , fj ∈ Lp (Rd ) ∩ C(Rd ), j = 1, . . . , M , such that

M
u= ∂ βj fj .
j=1
2.2. SG-symbols and operators

We recall the basic definitions concerning the so-called SG-classes, relying on ma-
terials appeared elsewhere, namely [1, 5], to which we refer the reader for further
details. Here and in what follows, A B means that A B and B A, where
A B means that A ≤ c · B, for a suitable constant c > 0.
For every m, μ ∈ R, the standard class S m,μ (R2d ) of SG symbols consists
of all functions a ∈ C ∞ (Rd × Rd ) with the property that, for any multiindices
α, β ∈ Zd+ , the estimates
|Dxα Dξβ a(x, ξ)| xm−|α| ξμ−|β| , (x, ξ) ∈ R2d , (2.1)
hold true.
More generally, we can consider smooth functions of SG type, namely, b ∈
C ∞ (Rkd ), k ∈ N, such that, for some m1 , . . . , mk ∈ R, and all x1 , . . . , xk ∈ Rd ,

k
|Dxα11 · · · Dxαkk b(x1 , . . . , xk )| xj mj −|αj | , xj ∈ Rd , j = 1, . . . , k.
j=1
Below we will need only the case k = 1 and the standard case k = 2. In the
case k = 1 we denote the corresponding class by S m (Rd ), m ∈ R. The above symbol
classes were first introduced in the ’70s by H.O. Cordes [3] and C. Parenti [10],
see also R. Melrose [8]. The corresponding classes of pseudo-differential operators
Op(S m,μ ) consist of the operators Op(a), a ∈ S m,μ (R2d ), which, for f ∈ S (Rd ),
have the form

(Op(a)f )(x) = (2π)−d eix,ξ a(x, ξ)f(ξ) dξ. (2.2)
Rd
They form a graded algebra, that is,
Op(S m1 ,μ1 ) ◦ Op(S m2 ,μ2 ) ⊆ Op(S m1 +m2 ,μ1 +μ2 ),
whose residual elements are the so-called smoothing operators in this context, that
is, those with symbols in

S −∞,−∞ (R2d ) = S m,μ (R2d ) = S (R2d ).
(m,μ)∈R2
Such operators have kernel in S (R2d ), and continuously map S (Rd ) to S (Rd ).
Operators in Op(S m,μ ) are linear and continuous from S (Rd ) to itself, and
extend uniquely to linear continuous operators from S (Rd ) to itself and from
H s,σ (Rd ) to H s−m,σ−μ (Rd ). Here H t,τ (Rd ), t, τ ∈ R, denotes the Sobolev–Kato
(or weighted Sobolev) space
H t,τ (Rd ) = {u ∈ S (Rd ) : ·t Dτ u = Op(ϑt,τ )u ∈ L2 (Rd )},
(2.3)
ϑt,τ (x, ξ) = xt ξτ ,
endowed with the natural norm u t,τ = Op(ϑt,τ )u L2 . Notice that s ≥ t and
σ ≥ τ imply H s,σ (Rd ) ⊆ H t,τ (Rd ), with continuous embedding (compact when
both order components inequalities are strict). We also recall that

H −∞,−∞ (Rd ) = H s,σ (Rd ) = S (Rd ),
s,σ∈R
(2.4)
H +∞,+∞ (Rd ) = H s,σ (Rd ) = S (Rd ).
s,σ∈R
An operator A = Op(a), is called elliptic (or S m,μ -elliptic) if a ∈ S m,μ (R2d )

and there exists R ≥ 0 such that
Cxm ξμ ≤ |a(x, ξ)|, |x| + |ξ| ≥ R,
for some constant C > 0. An elliptic SG operator A ∈ Op(S m,μ ) admits a
parametrix P ∈ Op(S −m,−μ ) such that
P A = I + K1 , AP = I + K2 ,
for suitable K1 , K2 , smoothing operators with symbols in S −∞,−∞ (R2d ), and A
turns out to be a Fredholm operator on the scale of functional spaces H t,τ (Rd ),
t, τ ∈ R.
In 1987, E. Schrohe [12] introduced a class of non-compact manifolds, the

so-called SG manifolds, on which it is possible to transfer from Rd the whole
SG calculus. These are manifolds which admit a finite atlas whose changes of
coordinates behave like symbols of order (1, 0) (see [12] for details and additional
technical hypotheses). An especially interesting example of SG manifolds are the
manifolds with cylindrical ends, where also the concept of classical SG operator
makes sense, see, e.g., [7, 8].
The calculus of corresponding classes of Fourier integral operators, in the
forms

f → (Opϕ (a)f )(x) = (2π)−d
eiϕ(x,ξ) a(x, ξ)f(ξ) dξ,
Rd
and
f → (Opϕ (a)∗ f )(x) = (2π)−d ei(x,ξ−ϕ(y,ξ)) a(y, ξ)f (y) dydξ,
R2d
when f ∈ S (Rd ), started in [4]. Here the operators Opϕ (a) and Op∗ϕ (a) = Opϕ (a)∗
are sometimes called Fourier integral operators of type I and type II, respectively,
with symbol (amplitude) a and phase function ϕ. Note that the type II operator
Opϕ (a)∗ is the formal L2 -adjoint of the type I operator Opϕ (a). We assume that
the phase function ϕ is real-valued, belongs to S 1,1 (Rd ) and satisfies
ϕx (x, ξ) ξ and ϕξ (x, ξ) x, (2.5)
if nothing else is stated. In this case, we say that ϕ is a SG-phase function. In
many cases, especially when studying the mapping properties of such operators,
ϕ should also fulfill the usual (global) non-degeneracy condition
| det(ϕxξ (x, ξ))| ≥ c, x, ξ ∈ Rd , (2.6)
for some constant c > 0. In such case, we say that ϕ is a regular SG-phase function,
and the corresponding Opϕ (a), Op∗ϕ (a), a ∈ S m,μ (R2d ), regular Fourier integral
operators. We have the following mapping properties for the SG Fourier integral
operators.
Theorem 2.5. Let ϕ be a SG-phase function, that is, ϕ is real-valued, ϕ ∈ S 1,1 (R2d )
and satisfies (2.5), a ∈ S m,μ (R2d ), m, μ ∈ R.
1. Opϕ (a) and Op∗ϕ (a) are linear continuous maps from S (Rd ) to itself, ex-
tendable to linear continuous maps from S (Rd ) to itself.
2. If ϕ satisfies also (2.6), then, for any s, σ ∈ R, Opϕ (a) and Op∗ϕ (a) are linear
continuous maps from H s,σ (Rd ) to H s−m,σ−μ (Rd ).
Point (2) in Theorem 2.5 is obtained as a corollary of a version of the classical
Asada–Fujjwara theorem on the L2 -boundedness for Fourier integral operators
with regular phase function and uniformly bounded symbol, see, e.g., [4, 11]. From
now on we will only deal with Type I operators. Similar results can be obtained for
Type II operators by formal L2 (Rd )-adjunction, and using the relations between
Type I and Type II operators, see, e.g., [4]. We will need, in particular, the following
composition theorem.
Theorem 2.6. Let ϕ be a SG-phase function, that is, ϕ is real-valued, ϕ ∈ S 1,1 (R2d )
and satisfies (2.5), and assume p ∈ S t,τ (R2d ), a ∈ S m,μ (R2d ), m, μ, t, τ ∈ R. Then,
Op(p) ◦ Opϕ (a) = Opϕ (h + r),
for h ∈ S m+t,μ+τ (R2d ), r ∈ S −∞,−∞ (R2d ).
Remark 2.7. For the symbol h in Theorem 2.6 it is possible to obtain an asymptotic
expansion. Since we will not use such result, and we will not need to separate the
smoothing part of the operator, we omit it here. In view of the inclusions among
the SG symbol spaces, below we will then simply write h in place of h + r. Recall
that
r ∈ S −∞,−∞ (R2d ) ⇒ Opϕ (r) = Op(S −∞,−∞ ),
that is, it is a smoothing operator.
3. Temperate distributions decaying at infinity and

mapping properties
3.1. Rapidly decreasing distributions
The next Definition 3.1 and Theorem 3.3 can be found in [13, Chapter VII, §5].
Definition 3.1. u ∈ S (Rd ) is a rapidly decreasing distribution if, for all k ∈ Z+ ,

·k u ∈ DL ∞ (Rd ). We denote by S (Rd )∞ the subspace of the rapidly decreasing
distributions.
Remark 3.2.
1. The space S (Rd )∞ is denoted by OC in [13]. It is a convolution algebra,
that is, u, v ∈ OC ⇒ u ∗ v ∈ OC .
2. Recall the definition of the space of smooth functions with slow growth, namely

OM = f ∈ E (Rd ) : ∀α ∈ Zd+ ∃mα ∈ Z+ such that
sup (1 + |x|)−mα |∂ α f (x)| < +∞ ⊂ S (Rd ).
x∈Rd
Denoting by F the Fourier transform on S (Rd ), one finds F OC = OM .

Moreover, F is an algebraic isomorphism of the convolution algebra OC onto the
function algebra OM , that is, for all u, v ∈ OC , u
∗v =u
· v.
Theorem 3.3. The following three properties are equivalent:

1. u ∈ S (Rd )∞ ;
2. for all χ ∈ D(Rd ), u ∗ χ ∈ S (Rd );
3. for any k ∈ Z+ there exist Mk ∈ N, βjk ∈ Zd+ , fjk ∈ L∞ (Rd ) ∩ C(Rd ),

j = 1, . . . , Mk , and Mk ∈ N, βjk ∈ Zd+ , fjk ∈ L∞ (Rd )∩C(Rd ), j = 1, . . . , Mk
such that
k

M

Mk
u= ·−k ∂ βjk fjk = ∂ βjk (·−k fjk ). (3.1)
j=1 j=1
Remark 3.4. The first equality in (3.1) is clear, from Definition 3.1 and Theorem
2.4. The second equality follows easily from the first and Leibniz rule for the
derivatives of products of smooth functions and distributions. Indeed, consider a
single term of the first summation in (3.1). If βjk = 0 for all j = 1, . . . , Mk , there
is nothing to prove. On the other hand, if |βjk | > 0, writing g in place of fjk and
γ in place of βjk , we find
γ
·−k ∂ γ g = ∂ γ (·−k g) − (∂ γ−α ·−k ) ∂ α g
α
α≤γ
|α|<|γ|
(3.2)
= ∂ γ (·−k g) + gαγk ∂ α g,
α≤γ
|α|<|γ|
with gαγk ∈ S −k−1 (Rd ). The term ∂ γ (·−k g) of the final expression in (3.2) has of
course the form of those appearing in the second summation of (3.1). If |γ| = 1, this
is true also for the summation in the final expression of (3.2). In fact, in this case
α = 0, and the summation is reduced to a single term such that g0γk = ·−k g0γk ,
g0γk ∈ S −1 (Rd ) ⊂ C ∞ (Rd ) ∩ L∞ (Rd ), which implies
g0γk g = ·−k g, with g ∈ C(Rd ) ∩ L∞ (Rd ).
If instead |γ| > 1, we iterate the procedure above on each term of the last
summation in (3.2), further reducing by one unit for each step the total number
of derivatives acting on g, and again obtaining terms of the desired form.
Indeed, for the characterization of S (Rd )∞ in terms of convolution, we can
equivalently check the behavior of u ∗ χ when χ ∈ S (Rd ), as it is shown in the
next Lemma 3.5.
Lemma 3.5. u ∈ S (Rd )∞ ⇔ ∀χ ∈ S (Rd ), u ∗ χ ∈ S (Rd ).
Proof. ⇐) This implication is obviously true, by (2) of Theorem 3.3, in view of
the embedding D(Rd ) → S (Rd ).
⇒) From the characterization of S (Rd )∞ in (3) of Theorem 3.3, for any
χ ∈ S (Rd ) we have, for arbitrary k ∈ Z+ ,

Mk
u∗χ= (∂ βjk χ) ∗ (·−k fjk ) ∈ E (Rd ),
j=1
in view of the properties of the convolution. Recalling Peetre’s inequality and that
fjk ∈ L∞ (Rd ), j = 1, . . . , Mk , for any α, β ∈ Zd+ , |α| ≤ k,
M
k

sup |xα ∂ β (u ∗ χ)(x)| = sup xα [(∂ β+βjk χ) ∗ (·−k fjk )](x)
x∈R d x∈R d
j=1

Mk
≤ sup |x |α
|(∂ β+βjk χ)(x − y)| y−k |fjk (y)| dy
d
j=1 x∈R

Mk
sup |x |α
|(∂ β+βjk χ)(x − y)| y−k dy
d
j=1 x∈R

Mk
= sup |xα | |(∂ β+βjk χ)(z)| x − z−k dz
d
j=1 x∈R
Mk

sup (|xα | x−k ) |(∂ β+βjk χ)(z)| zk dz 1.
x∈Rd j=1
Since k is arbitrary, this proves the claim.
Corollary 3.6. For any u ∈ S (Rd )∞ and any seminorm pl of S (Rd ) there exist
a seminorm pn of S (Rd ) and a constant q(u), depending only on u, l and d, such
that pl (u ∗ χ) pn (χ) q(u).
Proof. The result follows immediately by the computations in the proof of Lem-
ma 3.5.
The next Theorem 3.7 is a first mapping property for the regular SG Fourier
integral operators on temperate distributions with decay at infinity, in this case,
for the rapidly decreasing ones.
Theorem 3.7. Let ϕ be a regular SG-phase function, that is, ϕ is real-valued,
ϕ ∈ S 1,1 (R2d ) and satisfies (2.5), (2.6), a ∈ S m,μ (R2d ), m, μ ∈ R. Then,
u ∈ S (Rd )∞ ⇒ Opϕ (a)u ∈ S (Rd )∞ .
The proof will follow easily from Theorem 2.5, once we establish the next
Lemma 3.8, which characterizes S (Rd )∞ in terms of the Sobolev–Kato spaces.
Lemma 3.8. We have the following further characterization of S (Rd )∞ :

S (Rd )∞ = H +∞,−∞ (Rd ) = H s,σ (Rd ).
s∈R σ∈R

That is, u ∈ S (R )∞ ⇔ ∀s ∈ R ∃σ ∈ R : u ∈ H s,σ (Rd ).
d
Proof. ⇐) By hypothesis, for any s ∈ R we can find σ ∈ R such that u ∈ H s,σ (Rd ).
Then, for any χ ∈ S (Rd ), setting c(x, ξ) = χ
(ξ) ∈ S 0,−∞ (R2d ), we find, by the
mapping properties of the SG pseudo-differential operators on the Sobolev–Kato

spaces,

∀s ∈ R, u ∗ χ = Op(c)u ∈ H s,+∞ (Rd ) = H s,σ (Rd )
σ∈R

⇒u∗χ∈ H s,σ (Rd ) = S (Rd ).
s,σ∈R
By Lemma 3.5, we can then conclude that u ∈ S (Rd )∞ , as claimed.

⇒) Recalling point (3) of Theorem 3.3, choosing, for an arbitrary s ∈ R,
k ∈ Z+ such that k > d2 + s, we find Mk ∈ N, βjk ∈ Zd+ , fjk ∈ L∞ (Rd ) ∩ C(Rd ),
j = 1, . . . , Mk , such that

Mk
u= ∂ βjk (·−k fjk ).
j=1
Then, setting σ = − maxj |βjk |, ϑs,σ (x, ξ) = xs ξσ , cjk (x, ξ) = (iξ)βjk ∈
0,−σ
S (R2d ), we find

Mk
Op(ϑs,σ )u = (·s Dσ Op(cjk )·−s )(·s−k fjk )
j=1

Mk
= Op(ajk )gjk ∈ L2 (Rd ) ⇔ u ∈ H s,σ (Rd ),
j=1
in view of the definition (2.3) of the Sobolev–Kato spaces and the mapping prop-
erties of SG pseudo-differential operators, observing that, for all j = 1, . . . , Mk ,
ajk ∈ S 0,0 (R2d ), and gjk = ·s−k fjk ∈ L2 (Rd ) = H 0,0 (Rd ), since fjk ∈ L∞ (Rd )
and s − k < − d2 .
Proof of Theorem 3.7. By Lemma 3.8,

u ∈ S (Rd )∞ ⇔ u ∈ H +∞,−∞ (Rd ) = H s,σ (Rd ).
s∈R σ∈R
The statement then follows from the mapping properties of regular SG Fourier
integral operators on the H s,σ (Rd ) spaces, recalled in Theorem 2.5. In fact,
u ∈ S (Rd )∞ ⇔ ∀s ∈ R ∃σ ∈ R : u ∈ H s+m,σ+μ (Rd )
⇒ ∀s ∈ R ∃σ ∈ R : Opϕ (a)u ∈ H s,σ (Rd ) ⇔ Opϕ (a)u ∈ S (Rd )∞ .

Remark 3.9. Notice that in the proof of Theorem 3.7 we relied on the essential
hypothesis of regularity of the phase function ϕ, since we used the mapping prop-
erties of the SG Fourier integral operators on the Sobolev–Kato spaces H s,σ (Rd ),
which holds true under such more restrictive hypothesis. Actually, the result holds
true also for SG phase functions which do not fulfill (2.6), but we need a more sub-
tle argument to achieve that, as illustrated in the proof of our next main Theorem
3.10.
Theorem 3.10. Let ϕ be a SG-phase function, that is, ϕ is real-valued, ϕ ∈
S 1,1 (R2d ) and satisfies (2.5), a ∈ S m,μ (R2d ), m, μ ∈ R. Then,
u ∈ S (Rd )∞ ⇒ Opϕ (a)u ∈ S (Rd )∞ .
For the proof we need the next Lemma 3.11, which involves the so-called
completed π-products of two Fréchet spaces, whose definition we now briefly recall.
First remember that an absolutely convergent series in a Fréchet space E is the

N
limit of the partial sums zj , as N → +∞, of an absolutely summable sequence
j=1
{zj }j∈N ⊂ E. Namely, a sequence {zj }j∈N ⊂ E is absolutely summable if, for every
continuous seminorm p on E, the sequence of non-negative real number {p(zj )}j∈N
is summable. Then, given two Fréchet spaces E, F , every element θ ∈ E ⊗ π F is
the sum of an absolutely convergent series
∞

θ= λj (ej ⊗ fj ),
j=1
where λ = {λj }j∈N ∈ , and {ej }j∈N ⊂ E, {fj }j∈N ⊂ F are sequences converging
1
to zero in E and F , respectively, see, e.g., F. Trèves [14].

π S (Rd ).
Lemma 3.11. For any m ∈ R, S m,−∞ (R2d ) = S m (Rd ) ⊗
Proof. This follows from general results on topological vector spaces, see [14], in
view of the fact that S m,−∞ (R2d ) = S m (Rd , S −∞ (Rd )) = S m (Rd , S (Rd )), and
that S (Rd ) is a nuclear space.
Remark 3.12. By the definition of completed π-product of Fréchet spaces recalled
above, Lemma 3.11 implies that, for any a ∈ S m,−∞ (R2d ), there exist sequences
{bj }j∈N ⊂ S m (Rd ), {cj }j∈N ⊂ S (Rd ), λ ∈ 1 ,

N
such that {bj }j∈N and {cj }j∈N are bounded and aN = λj (bj ⊗ cj ) → a for
j=1
N → +∞ in the topology of S m,−∞ (R2d ).
The next Theorem 3.13 is the key result for the proof of Theorem 3.10.
Theorem 3.13. Let m ∈ R and ϕ be a SG phase-function.
(i) Assume b ∈ S m (Rd ), χ ∈ S (Rd ), c = χ
, and set a(x, ξ) = (b ⊗ c)(x, ξ) =
b(x) · c(ξ) ∈ S m,−∞ (R2d ). Then,
u ∈ S (Rd )∞ ⇒ Opϕ (a)u ∈ S (Rd ).
(ii) The same result holds true for any a ∈ S m,−∞ (R2d ).
Proof. (i) Under the hypotheses, for any u ∈ S (Rd )∞ , ψ ∈ S (Rd ), in view of
the definition (by duality) of the action of a SG Fourier integral operator of Type
I on temperate distributions, we have (Opϕ (a)u)(ψ) = u(ψ), where

ψ(y) = (2π)−d ei(−y,ξ+ϕ(x,ξ)) a(x, ξ)ψ(x) dxdξ

= (2π)−d Fξ→y χ (ξ) eiϕ(x,ξ) b(x)ψ(x) dx
= (2π)−d F·→y (
χ g),
and
g(ξ) = eiϕ(x,ξ) b(x)ψ(x) dx ∈ S (Rd ),
see [4]. Recalling the definition of Fourier transform on S (Rd ), and that u ∗ χ ∈
S (Rd ), which implies χu
= u ∗ χ ∈ S (Rd ), it follows
(Opϕ (a)u)(ψ) = u(ψ) = (2π)−d u χ g) = (2π)−d (

( χu)(g)

= (2π)−d ( )(ξ) g(ξ) dξ
χu

= (2π)−d (
χu)(ξ) eiϕ(x,ξ) b(x)ψ(x) dxdξ

= b(x)(2π)−d eiϕ(x,ξ) u ∗ χ(ξ) dξ ψ(x) dx
= (b · Opϕ (1)(u ∗ χ))(ψ),

which, by the continuity properties of the Type I SG Fourier integral operators,
implies
Opϕ (a)u = b · Opϕ (1)(u ∗ χ) ∈ S (Rd ).
(ii) By Lemma 3.11 and Remark 3.12, there exists a sequence {aN }N ∈N ⊂

N
S m,−∞ (R2d ), of finite linear combinations aN = λj aj of tensor products of the
j=1
form aj (x, ξ) = (bj ⊗ χ
j )(x, ξ), bj ∈ S m (Rd ), χj ∈ S (Rd ), j ∈ N, such that we
have aN → a in the S m,−∞ (R2d ) topology, with λ = {λj } ∈ 1 , {bj } bounded in
S m (Rd ), {χj } bounded in S (Rd ). This implies
Opϕ (aN ) → Opϕ (a), N → +∞, (3.3)
as linear operators from S (Rd ) to itself, as well as linear operators from S (Rd )
to itself. Moreover, by point (i) above, we also have, for any N ∈ N,

N
vN = Opϕ (aN )u = λj bj · Opϕ (1)(u ∗ χj ) ∈ S (Rd ). (3.4)
j=1
We will now show that {vN } is actually a Cauchy sequence in S (Rd ), thus con-
verging to v ∈ S (Rd ). This, together with (3.3), proves the claim, since
Opϕ (a)u = lim Opϕ (aN )u = lim vN = v ∈ S (Rd ).
N →+∞ N →+∞
Indeed, in view of (3.4), of the boundedness of {bj } and {χj }, of the continuity on
S (Rd ) of bj · = Op(bj ⊗ 1) and Opϕ (1), and of Corollary 3.6, for any seminorm pk
on S (Rd ), there exist a seminorm pl on S (Rd ) and a constant q(u), such that,
for any M ≥ N
M

M
pk (vM − vN ) = pk λj bj · Opϕ (1)(u ∗ χj ) ≤ |λj | pk (bj · Opϕ (1)(u ∗ χj ))
j=N j=N

M
M
|λj | pl (u ∗ χj ) q(u) |λj | < ε,
j=N j=N
provided N is chosen large enough, recalling that {λj } ∈ 1 .

Proof of Theorem 3.10. By the calculus of SG Fourier integral operators, for any
χ ∈ S (Rd ), m, μ ∈ R, a ∈ S m,μ (R2d ), u ∈ S (Rd )∞ , and ϕ as above, setting
(ξ) ∈ S 0,−∞ (R2d ),
p(x, ξ) = χ
(Opϕ (a)u) ∗ χ = (Op(p) ◦ Opϕ (a))u = Opϕ (h)u,
with h ∈ S m,−∞ (R2d ). We then have, by Theorem 3.13, (ii), and Lemma 3.5,
∀χ ∈ S (Rd ) (Opϕ (a)u) ∗ χ ∈ S (Rd ) ⇔ Opϕ (a)u ∈ S (Rd )∞ ,
as claimed.
3.2. s-decreasing distributions, s ∈ R
In the next Definition 3.14 we introduce, in analogy with Definition 3.1, a class of
temperate distributions which have a finite decay rate at infinity.
Definition 3.14. Given s ∈ R, we say that u ∈ S (Rd ) is an s-decreasing distribu-
tion if ·s u ∈ DL ∞ (Rd ). We denote by S (Rd )s the subspace of the s-decreasing
distributions.
We have the following analogy of the characterization in Theorem 3.3 for the
s-decreasing distributions.
Theorem 3.15. The following three properties are equivalent:
1. u ∈ S (Rd )s , s ∈ R;
2. for all χ ∈ S (Rd ), ·s (u ∗ χ) ∈ L∞ (Rd ) ∩ C ∞ (Rd );
3. there exist M ∈ N, βj ∈ Zd+ , fj ∈ L∞ (Rd ) ∩ C(Rd ), j = 1, . . . , M , and
M ∈ N, βj ∈ Zd+ , fj ∈ L∞ (Rd ) ∩ C(Rd ), j = 1, . . . , M such that

M
j
M
−s β
u= · ∂ fj = ∂ βj (·−s fj ). (3.5)
j=1 j=1
Let us introduce, for r ∈ R, the weighted space

Lr (R ) = f : R → C measurable: f L1r (Rd ) := |f (x)|x dx < +∞
1 d d r
= ·−r L1 (Rd ),
and correspondingly the space
DL1r (Rd ) = {ψ ∈ E (Rd ) : ∀α ∈ Zd+ ∂ α ψ ∈ L1r (Rd )}, (3.6)
with the natural notion of convergence, namely, the one in Definition 2.1, (1),
with the L1r norm in place of the Lq norm. Notice that, of course, (L1r (Rd )) ≡
L∞ r ∞
−r (R ) = · L (R ). For the proof of Theorem 3.15 we need to show that
d d
r
we can identify · DL∞ (Rd ) and the topological dual space of DL1r (Rd ). The de-
sired identification, proved in Proposition 3.17, will follow from the next technical
Lemma 3.16.
Lemma 3.16. We have, for any r ∈ R,
ψ ∈ DL1r (Rd ) ⇔ xr ψ ∈ DL1 (Rd ).
Proof. We have to check that, for any ψ ∈ E (Rd )
ψ ∈ DL1r (Rd ) ⇔ ∀α ∈ Zd+ ·r ∂ α ψ ∈ L1 (Rd )
⇔ ∀α ∈ Zd+ ∂ α (·r ψ) ∈ L1 (Rd ) (3.7)
⇔ · ψ ∈ DL1 (R ).
r d
The equivalence (3.7) is trivially true for α = 0. It also holds true for |α| ≤ 1.
In fact, for |α| = 1,
·r ∂ α ψ = ∂ α (ψ·r ) − ψ(∂ α ·r ) = ∂ α (ψ·r ) − (ψ·r ) · gα ,
gα ∈ S −1 (Rd ), where, of course, S −1 (Rd ) ⊂ S 0 (Rd ) ⊂ L∞ (Rd ) ∩ C ∞ (Rd ) ⇒
S −1 (Rd ) · L1 (Rd ) ⊂ L1 (Rd ). We then find, on one hand,
·r ∂ α ψ ∈ L1 (Rd ), |α| ≤ 1 ⇒ ∂ α (·r ψ) = ·r ∂ α ψ + (ψ·r ) · gα ∈ L1 (Rd ), |α| ≤ 1,
and, on the other hand,
∂ α (·r ψ) ∈ L1 (Rd ), |α| ≤ 1 ⇒ ·r ∂ α ψ = ∂ α (ψ·r )−(ψ·r )·gα ∈ L1 (Rd ), |α| ≤ 1.
We now proceed by induction on |α|, that is, we assume that (3.7) holds
true for all ψ ∈ E (Rd ) and all α ∈ Zd+ such that |α| ≤ k, k ≥ 1. Let now
|α| = k + 1 ⇒ α = β + γ, |β| = k, |γ| = 1. Then,
(3.7) for |α| ≤ k and ·r ∂ α ψ ∈ L1 (Rd ), |α| ≤ k + 1
⇒ ∂ α (·r ψ) = ∂ β (·r ∂ γ ψ) + ∂ β ((ψ·r ) · gγ ) ∈ L1 (Rd ), |α| ≤ k + 1.

(I) (II)
In fact, by assumption, · ∂ ψ ∈ L (Rd ) ∩ C ∞ (Rd ), and, by the inductive

r γ 1
hypothesis,
∂ β (·r ∂ γ ψ) ∈ L1 (Rd ) ⇔ ·r (∂ β (∂ γ ψ)) ∈ L1 (Rd ),
which holds true, since |β + γ| = |α| = k + 1. Thus, (I) belongs to L1 (Rd ). This is
true also for (II), since, by the Leibniz formula and the inductive hypotheses1 , it
is a finite linear combination of terms belonging to L1 (Rd ). Conversely,
(3.7) for |α| ≤ k and ∂ α (·r ψ) ∈ L1 (Rd ), |α| ≤ k + 1

⇒ ·r ∂ α ψ = ∂ α (·r ψ) − (·r ∂ β ψ) · gαβ ∈ L1 (Rd ), |α| ≤ k + 1.

β≤α
(I) |β|<|α|

(II)
In fact, (I) belongs to L1 (Rd ) by hypothesis. Moreover, as above, we have

gαβ ∈ S −1 (Rd ) ⊂ L∞ (Rd ). This implies that (II) is also in L1 (Rd ), since in each
term of the sum we have |β| ≤ k, so that ·r ∂ β ψ ∈ L1 (Rd ) ⇔ ∂ β (·r ψ) ∈ L1 (Rd )
holds true by the inductive hypothesis.

Proposition 3.17. For any r ∈ R we have

(DL1r (Rd )) = ·r DL ∞ (Rd ).
Proof. 1. We start by proving that v ∈ ·r DL ∞ (Rd ) ⇒ v ∈ (DL1r (Rd )) . If v ∈

M
·r DL ∞ (Rd ), then by Theorem 2.4, (3), we have v = ·r ∂ βj fj , for suitable
j=1
M ∈ N, fj ∈ L∞ (Rd ), βj ∈ Zd+ , j = 1, . . . , M . Let {ψk } ⊂ DL1r (Rd ) be a sequence
converging to 0 in DL1r (Rd ), that is, for any α ∈ Zd+ , ·r ∂ α ψk L1 (Rd ) → 0 as
k → +∞. By Lemma 3.16, this is the same as saying that, for any α ∈ Zd+ ,
∂ α (·r ψk ) L1 (Rd ) → 0 as k → +∞. Then,

M M
|βj |
|v(ψk )| = (· ∂ fj )(ψk ) = (−1) fj (∂ (· ψk ))
r βj βj r

j=1 j=1
M

= (−1)|βj | fj · ∂ βj (·r ψk )

j=1

M
≤ fj L∞ (Rd ) ∂ βj (·r ψk ) L1 (Rd ) → 0, k → +∞,
j=1
which implies v ∈ (DL1r (Rd )) , as claimed.

2. We now show that v ∈ (DL1r (Rd )) ⇒ v ∈ ·r DL ∞ (Rd ), that is, equiva-
lently, ·−r v ∈ DL ∞ (Rd ). Let {ψk } ⊂ DL1 (Rd ) be a sequence converging to 0 in
DL1 (Rd ), that is, for any α ∈ Zd+ , ∂ α ψk L1 (Rd ) → 0 for k → +∞. By Lemma
1 Clearly,
only derivatives of ψ·r of total order not higher then |β| = k appear in (II), multiplied
by smooth functions belonging to ∂ κ S −1 (Rd ) ⊂ L∞ (Rd ).
3.16, this is equivalent to saying that {·−r ψk } ⊂ DL1r (Rd ) and ·−r ψk → 0 in
DL1r (Rd ). Then, we immediately have
(·−r v)(ψk ) = v(·−r ψk ) → 0 for k → +∞.
Proof of Theorem 3.15. The equivalence between (1) and (3) is clear from Defini-
tions 2.1 and 3.14, Theorem 2.4, and Remark 3.4. The implication (3)⇒(2) also
follows easily, in view of the properties of the convolution, writing, for an arbitrary
χ ∈ S (Rd ),

M
M
·s (u ∗ χ) = ·s ∂ βj (·−s fj ) ∗ χ = ·s [(·−s fj ) ∗ (∂ βj χ)].
j=1 j=1
Let us examine each summand, by considering functions of type ·s ·[(·−s f )∗

ψ] with ψ ∈ S (Rd ), f ∈ L∞ (Rd )∩C(Rd ). Of course, ·s ·[(·−s f )∗ ψ] ∈ C ∞ (Rd ).
Let us now check that it is uniformly bounded on Rd . We compute, using Peetre’s
inequality,

−s
[· · (· f ) ∗ ψ)](x) = x
s s
x − y−s f (x − y)ψ(y) dy

xs x−s y|s| f (x − y)ψ(y) dy

f L∞ (Rd ) y|s| ψ(y) dy 1,
which is the claimed estimate. For what concerns the implication (2)⇒(3), we can
use a variant of the argument in the proof of Theorem XXV in [13, Chapter VI].
Consider any χ ∈ D(Rd ) ⊂ S (Rd ) and ϕ ∈ DL1−s (Rd ) with ϕ varying in the unit
ball B of L1−s (Rd ). Then, for any u ∈ S (Rd )s we find that2
{(u ∗ χ)(ϕ) = (u ∗ ϕ̌)(χ̌)}ϕ∈B
is a bounded set of C. It follows that {u ∗ ϕ̌}ϕ∈B , is a bounded set in D (Rd ), and
this implies (see [13, Chapter VI, §7], Theorem XXII) that, for any K compact
subset of Rd , there exists σ ∈ Z+ such that, for3 χ ∈ D σ (K), both (u ∗ χ)(ϕ) and
(u ∗ ϕ̌)(χ̌) give again a bounded subset of C for ϕ ∈ B. This implies that u ∗ χ
is still in ·−s L∞ (Rd ) = (L1−s (Rd )) for χ ∈ D σ (K). Then, choosing χ = γE,
where γE is a parametrix of Δk , k large enough (see [13], (VI, 6; 22)), we find
Δk (γE) − ζ = δ with ζ ∈ D(Rd ), so that
u = u ∗ δ = u ∗ Δk (γE) − u ∗ ζ = Δk (u ∗ (γE)) − u ∗ ζ.
2 In fact, by the definition of convolution of distributions and test functions,

(u ∗ χ)(ϕ) = ((u ∗ χ) ∗ ϕ̌)(0) = (u ∗ (χ ∗ ϕ̌))(0) = (u ∗ (ϕ̌ ∗ χ))(0)
= ((u ∗ ϕ̌) ∗ χ)(0) = (u ∗ ϕ̌)(χ̌).
3 Recall that D σ (K) = {ψ ∈ C σ (Rd ) : supp ψ ⊆ K}.

This implies the desired claim, since we have written u as a sum of derivatives
of functions belonging to ·−s (L∞ (Rd ) ∩ C(Rd )).
We conclude with our last main result, the next Theorem 3.18. Here we
show that regular SG Fourier integral operators map S (Rd )s to S (Rd )s−m− d −ε ,
2
∈ (0, δ], for a suitable δ > 0, provided the decay index s is large enough.
Theorem 3.18. Let m, μ ∈ R, a ∈ S m,μ (R2d ), ϕ a regular SG-phase function, that
is, ϕ is real-valued, ϕ ∈ S 1,1 (R2d ) and satisfies (2.5), (2.6). Moreover, assume
s > d2 and ε ∈ (0, δ], δ = s − d2 . Then,
Opϕ (a) : S (Rd )s → S (Rd )s−m− d −ε .
2
Proof. Recall that, from Theorem 3.15, (3), for any u ∈ S (Rd )s there exist M ∈

M
N, β ∈ Zd+ , fj ∈ L∞ (Rd ) ∩ C(Rd ), j = 1, . . . , M , such that u = ∂ βj (·−s fj ).
j=1
Set δ = s − d2 > 0, and fix ε ∈ (0, δ]. Then, gj = ·− 2 −ε fj ∈ L2 (Rd ), since
d
fj ∈ L∞ (Rd ), j = 1, . . . , M , and ·− 2 −ε ∈ L2 (Rd ) = H 0,0 (Rd ). It also follows,

d
setting k = s − d2 − ε ≥ 0, κ = − maxj |βj | ≤ 0,

M
M
i|βj | Dβj ·−s+ 2 +ε (·− 2 −ε fj ) = i|βj | Dβj ·−k gj ∈ H k,κ (Rd ).
d d
u=
j=1 j=1
In fact, by the mapping properties of the SG pseudo-differential operators on

the Sobolev–Kato spaces, as well as their inclusions when the two indices increase,
gj ∈ H 0,0 (Rd ) ⇒ i|βj | Dβj ·−k gj ∈ H k,−|βj | (Rd ) ⊆ H k,κ (Rd ),
j = 1, . . . , M ⇒ u ∈ H k,κ (Rd ).
By Theorem 2.5, (2), we then have
Opϕ (a)u ∈ H k−m,κ−μ (Rd ).
It now follows that, for an arbitrary χ ∈ S (Rd ), setting c(x, ξ) = χ
(ξ) ∈
S 0,−∞ (Rd ), again by by the mapping properties of the SG pseudo-differential op-
erators on the Sobolev–Kato spaces,
(Opϕ (a)u) ∗ χ = Op(c)(Opϕ (a)u) ∈ H k−m,+∞ (Rd ) = ·m−k H 0,+∞ (Rd )
⇒ ·s−m− 2 −ε ((Opϕ (a)u) ∗ χ) ∈ ·s−m− 2 −ε+m−s+ 2 +ε H 0,+∞ (Rd )
d d d
= H 0,+∞ (Rd ).
Notice that, by Sobolev’s Lemma, H 0,+∞ (Rd ) ⊂ C ∞ (Rd ) and w ∈ H 0,+∞ (Rd ) ⇒
∂ α w ∈ L∞ (Rd ) for any α ∈ Zd+ . We can then conclude
·s−m− 2 −ε ((Opϕ (a)u) ∗ χ) ∈ L∞ (Rd ) ∩ C ∞ (Rd ),
d
which, by Theorem 3.15, (2), gives the claim.

Remark 3.19. In the case of S (Rd )s , s ∈ R, the argument used in the proof of
u
Theorem 3.10 fails in general, since χ does not need to be a function, given the fact
that we can only claim the much weaker property u ∗ χ ∈ ·−s (L∞ (Rd ) ∩ C(Rd )),
compared to the case of the rapid decrease in S (Rd )∞ , where u∗χ was a Schwartz
function. So, for convenience, here we kept the assumption on the regularity of the
phase function of Opϕ (a), which is anyway often fulfilled in the applications.
Of course, one could have chosen to estimate the decay at infinity in a dif-
s,−∞ d
2 d
ferent way, namely, through the L (R )-based space H (R ) = H s,σ (Rd ).
σ∈R
In such case, under the regularity assumption on the phase functions, we clearly
have the sharp result Opϕ (a) : H s,−∞ (Rd ) → H s−m,−∞ (Rd ). However, here we
were interested in an environment obtained through the bounded distributions, in
analogy with the definition of S (Rd )∞ , which is instead modeled on L∞ (Rd ).
Already in the cases p ∈ (1, ∞), p = 2, losses of regularity and decay for the
maps generated by SG Fourier integral operators on Lp (Rd ) and the corresponding
Lp (Rd )-modeled weighted Sobolev spaces are expected and known, see, e.g., [6]
and the references quoted therein. So, the result in Theorem 3.18 is not surprising,
given the fact that, essentially, we are treating the even more complicated situation
of p = ∞.
The condition s > d2 appears to be a quite natural one (comparable, in spirit,
with the often assumed regularity one σ > d2 ), with which we ask to be indeed in a
case of sufficiently strong decay. Since here we are not interested in an estimate of
the regularity (which is anyway given by the multiindices βj appearing in (3.5) and
the order component μ of the symbol a), we focus on the decay at infinity. Note
that, when s > d2 , for u ∈ S (Rd )s we also have u ∗ χ ∈ ·−s (L∞ (Rd ) ∩ C(Rd )) ⇒
u ∗ χ ∈ L2 (Rd ), so an adaptation of the argument in the proof of Theorem 3.10,
leading to the removal of the assumption of regularity of the phase function, should
also be possible.
This last point, as well as the possibility of improving Theorem 3.18 to
a sharper result (for instance, to prove that Opϕ (a) indeed maps S (Rd )s to
S (Rd )s−m− d , or to get an even better decay estimate), are still open problems,
2
as well as the continuity properties of the maps studied in our main Theorems 3.7
and 3.10.
Acknowledgement
We wish to thank J. Seiler and P. Wahlberg, for useful hints and discussions. We
also wish to thank an anonymous referee, for the constructive criticism and the
suggestions, aimed at improving the overall quality of the paper.
References
[1] A. Ascanelli, S. Coriasco, Fourier integral operators algebra and fundamental solu-
tions to hyperbolic systems with polynomially bounded coefficients on Rn . J. Pseudo-
Differ. Oper. Appl. 6, 4 (2015), 521–565.
[2] A. Ascanelli, S. Coriasco, A. Süß, Solution theory of hyperbolic stochastic partial dif-
ferential equations with polynomially bounded coefficients (2016), arXiv:1610.01208.
[3] H.O. Cordes. The Technique of Pseudodifferential Operators. Cambridge Univ. Press,
1995.
[4] S. Coriasco, Fourier Integral Operators in SG Classes I: Composition Theorems and
Action on SG Sobolev Spaces. Rend. Sem. Mat. Univ. Politec. Torino 57, 4 (1999),
249–302 (2002).
[5] S. Coriasco, K. Johansson, J. Toft, Global wave-front sets of Banach, Fréchet and
Modulation space types, and pseudo-differential operators, J. Differential Equations
254 (2013), 3228–3258.
[6] S. Coriasco, M. Ruzhansky, Global Lp -continuity of Fourier Integral Operators.
Trans. Amer. Math. Soc. 366, 5 (2014), 2575–2596.
[7] Y.V. Egorov, B.-W. Schulze. Pseudo-differential operators, singularities, applica-
tions, Operator Theory: Advances and Applications 93 (1997), Birkhäuser Verlag,
Basel.
[8] R. Melrose. Geometric scattering theory. Stanford Lectures. Cambridge University
Press, Cambridge, 1995.
[9] M. Oberguggenberger, M. Schwarz. Fourier Integral Operators in Stochastic Struc-
tural Analysis. Proceedings of the 12th International Probabilistic Workshop, 2014.
[10] C. Parenti, Operatori pseudodifferenziali in Rn e applicazioni, Ann. Mat. Pura Appl.
93 (1972), 359–389.
[11] M. Ruzhansky, M. Sugimoto, Global L2 boundedness theorems for a class of Fourier
integral operators. Comm. Partial Differential Equations 31 (2006), 547–569.
[12] E. Schrohe, Spaces of weighted symbols and weighted Sobolev spaces on manifolds.
In: H.O. Cordes, B. Gramsch, and H. Widom (eds.), Proceedings, Oberwolfach, 1256
Springer LMN, New York, 1986, pp. 360–377.
[13] L. Schwartz. Théorie des Distributions. Hermann, 2nd edition, 2010.
[14] F. Trèves. Topological Vector Spaces, Distributions, and Kernels. Academic Press
(1967).
Alessia Ascanelli and André Süß

Dipartimento di Matematica ed Informatica
Università di Ferrara
Via Machiavelli n. 30
I-44121 Ferrara, Italy
e-mail: [email protected]
[email protected]
Sandro Coriasco
Dipartimento di Matematica “G. Peano”
Università degli Studi di Torino
via Carlo Alberto n. 10
I-10123 Torino, Italy
Operator Theory:

Transport in a Stochastic Goupillaud Medium

Florian Baumgartner, Michael Oberguggenberger
and Martin Schwarz
Abstract. This paper is part of a project that aims at modelling wave prop-
agation in random media by means of Fourier integral operators. A partial
aspect is addressed here, namely explicit models of stochastic, highly irregular
transport speeds in one-dimensional transport, which will form the basis for
more complex models. Starting from the concept of a Goupillaud medium (a
layered medium in which the layer thickness is proportional to the propaga-
tion speed), a class of stochastic assumptions and limiting procedures leads to
characteristic curves that are Lévy processes. Solutions corresponding to dis-
cretely layered media are shown to converge to limits as the time step goes to
zero (almost surely pointwise almost everywhere). This translates into limits
in the Fourier integral operator representations.
1. Introduction
This contribution is part of a long-term project that aims at modelling wave prop-
agation in random media by means of Fourier integral operators. The intended
scope includes, for example, the equilibrium equations in linear elasticity theory
∂ 2 ui ∂ ∂ul
ρ − cijkl = fi , i = 1, 2, 3
∂t2 ∂xj ∂xk
j,k,l
or, more generally, hyperbolic systems of the form

∂u
m
∂u
= Aj (t, x) + B(t, x)u + f (t, x), (1)
∂t j=1
∂xj
to be solved for the unknown functions u(t, x) = (u1 (t, x), . . . , un (t, x)). Here t
denotes time, x is an m-dimensional space variable, and Aj , B are (n×n)-matrices.
Our specific interest is in the situation where the coefficient matrices are
random functions of the space variable x, i.e., random fields. Such a situation
20 F. Baumgartner, M. Oberguggenberger and M. Schwarz
typically arises in seismology (propagation of acoustic waves) or in material sci-

ence (damage detection). There are many ways of setting up models for random
fields (see, e.g., [9, 14]), but typically random fields describing randomly perturbed
media have continuous, but not differentiable paths. As coefficients in hyperbolic
equations such as (1), this degree of regularity is too low and does not allow one
to apply the classical solution theory for hyperbolic equations. In addition, the
solution depends nonlinearly on the coefficients, so it is generally impossible to
directly calculate the stochastic properties of the solution from knowledge of the
distribution of the coefficients.
The main thrust of the project will be to write the solution to equations like
(1) as a sum of Fourier integral operators

1
u(t, x) = e iϕ(t,x,y,η) a(t, x, y, η)u0 (y) dy dη (2)
(2π)m
applied to the initial data u0 . This is possible in the case of deterministic, smooth
coefficients (up to a smooth error). The ultimate goal of the project will be to
set up the stochastic model of the medium through the phase function ϕ and
the amplitude a of the Fourier integral operator, rather than through a direct
stochastic model of the coefficients, as described in [16].
A second thrust is in understanding wave propagation in strongly irregular
stochastic media with a sufficiently simple structure and tractable properties, in
order to get insight into what stochastic processes are suitable to be entered as
phase functions and amplitudes. This brings us to the topic of this paper, namely,
wave propagation in a Goupillaud medium (the name goes back to [10]). In this
contribution, we will work out the case of one-dimensional transport under assump-
tions that will lead to characteristic curves given by an increasing Lévy process
with possibly infinitely many jumps on each subinterval.
One-dimensional transport is described by the equation
∂ ∂
u(t, x) + c(x) u(t, x) = 0,
∂t ∂x (3)
u(0, x) = u0 (x).
The material properties of the medium are encoded in the transport speed c(x).
The Goupillaud assumption is that c(x) is a piecewise constant function so that
the travel time in each layer is the same. That is, the thickness of layer number k
is proportional to the propagation speed ck in that layer.
Further, the propagation speeds ck will be given by independent, identically
distributed random variables. At this stage, various choices of the type of random
variables as well as scalings are possible. For the wave equation, such scalings lead-
ing to fairly regular limiting processes have been introduced in [5] and studied in
[8, 15]. Our procedure of dyadic refinements on the time axis will lead to infinitely
divisible, positive random variables. It turns out that they can be constructed as
increments of a strictly increasing Lévy process. As the time step goes to zero, the
characteristic curve of (3) passing through the origin is a path of a Lévy process.
Transport in a Stochastic Goupillaud Medium 21
We will show that the characteristic curves of the discrete Goupillaud medium
converge (almost surely at almost every (t, x)) to limiting curves (actually trans-
lates of the obtained Lévy process), and that the corresponding solutions and their
Fourier integral operator representations converge as well.
The limiting function u(t, x) is constant along the limiting characteristic
curves, as in the case of classical transport. However, the limiting characteris-
tics may possibly have infinitely many jumps on each interval. Due to this high
degree of singularity, we cannot give a meaning to the limiting function u(t, x) as
a solution to (3); it is just a limit of piecewise classical solutions. This situation
is quite common in the theory of singular stochastic partial differential equations,
see, e.g., [11].
A few remarks about the regularity of the coefficient c(x) in (3) is in or-
der. If the coefficient is Lipschitz continuous, classical solutions can be readily
constructed. If the coefficient is a piecewise constant, positive function, piecewise
classical solutions are obtained easily. In case of lower regularity of the coefficient,
various approaches have been proposed in the literature. We mention the work
of DiPerna and Lions [6], Bouchut and James [4], and Ambrosio et al. [1, 2] in
the deterministic, x-dependent case; for a recent survey, see Haller and Hörmann
[12]. In the stochastic case, recent work of Flandoli [7] shows how solutions can
be constructed adding noise in the transport term. Finally, another different line
of development is constituted by extending the reservoir of generalized functions,
either in the direction of white noise analysis or in the direction of Colombeau
theory. For a representative article pursuing and comparing both approaches, see
Pilipović and Seleši [17, 18].
The plan of the paper is as follows: In the first part, the stochastic Goupillaud
medium is set up and analyzed. In the second part, the limiting behavior as the
time step goes to zero is established. The paper ends with some conclusions and
open questions.
2. Setting up the Goupillaud medium

If the initial data u0 is differentiable and the propagation speed c is Lipschitz
continuous, classical solutions to the transport equation (3) can be readily obtained
by the method of characteristics. The characteristic curves are the integral curves
of the vector field ∂/∂t + c(x)∂/∂x passing through the point x at time t, that is,
the solutions to the ordinary differential equation
d
γ(τ ; t, x) = c(γ(τ ; t, x)), γ(t; t, x) = x.
dτ
Then the solution to (3) is given by
u(t, x) = u0 (γ(0; t, x)).
Under the mentioned assumptions, the function u is continuously differentiable,
and the solution is unique in this class.
If the speed parameter c is constant the characteristic curves are simply given
by γ(τ ; t, x) = x + c(τ − t). If the parameter is piecewise constant one can compute
the characteristic curves as polygons. Assuming continuity across interfaces, the
solution u is given as a continuous, piecewise differentiable function, which solves
(3) in the weak sense.
2.1. Dyadic deterministic structure

We begin by setting up the discrete, deterministic Goupillaud medium. Take an
equidistant sequence (tj )j∈N of points of time with t0 = 0 and time step Δt ≡
tj − tj−1 for all j = 1, 2, . . .. Furthermore, take a strictly increasing sequence
(xk )k∈Z with x0 = 0 and xk → ±∞ as k → ±∞ and let Δxk = xk − xk−1 . The
coefficient c(x) is defined as
∞
Δxk
c(x) = 1[xk−1 ,xk ) (x), (4)
Δt
k=−∞
where 1A denotes the characteristic function of the set A. In other words, the time
for passing a layer Δxk is constant, namely Δt. For an illustration see Figure 1.
Figure 1. Illustration of τ → γ(τ ; 0, 0) in the Goupillaud medium

(left). Refining the grid on the right preserves existing grid points
γ (1) (·; 0, 0) (dashed) and refinement γ (2) (·; 0, 0).
Call ck the value of c(x) in the kth layer, that is, xk−1 ≤ x < xk . Then the
Goupillaud relation
Δxk = ck Δt (5)
holds for all k, with constant Δt. The structure of the Goupillaud medium makes
computing the values of the characteristic curves γ(τ ; t, x) in the grid points very
simple. In fact,
γ (tj ; tl , xk ) = xj+k−l (6)
for all integers j, k, l. Since every point (t, x) is just a convex combination of the
neighboring grid points, the values γ(τ ; t, x) can be easily obtained anywhere.
We now set up a dyadic refinement of the initial grid. Define
j
Δt(N ) = 2−N , tj = N
(N )
2
(N )
and let xk ∈ R, k ∈ Z, be a strictly increasing sequence of spatial points (or
(N ) (N ) (N )
equivalently, propagation speeds ck > 0 satisfying Δxk = ck Δt(N ) ). We
require that each resulting grid is a dyadic refinement of the previous one, that is
(N +1) (N +1) (N ) (N )
t2j , x2k = tj , xk , (7)
as illustrated in Figure 1. This condition implies
(N ) (N +1) (N +1)
Δxk = Δx2k−1 + Δx2k .
Inductively, one obtains
M
(N )

2
(N +M)
Δxk = Δx(k−1)2M +i (8)
i=1
for all N, M ∈ N, k ∈ Z. The value of the characteristic curve γ (N ) in the grid

points is readily obtained according to (6). For any integers N, j, k, l we have

(N ) (N ) (N ) (N )
γ (N ) tj ; tl , xk = xj+k−l . (9)
(N ) (N )
For any N ∈ N and τ ∈ [tk−1 , tk ), the characteristic curve through the origin
ξ (N ) (τ ) = γ (N ) (τ ; 0, 0) can be represented as
(N ) (N )
ξ (N ) (τ ) = α(N ) (τ )ξ (N ) (tk−1 ) + 1 − α(N ) (τ ) ξ (N ) (tk ), (10)
where
(N )
tk −τ (N )
α(N ) (τ ) = (N ) (N )
= tk − τ 2N ∈ [0, 1]
tk − tk−1
(N ) (N ) (N )
and ξ (tk )
= xk by (9).
(N ) (N )
In other words, ξ (N )
is an increasing polygon through (tk , xk ), k ∈ Z. For
(t, x) ∈ R one obtains the characteristic curve through (t, x) by
2

γ (N ) (τ ; t, x) = ξ (N ) τ + (ξ (N ) )−1 (x) − t , τ ∈ R, (11)
i.e., by shifting ξ (N ) in time direction such that it passes through (t, x).
2.2. The stochastic model
In this subsection, we formulate the stochastic assumptions underlying our model
(N )
of a randomly layered medium, in which Δxk is random. In the sequel, we will
denote random elements by capital letters and realizations by the corresponding
small ones. Let (Ω, F , P) be a probability space which is rich enough. Our decisive
assumption is that for each N ∈ N, the increments are positive, independent and
(N )
identically distributed random variables ΔXk , k ∈ Z.
Together with our previous consistency assumption (8), this implies that
(N )
ΔXk is infinitely divisible for every k ∈ Z and N ∈ N (using, e.g., [13, Thm.
15.12]). Let μ = P ◦ (ΔX1 )−1 be the distribution of ΔX1 . Then ΔXk
(0) (0) (N )
∼
μ∗1/2 , k ∈ N, the 2N th unique root of μ, cf. [19, p. 34]. Again by [13, Thm.
N
15.12], there exists a Lévy process X = (X(t))t∈R on a probability space, w.l.o.g.

say (Ω, F , P), with P ◦ X(1)−1 = μ, that is, ΔX1 has the same distribution as
(0)
X(1). These conditions are met, e.g., by Poisson processes or Gamma processes
with positive drift.
Having derived the Lévy process, we may use it as a starting point for defining
(N )
the stochastic Goupillaud medium. We let tk = k/2N as in Subsection 2.1 and
define
(N ) (N ) (N )
ΔXk = X tk − X tk−1
and
(N )

k
(N ) (N )
Xk = ΔXi = X tk
i=1
for k > 0 and similarly for k ≤ 0. The consistency condition (8) is clearly satisfied.
Let furthermore X (N ) (ω, ·) be the piecewise affine interpolation of X(ω, ·) through
(N ) (N )
the grid points (tk , Xk ) as in (10). This construction is carried out pathwise
for fixed ω ∈ Ω.
3. Limits as the time step goes to zero

The main result of this section is that the characteristic curves of the discrete
Goupillaud medium converge to limiting curves (almost surely almost everywhere).
This will imply that the solutions to the transport equation converge to a limit as
well (in a sense to be made precise). The crucial observation is that the paths of
a Lévy process are càdlàg (continue à droite, limite à gauche) almost surely, i.e.,
they are continuous from the right and have left-hand limits.
3.1. A convergence result for càdlàg functions
The first convergence result holds generally for càdlàg functions. Thus let t → ξ(t)
be an increasing càdlàg function with ξ(t) → ±∞ for t → ±∞ and let (t, x) ∈ R2 .
Set
ξ ∗ (x) = inf {t ∈ R : ξ(t) ≥ x}
which is Borel measurable, and
γ(τ ; t, x) = ξ(τ + ξ ∗ (x) − t), τ ∈ R.
(N )
Further, let ξ be a piecewise linear interpolation of ξ that coincides with ξ at
(N )
the grid points tk = k/2N , k ≥ 0, and define γ (N ) (τ ; t, x) by formula (11).
Lemma 1. Let (t, x) ∈ R2 , ξ, ξ (N ) , γ, γ (N ) as described above. If the function
τ → γ(τ ; t, x) does not have a jump in τ0 , then
lim γ (N ) (τ0 ; t, x) = γ(τ0 ; t, x),
N →∞
i.e., γ (N ) ( · ; t, x) converges pointwise to γ( · ; t, x) at the points of continuity of

γ( · ; t, x).
Proof. Fix ε > 0 and R > |τ0 | + |ξ ∗ (x)| + |t| and define
s := τ0 + ξ ∗ (x) − t,
s(N ) := τ0 + (ξ (N ) )−1 (x) − t.
As ξ is càdlàg there exist finitely many (t1 , . . . , t ) ∈ [−R, R] such that
ε
∀r1 , r2 ∈ [ti , ti+1 ) : |ξ(r1 ) − ξ(r2 )| < , (12)
3
see, e.g., [3, Lemma 1, p. 110]. Since γ is continuous in τ0 , we can assume without
loss of generality that s = ti for all i.
Since ξ and ξ (N ) coincide at the grid points and both are increasing,
(ξ ) (x) and ξ ∗ (x) belong to the same interval of length 1/2N , for every N . It
(N ) −1
follows that |s(N ) − s| < 1/2N as well. We can choose N large enough, such that
both s and s(N ) belong to (ti + 1/2N , ti+1 − 1/2N ) for some i ∈ {1, . . . , − 1}.
From (12) we get that
ε
(N )
γ (τ0 ; t, x) − γ(τ0 ; t, x) = ξ (N ) (s(N ) ) − ξ(s) ≤ ξ (N ) (s(N ) ) − ξ(s(N ) ) + .
3
Now choose k such that
(N ) (N )
ti < tk−1 ≤ s(N ) ≤ tk < ti+1 .
(N )
Using (10) and the fact that ξ and ξ (N ) coincide at all grid points tj we can
write
(N ) (N )
ξ (N ) (s(N ) ) − ξ(s(N ) ) = α(N ) (s(N ) )ξ(tk−1 ) + 1 − α(N ) (s(N ) ) ξ(tk ) − ξ(s(N ) )
(N ) (N ) (N )
= 1 − α(N ) (s(N ) ) ξ(tk ) − ξ(tk−1 ) + ξ(tk−1 ) − ξ(s(N ) .
Recalling α(N ) ∈ [0, 1] for all N and invoking again (12), we conclude that

(N )
γ (τ0 ; t, x) − γ(τ0 ; t, x) < ε,
which implies the desired convergence.
Denote by T = (Ti )i∈N the (countable) set of jump points of the càdlàg
function ξ. At fixed (t, x), convergence may fail at those values τ for which τ +
ξ ∗ (x) − t ∈ T . This exceptional set is countable, but may be different for every
(t, x). Next, we fix τ0 and determine a set of all (t, x) for which convergence fails.
We are going to show that its two-dimensional Lebesgue measure λ is zero.
Lemma 2. Let τ0 ≥ 0; ξ, γ as in Lemma 1. The set
M = {(t, x) : γ(τ ; t, x) jumps at τ0 }
has Lebesgue measure zero.
Proof. Letting

M = {(t, x) : τ0 + ξ ∗ (x) − t ∈ T } = Mi
i∈N
where Mi = {(t, x) : τ0 + ξ ∗ (x) − t = Ti }, it suffices to check that each Mi has

Lebesgue measure zero. But each Mi is jointly measurable, and for each x the set
Mi (x) = {t : τ0 + ξ ∗ (x) − t = Ti } is a singleton. Hence

λ(Mi ) = λ(Mi (x)) dx = 0
R
by Fubini’s theorem.
3.2. Convergence of characteristic curves
We now apply Lemma 1 to a path t → X(ω, t) = ξ(t) of the Lévy process X
constructed in Subsection 2.2. Since Lévy processes are càdlàg almost surely, there
is Ω0 ∈ F with P(Ω0 ) = 1 such that X(ω, ·) is càdlàg for all ω ∈ Ω0 . With the
notation of Subsections 2.1 and 2.2, let

Γ(N ) (ω; τ ; t, x) = X (N ) ω; τ + (X (N ) (ω))−1 (x) − t
and

Γ(ω; τ ; t, x) = X ω; τ + X(ω)∗ (x) − t .
Proposition 1. Let ω ∈ Ω0 .
(1) For a.e. (t, x) ∈ R2 it holds that
Γ(N ) (ω; 0; t, x) → Γ(ω; 0; t, x) as N → ∞.
(2) Let u0 be bounded and continuous, 1 ≤ p < ∞ and K ⊆ R2 compact. Then

(N ) p
u0 Γ (ω; 0; t, x) − u0 Γ(ω; 0; t, x) dλ(t, x) → 0 (13)
K
as N → ∞.
(3) Γ(N ) (·; 0; ·, ·) → Γ(·; 0; ·, ·) as N → ∞ with convergence P ⊗ λ-a.e.
(4) Let u0 be bounded and continuous, 1 ≤ p < ∞ and K ⊆ R2 compact. Then

p
u0 Γ(N ) (ω; 0; t, x) − u0 Γ(ω; 0; t, x) d(P ⊗ λ)(ω, t, x) → 0 (14)
Ω×K
as N → ∞.

Proof. First we notice that Γ(ω; τ ; t, x) = X ω; τ + X(ω)∗ (x) − t is jointly mea-
surable in all variables. This follows by measurability of the mapping R × D →
R, (t, ξ) → ξ(t), cf. [19, p. 132], where D is the space of càdlàg functions endowed
with the σ-algebra generated by the coordinate mappings ξ → ξ(t), t ∈ R.
(1) follows from Lemma 2, and (2) follows from Lebesgue’s convergence the-
orem.
(3) We first convince ourselves that the exceptional set of those (ω, t, x) at
which Γ(ω; τ ; t, x) jumps at τ = 0 is jointly measurable, that is

M = (ω,t,x) ∈ Ω0 × R2 : Γ(ω;0;t,x) − Γ(ω;0−;t,x) = 0

∗
∗

= (ω,t,x) ∈ Ω0 × R : X ω;X(ω) (x) − t = lim X ω;X(ω) (x) − t + τ
2
τ →0−

is F ⊗ B(R2 )-measurable.
It follows from the joint measurability of X ω; τ +
∗
X(ω) (x) − t that the function d : Ω0 × R → R,
2

d(ω, t, x) = X ω; X(ω)∗ (x) − t − lim X ω; X(ω)∗ (x) − t + τ
τ →0−
is measurable and therefore M = d−1 (R \ {0}) ∈ F ⊗ B(R2 ).

The fact that M has measure zero, hence (3), is a consequence of Fubini’s
theorem. Indeed, for ω ∈ Ω0 , let M (ω) = M ∩ ({ω} × R2 ). By Lemma 2 applied
that M (ω) has Lebesgue measure zero. Also, P(Ω0 ) = 1.
with τ0 = 0 it follows
Thus P ⊗ λ(M) = Ω λ(M (ω)) dP(ω) = 0.
(4) follows immediately from (2) and (3).
3.3. Convergence of approximate solutions
We return to the transport equation in the discrete stochastic Goupillaud medium
∂ (N ) ∂
U (ω; t, x) + C (N ) (ω; x) U (N ) (ω; t, x) = 0
∂t ∂x (15)
U (N ) (0, x) = u0 (x)
with
∞ (N )
ΔXk (ω)
C (N ) (ω; x) = 1[X (N ) (ω),X (N ) (ω)) (x) (16)
k=−∞
Δt(N ) k−1 k
(N )
where ΔXk is derived from the Lévy process X as in Subsection 2.2. To be
precise about the solution concept, assume that u0 belongs to the Sobolev space
1,1
Wloc (R). Note that this implies that u0 is a continuous function. At fixed ω, the
transport coefficient C (N ) (ω; ·) is a piecewise constant, locally bounded function,
and the characteristic curves Γ(N ) (ω; τ ; t, x) are piecewise linear, continuous func-
tions. We put
U (N ) (ω; t, x) = u0 Γ(N ) (ω; 0; t, x) .
1,1
It is straightforward to check that U (N ) (ω; ·, ·) belongs to Wloc (R2 ) and is con-
1,1 2
tinuous. Taking weak derivatives in the sense of Wloc (R ) and performing the
multiplication with the L∞loc -function C
(N )
(ω; ·) in L1loc (R2 ) shows that U (N ) (ω; ·, ·)
satisfies the equation (15) in the sense of the latter space. Further, the initial data
are taken as continuous functions. In this sense, U (N ) (ω; ·, ·) is a pathwise solution
to (15). Define
U (ω; t, x) = u0 Γ(ω; 0; t, x) .
With the results from Subsection 3.2 we are now in the position to formulate
convergence of the approximate solutions U (N ) to U .
1,1
Proposition 2. Let u0 ∈ Wloc (R). Then
(1) limN →∞ U (N ) (ω; t, x) = U (ω; t, x) pointwise P ⊗ λ-a.e. and
lim EP U (N ) − U Lp (K) =0
N →∞
whenever K is a compact subset of R2 and 1 ≤ p < ∞.

(2) If the Fourier transform of u0 belongs to L1 (R), then U has the Fourier
integral operator representation

1
U (ω; t, x) = e i(Γ(ω;0;t,x)−y)η u0 (y) dy dη.
2π
Proof. (1) is evident from Proposition 1. Concerning (2), observe that

(N )
(N ) 1 (N )
U (ω; t, x) = u0 Γ (ω; 0; t, x) = e i(Γ (ω;0;t,x)−y)η u0 (y) dy dη
2π
as shown by taking Fourier transforms, where the double integral converges as an
iterated integral. Proposition 1 allows us to take the limit as N → ∞ inside the
integral, whence the assertion follows.
Note that a priori there is no meaning for u to be a solution of the transport
equation (3) other than being a limit of approximate solutions.
For the sake of illustration, we show two realizations of the limiting solutions.
The initial value u0 is taken as a triangular function, the realizations of U are
shown at times t = 1, 2, 3. We use two different Lévy processes as drivers X (cf.
Subsection 2.2). In the first picture in Figure 2, X is taken as a Gamma process, in
the second picture, X is a Poisson process, both with positive drift. The solutions
have constant parts, which are created if the Lévy process jumps at this point.
u0 (x)
1
U t=1

U t=2

U
0.5 t=3
x
−1 1 2 3 4 5 6 7 8 9
u0 (x)
1
U t=1

U t=2

U
0.5 t=3
x
−1 1 2 3 4 5 6 7 8 9
Figure 2. A trajectory of the solution U (x, t) at several time points.

Above: The generating Lévy process X is Gamma distributed with scale
parameter k = 1, shape parameter θ = 1 and drift d = 1. Below:
The generating Lévy process X is Poisson distributed with jump size 1,
intensity c = 1 and drift d = 1.
4. Conclusion
A Goupillaud medium is a piecewise constant layered medium such that the thick-
ness of each layer is proportional to the corresponding propagation speed. We
have developed a set-up for a specific stochastic Goupillaud medium in which
the propagation speeds (or equivalently the thickness of the layers) are given by
infinitely divisible random variables. Using a dyadic refinement, these random vari-
ables could be constructed as increments of a strictly increasing Lévy process. We
have shown that the one-dimensional transport equation can be solved in such
a medium, and that the characteristic curves converge to shifted trajectories of
the underlying Lévy process as the time step goes to zero. If the initial data are
sufficiently regular, the corresponding solutions converge pathwise and in the pth
mean to a limiting function, which in addition can be computed by means of a
Fourier integral operator.
At this stage, several questions remain open. The first issueis the probability
distribution
of the limiting characteristic curves Γ(ω; τ ; t, x) = X ω; τ +X(ω)∗ (x)−
t , and subsequently of the limiting solution U (ω; t, x) = u0 Γ(ω; 0; t, x) . The
second question is how one can give a meaning to the limiting propagation speed
c(x) as a (generalized) function of x. Given a positive answer to this question, one
may finally ask if there is a solution concept that would allow one to interpret
U (ω; t, x) as a solution in some sense. All these issues are the subject of ongoing
research.
Acknowledgement
The second author acknowledges support through the research project P-27570-
N26 “Stochastic generalized Fourier integral operators” of FWF (The Austrian
Science Fund). The third author acknowledges support through the Bridge Project
No. 846038 “Fourier Integral Operators in Stochastic Structural Analysis” of FFG
(The Austrian Research Promotion Agency).
References
[1] Ambrosio, L.: Transport equation and Cauchy problem for BV vector fields. Invent.
Math. 158(2), 227–260 (2004)
[2] Ambrosio, L.: Transport equation and Cauchy problem for non-smooth vector fields.
In: L. Ambrosio, L. Cafarelli, M.G. Crandall, L.C. Evans, N. Fusco (eds.) Calculus of
variations and nonlinear partial differential equations, Lecture Notes in Math., vol.
1927, pp. 1–41. Springer, Berlin (2008)
[3] Billingsley, P.: Convergence of probability measures, 2 edn. John Wiley & Sons, New
York (1999)
[4] Bouchut, F., James, F.: One-dimensional transport equations with discontinuous
coefficients. Nonlinear Anal. 32(7), 891–933 (1998)
[5] Burridge, R., Papanicolaou, G.S., White, B.S.: One-dimensional wave propagation
in a highly discontinuous medium. Wave Motion 10(1), 19–44 (1988)
[6] DiPerna, R.J., Lions, P.L.: Ordinary differential equations, transport theory and
Sobolev spaces. Invent. Math. 98(3), 511–547 (1989)
[7] Flandoli, F.: Random perturbation of PDEs and fluid dynamic models, Lecture Notes
in Mathematics, vol. 2015. Springer, Heidelberg (2011). Lectures from the 40th Prob-
ability Summer School held in Saint-Flour, 2010
[8] Fouque, J.P., Garnier, J., Papanicolaou, G., Sølna, K.: Wave propagation and time
reversal in randomly layered media, Stochastic Modelling and Applied Probability,
vol. 56. Springer, New York (2007)
[9] Ghanem, R.G., Spanos, P.D.: Stochastic finite elements: a spectral approach.
Springer-Verlag, New York (1991)
[10] Goupillaud, P.L.: An approach to inverse filtering of near-surface layer effects from
seismic records. Geophysics 26, 754–760 (1961)
[11] Hairer, M.: A theory of regularity structures. Invent. Math. 198(2), 269–504 (2014)
[12] Haller, S., Hörmann, G.: Comparison of some solution concepts for linear first-order
hyperbolic differential equations with non-smooth coefficients. Publ. Inst. Math.
(Beograd) (N.S.) 84(98), 123–157 (2008)
[13] Kallenberg, O.: Foundations of modern probability, second edn. Probability and its
Applications (New York). Springer-Verlag, New York (2002)
[14] Matthies, H.G.: Stochastic finite elements: computational approaches to stochastic
partial differential equations. Z. Angew. Math. Mech. 88(11), 849–873 (2008)
[15] Nair, B., White, B.S.: High-frequency wave propagation in random media – a unified
approach. SIAM J. Appl. Math. 51(2), 374–411 (1991)
[16] Oberguggenberger, M., Schwarz, M.: Fourier integral operators in stochastic struc-
tural analysis. In: F. Werner, M. Huber, T. Lahmer, T. Most, D. Proske (eds.)
Proceedings of the 12th International Probabilistic Workshop, Schriftenreihe des
DFG Graduiertenkollegs 1462 Modellqualitäten, vol. 11, pp. 250–257. Bauhaus-
Universitätsverlag, Weimar (2014)
[17] Pilipović, S., Seleši, D.: On the generalized stochastic Dirichlet problem – Part II:
solvability, stability and the Colombeau case. Potential Anal. 33(3), 263–289 (2010)
[18] Pilipović, S., Seleši, D.: On the generalized stochastic Dirichlet problem. I. The
stochastic weak maximum principle. Potential Anal. 32(4), 363–387 (2010)
[19] Sato, K.: Lévy Processes and Infinitely Divisible Distributions. Cambridge University
Press, Cambridge (1999)
Florian Baumgartner Michael Oberguggenberger

Institute of Mathematics and Martin Schwarz
University of Innsbruck Unit of Engineering Mathematics
Technikerstraße 13 University of Innsbruck
A-6020 Innsbruck, Austria Technikerstraße 13
e-mail: A-6020 Innsbruck, Austria
[email protected] e-mails:
[email protected]
[email protected]
Operator Theory:

Hilbert Space Embeddings for Gelfand–Shilov

and Pilipović Spaces
Yuanyuan Chen, Mikael Signahl and Joachim Toft
Abstract. We consider quasi-Banach spaces that lie between a Gelfand–Shilov

space, or more generally, Pilipović space, H, and its dual, H . We prove that for
such quasi-Banach space B, there are convenient Hilbert spaces, Hk , k = 1, 2,
with normalized Hermite functions as orthonormal bases and such that B lies
between H1 and H2 , and the latter spaces lie between H and H .
0. Introduction
A common issue consists in replacing complicated topological spaces with topo-
logical spaces having less complicated structures. For example, when investigating
local properties of solutions to differential equations, it might be convenient to use
the Sobolev spaces Hs2 , instead of C0∞ and E , the spaces of compactly supported
smooth functions, and of compactly supported distributions, respectively. Due to
the identities

C0∞ = (Hs2 E ) and E = (Hs2 E ),
s∈R s∈R
it follows that several deduced properties in the framework of Sobolev spaces, carry
over to similar properties in the framework of compactly supported functions and
distributions. If instead global properties of solutions are in the spotlight, then the
2
Kato–Sobolev spaces Hs,t might be useful (see, e.g., [7]). Due to the relations

S = 2
Hs,t and S = 2
Hs,t ,
s,t∈R s,t∈R
it again follows that several properties, deduced in the framework of the Hilbert
2
spaces Hs,t , can be carried over to the frameworks of S and S , the sets of
Schwartz functions and tempered distributions, respectively.
In other problems it might be more suitable to use Hilbert spaces of Hermite
series expansions, using the fact that S and S are the intersection and union,
32 Y. Chen, M. Signahl and J. Toft
respectively, of Hilbert spaces

{ f ; {cα (f )αr }α∈Nd ∈ 2 },
parameterized by r > 0 (r < 0). Here cα (f ) = cα is the Hermite coefficient of
order α ∈ Nd in the expansion

f= cα h α . (0.1)
α
In [16], a different situation is considered, where S and S above are replaced

by Fourier invariant Gelfand–Shilov spaces and their duals. In view of [5, 13], these
spaces may in convenient ways be described by Hermite series expansions with
suitable conditions on the Hermite coefficients. More precisely, assume that s > 0
is real. Let Hs (Rd ) (H0,s (Rd )) be the set of all formal Hermite series expansions
in (0.1) such that
1
|cα | e−c|α| 2s for some (every) c > 0,
and Hs (Rd )
(H0,s (Rd )) the set of formal expansions (0.1) such that
1
|cα | ec|α| 2s for every (some) c > 0.
By [5, 13], it follows that Hs and Hs (H0,s and H0,s
) agree with the Roumeu
(Beurling) type Gelfand–Shilov spaces of functions and distributions, respectively
when s ≥ 12 (s > 12 ).
We remark that Hs and H0,s are often called Pilipović spaces (see Section 1
for details and motivations).
For any quasi-Banach space B which lies between a Gelfand–Shilov space and
its dual it is proven in [16] that there are Hilbert spaces Hj , which lie between
the Gelfand–Shilov space and its dual, such that
H1 ⊆ B ⊆ H2 .
More precisely, in [16, Proposition 3.8] it is proved that if
Hs (Rd ) ⊆ B ⊆ Hs (Rd ),
for some s ≥ 12 , then there are Hilbert spaces H1 and H2 such that
Hs (Rd ) ⊆ H1 ⊆ B ⊆ H2 ⊆ Hs (Rd ), (0.2)
and that the same holds true with H0,s and its dual in place of Hs and its dual.
In [16], these embedding results are used to extend Schatten–von Neumann prop-
erties for certain operators acting between the Hilbert spaces in (0.2) into similar
properties when these Hilbert spaces are replaced by suitable quasi-Banach spaces.
In Section 2 we extend these properties to the case s ≥ 0.
We also remark that in [13], Pilipović shows that the Fourier invariant Gel-
fand–Shilov spaces can be characterized by suitable estimates on the involved
functions, after powers of the harmonic oscillator have been applied. (See Section
1 for more details.) Here Pilipović also gives an example of a non-trivial and
Fourier invariant space, defined by such kind of estimates and which is smaller
Hilbert Space Embeddings for Gelfand–Shilov and Pilipović Spaces 33
than any non-trivial and Fourier invariant Gelfand–Shilov space. For these reasons
we call spaces deduced by estimates involving powers of the harmonic oscillator
as Pilipović spaces. In [17] it is proved that Hs and H0,s can be described by such
estimates. Hence they are examples on Pilipović spaces.
1. Preliminaries
In this section we recall some basic facts. We start by discussing Pilipović spaces
and their properties. Thereafter we consider suitable spaces of formal Hermite
series expansions, and discuss their links with Pilipović spaces.
1.1. The Pilipović spaces
We start to consider spaces which are obtained by suitable estimates of Gelfand–
Shilov or Gevrey type when using powers of the harmonic oscillator H = |x|2 − Δ,
x ∈ Rd .
Consider h > 0, s ≥ 0 and let Sh,s (Rd ) be the set of all f ∈ C ∞ (Rd ) such
that
H N f L∞
f Sh,s ≡ sup N 2s
< ∞. (1.1)
N ≥0 h (N !)
We set

Σs (Rd ) ≡ Sh,s (Rd ) and Ss (Rd ) ≡ Sh,s (Rd ),
h>0 h>0
and equip these spaces by projective and inductive limit topologies, respectively,
of Sh,s (Rd ), h > 0. (Cf. [5, 12, 13, 17].)
In [12, 13], Pilipović proved that if s1 ≥ 12 and s2 > 12 , then Ss1 (Rd ) and
Σs2 (Rd ) agree with the Gelfand–Shilov spaces Ss1 (Rd ) and Σs2 (Rd )1, respectively,
but Σ 12 (Rd ) = Σ 12 (Rd ) = {0}. (See, e.g., [17] for notations.) In [17], the latter
relations were extended into
1 1
Ss1 = Ss1 , Σs2 = Σs2 , s1 ≥ , s2 >
2 2
and
1 1
Ss1 = Ss1 = {0}, Σs2 = Σs2 = {0}, s1 < , 0 < s2 ≤ .
2 2
The space Σs (R ) is called the Pilipović space (of Beurling type) of order s ≥ 0
d
on Rd . Similarly, Ss (Rd ) is called the Pilipović space (of Roumieu type) of order
s ≥ 0 on Rd .

The dual spaces of Sh,s (Rd ), Σs (Rd ) and Ss (Rd ) are denoted by Sh,s (Rd ),

Σs (R ) and Ss (R ), respectively. We have
d d

Σs (Rd ) =
Sh,s (Rd )
h>0
1 Note that Pilipović spaces are denoted by boldface characters, S and Σ, while related Gelfand–
Shilov spaces, are denoted by non-boldface characters, S and Σ, respectively.

when s > 0 and

Ss (Rd ) =
Sh,s (Rd )
h>0

when s ≥ 0, with inductive respective projective limit topologies of Sh,s (Rd ),
h > 0 (cf. [17]).
1.2. Spaces of Hermite series expansions
Next we recall the definitions of topological vector spaces of Hermite series expan-
sions, given in [17]. As in [17], it is convenient to use the sets R and R when
indexing our spaces.
Definition 1.1. The sets R and R are given by

R = R+ {σ } and R = R {0}.
σ>0
Moreover, beside the usual ordering in R, the elements σ in R and R are

ordered by the relations x1 < σ1 < σ2 < x2 , when σ1 < σ2 , x1 < 12 and x2 ≥ 12
are real.
Definition 1.2. Let p ∈ (0, ∞], s ∈ R , r ∈ R, ϑ be a weight on Nd , and let
⎧
⎨er|α| 2s
1
, when s ∈ R+ ,
ϑr,s (α) ≡
⎩r|α| (α!) 2σ
1
, when s = , α ∈ Nd .
σ
Then,
1. 0 (Nd ) is the set of all sequences {cα }α∈Nd ⊆ C on Nd ;
2. 0,0 (Nd ) ≡ {0}, and 0 (Nd ) is the set of all sequences {cα }α∈Nd ⊆ C such
that cα = 0 for at most finitely many α;
3. p[ϑ] (Nd ) is the quasi-Banach space which consists of all sequences {cα }α∈Nd ⊆
C such that
{cα }α∈Nd p[ϑ] ≡ {cα ϑ(α)}α∈Nd p
is finite;
p p
4. 0,s (Nd ) ≡ [ϑr,s ] (Nd ) and s (Nd ) ≡ [ϑr,s ] (Nd ), with projective re-
r>0 r>0
spective inductive limit topologies of p[ϑr,s ] (Nd ) with respect to r > 0;
p p
5. 0,s (Nd ) ≡ [1/ϑr,s ] (Nd ) and s (Nd ) ≡ [1/ϑr,s ] (Nd ), with inductive
r>0 r>0
respective projective limit topologies of p[1/ϑr,s ] (Nd ) with respect to r > 0.
Next we introduce spaces of formal Hermite series expansions

f= cα hα , {cα }α∈Nd ∈ 0 (Nd ), (1.2)
α∈Nd
which correspond to
0,s (Nd ), s (Nd ), p[ϑ] (Nd ), s (Nd ) and 0,s (Nd ). (1.3)
Here and in what follows we let hα , the Hermite function of order α ∈ Nd ,

be defined by
1 |x|2 2
hα (x) = π − 4 (−1)|α| (2|α| α!)− 2 e (∂ α e−|x| ).
d
2
For that reason we consider the mappings

T : {cα }α∈Nd → cα h α (1.4)
α∈Nd
between sequences and formal Hermite series expansions.
Definition 1.3. Let p ∈ (0, ∞], ϑ be a weight on Nd , and let s ∈ R .

• the spaces
H0,s (Rd ), Hs (Rd ), p
H[ϑ] (Rd ), Hs (Rd ) and H0,s

(Rd ) (1.5)
are the images of T in (1.4) under corresponding spaces in (1.3). Furthermore,

the topologies of the spaces in (1.5) are inherited from corresponding spaces
in (1.3).
• the quasi-norm f Hp[ϑ] of f ∈ H0 (Rd ), is given by {cα }α∈Nd p[ϑ] , when f
is given by (1.2).
We remark that the spaces Hs , Hs , H0,s , and H0,s

are independent of the
p
particular choice of p in the spaces.
By the definitions it follows that the inclusions
H0 (Rd ) ⊆ H0,s (Rd ) ⊆ Hs (Rd ) ⊆ H0,t (Rd )
⊆ S (Rd ) ⊆ S (Rd ) ⊆ H0,t

(Rd ) ⊆ Hs (Rd )

⊆ H0,s (Rd ) ⊆ H0 (Rd ), when s, t ∈ R , s < t (1.6)
hold true.
The next result shows that the spaces in Definition 1.3 for s, t ∈ R+ agrees
with Pilipović spaces. We refer to [17] for the proof.
Proposition 1.4. Let 0 ≤ s ∈ R. Then H0,s (Rd ) = Σs (Rd ) and Hs (Rd ) = S s (Rd ).
2. Embedding properties for quasi-Banach spaces contained in

Pilipović distribution spaces
In this section we deduce embedding properties for quasi-Banach spaces which
contain H0 (Rd ) and are continuously embedded in H0 (Rd ). For such spaces we
show that there are convenient Hilbert spaces of Hermite series expansions which
contains, or is contained in the given quasi-Banach space.
We start by introducing some notations. A quasi-norm · B on a vector

space B (over C) is a non-negative real-valued function · B on B which fulfills
f B =0 ⇐⇒ f = 0, f ∈ B,
αf B = |α| · f B, f ∈ B, α ∈ C (2.1)
and f +g B ≤ D( f B + g B ), f, g ∈ B,
for some constant D ≥ 1 which is independent of f, g ∈ B. The vector space B is
a quasi-Banach space when the topology for B is defined by a quasi-norm, · B ,
and complete under this topology.
Let B be a quasi-Banach space such that
H0 (Rd ) → B → H0 (Rd ). (2.2)
The L2 -dual B of B is the set of all ϕ ∈ H0 (Rd ) such that
ϕ B ≡ sup |(ϕ, f )L2 (Rd ) |
is finite. Here the supremum is taken over all f ∈ H0 (Rd ) such that f B ≤ 1.
Evidently, if B0 is the completion of H0 (Rd ) under · B , then the map (f, ϕ) →
(ϕ, f )L2 from H0 (Rd ) × B to C extends uniquely to a continuous mapping from
B0 × B to C. Furthermore, if B is a Banach space, then Hahn–Banach’s theorem
implies that the latter map is extendable to a continuous mapping from B × B
to C.
The following definition is analogous to Definition 3.1 in [16].
Definition 2.1. Let B be a quasi-Banach space such that (2.2) is fulfilled, and let
s ∈ R .
(1) B is called Pilipović–Beurling tempered, or PB-tempered (of order s) on Rd ,
if B, B → H0,s

(Rd );
(2) B is called Pilipović–Roumieu tempered, or PR-tempered (of order s) on Rd ,
if B, B → Hs (Rd ).
A Hilbert space H is called Hermite type, if {hα / hα H }α is an orthonormal
basis for H ,
Sπ f (x) ∈ H when f ∈ H
for every permutation π on {1, . . . , d}, and that Sπ f H = f H for every f ∈
H . Here and in what follows we let Sπ for any permutation π on {1, . . . , d} be
the operator on H0 (Rd ), defined by the formula
(Sπ f )(x) = f (y), yj = xπ(j) ,
for every j.
Theorem 2.2. Let s ∈ R , and let B1 , B2 be quasi-Banach spaces which are con-
tinuously embedded in H0 (Rd ). Then the following is true:
(1) if Hs (Rd ) → B1 and B2 → Hs (Rd ), then there are PR-tempered Hilbert
spaces H1 and H2 of order s and of Hermite type such that
Hs (Rd ) → H1 → B1 , B2 → H2 → Hs (Rd ),

H1 → H0,t1 (Rd ) and H0,t1
(Rd ) → H2 (2.3)
hold true for every t1 ∈ R such that t1 > s;

(2) if H0,s (Rd ) → B1 and B2 → H0,s (Rd ), then there are PB-tempered Hilbert
spaces H1 and H2 of order s and of Hermite type such that

H0,s (Rd ) → H1 → B1 , B2 → H2 → H0,s (Rd ),
H1 → Ht2 (Rd ) and Ht 2 (Rd ) → H2 (2.4)

hold true when t2 ∈ R satisfies t2 ≥ s;.
We note that if (2.3) holds true, then H1 → Ht1 (Rd ) and Ht 1 (Rd ) → H2 .
The assertion (2) in Theorem 2.2 is a consequence of the following result.
Proposition 2.3. Let s ∈ R , B1 , B2 be quasi-Banach spaces such that H0,s (Rd ) →

B1 and B2 → H0,s (Rd ). Then H[ϑ2
r,s ]
(Rd ) → B1 and B2 → H[1/ϑ
2
r,s ]
(Rd ) for
some r > 0.
Proof. First we prove H[ϑ2
r,s ]
(Rd ) → B1 for some r > 0. Since the topology of
H0,s (Rd ) is defined by the semi-norms · H2[ϑ ] , it follows that
r,s
f B1 ≤C f H2[ϑ , f ∈ H0,s (Rd ),

r,s]
for some positive constants C and r.

The extension of this to the inclusion H[ϑ
2
r,s ]
(Rd ) → B1 now follows from
the fact that H0,s is dense in H[ϑr,s ] .
2
In order to prove B2 → H[1/ϑ

2
r,s ]
(Rd ) for some r > 0, we assume that the
statement is false. Then B2 → H[1/ϑ
2
r,s ]
, for every r > 0. We need to show that
this leads to a contradiction.
By the assumption it follows that for every k ≥ 1, there is an element fk ∈
H[1/ϑ
2
k,s ]
such that
fk H2[1/ϑ ≥ fk B2 .
k,s ]
By modifying fk , it follows that for every R ≥ 1, there is a sequence {fk }k≥1 such
that
fk H2[1/ϑ ] = R and fk B2 ≤ (6k Dk )−1 ,
k,s
where D ≥ 1 is the same as in (2.1). Also let

N
gN = 3k fk .
k=1
Then

N
gN B2 ≤ 3k D k fk B2 ≤1
k=1
and

N −1
gN H2[1/ϑ ≥3 N
fN H2[1/ϑ − 3k fk H2[1/ϑ
N,s ] N,s ] N,s ]
k=1

N −1
N −1
(3N + 3)R
≥ 3N fN H2[1/ϑ − 3k fk H2[1/ϑ ] = 3N R − 3k R = .
N,s] k,s 2
k=1 k=1

Since H0,s (Rd ) is the inductive limit of H[1/ϑ
2
r,s ]
(Rd ) with respect to r > 0,
and that R can be chosen arbitrarily large, it follows that the ball
{ f ∈ H0 (Rd ) ; f B2 ≤ 1}

is unbounded in H0,s (Rd ). This contradicts the fact that B2 → H0,s (Rd ), and
the result follows.
Corollary 2.4. Let s ∈ R , and let B1 , B2 be quasi-Banach spaces such that

H0,s (Rd ) → B1 and B2 → H0,s (Rd ). Then there are positive constants r and C
such that
1 1
hα B1 ≤ Cer|α| 2s , hα B2 ≥ C −1 e−r|α| 2s when s ∈ R+ ,
1 1
hα B1 ≤ Cr|α| (α!) 2σ , hα B2 ≥ C −1 r−|α| (α!) − 2σ
when s = σ .
Proof. By Proposition 2.3 it follows that B2 → H[1/ϑ

2
r,s ]
(Rd ) for some r > 0,
which in particular implies
hα B2 ≥ C −1 hα H2[1/ϑ = C −1 ϑr,s (α)−1 ,
r,s ]
for some constant C > 0. In the same way it follows that hα B1 ≤ Cϑr,s (α), and
the result follows.
Proof of Theorem 2.2 in the case s ∈ R+ . As remarked above, (2) is an immedi-
ate consequence of Proposition 2.3. We have to prove (1).
By replacing B1 with the completion of Hs , it follows that it is no restriction
to assume that Hs is dense in B1 under the quasi-norm · B1 . Let f ∈ B1 . Since
B1 → H0 , it follows that f is given by (1.2), where hα is the Hermite function of
order α and
cα = cα (f ) = (f, hα )L2 .
The fact that Hs is continuously embedded in B1 implies that for every
integer j > 0 we have
1
1
f 2B1 ≤ Cj D−2j |cα |2 e j |α| ,
2s
α
where the constant D ≥ 1 is the same as in (2.1), and the constant Cj ≥ 1 is

independent of f (cf. formula (2.12) in [5]).
For every integer j ≥ 1, let
1
j
−1 − 1j |α| 2s
Nj = sup{ |α| ; Cj j 2 ej e ≥ 1 },
and define inductively
R1 = N1 and Rj = max(Rj−1 + 1, Nj ), j ≥ 2.
Furthermore, let
I0 = { α ; |α| ≤ R1 } and Ij = { α ; Rj < |α| ≤ Rj+1 }, j ≥ 1,
⎧ 1
⎪
⎨ sup C1 e|α| 2s when α ∈ I0
m(α) = α∈I0
⎪ 1
⎩ej j −1 e 2j |α| 2s when α ∈ Ij , j ≥ 1,
and let H1 be the Hilbert space which consists of all f ∈ H0 (Rd ) such that
12
f H1 ≡ |cα (f )|2 m(α)
α
is finite. We shall prove that H1 satisfies the requested properties. Since
1
lim m(α)e−c|α| 2s = 0
|α|→∞
when c > 0, it follows that Hs is continuously embedded in H1 . Furthermore, the

fact that m(α) = m(β) when |α| = |β| implies that f → Sπ f is a unitary map on
H1 , for every permutation π on {1, . . . , d}.
It remains to prove that H1 is continuously embedded in B1 and in H0,t
when t > s. Let f ∈ H1 and

fj = cα (f )hα , j ≥ 0.
α∈Ij
Then
#
cα (f ), α ∈ Ij
2 2
f= fj , cα (fj ) = and f H1 = fj H1 .
j≥0 0, α∈
/ Ij j≥0
This gives
1 1
2 1
1 2
|cα |2 e j |α|
2s
f B1 ≤ Dj+1 fj B1 ≤D |cα |2 m(α) +D Cj
j α∈I0 j≥1 α∈Ij
1 1
2 1 2
1 2
|cα |2 e j |α|
2s
≤D |cα |2 m(α) +D
j
α∈I0 j≥1 α∈Ij

1
≤D f0 H1 + fj H1 .
j
j≥1
Hence, by the Cauchy–Schwarz inequality we get

1 12 12
1 2
f B1 f0 H1 + 2
2
fj H1 fj 2
H1 = f H1 ,
j
j≥1 j≥1 j≥0
which proves that H1 → B1 .

The inclusion H1 → H0,t1 when t1 > s follows if we prove that for every
c > 0, there is a constant C0 which is independent of j ≥ 1 and α such that
2
e j |α| ,
r
θr j
−1
ec|α| ≤ C0 ej (2.5)
where θ = ts1 < 1 and r = 2s
1
.
By applying the logarithm, (2.5) follows if we prove that for some constants
m1 , m2 > 0, the function
h(u, v) = m1 uu + u−1 v r − m2 v θr
is bounded from below, when u, v ≥ c and c > 0. Furthermore, by taking (u, v r )
as new variables, we may assume that r = 1.
In order to prove this, let 0 < θ1 , θ2 < 1 be chosen such that θ1 > θ and
θ1 + θ2 = 1. Then the inequality on arithmetic and geometric mean-values gives
that h0 (u, v) h(u, v), where
h0 (u, v) = uθ2 u−θ1 v θ1 − m2 v θ = v θ (uθ2 u−θ1 v (θ1 −θ) − m1 ),
for some m2 > 0. Since θ1 > θ, it follows that h0 (u, v) tends to infinity when
u + v → ∞ and u, v ≥ c. The fact that h0 is continuous then implies that h0 (u, v)
and thereby h(u, v) is bounded from below when u, v ≥ c, which proves that (2.5)
holds.
This gives
θ 1
2
−1
e j |α|
j 2s
ec|α| 2s ≤ C0 ej = C0 m(α), α ∈ Ij ,
which proves the first part of (1).
It remains to prove that H2 exists with the asserted properties. The fact
that B2 is continuously embedded in Hs implies that for every j ≥ 1, there is a
constant Cj ≥ 1 such that
1
1
|cα |2 Cj−1 e− j |α| ≤ f 2B2 .
2s
Let
1
1
j −2 e−j Cj−1 e− j |α|
j 2s
m(α) = ,
j≥1
$ 12
and let H2 be the set of all f ∈ H0 such that f H2 ≡ α |cα | m(α)
2
is finite.
By the definition it follows that H2 → Hs (Rd ), and that f H2 f B2
when f ∈ B2 , giving that B2 is continuously embedded in H2 .

The inclusion H0,t1
→ H2 when t1 > s follows if we prove that
θ
m(α) e−c|α| 2s , (2.6)
s
for every c > 0, where θ = t1 < 1 as before. Now, (2.5) is equivalent to
1 θ
2
e−j e− j |α| ≤ C0 e−c|α| 2s ,
j 2s
for some constant C0 which neither depends on j nor on α. This gives

1
1 θ
1 θ
j −2 e−j e− j |α| e−c|α| 2s e−c|α| 2s ,
j 2s
m(α) ≤ 2
j
j≥1 j≥1
and (2.6) follows. The proof is complete.

Proof of Theorem 2.2 in the case when s = σ . Again, the assertion (2) is an im-
mediate consequence of Proposition 2.3.
As in the previous case we may assume that Hs is dense B1 , when proving (1).
Let f ∈ B1 be as in the previous proof.
The fact that Hσ is continuously embedded in B1 implies that for every
integer j > 0 we have
1
f 2B1 ≤ Cj D−2j |cα |2 j −4|α| (α!) σ ,
α
where the constant D ≥ 1 is the same as in (2.1), and the constant Cj ≥ 1 is
independent of f (cf. formula (2.12) in [5]).
For every integer j ≥ 1, let
j
−1 −|α|
Nj = sup{ |α| ; Cj j 2 ej j ≥ 1 },
and define inductively
R1 = N1 and Rj = max(Rj−1 + 1, Nj ), j ≥ 2.
Furthermore, let
I0 = { α ; |α| ≤ R1 } and Ij = { α ; Rj < |α| ≤ Rj+1 }, j ≥ 1,
⎧ 1
⎨ sup C1 (|α|!) σ when α ∈ I0
m(α) = α∈I0
⎩ j j −1 −2|α| 1
e j (|α|!) σ when α ∈ Ij , j ≥ 1,
and let H1 be the Hilbert space which consists of all f ∈ Hs (Rd ) such that
1/2
f H1 ≡ |cα (f )|2 m(α)
α
is finite. We shall prove that H1 satisfies the requested properties. Since
1
lim m(α)h|α| (α!)− σ = 0
|α|→∞
when c > 0, it follows that Hs is continuously embedded in H1 . Furthermore, the

fact that m(α) = m(β) when |α| = |β| implies that f → Sπ f is a unitary map on
H1 , for every permutation π on {1, . . . , d}.
It remains to prove that H1 is continuously embedded in B1 and in H0,t1

when t1 > σ . Let f ∈ H1 , and write

fj = cα (f )hα , j ≥ 0.
α∈Ij
Then
#
cα (f ), α ∈ Ij
2 2
f= fj , cα (fj ) = and f H1 = fj H1 .
j≥0 0, α∈
/ Ij j≥0
This gives

f B1 ≤ Dj+1 fj B1
j
1 1
2 1
2
≤D |cα |2 m(α) +D Cj |cα |2 j −4|α| (|α|!) σ
1 1
2 1 1
2
2 −3|α|
≤D |cα | m(α)
2
+D |cα | j (|α|!) σ
j
1
=D f0 H1 + fj H1
j
j≥1
where the third inequality follows from the fact that Cj j 2 j −|α| ≤ 1 when α ∈ Ij .
Hence, by the Cauchy–Schwarz inequality we get
1 12 12
1 2
f B1 f0 H1 + 2
fj H1 2
fj H1 = f H1 ,
j2
j≥1 j≥1 j≥0
which proves that H1 → B1 .

Since H0,τ → H0,t when t ≥ τ it suffices to prove that H1 → H0,t when
t = τ and τ > σ, which is the same as
1
R|α| (|α|!) τ m(α), |α| ≥ 1, (2.7)
holds true for every fixed R > 0. By letting θ = 1
σ − 1
τ > 0, the latter inequality
follows if we prove that
(α, j) → R−|α| ej −1 −2|α|
j
j |α|θ|α|
is lower bounded by a positive constant, and by applying the logarithm, (2.7)
follows if we can verify that
h(u, v) = −C1 u + v v − u ln v + C2 u ln u, u, v ≥ 1, (2.8)
is lower bounded when C1 and C2 are positive constants.
C2
If ueC1 ≤ v, then the dominating term in h(u, v) is v v , and it follows that
h(u, v) is lower bounded for such choices of (u, v).
C2
If instead v ≤ ueC1 , then −C1 u − u ln v + C2 u ln u ≥ 0, giving that h(u, v) is
positive and thereby lower bounded, in this case as well. Hence, H1 → H0.t (Rd ).
It remains to prove that H2 exists with the asserted properties. The fact
that B2 is continuously embedded in H σ implies that for every j ≥ 1, there is a
constant Cj ≥ 1 such that
1
|cα |2 Cj−1 j |α| (α!)− σ ≤ f 2B2 .
α
Let 1
j −2 e−j Cj−1 j |α| (|α|!)− σ ,
j
m(α) =
j≥1
and let H2 be the set of all f ∈ Hs such that
1
2
f H2 ≡ |cα | m(α)
2
α
is finite.
By the definition it follows that f H2 f B2 when f ∈ B2 , giving that
B2 is continuously embedded in H2 .
It remains to prove that H2 → H σ and H0,

τ
→ H2 when τ > σ. The first
embedding is equivalent to

k |α| j −2 e−j Cj−1 j |α|
j
j≥1
uniformly in α for every k ≥ 1 with implied constant depending on k. This is

immediate since
j
|α|
j −2 e−j Cj−1 ≥ k −2 e−k Ck−1 .
j k
k
j≥1
The second embedding follows if we prove that
1
m(α) j −|α| (α!)− τ
for every j ≥ 1 where the implied constant may vary with j. This is fulfilled
whenever
1 1
e−k k |α| j −|α| (α!) σ − τ
k
for every fixed j. Taking logarithms and choosing u = |α|, v = k, we see that this
holds since the function h = h(u, v) in (2.8) is bounded from below. The proof is
complete.
References
[1] R. Beals, Characterization of pseudodifferential operators and applications, Duke
Math. J. 44 (1977), 45–57.
[2] J. Chung, S.-Y. Chung, D. Kim, Characterizations of the Gelfand–Shilov spaces via
Fourier transforms, Proc. Amer. Math. Soc. 124 (1996), 2101–2108.
[3] E. Cordero, S. Pilipović, L. Rodino, N. Teofanov, Quasianalytic Gelfand–Shilov
spaces with applications to localization operators, Rocky Mt. J. Math. 40 (2010),
1123–1147.
[4] I.M. Gelfand, G.E. Shilov, Generalized functions, I–III, Academic Press, New York
London, 1968.
[5] T. Gramchev, S. Pilipović, L. Rodino, Classes of degenerate elliptic operators in
Gelfand–Shilov spaces in: L. Rodino, M.W. Wong (eds.) New developments in
pseudo-differential operators, Operator Theory: Advances and Applications 189,
Birkhäuser Verlag, Basel 2009, pp. 15–31.
[6] L. Hörmander, The Analysis of Linear Partial Differential Operators, vol I–III,
Springer-Verlag, Berlin Heidelberg New York Tokyo, 1983, 1985.
[7] T. Kato, On the Cauchy problem for the (generalized) Korteweg–de Vries equation,
Adv. Math. Suppl. Stud. 8, Academic Press, New York 1983, pp. 93–128.
[8] Czeslaw Klis, A simple proof of the identity S ◦ S = S , in: Proc. of the Conference
on Convergence and Generalized Functions, Katowice 1983, preprint of the Institute
of Math., Polish Academy of Sci., pp. 71–74.
[9] Z. Lozanov-Crvenković, D. Perišić, Kernel theorems for the spaces of tempered ultra-
distributions, Integral Transform Spec. Funct. 18 (2007), 699–713.
[10] Z. Lozanov-Crvenković, D. Perišić, M. Tasković, Gelfand–Shilov spaces structural
and kernel theorems, (preprint), arXiv:0706.2268v2.
[11] A. Pietsch, Operator ideals, Mathematische Monographien, vol. 16., VEB Deutscher
Verlag der Wissenschaften, Berlin, 1978.
[12] S. Pilipović, Generalization of Zemanian spaces of generalized functions which have
orthonormal series expansions, SIAM J. Math. Anal. 17 (1986), 477–484.
[13] S. Pilipović, Tempered ultradistributions, Boll. U.M.I. 7 (1988), 235–251.
[14] Z. Sadlok, On uniform convergence of Hermite series, Ann. Polon. Math. 43 (1983),
207–210.
[15] B. Simon, Trace ideals and their applications, I, London Math. Soc. Lecture Note
Series, Cambridge University Press, Cambridge London New York Melbourne, 1979.
[16] J. Toft, Multiplication properties in Gelfand–Shilov pseudo-differential calculus in:
S. Molahajlo, S. Pilipović, J. Toft, M.W. Wong, H. Zhu (eds.) Pseudo-Differential
Operators, Generalized Functions and Asymptotics, Operator Theory: Advances and
Applications, Birkhäuser Verlag, Basel, 231, pp. 117–172.
[17] J. Toft, Images of function and distribution spaces under the Bargmann transform,
J. Pseudo-Differ. Oper. Appl. (appeared online 2016).
[18] J. Toft, B. Nilsson, S. Nordebo, A. Khrennikov, Decompositions of Gelfand–Shilov
kernels into kernels of similar class, J. Math. Anal. Appl. 396 (2012), 315–322.
[19] J. Voigt, Factorization in some Fréchet algebras of differentiable functions, Studia
Math. 77 (1984), 333–348.
Yuanyuan Chen and Joachim Toft Mikael Signahl

Department of Mathematics Department of Mathematical Sciences
Linnæus University University of Agder
Växjö, Sweden Kristiansand, Norway
e.mails: [email protected] e-mail: [email protected]
[email protected]
Operator Theory:

Blow-up Phenomena for Solutions

of Discrete Nonlinear p-Laplacian
Parabolic Equations on Networks
Soon-Yeong Chung
This work is dedicated to Prof. Stevan Pilipovic for his 65th birthday
Abstract. This is an article to introduce discrete nonlinear p-Laplacian para-

bolic equations on networks and discuss the conditions under which blow-up
occurs for the solutions. We first deal with the case p = 2, introducing a re-
cent result about the blow-up phenomena for the solutions. Secondly, we deal
with the general p-Laplacian case. In each case, we classify the parameters
depending on the equations so that we can see when the solutions blow up or
globally exist. Moreover, the blow-up time and blow-up rate are introduced
for the blow-up solutions. The last part is devoted to the blow-up of Fujita
type.
1. Introduction
Discrete p-Laplacian parabolic equations (reaction-diffusion systems) have found
many applications in chemical reactions and biological phenomena. A typical ex-
ample is an autocatalytic chemical reaction between several chemicals in which the
concentration of each chemical grows (or decay) due to diffusion and difference of
concentration. In general, many of such phenomena are modeled by
q−1
ut (x, t) = Δp,ω u (x, t) + μ |u (x, t)| u (x, t) , x ∈ S
with some boundary and initial conditions where S is the set of chemicals and
p > 1, q > 0, μ > 0. Here, Δp,ω is the discrete p-Laplace operator on a network S,
defined by
p−2
Δp,ω f (x) := |f (y) − f (x)| [f (y) − f (x)] ω (x, y) .
y∈S
46 S.-Y. Chung
The goal of this article is to introduce recent results about the blow-up phe-
nomena arising from the following discrete nonlinear p-Laplacian parabolic equa-
tion in S × (0, +∞)
⎧ q−1
⎨ ut (x, t) = Δp,ω u(x, t) + μ |u(x, t)|
⎪ u(x, t), (x, t) ∈ S × (0, +∞)
u(x, t) = 0, (x, t) ∈ ∂S × (0, +∞) (1)
⎪
⎩
u(x, 0) = u0 (x) ≥ 0, x ∈ S,
where p > 1, q > 0, μ > 0 and u0 is nontrivial. Here S is a finite network and the
operator Δp,ω is the discrete p-Laplacian on the network S with boundary ∂S.
We say that a solution u to the equation blows up at time T , if |u (xn , tn )| →
+∞ for some sequence (xn , tn ) → (a, T ). Here, T is called a blow-up time and a
is called a blow-up point.
On the other hand, there have been many papers in which they study the
blow-up phenomenon for the solution to the nonlinear p-Laplacian parabolic equa-
tions defined on the continuous domain. In fact, they show that the solution may
or may not blow up in finite time, depending on the exponent q and the magnitude
of the initial data (see [1], [7], [10], [13], [12], [18], [20]). In particular, the equation
(1) on a continuous domain Ω ⊂ RN has also been studied even until these days
(see [6] and [15]). For example, in order to get a blow-up solution to equation (1)
with p = 2 they adopted the condition such as
1 < q < (3N + 8)/(3N − 4)
in the paper [15] and
1 < q < (N + 2)/(N − 2)
in the paper [6], respectively, where N denotes the space dimension of RN . Here
we have to mention that conditions to obtain a blow-up solution should include
something related to the domain Ω ⊂ RN . From a similar point of view, it is
quite natural in the discrete case to find something related to the networks which
corresponds to N . We believe that a number representing the internal topology of
the network S plays a similar role as the dimension N .
2. Preliminaries
In this section, we start with the theoretic graph notions frequently used through-
out this paper. For more detailed information on notations, notions, and conven-
tions, we refer the reader to [3].
By a graph G = G(V, E) we mean a finite set V of vertices (or nodes) with a
set E of edges (two-element subsets of V ). The set of vertices and edges of a graph
G are sometimes denoted by V (G) and E(G), or simply V and E, respectively.
A graph G is said to be simple if it has neither multiple edges nor loops,
and G is said to be connected if for every pair of vertices x and y, there exists a
sequence (called a path) of vertices x = x0 , x1 , . . . , xn−1 , xn = y such that xj−1
and xj are connected by an edge (called adjacent ) for j = 1, . . . , n.
Nonlinear p-Laplacian Parabolic Equations 47
A graph G = G (V , E ) is said to be a subgraph of a graph G = G(V, E)

if V ⊂ V and E ⊂ E. In this case, G is a host graph of G . If E consists of all
the edges from E which connect the vertices of V in its host graph G, then G is
called an induced subgraph. It is noted that an induced subgraph of a connected
host graph may not be connected.
A weight on a graph G is a symmetric function ω : V × V → [0, ∞) satisfying
that
(i) ω(x, x) = 0, x ∈ V,
(ii) ω(x, y) = ω(y, x) if x ∼ y,
(iii) ω(x, y) > 0 if and only if {x, y} ∈ E.
Here, {x, y} denotes the edge connecting the vertices x and y. Then we call
a graph G with a weight ω a network.
For an induced subgraph S of a G = G(V, E), the (vertex) boundary ∂S of
S is the set of all vertices z ∈ V \ S but are adjacent to some vertex in S, i.e.,
∂S := {z ∈ V \ S | z ∼ y for some y ∈ S}.
By S, we denote a subgraph of G whose vertices are consisting of those in S
or ∂S and whose edges are formed by the edges between vertices in S and edges
between a vertex in S and a vertex in ∂S.
Throughout this paper, a subgraph S in our concern is assumed to be an
induced subgraph which is simple and connected. From now on, for simplicity, by
a network S with boundary ∂S we mean a subgraph S of G, associated with the
weight ω.
The degree dω x of a vertex x in a network S (with boundary ∂S) is defined
to be

dω x := ω(x, y).
y∈S
For a number p with p > 1, the discrete p-Laplacian Δp,ω on a network S of

a function u : S → R is defined by

Δp,ω u(x) := |u(y) − u(x)|p−2 [u(y) − u(x)]ω(x, y)
y∈S
for each x ∈ S.
The following lemmas are useful throughout this article.
Lemma 1 ([16, 17]). For p > 1, and functions f, g : S → R, the discrete Laplacian
Δp,ω satisfies that

2 g (x) [−Δp,ω f (x)]
x∈S
p−2
= |f (y) − f (x)| [f (y) − f (x)] · [g (y) − g (x)] ω (x, y) .
x,y∈S
48 S.-Y. Chung
In particular, in the case that g = f , we have

2 f (x) [−Δp,ω f (x)] = [f (x) − f (y)]p ω (x, y) .
x∈S x,y∈S
Lemma 2 ([16, 17]). For p > 1, there exist λp,0 > 0 and φp,0 (x) > 0, x ∈ S such
that ⎧
⎨ −Δp,ω φp,0 (x) = λp,0 φp,0 (x) , x ∈ S,
p−1
⎪
φp,0 (x) = 0, x ∈ ∂S,
⎪
⎩$ p
x∈S φp,0 (x) = 1.
Moreover,
$ p
x,y∈S [u (x) − u (y)] ω (x, y)
1
λp,0 = min 2
$ p
u∈A,u≡0 x∈S [u (x)]

where A := u : S → R | u = 0 on ∂S .
In the above, the number λp,0 is called the first eigenvalue Δp,ω on a network
S with corresponding eigenfunction φp,0 (see [2] and [5] for more details).
In order to discuss the blow-up phenomena for the solutions to the equation
(1), we need some preparation. In fact, the local existence can be proved by the
usual method, for example, the contractive method.
Lemma 3 ([4]). [Comparison Principle] Let T > 0 (T may be +∞), μ > 0, q ≥ 1
and p > 1. Suppose that the real-valued functions u(x, ·), v(x, ·) ∈ C[0, T ) are
differentiable in (0, T ) for each x ∈ S and satisfy
⎧
⎪ut (x, t) − Δp,ω u (x, t) − μ|u(x, t)| u(x, t)
⎪ q−1
⎪
⎨ ≥ v (x, t) − Δ v (x, t) − μ|v(x, t)|q−1 v(x, t), (x, t) ∈ S × (0, T ) ,
t p,ω
(2)
⎪u (x, t) ≥ v (x, t) ,
⎪ (x, t) ∈ ∂S × [0, T ),
⎪
⎩
u (x, 0) ≥ v (x, 0) , x ∈ S.
Then u (x, t) ≥ v (x, t) for all (x, t) ∈ S × [0, T ).
Proof. Let T > 0 be arbitrarily given with T < T . Then by the mean value
theorem, for each x ∈ S and 0 ≤ t ≤ T ,
|u (x, t)|q−1 u (x, t) − |v (x, t)|q−1 v (x, t) = q |ξ (x, t)|q−1 [u (x, t) − v (x, t)]
for some ξ (x, t) lying between u (x, t) and v (x, t). Then it follows from (2) that
we have
q−1 q−1
ut − Δp,ω u − μq |ξ (x, t)| u (x, t) ≥ vt − Δp,ω v − μq |ξ (x, t)| v (x, t) (3)

for all (x, t) ∈ S × (0, T ]. Let ũ, ṽ : S × [0, T ] → R be the functions defined by
ũ (x, t) := e−2μqLt u (x, t) and ṽ (x, t) := e−2μqLt v (x, t).

where L := max|r|≤M rq−1 and M := maxx∈S, t∈[0,T ] {|u (x, t)| , |v (x, t)|}.
Then the inequality (3) can be written as
ũt (x, t) − ṽt (x, t) − e2qμL(p−2)t [Δp,ω ũ (x, t) − Δp,ω ṽ (x, t)]
% & (4)
q−1
+ μq 2L − |ξ (x, t)| [ũ (x, t) − ṽ (x, t)] ≥ 0
for all (x, t) ∈ S × (0, T ]. Since S × [0, T ] is compact, there exists (x0 , t0 ) ∈
S × [0, T ] such that
(ũ − ṽ)(x0 , t0 ) = min min (ũ − ṽ) (x, t).
x∈S 0≤t≤T
Then we only have to show that (ũ − ṽ)(x0 , t0 ) ≥ 0. Suppose that (ũ −
ṽ)(x0 , t0 ) < 0, on the contrary. Since (ũ − ṽ)(x, t) ≥ 0 on both ∂S × [0, T ] and
S × {0}, we have (x0 , t0 ) ∈ S × (0, T ]. Then, we have
ũt (x0 , t0 ) ≤ ṽt (x0 , t0 ) and Δp,ω ũ(x0 , t0 ) ≥ Δp,ω ṽ(x0 , t0 ). (5)
q−1
Since ξ (x, t) ≤ max|r|≤M rq−1 = L, we have
[2L − |ξ(x0 , t0 )|q−1 ](ũ − ṽ)(x0 , t0 ) ≤ L(ũ − ṽ)(x0 , t0 ) < 0. (6)
Combining (5) and (6), we obtain
ũt (x0 , t0 ) − ṽt (x0 , t0 ) − e2qλL(p−2)t0 [Δp,ω ũ (x0 , t0 ) − Δp,ω ṽ (x0 , t0 )]

% &
q−1
+ μq 2L − |ξ (x0 , t0 )| [ũ (x0 , t0 ) − ṽ (x0 , t0 )] < 0,
which contradicts (4). Therefore, ũ (x, t) − ṽ (x, t) ≥ 0 for all (x, t) ∈ S × (0, T ] so
that we get u (x, t) ≥ v (x, t) for all (x, t) ∈ S × [0, T ), since T < T is arbitrarily
given.
When p ≥ 2, we obtain a strict comparison principle as follows:
Lemma 4 ([4]). [Strict Comparison Principle] Let T > 0 (T may be +∞), μ > 0,
q ≥ 1 and p ≥ 2. Suppose that real-valued functions u(x, ·), v(x, ·) ∈ C[0, T ) are
differentiable in (0, T ) for each x ∈ S and satisfy
⎧
⎪
⎪ ut (x, t) − Δp,ω u (x, t) − μ|u(x, t)|q−1 u(x, t)
⎪
⎨ ≥ v (x, t) − Δ v (x, t) − μ|v(x, t)|q−1 v(x, t), (x, t) ∈ S × (0, T ) ,
t p,ω
(7)
⎪
⎪ u (x, t) ≥ v (x, t) , (x, t) ∈ ∂S × [0, T ),
⎪
⎩
u (x, 0) ≥ v (x, 0) , x ∈ S.
If u0 (x∗ ) > v0 (x∗ ) for some x∗ ∈ S, then u (x, t) > v (x, t) for all (x, t) ∈
S × (0, T ).
Lemma 5 ([4]). Let T > 0 (T may be +∞), p > 1, q > 0, and μ > 0. Suppose that
real-valued functions u (x, ·), v (x, ·)∈ C [0, T ) are differentiable in (0, T ) for each
50 S.-Y. Chung
x ∈ S and satisfy
⎧
⎪
⎪ ut (x, t) − Δp,ω u (x, t) − μ|u(x, t)|q−1 u(x, t)
⎪
⎨ ≥ v (x, t) − Δ v (x, t) − μ|v(x, t)|q−1 v(x, t),
t p,ω (x, t) ∈ S × (0, T ) ,
(8)
⎪
⎪ u (x, t) > v (x, t) , (x, t) ∈ ∂S × [0, T ),
⎪
⎩
u (x, 0) > v (x, 0) , x ∈ S.
Then u (x, t) ≥ v (x, t) for all (x, t) ∈ S × (0, T ).
Remark 1. We end this section with two remarks.
(i) In the above, Lemma 3 is a general form of the comparison principle, in which
we assume p > 1 and q ≥ 1 to get u ≥ v on S × (0, T ). But if we assume
p ≥ 2, particularly, then we get the strict inequality u > v on S × (0, T ), as
seen in Lemma 4.
(ii) On the other hand, it is well known that Lemma 3 is no longer true for the
case 0 < q < 1. Instead, if we assume that u > v on the boundary and at the
initial time as seen in Lemma 5, then we can get a similar result as Lemma 3.
3. Blow-up for the case p = 2

In this section, we discuss the long time behavior of equation (1) with the case
p = 2. Then equation (1) can be written as follows:
⎧ q−1
⎨ ut (x, t) = Δ2,ω u (x, t) + |u (x, t)|
⎪ u (x, t) , (x, t) ∈ S × (0, +∞) ,
u (x, t) = 0, (x, t) ∈ ∂S × (0, +∞) , (9)
⎪
⎩
u (x, 0) = u0 (x) ≥ 0, x ∈ S,
where q > 0 and the initial data u0 is nontrivial.
$
Theorem 1 ([9]). Let
$ u be a solution of equation (9) and consider y0 = x∈S u0 (x)
and K = maxx∈S y∈∂S ω(x, y). Then
(i) If 0 < q ≤ 1, then the solution u is global.
1
(ii) If q > 1 and y0 > K q−1 · |S|, then the solution u blows up.
Moreover, the blow-up time T in (ii) satisfies
' (
q−1 −1
1 |S|
T ≤ ln 1 − K · .
K(q − 1) y0
Proof. We here give only the sketch of proof. For the case (i), we consider the
following ODE problem
⎧
⎨ dt
d
z (t) = z q (t) , t > 0,
⎩ z (0) = max u0 (x) + 1.
x∈S
Then it is easy to see that a solution z (t) globally exists for all t > 0, so
that we have 0 ≤ u (x, t) < v (x, t) = z (t), (x, t) ∈ S × (0, +∞) by the comparison
principle (Lemma 5). This implies that u must be global.
Secondly, we assume that q > 1. Then summing on S equation (9), we get
' (q
1−q

ut (x, t) ≥ −K u (x, t) + |S| u (x, t)
x∈S x∈S x∈S
and
η (t) ≤ K (q − 1) η (t) + (1 − q) |S|
1−q
,
)$ *1−q
where η (t) := x∈S u (x, t) . The second inequality gives
' (
1−q
|S|
η (t) ≤ e K(q−1)t
η (0) + e K(1−q)t
−1 ,
K
which is equivalent to
1
y q−1 (t) ≥ , (10)
|S|1−q |S|1−q
y01−q − K eK(q−1)t + K
$ q−1
where y (t) := x∈S u (x, t). Thus, if y0q−1 > K · |S| , the solution blows up and
we have the following estimate for the blow-up time T:
' (
q−1 −1
1 |S|
T ≤ ln 1 − K · .
K (q − 1) y0
Now, we introduce the blow-up rate for the solution of equation (9).
Theorem 2 ([9]). Let q > 1 and u be a solution to equation (9) blowing up at finite
time T . Then:
(i) (The lower bound)
1
1 q−1
1
max u(x, t) ≥ (T − t)− q−1 , t > 0.
x∈S q−1
(ii) (The upper bound)
+ ,− q−1
1
1
max u(x, t) ≤ (q − 1)(T − t) − d(q − 1)2 (T − t)2 , t > 0,
x∈S 2
where d = maxx∈S dω x.
(iii) (The blow-up rate)
1
1 1 q−1
lim (T − t) q−1 max u(x, t) = .
t→T − x∈S q−1
52 S.-Y. Chung
4. Blow-up for the general case

In this section, we discuss the blow-up phenomena for the general case p > 1. We
recall the equation in our concern as follows:
⎧ q−1
⎨ ut (x, t) = Δp,ω u (x, t) + μ |u (x, t)|
⎪ u (x, t) , (x, t) ∈ S × (0, +∞) ,
u (x, t) = 0, (x, t) ∈ ∂S × (0, +∞) , (11)
⎪
⎩
u (x, 0) = u0 (x) ≥ 0, x ∈ S,
where p > 1, q > 0, μ > 0 and the initial data u0 is nontrivial on S.
Theorem 3 ([4]). Let u be a solution to equation (11). Then, we have the following:
(i) If 0 1, then the solution u blows up in a finite time,
provided
1
u0 > ( ω0 / μ) q−p+1 ,
$
where ω0 := maxx∈S y∈S ω(x, y) and u0 := maxx∈S u0 (x).
(ii) If 0 < q ≤ 1, then the nonnegative solution u is global.
(iii) If 1 < q 0, let xt ∈ S be a node such that u (xt , t) := maxx∈S u (x, t).
Then, it follows that u(xs , s) > u0 , s ∈ (0, +∞).
Let ζ : [u0 , +∞) → (0, ζ (u0 )] be a function defined by
+∞
ds
ζ (y) := < +∞, y ≥ u0 .
y −ω0 s p−1 + μsq
Then ζ is a decreasing continuous function from [u0 , +∞) onto (0, ζ (u0 )]
with its inverse function ζ −1 . Moreover, we have
t u(xt ,t)
ut (xs , s) ds
t≤ ds = .
0 −ω0 u −ω0 s
p−1 q
(xs , s) + μu (xs , s) p−1 + μsq
u0
This can be written as
u (xt , t) ≥ ζ −1 [ζ (u0 ) − t] ,
which implies that u (xt , t) blows up, as t → ζ (u0 ).
(ii) Consider the following ODE problem
#
d q
dt z (t) = μz (t) , t > 0,
z (0) = u0 + 1.
Then it is easy to see that a solution z (t) globally exists for all t > 0, so
that we have 0 ≤ u (x, t) < v (x, t) = z (t), (x, t) ∈ S × (0, +∞) by the comparison
principle (Lemma 5). This implies that u must be global.
(iii) Consider the following eigenvalue problem
# p−2
−Δp,ω φp,0 (x) = λp,0 |φp,0 (x)| φp,0 (x) , x ∈ S,
φp,0 (x) = 0, x ∈ ∂S.
Note that it is well known that λp,0 > 0 and φp,0 (x) > 0 for all x ∈ S (see
[16], [17]).
Now, for a sufficiently large k, take v(x, t) := kφp,0 (x), x ∈ S, t ≥ 0. Then
0 ≤ u(x, t) ≤ v(x, t) = kφp,0 (x), (x, t) ∈ S × (0, +∞) by comparison principle
(Lemma 3), which is required.
Remark 2.
(i) When the solution blows up in Theorem 3, the blow-up time T can be esti-
mated as
+∞
u1−q ds
0
≤ T ≤ ζ(u0 ) = .
μ(q − 1) u0 −ω0 s p−1 + μsq
(ii) In Theorem 3, if u0 = maxx∈S u0 (x) is not sufficiently large, then the solution
u may be global.
(iii) In the case 1 < p − 1 = q, which was not discussed in Theorem 3, the long
time behavior of the solution u is determined by the magnitude of parameter
μ. This case will be dealt with in Theorem 5, as the critical case.
We now introduce the lower bound, the upper bound and the blow-up rate
for the maximum function of blow-up solutions to (11).
Theorem 4 ([4]). Let u be a solution to equation (11), which blows up at a finite
time T , q > p − 1 > 0, and q > 1. Then, we have the following:
(i) (The lower bound)
1
− q−1
max u(x, t) ≥ [μ(q − 1)(T − t)] , 0 < t < T.
x∈S
(ii) (The upper bound)

% & 1
2q−p − q−1
max u(x, t) ≤ μ(q − 1)(T − t) − α(T − t) q−1 , 0 < t < T,
x∈S
q−p+1 2q−p $
where α := ω0 μ q−1 (q − 1) q−1 and ω0 = maxx∈S y∈S ω (x, y).
(iii) (The blow-up rate)
+ , q−1
1
1 1
lim (T − t) q−1 max u(x, t) = , 0 < t < T.
t→T − x∈S μ(q − 1)
We now discuss the case p − 1 = q, which is the critical case. In this case,
equation (11) can be written as follows:
⎧ p−2
⎨ ut (x, t) = Δp,ω u (x, t) + μ |u (x, t)|
⎪ u (x, t) , (x, t) ∈ S × (0, ∞) ,
u (x, t) = 0, (x, t) ∈ ∂S × (0, ∞) , (12)
⎪
⎩
u (x, 0) = u0 (x) ≥ 0, x ∈ S,
where μ > 0, p > 1, and the initial data u0 is nontrivial on S.
Theorem 5 ([8]). Let u(x, t) be a solution to equation (12). Then, we have the
following:
54 S.-Y. Chung
(i) For p > 2 and μ > 0, the solution u blows up at some T , provided that
1 p
p
− |u0 (x) − u0 (y)| ω (x, y) + μ u0 (x) > 0. (13)
2
x,y∈S x∈S
(ii) For p > 2 and μ > λp,0 , the solution u blows up at some T , for every
nonnegative and nontrivial initial data u0 .
(iii) For p > 2 and μ ≤ λp,0 , the solution u is global for every nonnegative initial
data u0 .
(iv) For 1 0. In
particular, when 1 2 $ p ≥ λp,0 .
x∈S u0 (x)
Hence, (i) can be considered as a special case of (ii).
5. Blow-up of the Fujita type

Long time ago, Fujita [11] introduced and discussed the so-called critical exponent
q ∗ := 1 + 2/N , for the first time as far as authors know, for the Cauchy problem
#
ut = Δu + uq ,x ∈ RN , t > 0,
(14)
u (x, 0) = u0 (x), x ∈ RN ,
in a sense that for 1 < q < q ∗ the solution for any nonnegative and non-trivial
initial data blows up in finite time, whereas for q > q ∗ the solution is global for suf-
ficiently small nonnegative initial data. In a similar context, Meier [14] determined
the critical exponent for the problem
⎧
⎨ ut (x, t) = Δu(x, t) + h(t)u (x, t), (x, t) ∈ D × (0, T ],
q
⎪
u(x, t) = 0, (x, t) ∈ ∂D × (0, T ], (15)
⎪
⎩
u (x, 0) = u0 (x) ≥ 0, x ∈ D,
where D ⊂ RN , q > 1 and the function h is identical to tq or eβt , β > 0 (see the
paper [14] for more details).
For discrete equations, Zhou et al. [19] recently considered the discrete ω-
diffusion equations as follows:
⎧
⎨ ut (x, t) = Δ2,ω u (x, t) + e u (x, t) , (x, t) ∈ S × (0, +∞) ,
βt q
⎪
u (x, t) = 0, (x, t) ∈ ∂S × (0, +∞) , (16)
⎪
⎩
u (x, 0) = u0 ≥ 0, x ∈ S,
where q > 1, β > 0, and the initial data u0 is nontrivial.
A nonnegative function u(x, t) ∈ C 1 (S × [0, T )) is a supersolution of equation

(16) if it satisfies
⎧
⎨ ut (x, t) ≥ Δ2,ω u (x, t) + e u (x, t) , (x, t) ∈ S × (0, +∞) ,
βt q
⎪
u (x, t) ≥ 0, (x, t) ∈ ∂S × (0, +∞) ,
⎪
⎩
u (x, 0) ≥ u0 , x ∈ S.
Similarly, we can define the subsolution u(x, t) by reversing the inequalities.
Now, we introduce the comparison principle for the nonnegative solutions of
equation (16) which plays an important role in the proof of the existence of the
critical exponent.
Lemma 6. Let u(x, t) and u(x, t) be super solution and subsolution of equation
(16), respectively. Meanwhile, there exists a point y ∈ S such that ω(x, y) = 0 for
any x ∈ S. Then, for any (x, t) ∈ S × [0, T ), we have u(x, t) ≥ u(x, t).
Now, we discuss the existence of the critical exponent for problem (16). The
number qβ is called the critical exponent of equation (16), if it satisfies:
(i) When q > qβ , there is a nonnegative and nontrivial global solution u of
equation (16).
(ii) When 1 < q < qβ , every nontrivial solution u of equation (16) blows up in
finite time.
The following theorem implies that if q > 1, equation (16) admits a critical
exponent qβ = 1 + λβ2,0 , where λ2,0 is the first eigenvalue of eigenvalue problem
⎧
⎪
⎨ −Δω ψ2,0 (x) = λ2,0 ψ2,0 (x) ,x ∈ S,
ψ2,0 (x) = 0, x ∈ ∂S,
⎪
⎩
maxx∈S̄ ψ2,0 (x) = 1.
Theorem 6 ([19]). Let u be a solution to equation (16). Then, we have the following:
(i) If q > qβ = 1 + λβ2,0 and the initial value u0 (x) ≤ z0 ψ2,0 (x), x ∈ S, then the
solution is global and positive. Here z0 is a constant satisfying
1
β q−1
0 < z0 < λ2,0 − .
q−1
(ii) If 1 < q < qβ = 1 + λβ2,0 and u0 is nonnegative and nontrivial, then the solu-
$
tion blows up in finite time in the sense that limt→T − x∈S u (x, t) ψ2,0 (x) =
+∞ and the blow-up time T satisfies

β−(q−1)λ2,0 1−q
ln 1 − 1−q G 0
T = .
β − (q − 1) λ2,0
Moreover, we have
1
1 e−βT q−1
lim (T − t) q−1
max u (x, t) = .
t→T − x∈S q−1
56 S.-Y. Chung
Proof. (i) Let v (x, t) = ψ2,0 (x) e−λ2,0 t and z (t) be the solution of the initial-value
problem ⎧
⎨ dt d
z (t) = eβt max v q−1 (x, t) z q (t) , t > 0,
x∈S (17)
⎩ z (0) = z > 0,
0
q−1
1
where z0 is a constant satisfying z0 < λ2,0 − q−1 β

. Then it is easy to see that
+ , 1−q
1
q−1
z (t) = z01−q − 1−e (β−(q−1)λ2,0 )t
,
(q − 1) λ2,0 − β
which is bounded uniformly for t ∈ [0, +∞). By comparison principle (Lemma 6),
we obtain that
u (x, t) ≤ z (t) v (x, t)
on S × [0, +∞), which implies that u (x, t) must be global.
(ii) Multiplying ψ2,0 (x) on the both sides of equation (16) and summing on
S, we have

ut (x, t) ψ2,0 (x) + λ2,0 u (x, t) ψ2,0 (x) = eβt uq (x, t) ψ2,0 (x) .
x∈S x∈S x∈S
$
Taking G (t) := x∈S u (x, t) ψ2,0 (x) and applying Jensen’s inequality, we
obtain
' (q

eβt uq (x, t) ψ2,0 (x) ≥ eβt u (x, t) ψ2,0 (x) = eβt Gq (t) .
x∈S x∈S
So that we have
G (t) ≥ −λ2,0 G (t) + eβt Gq (t) .
Thus, using the comparison for the linear ODE, we have
1
Gq−1 (t) ≥ % & . (18)
1−q
G0 − β−λ2,0 (q−1) + β−λ2,0 (q−1) e(β−(q−1)λ2,0 )t e(q−1)λ2,0 t
1−q 1−q
Since 1 < q < qβ = 1 + λβ2,0 , we have β−(q−1)λ 1−q

< 0 and G1−q (0) −
1−q $ 2,0
β−(q−1)λ2,0 > 0. Thus, G (t) = x∈S u (x, t) ψ2,0 (x) cannot be global. From the
right-hand side of (18), we have the blow-up time

β−(q−1)λ2,0 1−q
ln 1 − 1−q G 0
T = .
β − (q − 1) λ2,0
To prove the last part, let u (xt , t) = maxx∈S u (x, t) for each t > 0. Then the
equation (16) can be written as

ut (xs , s) = [u (y, s) − u (xs , s)] ω (xs , y) + eβs uq (xs , s)
y∈S̄
and we obtain
1
β q−1 βT − 1
u (xs , s) ≥ e − eβs q−1 .
q−1
This implies
q−1
1
1
1 (T − t) q−1
β
e−βT q−1
lim (T − t) q−1
u (xt , t) ≥ lim = . (19)
t→T − t→T − eβT − eβt q−1
On the other hand, it follows that

ut (xs , s) = [u (y, s) − u (xs , s)] ω (xs , y) + eβs uq (xs , s)
y∈S̄
+ ,
q − 1 βT
≥ uq (xs , s) eβs − d e − eβs ,
β
where d = maxx∈S̄ dω x. Then we obtain
' ( 1−q
1
(q − 1) (β + d (q − 1)) βT d (q − 1)2
u (xt , t) ≤ e −e βt
− (T − t) e βT
β2 β
so that
1
1 e−βT q−1
lim (T − t) q−1
u (xt , t) ≤ , (20)
t→T − q−1
which completes the proof.
Remark 4. The case 0 < q ≤ 1 was not discussed in the above Theorem 6. In fact,
this case will be treated in a forth coming paper.
References
[1] C. Bandle and H. Brunner: Blow-up in diffusion equations, a survey. J. Comput.
Appl. Math. 97, 3–22 (1998)
[2] F.R.K. Chung: Spectral graph theory. BMS Regional Conference Series in Math. 92,
Amer. Math. Soc. (1997)
[3] S.-Y. Chung and C.A. Berenstein: ω-harmonic functions and inverse conductivity
problems on network. SIAM J. Appl. Math. 65, no. 4, 1200–1226 (2005)
[4] S.-Y. Chung, M.-J. Choi: Blow-up solutions and global solutions to discrete p-
Laplacian parabolic equations. Abstr. Appl. Anal. 2014, Art. ID 351675, 11 pp.
[5] D.M. Cvetkovic, M. Doob, H. Sachs: Spectra of graphs, Theory and applications.
Acad. Press, New York, (1980)
[6] C. Cortazar, M. Elgueta, J.D. Rossi: The blow-up problem for a semilinear parabolic
equation with a potential. J. Math. Anal. Appl. 335(1), 418–427 (2007)
[7] L.A. Caffarelli, A. Friedman: Blow-up of solutions of nonlinear heat equations. J.
Math. Anal. Appl. 129, 409–419 (1988)
58 S.-Y. Chung
[8] S.-Y. Chung: Critical Blow-Up and Global Existence for Discrete Nonlinear p-
Laplacian Parabolic Equations, Discrete Dyn. Nat. Soc. 2014, Art. ID 716327, 10
pp.
[9] S.-Y. Chung and J.-H. Lee: Blow-up for Discrete Reaction-Diffusion Equations on
Networks, Appl. Anal. Discrete Math. 9, 103–119 (2015)
[10] B. Hu: Blow-up Theories for Semilinear Parabolic Equations. Lecture Notes in Math-
ematics Vol. 2018. Springer, Heidelberg, (2011)
[11] H. Fujita: On the blowing up of solutions of the Cauchy problem for ut = Δu+u1+α ,
J. Fac. Sci. Univ. Tokyo Sect. I 13 109–124 (1966).
[12] M.P. Llanos and J.D. Rossi: Blow-up for a non-local diffusion problem with Neumann
boundary conditions and reaction term. Nonlinear Anal. 70, 1629–1640 (2009)
[13] Y. Li and C. Xie: Blow-up for p-Laplacian parabolic equations. Electron. J. Diff.
Equa. no. 20, 1–12 (2003)
[14] P. Meier: On the critical exponent for reaction-diffusion equations, Arch. Rational
Mech. Anal. 109, no. 1, 63–71 (1990)
[15] F. Merle, H. Zaag: O. D. E. type behavior of blow-up solutions of nonlinear heat
equations. Discrete contin. Dyn. Syst. 8(2), 435–450 (2002)
[16] J.-H. Park and S.-Y. Chung: The Dirichlet boundary value problems for p-
Schrödinger operators on finite networks. J. Difference Equ. Appl. 17, no. 5, 795–811
(2011)
[17] J.-H. Park and S.-Y. Chung: Positive solutions for discrete boundary value problems
involving the p-Laplacian with potential terms. Comp. Math. Appl. 61, no. 1, 17–29
(2011)
[18] J. Yin and C. Jin: Critical extinction and blow-up exponents for fast diffusive p-
Laplacian with sources. Math. Meth. appl. Sci. 30 no. 10 1147–1167 (2007)
[19] W. Zhou, M. Chen and W. Liu: Critical exponent and blow-up rate for the ω-
diffusion equations on graphs with Dirichlet boundary conditions, Electron. J. Diff.
Equa. (2014), no. 263, 13 pp.
[20] J. Zhao and Z. Liang: Blow-up rate of solutions for p-Laplacian equation. J. Part.
Diff. Equa. 21 no. 2, 134–140 (2008)
Soon-Yeong Chung
Department of Mathematics
Sogang University
Seoul 04107, Republic of Korea
Operator Theory:

Generalized Function Algebras Containing

Spaces of Periodic Ultradistributions
Andreas Debrouwere
Abstract. We construct differential algebras in which spaces of (one-dimen-

sional) periodic ultradistributions are embedded. By proving a Schwartz im-
possibility type result, we show that our embeddings are optimal in the sense
of being consistent with the pointwise multiplication of ordinary functions.
In particular, we embed the space of hyperfunctions on the unit circle into
a differential algebra in such a way that the multiplication of real analytic
functions on the unit circle coincides with their pointwise multiplication. Fur-
thermore, we introduce a notion of regularity in our newly defined algebras
and show that an embedded ultradistribution is regular if and only if it is an
ultradifferentiable function.
1. Introduction
Differential algebras containing the space of distributions were first introduced
and studied by J.F. Colombeau [5, 6]. It was the starting point of the by now
well-established nonlinear theory of generalized functions. For a clear exposition
of this theory and its applications to various branches of analysis we refer to the
monographs [7, 14].
In this paper we contribute to this program by developing a nonlinear the-
ory of (one-dimensional) periodic ultradistributions. The problem of embedding
the space of hyperfunctions on the unit circle [13] into a differential algebra has
attracted the attention of various authors [3, 4, 22]. However, the embeddings pro-
posed so far do not preserve the multiplication of all real analytic functions. In this
article we construct a new algebra, containing the space of hyperfunctions on the
unit circle, which does enjoy this property and we show that it is optimal in this
respect. In fact, we consider the aforementioned embedding problem for general
classes of periodic ultradistributions, both of Beurling and Roumieu type, defined
Supported by Ghent University, through a BOF Ph.D.-grant.

60 A. Debrouwere
via a weight sequence Mp on which only very mild growth conditions are imposed.
The Roumieu case of Mp = p! corresponds with the hyperfunction case.
In [3] (see also [4]) a differential algebra containing the space of hyperfunctions
on the unit circle (the topological dual of the space of real analytic functions on
the unit circle) is constructed and it is shown that the multiplication of functions
belonging to the quasianalytic class defined via the weight sequence p!s , for some
0 < s < 1, is preserved. A comparison with the situation for classical distributions
(the topological dual of the space of compactly supported smooth functions) shows
that this is not the most optimal result one would presume to be true. Namely,
recall that Schwartz’ impossibility result [20] asserts there is no associative and
commutative differential algebra, containing the space of distributions, in which
the multiplication of k-times differentiable functions, k ∈ N, coincides with their
pointwise multiplication. Nonetheless, in the Colombeau algebra the multiplication
of smooth functions is preserved. In analogy to Colombeau’s construction, it is
natural to expect that it is possible to embed the space of hyperfunctions on the
unit circle into a differential algebra in such a way that the multiplication of real
analytic functions is preserved and, moreover, that such an embedding is optimal.
In this paper, we construct such an algebra and show its optimality.
Stated in terms of a general weight sequence Mp , our results may be sum-
marized as follows: We show that it is possible to embed the space of periodic
ultradistributions of class Mp into a differential algebra in such a way that the
multiplication of periodic ultradifferentiable functions of class Mp is preserved.
By establishing an analogue of Schwartz’ impossibility result for periodic ultradis-
tributions, we show that our embedding is optimal. Furthermore, we introduce a
notion of regularity in our algebras of periodic generalized functions and show that
an embedded ultradistribution is regular if and only if it is an ultradifferentiable
function.
Finally, we would like to remark that in a forthcoming paper [2], jointly with
H. Vernaeve and J. Vindas, a nonlinear theory for (non-quasianalytic) ultradis-
tributions on the n-dimensional Euclidean space is developed (see also the earlier
works [1, 9, 18]). There we construct a sheaf of differential algebras in which the
sheaf of ultradistributions of class Mp is embedded and the multiplication of ultra-
differentiable functions of class Mp is preserved. Moreover, we define a notion of
regularity and lay the ground for microlocal analysis in these algebras. The results
in the present paper are the natural counterparts of these results in the periodic
case.
This paper is organized as follows. In Section 2 we introduce certain spaces of
periodic ultradifferentiable functions and ultradistributions, and recall their char-
acterization in terms of Fourier coefficients [8, 15, 16]. Section 3 is devoted to an
analogue of Komatsu’s second structure theorem for periodic ultradistributions,
this result is used to prove the Schwartz impossibility type result for periodic ul-
tradistributions in Section 4. The construction of our algebras is given in Section
5. Furthermore, we give an alternative projective description of the algebras of
Roumieu type (see [11, 17] for analogues in the theory of ultradistributions) and
Periodic Generalized Function Algebras 61
provide a null characterization of the so-called space of negligible elements, which

we apply to obtain a pointwise characterization of our generalized functions. We
remark that the definition of our algebras containing the space of hyperfunctions
on the unit circle significantly differs from the one considered in [3, 4, 22]. The
embedding of spaces of periodic ultradistributions of class Mp into our algebras
is discussed in Section 6. As usual, this will be achieved by means of convolution
with a suitable mollifier sequence. We also show that the multiplication of ultra-
differentiable functions of class Mp is preserved. It should be pointed out that in
the Roumieu case this result is valid precisely because of the (different) definition
of our algebras. In the last section we study regularity in our algebra of periodic
generalized functions and show that an embedded ultradistribution is regular if
and only if it is an ultradifferentiable function of class Mp .
2. Spaces of periodic ultradifferentiable functions

and their duals
In this preliminary section we introduce test function spaces of periodic ultra-
differentiable functions and recall the well-known characterization of these spaces
and their duals by means of Fourier coefficients [8, 15, 16]. For the sake of com-
pleteness, we discuss these results in quite some detail and provide proofs. We shall
work with the notion of ultradifferentiability through weight sequences [10]. Fix a
positive weight sequence (Mp )p∈N with M0 = 1. We always assume the following
conditions on Mp :
(M.1) Mp2 ≤ Mp−1 Mp+1 , p ∈ Z+ ,
(M.2) Mp+q ≤ AH p+q Mp Mq , p, q ∈ N , for some A, H ≥ 1 .
The associated function of Mp is defined as
tp
M (t) := sup log , t>0.
p∈N Mp
We extend M to the whole real line by setting M (t) = M (|t|) for t ∈ R.
Lemma 1 ([10, Prop. 3.6]). Let A, H be the constants occurring in (M.2). Then
2M (t) ≤ M (Ht) + log A , t>0.
As usual, the relation Mp ⊂ Np between two weight sequences means that
there are C > 0 and h > 0 such that Mp ≤ Chp Np for all p ∈ N. The stronger
relation Mp ≺ Np means that the latter inequality remains valid for every h > 0
and a suitable C = Ch > 0. A function f defined on R is said to be 2π-periodic if
f (t + 2π) = f (t), t ∈ R.
M ,h
We write E2πp , h > 0, for the Banach space of all smooth 2π-periodic functions
on R satisfying
hp Dp ϕ L∞ (R)
ϕ Mp ,h = ϕ h := sup <∞,
p∈N Mp
62 A. Debrouwere
where D = −id/dt. Furthermore, we define

(M ) M ,h {Mp } M ,h
E2π p = lim E2πp , E2π = lim E2πp .
←− −→+
h→∞ h→0
(M ) {M }
Elements of their dual spaces E2π p and E2π p are called periodic ultradistribu-
tions of class (Mp ) (of Beurling type) and periodic ultradistributions of class {Mp }
(of Roumieu type), respectively. In the sequel we shall write ∗ instead of (Mp ) or
{Mp } if we want to treat both cases simultaneously. In addition, we shall often first
state assertions for the Beurling case followed in parenthesis by the corresponding
statements for the Roumieu case.
Proposition 1. The sequence Mp satisfies 1 ≺ Mp (1 ⊂ Mp ) if and only if the space
(M ) {M }
E2π p (the space E2π p ) contains a function which is not identically zero.
(M ) {M }
Proof. If Mp satisfies 1 ≺ Mp (1 ⊂ Mp ) then eiλt ∈ E2π p (eiλt ∈ E2π p ) for each
(M ) {M }
λ ∈ R. Conversely, suppose that E2π p (E2π p ) contains a function ϕ which is not
identically zero. Then there is λ ∈ R such that
2π
ϕ(t)eiλt dt = 0.
0
Hence, for each h > 0 (some h > 0) it holds that
2π 2π
p 2π ϕ h Mp

0 < λ iλt
ϕ(t)e dt = D ϕ(t)e dt ≤
p iλt
, p∈N,
0 0 hp
from which the result follows.
We shall always assume 1 ≺ Mp and 1 ⊂ Mp in the Beurling and Roumieu
case, respectively.
Remark 1. We write Oλ , λ > 1, for the space of holomorphic functions on the
open annulus {z ∈ C | 1/λ < |z| < λ} and set
A(T) = lim Oλ ,
−→+
λ→1
{p!}
where T denotes the complex unit circle. The spaces A(T) and E2π are isomor-
phic as topological vector spaces (in the sequel abbreviated as t.v.s.) by means
of the mapping ϕ → ϕ(t) = ϕ(eit ). Elements of the dual space A (T) are called
hyperfunctions on the unit circle [13].
∗
Let ϕ ∈ E2π , the Fourier coefficient of index k of ϕ, k ∈ Z, is given by
2π
1

ϕ(k) := ϕ(t)e−ikt dt .
2π 0
In order to study the Fourier coefficients of periodic ultradifferentiable functions
we introduce the following sequence spaces
sMp ,λ (Z) = {(ck )k ∈ CZ | σλ ((ck )k ) := sup |ck |eM(λk) < ∞} , λ>0.
k∈Z
Furthermore, we set
s(Mp ) (Z) = lim sMp ,λ (Z) , s{Mp } (Z) = lim sMp ,λ (Z) .
←− −→+
λ→∞ λ→0
∗
Proposition 2. For ϕ ∈ E2π the following series expansion holds

ϕ(t) =
ϕ(k)e ikt
, t∈R, (1)
k∈Z
∗
with convergence in E2π . Moreover, the mapping ϕ → (ϕ(k))
k yields the (t.v.s.)
∗ ∼ ∗
isomorphism E2π = s (Z).
M ,h
Proof. Let ϕ ∈ E2πp , h > 0. For all p ∈ N we have

1 2π p ϕ h Mp
p

|k ϕ(k)| = D ϕ(t)e −ikt
dt ≤ , k∈Z.
2π 0 hp
Hence
Mp
|ϕ(k)|
≤ ϕ h inf = ϕ h e−M(hk) , k∈Z,
(h|k|)p
p∈N
which shows the continuity of the mapping ϕ → (ϕ(k)) k in both the Beurling and
Roumieu case. Conversely, for (ck )k ∈ sMp ,λ (Z), λ > 0, and N ∈ N we have

Dp ck eikt L∞ (R) ≤ σλ ((ck )k ) k p e−M(λk)
|k|≥N |k|≥N
Mp+2 1
≤ σλ ((ck )k ) p+2 ,
λ k2
|k|≥N
$
which, by (M.2), shows that the series k ck e , (ck )k ∈ s∗ (Z), converges in E2π
ikt ∗
$
and that the mapping (ck )k → k∈Z ck eikt is continuous in both cases. Since the
expansion (1) holds in the space of smooth 2π-periodic functions endowed with its
natural topology [13] the proof is completed.
Next, we consider Fourier series of periodic ultradistributions. Observe first

that a standard argument shows that the topological duals of the spaces s(Mp ) (Z)
and s{Mp } (Z) are given by
s(Mp ) (Z) = lim sMp ,−λ (Z) , s{Mp } (Z) = lim sMp ,−λ (Z) ,
−→ ←−+
λ→∞ λ→0
where
sMp ,−λ (Z) = {(ck )k ∈ CZ | σλ ((ck )k ) := sup |ck |e−M(λk) < ∞} , λ>0.
k∈Z
∗
Let f ∈ E2π , the Fourier coefficient of index k of f , k ∈ Z, is given by
1
f(k) := f (t), e−ikt .
2π
In view of the above remark, Proposition 2 implies the following:
64 A. Debrouwere
∗
Proposition 3. For f ∈ E2π the following series expansion holds

f (t) = f(k)eikt ,
k∈Z
with (strong) convergence in E2π ∗

. Moreover, the mapping f → (f(k))k yields the
∗ ∼ ∗
(t.v.s.) isomorphism E2π = s (Z).
$
An entire function P (z) = ∞ n=0 an z , z ∈ C, is said to be an ultrapolynomial
n
of class (Mp ) (class {Mp }) if the coefficients satisfy the estimate
CLn
|an | ≤ , n∈N,
Mn
for some L > 0 and C > 0 (for every L > 0 and a suitable C = CL > 0). The
associated infinite-order differential operator
∞

P (D) = an D n ,
n=0
is called an ultradifferential operator of class (Mp ) (class {Mp }). Condition (M.2)
∗
ensures that P (D) acts continuously on E2π [10, Thm. 2.12] and hence it can
be defined on the corresponding ultradistribution space by duality, namely, for
∗
f ∈ E2π
∗
P (D)f, ϕ = f, P (−D)ϕ , ϕ ∈ E2π .
One has the following relation between Fourier coefficients
P
(D)f (k) = P (k)f(k) , k∈Z.
Moreover,
∗
P (D)(eikt f ) = eikt P (D + k)f , f ∈ E2π ,k ∈ Z , (2)
where P (D + k) is the ultradifferential operator related to the symbol P (z +
k). Remark that this may be seen as a weak form of Leibniz’ rule. Finally, we
∗
briefly review the convolution of periodic ultradistributions: For f, g ∈ E2π their
convolution is defined as
∗
f ∗ g, ϕ = f (t), g(u), ϕ(t + u) , ϕ ∈ E2π .
One can readily check that f ∗ g ∈ E2π ∗

and that f∗ g(k) = 2π f(k)
g (k) for
∗ ∗
all k ∈ Z. If g ∈ E2π , then f ∗ g ∈ E2π and it is given by
f ∗ g(t) = f (u), g(t − u), t∈R.

3. Structure theorem for periodic ultradistributions

We present an analogue of Komatsu’s second structure theorem (as stated in
[21]) for periodic ultradistributions. In the next section we shall use it to prove a
Schwartz impossibility type result for periodic ultradistributions. In order to treat
the Beurling and Roumieu case uniformly we first give an alternative projective
{M }
description of the spaces E2π p and s{Mp } (Z) (an idea which goes back to Komatsu
[11]). We write R for the family of non-decreasing sequences (rj )j∈N with r0 = 1
which tend to infinity. This set is partially ordered and directed by the relation
rj sj , namely, there is j0 ∈ N such that rj ≤ sj for all j ≥ j0 . Let rj ∈-R. The
p
function Mrj denotes the associated function of the weight sequence Mp j=0 rj .
M ,rj
We write E2πp for the Banach space of all smooth 2π-periodic functions on R
such that
Dp ϕ L∞ (R)
ϕ Mp ,rj = ϕ rj := sup -p <∞. (3)
p∈N Mp j=0 rj
Lemma 2 ([19, Lemma 2.3]). Let rj ∈ R. Then, there is rj ∈ R such that rj ≤ rj ,
j ∈ N, and

p+q
p
q
rj ≤ 2p+q rj rj , ∀p, q ∈ N .
j=0 j=0 j=0
In view of Lemma 2 one can show, in the same way as in Proposition 2, that
the mapping ϕ → (ϕ(k))
k yields the (t.v.s.) isomorphism
lim E2πp
M ,rj ∼
= lim sMp ,rj , (4)
←− ←−
rj ∈R rj ∈R
where
sMp ,rj = {(ck )k ∈ CZ | σrj ((ck )k ) := sup |ck |eMrj (k) < ∞} .
k∈Z
Lemma 3. Let (an )n∈N be a sequence of positive reals. Then

(i) supn an e−M(λn) < ∞ for all λ > 0 if and only if supn an e−Mrj (n) < ∞ for
some sequence rj ∈ R.
(ii) supn an eMrj (n) < ∞ for all sequences rj ∈ R if and only if supn an eM(λn) <
∞ for some λ > 0.
Proof. We only need to show the direct implications, the “if” parts are clear.
(i) We first show that there is a subordinate function ε (which means that
ε is continuous, increasing, and satisfies ε(0) = 0 and ε(t) = o(t)) such that
supn an e−M(ε(n)) < ∞. Our assumption and Lemma 1 imply that for every λ >
0 there is N = Nλ ∈ N such that an ≤ AeM(λn) for all n ≥ N , where A is
the constant occurring in (M.2). Hence we can inductively determine a sequence
(Nj )j∈Z+ of positive natural numbers with N1 = 0 which satisfies
Nj Nj−1
an ≤ AeM(n/(j+1)) , n ≥ Nj , ≥ +1, j≥2.
j j−1
66 A. Debrouwere
Let lj denote the line through the points (Nj , Nj /j) and (Nj+1 , Nj+1 /(j + 1)),
and define
ε(t) = lj (t) , for t ∈ [Nj , Nj+1 ) .
The function ε is subordinate and, moreover, an ≤ AeM(ε(n)) for all n ≥ N2 . Thus,
it suffices to show that there is a sequence rj ∈ R such that M (ε(t)) ≤ Mrj (t) + C
for t > 0 and some C > 0. By [10, Lemma 3.12] there is a weight sequence Np
satisfying (M.1) with associated function N such that Mp ≺ Np and M (ε(t)) ≤
N (t) for t > 0. On the other
- hand, [11, Lemma 3.4] yields the existence of a
sequence rj ∈ R with Mp pj=0 rj ≺ Np . The result now follows from [10, Lemma
3.10].
(ii) By the last part of the proof of (i) and our assumption we have that
supn an eM(ε(n)) < ∞ for all subordinate functions ε. Suppose that supn an eM(λn) <
∞ does not hold for any λ > 0. We could therefore find a sequence (Nj )j∈Z+ of
positive natural numbers with N1 = 0 which satisfies
Nj Nj−1
aNj eM(Nj /j) ≥ j , ≥ +1, j≥2.
j j−1
Exactly as in the proof of (i), we define the subordinate function ε(t) as the line
through (Nj , Nj /j) and (Nj+1 , Nj+1 /(j + 1)) for t ∈ [Nj , Nj+1 ). We would then
have
aNj eM(ε(Nj )) = aNj eM(Nj /j) ≥ j , j ≥ N2 ,
contradicting supn an eM(ε(n)) < ∞.
Corollary 1. We have
{Mp }
s{Mp } (Z) = lim sMp ,rj (Z) ,
M ,rj
E2π = lim E2πp ,
←− ←−
rj ∈R rj ∈R
as topological vector spaces.
Proof. By Proposition 2 and the isomorphism (4) it suffices to show the second
equality. By Lemma 3(ii) the spaces coincide as sets. Moreover, it is clear that the
seminorms σrj , rj ∈ R, are continuous on s{Mp } (Z). Conversely, let us prove that
every seminorm p on s{Mp } (Z) is continuous on limr sMp ,rj (Z). Since s{Mp } (Z)
←− j
is reflexive, there is a bounded set B in s{Mp } (Z) such that p is bounded by the
seminorm

pB ((ck )k ) = sup b−k ck , (ck )k ∈ s{Mp } (Z) .
(bk )k ∈B
k∈Z
Lemma 3(i) implies that there is rj ∈ R such that
sup |bk | ≤ CeMrj (k) , k∈Z,

(bk )k ∈B
for some C > 0. Let rj ∈ R satisfy the conditions from Lemma 2. Hence, Lemma
1 yields that
Mr (k) −Mr (k)
pB ((ck )k ) ≤ C e j |ck | ≤ ACσrj ((ck )k ) e j , (ck )k ∈ s{Mp } (Z) ,
k∈Z k∈Z
where rj = rj /(2H), j ∈ N, and A,H are the constants occurring in (M.2).
Theorem 1. Let Np be a weight sequence satisfying (M.1) and Mp ≺ Np . For every
(M ) {M }
f ∈ E2π p (f ∈ E2π p ) there is an ultradifferential operator P (D) of class (Mp )
(N )
(of class {Mp }) and g ∈ E2π p such that f = P (D)g. In the Beurling case, one
{Mp }
can even choose g ∈ E2π .
Proof. By Proposition 2 and 3 it suffices to show that for every (ck )k ∈ s(Mp ) (Z)
((ck )k ∈ s{Mp } (Z)) there is an ultrapolynomial P of class (Mp ) (of class {Mp })
and (ak )k ∈ s(Np ) (Z) such that ck = P (k)ak for all k ∈ Z. In the Beurling case, we
must show that it is possible to choose (ak )k ∈ s{Mp } (Z). Let H be the constant
occurring in (M.2).
Beurling case: There is λ > 0 such that |ck | ≤ CeM(λk) for all k ∈ Z and some
C > 0. Define
∞
(λH 2 z)2p
P (z) = , z∈C.
p=0
M2p
It is clear that P is an ultrapolynomial of class (Mp ). Condition (M.2) implies that
P (x) ≥ C e2M(λx) for all x ∈ R and some C > 0. The ultrapolynomial P and the
sequence ak = ck /P (k) satisfy the requirements.
Roumieu case: Lemma 3(i) implies that there is rj ∈ R such that |ck | ≤ CeMrj (k)
for all k ∈ Z and some-C > 0. On the other hand, by [11, Lemma 3.4] there is
kj ∈ R such that Mp j=0 kj ≺ Np . Let rj , kj ∈ R satisfy the conditions from
p
Lemma 2 with respect to rj and kj , respectively. Define

∞
∞

(2Hz)2p (2Hz)2p
P1 (z) = -2p , P2 (z) = -2p , z∈C.
p=0 j=0 rj M2p p=0 j=0 kj M2p
Set P (z) = P1 (z)P2 (z). It is clear that P is an ultrapolynomial of class {Mp }.

M (x)+M (x)
Condition (M.2) implies that P (x) ≥ C e rj k
j for all x ∈ R and some C >
0. The ultrapolynomial P and the sequence ak = ck /P (k) satisfy the requirements.

4. Impossibility result on the multiplication

of periodic ultradistributions
In this section we show an analogue of Schwartz’ famous impossibility result [20] for
periodic ultradistributions. Recall that in the case of classical distributions it states
that it is impossible to embed the space of distributions into an associative and
68 A. Debrouwere
commutative differential algebra in such a way that both differentiation, the unity
function (= constant function 1) and the pointwise multiplication of continuous
functions is preserved – or more generally, k-times differentiable functions for any
k ∈ N, see [7, p. 7]. In our impossibility result the role of the continuous functions
is played by a class of ultradifferentiable functions which is less regular than the
∗
class E2π . We assume in the rest of this section that Np is a weight sequence
satisfying (M.1). Recall that ∗ stands for (Mp ) or {Mp }. In addition, we write †
∗
for (Np ) or {Np }. When embedding E2π into some associative and commutative
algebra (Λ∗,† , +, ◦) = Λ the following requirements seem to be natural:
∗
(i) E2π is linearly embedded into Λ and f (x) ≡ 1 is the unity in Λ.
(ii) For each ultradifferential operator P (D) of class ∗ there is a linear operator
P (D) : Λ → Λ satisfying (cf. (2))
P (D)(eikt ◦ f ) = eikt ◦ P (D + k)f , f ∈ Λ ,k ∈ Z .

Moreover, P (D)|E2π∗ coincides with the usual action of P (D) on periodic ul-
tradistributions of class ∗.
(iii) ◦|E † ×E † coincides with the pointwise product of functions.
2π 2π
Theorem 2. Suppose that Mp ≺ Np . There are no associative and commutative

algebras Λ(Mp ),(Np ) and Λ{Mp },(Np ) satisfying conditions (i)–(iii). In the Beurling
case, there is even no associative and commutative algebra Λ(Mp ),{Mp } satisfying
conditions (i)–(iii).
Proof. Suppose Λ∗,† = Λ is such an algebra (we treat all cases at once). One
can generalize Schwartz’ original idea by making use of the following observation:
∗
eikt ◦ f = eikt f for all f ∈ E2π and k ∈ Z; this readily follows from Theorem 1 and
conditions (ii) and (iii). Define
∞

cot(t) = i + 2i e−2ikt ,
k=1
it is a regularization of the cotangent function in the space of periodic (ultra)distri-

∗
butions. The following relations clearly hold in E2π
cos δ = δ , sin δ = 0 , cos = sin cot .

Since the sine and cosine are finite linear combinations of eit and e−it , the remark
at the beginning of the proof implies that
δ = cos δ = cos ◦δ = (sin cot) ◦ δ = (sin ◦cot) ◦ δ = (sin δ) ◦ cot = 0 ,
∗
contradicting the injectivity of E2π → Λ.
(M ) {M }
Despite this impossibility result, we shall construct algebras G2π p and G2π p
that do satisfy properties (i)–(iii) for ∗ = † = (Mp ) and ∗ = † = {Mp }, respectively.
5. Algebras of periodic generalized functions

of class (Mp ) and {Mp }
We now introduce algebras of periodic generalized functions as quotients of alge-
bras consisting of sequences of ultradifferentiable functions satisfying appropriate
growth conditions. Their construction is inspired by the theory of sequence space
algebras [4] (see also [3, 22]). In this section we present some basic properties
of these algebras. More concretely, we show a null characterization of the space
of so-called negligible elements, provide an alternative projective description of
the algebra of Roumieu type (cf. Cor. 1), define associated rings of generalized
numbers and obtain a pointwise characterization of our generalized functions. The
embedding of periodic ultradistributions is postponed to the next section.
5.1. Definition and basic properties

The spaces of moderate sequences of class ∗ are defined as
(M ) (M ) N
E2π,M
p
= {(fn )n ∈ E2π p | ∀h > 0, ∃λ > 0
−M(λn)
sup fn he < ∞} ,
n∈N
{M } {Mp } N
E2π,M
p
= {(fn )n ∈ E2π | ∀λ > 0, ∃h > 0
−M(λn)
sup fn he < ∞} ,
n∈N
and the spaces of negligible sequences of class ∗ as

(M ) (M ) N
E2π,N
p
= {(fn )n ∈ E2π p | ∀h > 0, ∀λ > 0
sup fn he
M(λn)
< ∞} ,
n∈N
{M } {Mp } N
E2π,N
p
= {(fn )n ∈ E2π | ∃λ > 0, ∃h > 0
sup fn he
M(λn)
< ∞} .
n∈N
Remark 2. In the Beurling case our definition fits into the general framework of
[4] while this is not so for the Roumieu case – the difference lies in the fact that
the choice and order of quantifiers in our definition is completely different. It is
important to point out that this will play an essential role when embedding the
space of periodic ultradistributions and preserving the product of periodic ultra-
differentiable functions (see Sect. 6). In particular, for the case of hyperfunctions
on the unit circle our definition differs from the one given in [3, 4, 22].
∗
Lemma 1 ensures that E2π,M is an algebra under pointwise operations of
∗ ∗
sequences and that E2π,N is an ideal of it. Hence, we can define the algebra G2π of
periodic generalized functions of class ∗ as the factor algebra
∗ ∗ ∗
G2π = E2π,M /E2π,N .
70 A. Debrouwere
∗ ∗
We denote the equivalence class of (fn )n ∈ E2π,M by [(fn )n ]. Observe that E2π can
∗
be regarded as a subalgebra of G2π via the constant embedding
∗
σ(f ) := [(f )n ] , f ∈ E2π .
∗
We also remark that G2π can be endowed with a canonical action of ultradifferential
∗ ∗
operators of class ∗: The spaces E2π,M and E2π,N are closed under ultradifferential
operators P (D) of class ∗ if we define their actions on sequences as P (D)((fn )n ) :=
(P (D)fn )n . Consequently, every ultradifferential operator P (D) of class ∗ induces
a linear operator
∗ ∗
P (D) : G2π → G2π ,
∗
which clearly satisfies the generalized Leibniz rule (2) for any f ∈ G2π and eikt ,
k ∈ Z, identified with σ(e ).
ikt
We can also define spaces of moderate and negligible sequences based on the
spaces s∗ (Z)
(M ) N
sM p (Z) = {(ck,n )k,n ∈ s(Mp ) (Z) | ∀h > 0, ∃λ > 0
sup σh ((ck,n )k )e−M(λn) < ∞} ,
n∈N
{M } N
sM p (Z) = {(ck,n )k,n ∈ s{Mp } (Z) | ∀λ > 0, ∃h > 0
sup σh ((ck,n )k )e−M(λn) < ∞} ,
n∈N
and
(Mp ) N
sN (Z) = {(ck,n )k,n ∈ s(Mp ) (Z) | ∀h > 0, ∀λ > 0
sup σh ((ck,n )k )eM(λn) < ∞} ,
n∈N
{Mp } N
sN (Z) = {(ck,n )k,n ∈ s{Mp } (Z) | ∃λ > 0, ∃h > 0
sup σh ((ck,n )k )eM(λn) < ∞} .
n∈N
Proposition 2 implies the following simple but useful lemma:

∗ N
Lemma 4. Let (fn )n ∈ E2π . Then
∗
(i) (fn )n ∈ E2π,M if and only if (fn (k))k,n ∈ s∗M (Z).
(ii) (fn )n ∈ E ∗ if and only if (fn (k))k,n ∈ s∗ (Z).
2π,N N
∗
We now show the null characterization of the ideal E2π,N .
∗ ∗
Proposition 4. Let (fn )n ∈ E2π,M . Then, (fn )n ∈ E2π,N if and only if
sup fn L∞ (R) e
M(λn)
<∞,
n∈N
for all λ > 0 (for some λ > 0).

Proof. The proof is based on the Landau–Kolmogorov inequality [12] on the real
line: For all 0 0 there is λ > 0 (for every
λ > 0 there is h > 0) such that
CeM(λn) Mp
D p fn L∞ (R) ≤ , p, n ∈ N ,
hp
for some C = Cλ,h > 0. The assumption yields that for every μ > 0 (some μ > 0)
fn L∞ (R) ≤ C e−M(μn) , n∈N,
for some C = Cμ > 0. By applying inequality (5) with m = 2p, p > 0, we obtain
that
1/2 1/2
D p fn L∞ (R) ≤ 2π fn L∞ (R) D2p f L∞ (R)
2π(CC )1/2 M2p

1/2
M (λn) − M (μn)
≤ exp , p, n ∈ N.
hp 2
1/2
The result follows from Lemma 1 and the fact that (M.2) implies that M2p ≤
A1/2 H p Mp , p ∈ N.
Corollary 2. Let (ck,n )k,n ∈ s∗M (Z). Then, (ck,n )k,n ∈ s∗N (Z) if and only if
sup sup |ck,n |eM(λn) < ∞ ,

n∈N k∈Z
for all λ > 0 (for some λ > 0).

{Mp }
5.2. Projective description of G2π
We give an alternative projective description of the algebra of periodic generalized
functions of Roumieu type. As in Section 3 this shall be done by using the family
R of all non-decreasing sequences tending to infinity. We define projective type
spaces of moderate and negligible sequences by means of the seminorms (3) as
{M } {Mp } N
Ẽ2π,M
p
= {(fn )n ∈ E2π | ∀rj ∈ R, ∃sj ∈ R
−Msj (n)
sup fn rj e < ∞} ,
n∈N
and
{M } {Mp } N
Ẽ2π,N
p
= {(fn )n ∈ E2π | ∀rj ∈ R, ∀sj ∈ R
Msj (n)
sup fn rj e < ∞} .
n∈N
72 A. Debrouwere
Likewise, we define
{M } N
s̃M p (Z) = {(ck,n )k,n ∈ s{Mp } (Z) | ∀rj ∈ R, ∃sj ∈ R
sup σrj ((ck,n )k )e−Msj (n) < ∞} ,
n∈N
and
{Mp } N
s̃N (Z) = {(ck,n )k,n ∈ s{Mp } (Z) | ∀rj ∈ R, ∀sj ∈ R
sup σrj ((ck,n )k )eMsj (n) < ∞} .
n∈N
The goal of this subsection is to show the following result:
Proposition 5. We have
{M } {M } {M } {M }
E2π,M
p
= Ẽ2π,M
p
, E2π,N
p
= Ẽ2π,N
p
,
and
{M } {M } {Mp } {Mp }
sM p (Z) = s̃M p (Z) , sN (Z) = s̃N (Z) .
{M }
Remark 3. Since the structure (choice and order of quantifiers) of Ẽ2π,M
p
and
{M }
Ẽ2π,N
p
coincides with the structure of the widely accepted definition for spaces of
moderate and negligible sequences based on an arbitrary locally convex space [4],
Proposition 5 may serve as a justification for our definitions in the Roumieu case.
{M } N
Proof. The isomorphism (4) yields the following: For (fn )n ∈ E2π p we have
˜{Mp } {Mp }
(fn )n ∈ E2π,M if and only if (fn (k))k,n ∈ s̃M (Z). A similar statement holds
for the null ideals. By Lemma 4 it therefore suffices to show the second pair of
{M } {M } {M }
equalities. We start by proving sM p (Z) = s̃M p (Z). Let (ck,n )k,n ∈ sM p (Z)
and fix rj ∈ R. Since the seminorm σrj is continuous on sMp ,h (Z) for each
h > 0, we obtain supn σrj ((ck,n )k )e−M(λn) < ∞ for all λ > 0. By applying
Lemma 3(i) to the sequence (σrj ((ck,n )k ))n , we find a sequence sj ∈ R such
{M }
that supn σrj ((ck,n )k )e−Msj (n) < ∞. Conversely, let (ck,n )k,n ∈ s̃M p (Z). The
{M }
definition of s̃M p (Z) and Lemma 3(i) imply that for all λ > 0 and rj ∈ R

sup sup |ck,n | exp Mrj (k) − M (λn) < ∞ .
n∈N k∈Z
Fix λ > 0. By applying Lemma 3(ii) to the sequence
max sup |ck,n |e−M(λn) ,

|k|=l n∈N l∈N
we find h > 0 such that supn σh ((ck,n )k )e−M(λn) < ∞. For the second equality,
{M } {M }
the inclusion sN p (Z) ⊆ s̃N p (Z) is clear, whereas the converse inclusion is a
consequence of Corollary 2 and Lemma 3(ii).
5.3. Generalized point values

In this subsection we introduce the ring of generalized numbers of class ∗ in order
to view periodic generalized functions of class ∗ as objects defined pointwise. We
introduce
CM p = {(zn )n ∈ CN | ∃λ > 0 sup |zn |e−M(λn) < ∞} ,
(M )
n∈N
{M }
CM p = {(zn )n ∈ C | ∀λ > 0 sup |zn |e−M(λn) < ∞} ,
N
n∈N
and
= {(zn )n ∈ CN | ∀λ > 0 sup |zn |eM(λn) < ∞} ,
(Mp )
CN
n∈N
{M } N
CN p = {(zn )n ∈ C | ∃λ > 0 sup |zn |eM(λn) < ∞} .
n∈N
Clearly C∗N is an ideal in the ring C∗M . The ring of generalized numbers of class ∗
is defined as the factor ring
C∗ = C∗M /C∗N .
Observe that C∗ is not a field. In fact, this follows from the same examples used
for the ring of Colombeau generalized numbers [7, Ex. 1.2.33, p. 32]. Furthermore,
the elements of C are canonically embedded into C∗ via the constant embedding
σ(z) := [(z)n ] , z∈C.
Likewise, one can define the subring R∗ . Let f = [(fn )n ] ∈ G2π
∗
and t = [(tn )n ] ∈
∗
R . The point value of f at t is defined as f (t) := [(fn (tn ))n ]. The point value
f (t) does not depend on the representative of f nor on the representative of t; the
former is clear while the latter follows from the mean value theorem.The induced
pointwise defined mapping f : R∗ → C∗ : t → f (t) is 2π-periodic, that is,
f (t + σ(2π)) = f (t) , t ∈ R∗ .
∗
The next proposition shows that every f ∈ G2π can be associated with the mapping
∗ ∗
f : R → C in a one-to-one fashion. We define
∗

[0, 2π] = {t ∈ R∗ | ∃(tn )n ∈ C∗M such that t = [(tn )n ] and tn ∈ [0, 2π], n ∈ N}.
∗ ∗
Proposition 6. Let f ∈ G2π . Then, f = 0 in G2π if and only if f (t) = 0 in C∗ for
∗

all t ∈ [0, 2π] .
Proof. The direct implication is clear. Conversely, suppose that f = [(fn )n ] = 0
∗
in G2π . By Proposition 4 there are λ > 0, mn ∞ and tn ∈ [0, 2π] (mn ∞ and
tn ∈ [0, 2π]) such that
|fmn (tn )| ≥ ne−M(λmn ) (|fmn (tn )| ≥ ne−M(mn /n) ) , n∈N.
∗

Define tl = tn if l = mn for some n ∈ N and as 0 otherwise. For t = [(tl )l ] ∈ [0, 2π]
we have f (t ) = 0 in C∗ .
74 A. Debrouwere
6. Embedding of periodic ultradistributions

In this section we embed the space of periodic ultradistributions of class ∗ into the
∗
algebra G2π in such a way that the multiplication of periodic ultradifferentiable
functions of class ∗ is preserved. As usual, we accomplish this by means of con-
volution with a suitable mollifier sequence. Let us start with defining the type of
mollifier sequences that will be employed in our embedding.
Definition 1. A sequence (ϕn )n∈N of smooth 2π-periodic functions is called a mol-
n (k) = ck,n , k ∈ Z, n ∈ N, satisfy the
lifier sequence if the Fourier coefficients ϕ
following conditions:
• There is C > 0 such that
|ck,n | ≤ C , k ∈ Z ,n ∈ N .
• There is R > 0 such that for all n ∈ N
ck,n = 0 , |k| ≥ Rn .
• There is r > 0 such that for all n ∈ N
1
ck,n = , |k| ≤ rn .
2π
Example 1. Define
1 ikt
ϕn (t) = e , n∈N,
2π
|k|≤n
then (ϕn )n is clearly a mollifier sequence. It is used in [22] to embed the space of
periodic hyperfunctions into some Colombeau type algebra.
Example 2. Let ψ be a compactly supported continuous function on R such that
ψ ≡ 1/(2π) in a neighbourhood of the origin. Let ψn = ψ(·/n) and define

ϕn (t) = ψn (ξ)eiξ(t+2πk) dξ , n∈N,
k∈Z R
then ϕn is a smooth 2π-periodic function with ϕ n (k) = ψ(k/n) for all k ∈ Z,

hence (ϕn )n is a mollifier sequence. These kinds of mollifier sequences are closely
related to the mollifiers usually used to embed the space of compactly supported
(ultra)distributions into some Colombeau type algebra (see the discussions at the
beginning of [7, Sect. 1.2.2] and [2, Sect. 5]).
Theorem 3. Let (ϕn )n be a mollifier sequence. Then
∗ ∗
ι : E2π → G2π : f → [(f ∗ ϕn )n ]
is a linear embedding that satisfies the following properties:
(i) ι commutes with ultradifferential operators of class ∗, that is, for all ultra-
differential operators P (D) of class ∗
∗
P (D)ι(f ) = ι(P (D)f ) , f ∈ G2π .
(ii) ι|E2π
∗ coincides with the constant embedding σ. Consequently,
∗
ι(f g) = ι(f )ι(g) , f, g ∈ E2π .
Proof. We use the same notation as in Definition 1. Let A, H be the constants
∗
occurring in (M.2). We start by proving that (f ∗ ϕn )n ∈ E2π,M . By Lemma 4 it
∗
suffices to show that (f (k)ck,n )k,n ∈ sM (Z). Proposition 3 implies that for some
λ > 0 (every λ > 0) K = σλ ((f(k))k ) < ∞. Hence for h > 0 it holds that
σh ((f(k)ck,n )k ) = sup |f(k)||ck,n |eM(hk)
k∈Z
≤ CK sup exp (M (λk) + M (hk))
|k|≤Rn
≤ ACKeM(μn) , n∈N,
∗
where μ = HR max(λ, h). This shows that (f ∗ ϕn )n ∈ E2π,M both in the Beurling
∗
and Roumieu case. The injectivity of ι follows from the fact that ϕn → δ in E2π
∗
as n → ∞. Property (i) is clear. Finally, we show (ii). Let f ∈ E2π . By Lemma 4
and Proposition 4 it suffices to show that
sup sup |f(k)||1 − 2πck,n |eM(λn) < ∞ ,
n∈N k∈Z
for all λ > 0 (for some λ > 0). Proposition 2 implies that for every λ > 0 (some
λ > 0) K = σλ ((f(k))k ) < ∞. Hence
|f(k)||1 − 2πck,n | ≤ K|1 − 2πck,n |e−M(λk)
≤ (1 + 2πC)Ke−M(λrn) , n ∈ N ,k ∈ Z .
∗ ∗
Remark 4. It is clear that the embedding ι : E2π → G2π satisfies the properties
(i)–(iii) from Section 4 with † = ∗. Hence, it is optimal in the sense discussed there.
7. Regular periodic generalized functions

of class (Mp ) and {Mp }
In this section we introduce a notion of regularity (with respect to ultradiffer-
∗
entiability of class ∗) in G2π . The definition given below is based on the regular
∞
algebra G frequently used in classical Colombeau theory [14]. Throughout this
section we fix a mollifier sequence (ϕn )n with r = 1 in Definition 1 and consider
its associated embedding ι. We define the algebra of regular periodic generalized
functions of class ∗ as
∗,∞ ∗,∞ ∗
G2π = E2π,M /E2π,N ,
where

(Mp ),∞ (M ) N −M(λn)
E2π,M = (fn )n ∈ E2π p | ∃λ > 0, ∀h > 0 sup fn he <∞ ,
n∈N
76 A. Debrouwere

{Mp },∞ {M } N −M(λn)
E2π,M = (fn )n ∈ E2π p | ∃h > 0, ∀λ > 0 sup fn he <∞ .
n∈N
∗ ∗
Lemma 5. Let f ∈ E2π and suppose (fn )n ∈ E2π,M is a representative of ι(f ).
Then,

sup σHλ ((f(k) − fn (k))k )eM(λn) < ∞ ,
n∈N
for some λ > 0, where H denotes the constant occurring in (M.2).

Proof. The assertion is independent of the chosen representative and therefore we
may assume that fn = f ∗ϕn , n ∈ N. Proposition 3 implies that K = σλ ((f(k))k ) <
∞ for some λ > 0. Hence

σHλ ((f(k)(1 − 2πck,n ))k ) = sup |f(k)||1 − 2πck,n |e−M(Hλk)
k∈Z
≤ (1 + 2πC)K sup exp (M (λk) − M (Hλk))
|k|≥n
≤ A(1 + 2πC)Ke−M(λn) , n∈N.

∗,∞
The next regularity theorem shows that the notion of G2π -regularity coin-
∗
cides with ultradifferentiability of class ∗ when restricted to the image of E2π under
the embedding ι.
Theorem 4. We have
∗,∞ ∗ ∗
G2π ∩ ι(E2π ) = ι(E2π ).
∗ ∗,∞ ∗ ∗
Proof. The inclusion ι(E2π ) ⊆ G2π ∩ ι(E2π ) is obvious. Conversely, let f ∈ E2π
and assume ι(f ) ∈ G2π . By Proposition 2 it suffices to show that (f(k))k ∈ s (Z).
∗,∞ ∗
∗
Let (fn )n ∈ E2π,M be a representative of ι(f ) and set gn = f − fn , n ∈ N. From
∗,∞
the definition of G2π it follows that
∃λ > 0, ∀h > 0 (∃h > 0, ∀λ > 0)
sup σh ((fn (k))k )e−M(λn) < ∞ .

n∈N
Furthermore, Lemma 5 implies that for some l > 0

sup σHl gn (k))k )eM(ln) < ∞ .
((
n∈N
Combining these two facts we obtain that

∃λ, l > 0, ∀h > 0 (∃h, l > 0, ∀λ > 0)
|f(k)| ≤ K (exp (M (λn) − M (hk)) + exp (M (Hlk) − M (ln))) , (6)

where K is a positive constant independent of k and n. In the remainder of the
proof we treat the Beurling and Roumieu case separately. For a positive real num-
ber x we write x for the smallest natural number n such that n ≥ x.
Beurling case: Let μ > 0 be arbitrary. Define m = H max(Hl, μ)/l. Inequality

(6) with n = m|k| and h = H max(λm, μ) implies that
|f(k)| ≤ 2AKe−M(μk) , k∈Z.
Roumieu case: Define m = H 2 . Inequality (6) with n = m|k| and λ = h/(Hm)
implies that
|f(k)| ≤ 2AKe−M(μk) , k∈Z,
where μ = min(h/H, Hl).
References
[1] Benmeriem, K., Bouzar, C.: Algebras of generalized ultradistributions. Novi Sad J.
Math. 41, 53–62 (2011)
[2] Debrouwere, A., Vernaeve, H., Vindas, J.: Optimal embeddings of ultradistributions
into differential algebras, preprint available at http://arxiv.org/abs/1601.03512
[3] Delcroix, A., Hasler, M.F., Pilipović, S., Valmorin, V.: Embeddings of ultradistri-
butions and periodic hyperfunctions in Colombeau type algebras through sequence
spaces. Math. Proc. Camb. Philos. Soc. 137, 697–708 (2004)
[4] Delcroix, A., Hasler, M.F., Pilipović, S., Valmorin, V.: Sequence spaces with expo-
nents weights. Realizations of Colombeau type algebras. Diss. Math. 447, 56 pp.
(2007)
[5] Colombeau, J.F.: New generalized functions and multiplication of distributions,
North-Holland Publishing Co., Amsterdam, (1984)
[6] Colombeau, J.F.: Elementary introduction to new generalized functions, North-
Holland Publishing Co, Amsterdam, (1985)
[7] Grosser, M., Kunzinger, M., Oberguggenberger, M., Steinbauer, M.: Geometric the-
ory of generalized functions, Kluwer Academic Publishers, Dordrecht (2001)
[8] Gorbachuk, V.I.: On Fourier series of periodic ultradistributions, Ukranian Math.
J., 34,144–150 (1982)
[9] Gramchev, T.: Nonlinear maps in spaces of distributions, Math. Z. 209, 101–114
(1992)
[10] Komatsu, H.: Ultradistributions, I. Structure theorems and a characterization. J.
Fac. Sci. Univ. Tokyo 20, 25–105 (1973)
[11] Komatsu, H.: Ultradistributions, III. Vector valued ultradistributions and the theory
of kernels. J. Fac. Sci. Univ. Tokyo 29, 653–717 (1982)
[12] Kolmogorov, A.F.: On inequalities between the upper bounds of successive deriva-
tives of an arbitrary function on an infinite interval, (translation) in: Amer. Math.
Soc. Translations, Ser. 1, Vol. 2: Number theory and analysis, pp. 233–243, Amer.
Math. Soc., Providence (1962)
[13] Morimoto, M.: Analytic functionals on the sphere. Amer. Math. Soc., Providence
(1998)
[14] Oberguggenberger, M.: Multiplication of distributions and applications to partial
differential equations. Longman Scientific & Technical, Harlow (1992)
78 A. Debrouwere
[15] Petzsche, H.J.: Die Nuklearität der Ultradistributionsräume und der Satz vom
Kern I. Manuscripta Math. 24, 133–171 (1978)
[16] Pilipović, S.: Structural theorems for periodic ultradistributions. Proc. Amer. Math.
Soc. 98, 261–266 (1986)
[17] Pilipović, S.: Characterization of bounded sets in spaces of ultradistributions, Proc.
Amer. Math. Soc. 120, 1191–1206 (1994)
[18] Pilipović, S., Scarpalézos, D.: Colombeau generalized ultradistributions, Math. Proc.
Camb. Phil. Soc. 130, 541–553 (2001)
[19] Prangoski, B.: Laplace transform in spaces of ultradistributions. Filomat 27, 747–760
(2013)
[20] Schwartz, L.: Sur l’impossibilité de la multiplication des distributions. C. R. Acad.
Sci. Paris 239, 847–848 (1954)
[21] Takiguchi, T.: Structure of quasi-analytic ultradistributions, Publ. Res. Inst. Math.
Sci. 43, 425–442 (2007)
[22] Valmorin, V.: Generalized hyperfunctions on the circle. J. Math. Anal. Appl. 261,
1–16 (2001)
Andreas Debrouwere
Ghent University
Krijgslaan 281 Gebouw S22
B-9000 Gent, Belgium
Operator Theory:

On General Prime Number Theorems

with Remainder
Gregory Debruyne and Jasson Vindas
Abstract. We show that for Beurling generalized numbers the prime number
theorem in remainder form

x
π(x) = Li(x) + O for all n ∈ N
log n x
is equivalent to (for some a > 0)

x
N (x) = ax + O for all n ∈ N,
logn x
where N and π are the counting functions of the generalized integers and
primes, respectively. This was already considered by Nyman (Acta Math. 81
(1949), 299–307), but his article on the subject contains some mistakes. We
also obtain an average version of this prime number theorem with remainders
in the Cesàro sense.
Keywords. The prime number theorem; zeta functions; Tauberian theorems
for Laplace transforms; Beurling generalized primes; Beurling generalized in-
tegers.
1. Introduction
Since the prime number theorem (PNT) was proved in 1896, independently by
Hadamard and de la Vallée-Poussin, mathematicians have wondered which condi-
tions on the primes (and the integers) were really necessary to prove this kind of
theorems. For this reason, Beurling introduced in [2] the idea of generalized prime
numbers. A real sequence {pk }k∈N is said to be a (Beurling) generalized prime
number system if it merely satisfies
1 < p1 ≤ p2 ≤ · · · ≤ pk → ∞.
The set of generalized integers [1, 2] is the semi-group generated by the gen-
eralized primes. We arrange the generalized integers in a non-decreasing sequence
1 = n0 < n1 ≤ n2 ≤ · · · ≤ nk → ∞,
80 G. Debruyne and J. Vindas
where one takes multiplicities into account. The central objects here are the count-
ing functions of the generalized primes and integers, denoted as

π(x) = 1 and N (x) = 1. (1)
pk ≤x nk ≤x
A typical question in this setting is to determine conditions on N , as mild as

possible, such that the PNT still holds. This question for the PNT in the form
x
π(x) ∼ (2)
log x
has been studied quite intensively, starting with the seminal work of Beurling [2].
We refer to [1, 2, 10, 14, 19] for results in this direction.
In this article we are interested in stronger PNT versions than (2) for Beurling
generalized primes. Our aim is to study the PNT with remainder
x
π(x) = Li(x) + On , for all n ∈ N , (3)
logn x
where Li stands for the logarithmic integral. Naturally (3) is equivalent to the
asymptotic expansion
∞
x n!
π(x) ∼ .
log x n=0 logn x
The following theorem will be shown:
Theorem 1. The PNT with remainder (3) holds if and only if the generalized
integer counting function N satisfies (for some a > 0)
x
N (x) = ax + On , for all n ∈ N. (4)
logn x
Nyman has already stated Theorem 1 in [12], but his proof contained some
mistakes [8]. It is not true that his condition [12, statement (B), p. 300], in terms
of the zeta function
∞ ∞
ζ(s) = n−s
k = x−s dN (x), (5)
k=0 1−
is equivalent to either (3) or (4) (see Examples 1–3 below) and his proof has several
gaps.
We will show a slightly more general version of Theorem 1 in Section 4 which
also applies to non-discrete generalized number systems (cf. Section 2). For it, we
first obtain a complex Tauberian remainder theorem in Section 3, and we then
give a precise translation of (3) and (4) into properties of the zeta function. In
Section 5 we provide a variant of Theorem 1 in terms of Cesàro–Riesz means of
the remainders in the asymptotic formulas (3) and (4).
On General Prime Number Theorems with Remainder 81
2. Preliminaries and notation

2.1. Beurling generalized number systems
We shall consider an even broader definition of generalized numbers [2], which
includes the case of non-necessarily discrete number systems.
A (Beurling) generalized number system is merely a pair of non-decreasing
right continuous functions N and Π with N (1) = 1 and Π(1) = 0, both having
support in [1, ∞), and linked via the relation
∞ ∞
ζ(s) := x−s dN (x) = exp x−s dΠ(x) , (6)
1− 1
on some half-plane where the integrals are assumed to be convergent. We refer
to N as the generalized number distribution function and call Π the Riemann
prime distribution function of the generalized number system. These functions
uniquely determine one another; in fact, dN = exp∗ (dΠ), where the exponential is
taken with respect to the multiplicative convolution of measures [4]. We are only
interested in generalized number systems for which the region of convergence of
the zeta function (6) is at least e s > 1, and hence we assume this condition in
the sequel1 . The latter assumption clearly implies that N (x) and Π(x) are both
O(x1+ε ), for each ε > 0.
If N is the counting function of a discrete number system with prime counting
function π, as defined in the Introduction via (1), the Riemann prime counting
function of the discrete generalized number system is given by
∞
π(x1/j )
Π(x) = . (7)
j=1
j
Since π(x) vanishes for x < p1 , the sum (7) is actually finite and in particular
convergent. It is not difficult to verify that (7) satisfies (6); indeed,
∞ ∞ ∞
−1
exp x−s dΠ(x) = exp − log(1 − x−s )dπ(x) = 1 − p−s
k ,
1 1 k=1
and thus (6) becomes in this case a restatement of the well-known Euler product
formula for the zeta function of a discrete generalized number system [1].
The function Π may be replaced by π in virtually any asymptotic formula
about discrete generalized primes. More precisely, we have that 0 ≤ Π(x) − π(x) ≤
π(x1/2 ) + π(x1/3 ) log x/ log p1 ; in particular, Π(x) = π(x) + O(x1/2+ε ), for each
ε > 0, for a discrete generalized number system satisfying our assumption that its
associated zeta function ζ(s) converges on e s > 1. Naturally, a Chebyshev type
bound π(x) = O(x/ log x) yields the better asymptotic relation Π(x) = π(x) +
O(x1/2 / log x). However, we mention that, in general, it is not always possible to
determine a non-decreasing function π satisfying (7) for (non-discrete) generalized
1 Thisassumption is actually no restriction at all. In fact, if the zeta only converges on es > α >
0, one may then perform a simple change of variables and replace N and Π by the generalized
number system αN (x1/α ) and αΠ(x1/α ).
number systems as defined above (cf. [6]). Therefore, we only work with Π in order
to gain generality.
2.2. Fourier transforms and distributions

∞
Fourier transforms, normalized as fˆ(t) = −∞ e−itx f (x) dx, will be taken in the
sense of tempered distributions; see [5, 18] for distribution theory. The standard
Schwartz spaces of compactly supported and rapidly decreasing smooth test func-
tions are denoted as usual by D(R) and S(R), while D (R) and S (R) stand for
their topological duals, the spaces of distributions and tempered distributions. The
dual pairing between a distribution f and a test function φ is denoted as f, φ,
or as f (u), φ(u) with the use of a variable of evaluation; when f is a regular
∞
distribution we of course have f (u), φ(u) = −∞ f (u)φ(u)du. For f ∈ S (R), its
Fourier transform fˆ ∈ S (R) is defined via duality as fˆ(t), φ(t) := f (u), φ̂(u),
for each φ ∈ S(R). If f ∈ S (R) has support in [0, ∞), its Laplace transform is
L {f ; s} := f (u), e−su , e s > 0, and its Fourier transform fˆ is the distributional
boundary value of L {f ; s} on e s = 0.
We also mention that asymptotic estimates O(g(x)) are meant for x ! 1
unless otherwise specified.
3. A Tauberian theorem with remainder

The following Tauberian remainder theorem for Laplace transforms will be our
main tool for translating information on the zeta function of a generalized number
system into asymptotic properties for the functions N and Π in the next section.
Theorem 2 extends a Tauberian result by Nyman (cf. [12, Lemma II]). We point
out that our O-constants hereafter depend on the parameter n ∈ N.
Theorem 2. Suppose S is non-decreasing and T is a function of (locally) bounded
variation such that it is absolutely continuous for large arguments and T (x) ≤ Aex
with A ≥ 0. Let both functions have support in [0, ∞). Assume that
∞
G(s) = e−su (dS(u) − dT (u)) is convergent for e s > 1
0−
and can be extended to a C ∞ -function on the line es = 1, admitting the following
bounds:
β
G(n) (1 + it) = O(|t| n ) for each n ∈ N, (8)
where the βn are such that
βn
lim= 0. (9)
n→∞ n
Then, the ensuing asymptotic formula holds:
ex
S(x) = T (x) + O , for all n ∈ N. (10)
xn
Proof. Clearly, by enlarging the exponents in (8) if necessary, we may assume that
βn is a non-decreasing sequence of positive numbers. Modifying T on finite intervals
does not affect the rest of the hypotheses, so we assume that T is locally absolutely
continuous on the whole [0, ∞) and that the upper bound on its derivative holds
globally. Furthermore, we may assume without loss of generality that T (x) ≥ 0.
Indeed, if necessary we may replace S by S + T− and T by T+ , where T (x) =
T+ (x) − T− (x) with T+ and T− the distribution functions of the positive and
negative parts of T . Since T (x) = O(ex ), the Laplace–Stieltjes transform of S
also converges on e s > 1. Thus,
x ∞
S(x) = dS(u) ≤ eσx e−σu dS(u) = Oσ (eσx ), σ > 1.
0− 0−
−x
Let us define Δ(x) = e (S(x) − T (x)) and calculate its Laplace transform,
∞ ∞
1
L{Δ; s} = e (−s−1)u
(S(u) − T (u))du = e(−s−1)u d(S − T )(u)
0 1 + s 0−
1 G(s + 1)
= L{dS − dT ; s + 1} = , e s > 0,
1+s s+1
where we have used that Δ(x) = o(eηx ) for each η > 0. Setting s = σ + it and
letting σ → 0+ in this expression in the space S (R), we obtain that the Fourier
transform of Δ is the smooth function
G(1 + it)
Δ̂(t) = .
1 + it
Since βn is non-decreasing, we obtain the estimates
Δ̂(n) (t) = O((1 + |t|)βn −1 ). (11)
We now derive a useful Tauberian condition on Δ from the assumptions on
S and T . If x ≤ y ≤ x + min{Δ(x)/2A, log(4/3)} and Δ(x) > 0, we find, by using
the upper bound on T ,
S(y) − T (y) S(x) − T (x) ex ex Δ(x)
Δ(y) = y
≥ x y
− A(y − x) ≥ Δ(x) y −
e e e e 2
Δ(x)
≥ .
4
Similarly one can show that
−Δ(y) ≥ −Δ(x)/2 if x + Δ(x)/2A ≤ y ≤ x and Δ(x) < 0.
We now estimate Δ(h) in the case Δ(h) > ∞0. Set ε = min{Δ(h)/2A, log(4/3)} and
choose φ ∈ D(0, 1) such that φ ≥ 0 and −∞ φ(x)dx = 1. We obtain
x
1 ε
Δ(h) = Δ(h)φ dx
ε 0 ε
ε x
4 2 ∞
≤ Δ(x + h)φ dx = Δ̂(t)eiht φ̂(−εt)dt
ε 0 ε π −∞
∞ (n)
2 iht
= e Δ̂(t)φ̂(−εt) dt
(ih)n π −∞
∞
2 n (j) t
n
≤ n Δ̂ φ̂(n−j)
(−t) εn−j−1 dt
h π j ε
j=0 −∞
1
=O ,
hn ε β n
where we have used φ̂ ∈ S(R) and (11). If Δ(x)< 0 one gets an analogous estimate
∞
by using a φ ∈ D(−1, 0) such that φ ≥ 0 and −∞ φ(x)dx = 1. If ε = log(4/3), it
clearly follows that Δ(h) = o(1) and we may thus assume that ε = Δ(h)/2A. This
gives that Δ(h) = On (h−n/(βn +1) ) which proves (10) because of (9).
We will also need a converse result, an Abelian counterpart. It is noteworthy
that the bounds for G(n) (1 + it) we get from the converse result are actually much
better than the ones needed for Theorem 2.
Proposition 1. Let S be a non-decreasing function, let T be of (locally) bounded
variation such that it is absolutely continuous for large arguments and T (x) ≤ Aex
for some positive A, and let both functions have support in [0, ∞). Suppose that
the asymptotic estimate (10) holds for all n. Then,
∞
G(s) = e−su (dS(u) − dT (u)) is convergent for e s ≥ 1.
0−
Furthermore, G is C ∞ on e s = 1 and for each ε > 0 and n ∈ N its nth derivative
satisfies the bound
G(n) (σ + it) = O((1 + |t|)ε ), σ ≥ 1, t ∈ R, (12)
−1 x
with global Oε,n -constants. Moreover, if T (x) ≤ Bx e for some positive B and
x ! 1, then the better asymptotic estimate
G(σ + it) = o(log |t|) (13)
is valid uniformly for σ ≥ 1 as |t| → ∞.
Proof. As in the proof of Theorem 2, we may assume that T is locally absolutely
continuous on [0, ∞) and 0 ≤ T (x) ≤ Aex . From the assumptions it is clear
that S as well as T are O(ex ). The asymptotic estimates (10) obviously give the
convergence of G(s) for e s ≥ 1 and the fact that G is C ∞ on e s = 1. Let us
now show the asymptotic bounds (12). It is clear that it holds with ε = 0 on the
half-plane σ ≥ 2. We thus restrict our attention to the strip 1 ≤ σ < 2. We keep
|t| ≥ 1. Let X ! 1 be a constant, which we will specify later. We have
X
X
−sx −sx
G(s) = e dS(x) − e T (x)dx + T (0)
0− 0
(14)
∞
+s e−sx (S(x) − T (x)) dx + e−sX (S(X) − T (X)) .
X
We differentiate the above formula n times and bound each term separately. The
first term can be estimated by

X X
−sx
e (−x) dS(x) ≤
n
e−x xn dS(x)
0− 0−
X X
= e−X X n S(X) + e−x xn S(x)dx − n e−x xn−1 S(x)dx
0 0
≤ CX n+1 ,
as S is non-decreasing and O(ex ). The second term from (14) can be bounded in
a similar way by this quantity, while the last term is even O(1). It thus remains
to bound the third term from (14). Suppose that S(x) − T (x) = O(ex x−γ ), where
γ > n + 1, then
∞ ∞

e x (S(x) − T (x)) dx ≤
−sx n
xn−γ dx ≤ C X n−γ+1 .

X X
1/γ
Combining these inequalities and choosing X = |t| , we obtain
n+1
(n)
G (σ + it) ≤ C X n+1 + C (2 + |t|)X n−γ+1 = O |t| γ .
Since γ can be chosen arbitrarily large, (12) follows.
The proof of (13) is similar if we work under the assumption T (x) ≤ Bx−1 ex .
This bound implies that T (x) " Li(ex ) = O(x−1 ex ), which gives S(x) = O(x−1 ex )
as well. The starting point is again the formula (14) for G. Via the same reasoning
as above, the bounds for the first and second term, in case n = 0, can be improved
to O(log X). Employing the same bound for the third term, we obtain the result
1/(γ−1)
after choosing X = |t| and letting γ → ∞.
4. The PNT with Nyman’s remainder

We establish in this section the following general form of Theorem 1:
Theorem 3. For a generalized number system, the following four statements are
equivalent:
(i) For some a > 0, the generalized integer distribution function N satisfies
x
N (x) = ax + O , for all n ∈ N. (15)
logn x
(ii) For some a > 0, the function
a
G(s) = ζ(s) − (16)
s−1
has a C ∞ -extension to e s ≥ 1 and there is some ε > 0 such that
G(n) (1 + it) = O(|t|ε ), for all n ∈ N. (17)
(iii) For some a > 0 and each ε > 0, the function (16) satisfies
G(n) (σ + it) = O((1 + |t|)ε ), σ > 1, t ∈ R, for all n ∈ N, (18)
with global Oε,n -constants.
(iv) The Riemann prime distribution function Π satisfies
Π(x) = Li(x) + O ( x/ (logn x)) , for all n ∈ N. (19)
Remark 1. The condition (iii) implies the apparently stronger assertion that G
has a C ∞ -extension to e s ≥ 1 and that (18) remains valid for σ ≥ 1, as follows
from a standard local L∞ weak∗ compactness argument.
Before giving a proof of Theorem 3, we make a comment on reference [12].
Therein, Nyman stated that the conditions (i) and (iv) from Theorem 3 were also
equivalent to: for each ε > 0 and n ∈ N
−1
ζ (n) (σ + it) = O(|t|ε ) and (ζ(σ + it)) = O(|t|ε ), (20)
uniformly on the region σ > 1 and |t| ≥ ε. It was noticed by Ingham in Mathe-
matical Reviews [8] that (20) fails to be equivalent to (15) and (19). In fact (20)
can hardly be equivalent to any of these two asymptotic formulas because it does
not involve any information about ζ near s = 1, contrary to our conditions (ii)
and (iii). A large number of counterexamples to Nyman’s statement can easily be
found among zeta functions arising as generating functions from analytic combi-
natorics and classical number theory. We discuss three examples here, the first of
them is due to Ingham [8], while the second one was suggested by W.-B. Zhang.
Example 1. Consider the generalized primes given by pk = 2k . The prime count-
ing function for these generalized primes clearly satisfies π(x) = log x/ log 2 + O(1)
and therefore (19) does not hold. The bound π(x) = O(log x) gives that its asso-
ciated zeta function is analytic on e s > 0 and satisfies ζ (n) (s) = O(1) uniformly
on any half-plane e s ≥ σ0 > 0. We also have the same bound for 1/ζ(s) be-
cause |ζ(σ)||ζ(σ + it)| ≥ 1, which follows from the trivial inequality 1 + cos θ ≥ 0
(see the 3-4-1 inequality in the proof of Lemma 1 below). In particular, Nyman’s
condition (20) is fulfilled. The generalized integer counting function N does not
satisfy (15), because, otherwise,
$ ζ would have a simple pole at s = 1. Interestingly,
in this example N (x) = 2k ≤x p(k), where p is the unrestricted partition func-
tion, which, according to the celebrated Hardy–Ramanujan–Uspensky formula,
has asymptotics √ 2n
eπ 3
p(n) ∼ √ . (21)
4n 3
From (21) one easily deduces

2 log x
π 3 log 2
e
N (x) ∼ A √ , (22)
log x
√ √
with A = (2π 2)−1 log 2, but (21) and (22) simultaneously follow from Ingham’s
theorem for abstract partitions [7].
Example 2. A simple example is provided by the generalized prime number system

2, 2, 3, 3, 5, 5, . . . , p, p, . . . ,
that is, the generalized primes consisting of ordinary rational primes p each taken
exactly twice. The set of generalized integers for this example then consists of
ordinary rational integers n, each repeated d(n) times, where d(n) is the classical
divisor function. In this case the associated zeta function to this number system is
the square of the Riemann zeta function, which clearly satisfies Nyman’s condition
(20). On the other hand, Dirichlet’s well-known asymptotic estimate for the divisor
summatory function and the classical PNT yield
√
N (x) = d(n) = x log x + (2γ − 1)x + O( x)
n≤x
and .
Π(x) = 2 Li(x) + O(x exp(−c log x)).
Example 3. This example and Example 1 are of similar nature. This time we use
generalized integers that arise as coding numbers of certain (non-planar) rooted
trees via prime factorization [11]. Consider the set of generalized primes given by
the subsequence {p2k }∞ ∞
k=0 of ordinary rational primes, where {pk }k=1 are all ra-
tional primes enumerated in increasing order. Using the classical PNT for rational
primes, one verifies that the prime counting function π of these generalized primes
satisfies
log x log log x
π(x) = − + O(1).
log 2 log 2
By the same reason as above, one obtains that the zeta function of these generalized
numbers satisfies Nyman’s condition (20). The generalized integers corresponding
to this example are actually the Matula numbers of rooted trees of height ≤2,
whose asymptotic distribution was studied in [15]; its generalized integer counting
function N satisfies
√
/
2
log(π/ 6 log 2) 2 log x (log log x)
N (x) ∼ A(log x) 2 log 2 exp π − ,
3 log 2 8 log 2
for a certain constant A > 0, see [15, Thm. 1].
The rest of this section is dedicated to the proof of Theorem 3. First we
derive some bounds on the inverse of the zeta function and the non-vanishing of ζ
on e s = 1.
Lemma 1. Suppose that condition (iii) from Theorem 3 holds. Then, (s − 1)ζ(s)
has no zeros on es ≥ 1 and, in particular, 1/ζ(s) has a C ∞ -extension to es ≥ 1
as well. Furthermore, for each ε > 0,
1
= O ((1 + |t|)ε ) , σ ≥ 1, t ∈ R, (23)
ζ(σ + it)
with a global Oε -constant.
Proof. We use (iii) in the form stated in Remark 1. The non-vanishing property
of ζ follows already from results of Beurling [2], but, since we partly need the
argument in the process of showing (23), we also prove this fact for the sake of
completeness. Let t = 0. We closely follow Hadamard’s classical argument [9] based
on the elementary 3-4-1 trigonometric inequality, that is,
P (θ) := 3 + 4 cos(θ) + cos(2θ) ≥ 0.
∞
Using the expression ζ(s) = exp 1− x−s dΠ(x) and the 3-4-1 inequality, one
derives, for 1 < η,
∞
3 log |ζ(η)| + 4 log |ζ(η + it)| + log |ζ(η + 2it)| = x−η P (t log x)dΠ(x) ≥ 0,
1−
namely, 3
ζ (η)ζ 4 (η + it)ζ(η + 2it) ≥ 1.
This 3-4-1 inequality for ζ already implies that 1/ζ(η + it) = O(1) uniformly on
η ≥ 2. We assume in the sequel that 1 < η < 2. Since ζ(η) ∼ a/(η − 1) as η → 1+ ,
we get

(η − 1)3 ≤ (η − 1)3 ζ 3 (η) ζ 4 (η + it) |ζ(η + 2it)|
4 ε
≤ A |ζ(η + it)| |t| . (24)
As is well known, (24) yields that ζ(1 + it) does not vanish for t = 0. Indeed, if
ζ(1 + it0 ) = 0, the fact that ζ(s) and ζ (s) have continuous extensions to e s = 1
η
would imply (η − 1)3 = O(|ζ(η + it0 )|4 ) = O(( 1 |ζ (λ + it0 )|dλ)4 ) = O((η − 1)4 ),
a contradiction. The assertions about the C ∞ -extensions of (s − 1)ζ(s) and 1/ζ(s)
must be clear, in particular 1/ζ(1) = 0.
Let us now establish the bound (23) on the range 1 ≤ σ ≤ 2. We keep here
|t| ! 1. If 1 ≤ σ ≤ η < 2, we find
η

|ζ(σ + it) − ζ(η + it)| = ζ (u + it)du ≤ A (η − 1) |t| ,
ε
where we have used the bound (17) for ζ . Combining this inequality with (24),
we find
|ζ(σ + it)| ≥ |ζ(η + it)| − A (η − 1) |t|ε
(η − 1)3/4
− A (η − 1) |t| .
ε
≥ ε/4
A1/4 |t|
Now choose η = η(t) in such a way that
(η − 1)3/4
= 2A (η − 1) |t| ,
ε
ε/4
A1/4 |t|
i.e.,
1 A
η =1+ = 1 + ,
A(2A )4 |t|5ε |t|5ε
assuming t large enough to ensure η < 2. Then, in this range,

−4ε
|ζ(σ + it)| ≥ A (η − 1) |t| = A A |t|
ε
.
−5ε
For the range 1 + A |t| ≤ σ ≤ 2, the estimate (24) with σ instead of η yields
exactly the same lower bound.
We now aboard the proof of Theorem 3.
Proof of Theorem 3. Upon setting S(x) = N (ex ) and T (x) = aex , so that
a
G(s) = L{dS − dT ; s} = ζ(s) − ,
s−1
Theorem 2 gives the implication (ii)⇒(i), Proposition 1 yields (i)⇒(iii), whereas
(iii)⇒(ii) follows from Remark 1. So, the first three conditions are equivalent and
it remains to establish the equivalence between any of these statements and (iv).
(iii)⇒(iv). We now set S1 (x) := Π(ex ) and
ex 1 − 1
y
T1 (x) := dy = Li(ex ) − log x + A, x ≥ 0.
1 log y
A quick calculation gives an explicit expression for G1 (s) := L{dS1 − dT1 ; s},
namely,
G1 (s) = log ζ(s) − log s + log(s − 1) = log((s − 1)ζ(s)) − log s, (25)
with the principal branch of the logarithm. By Remark 1, Lemma 1, and the
Leibniz rule, we obtain that G1 (1 + it) ∈ C ∞ (R) and bounds G1 (1 + it) =
(n)
ε
Oε,n (|t| ), |t| ! 1. Another application of Theorem 2 yields (19).
(iv)⇒(ii). Conversely, let (19) hold and retain the notation S1 , T1 , and G1
as above. We apply Proposition 1 to S1 and T1 to get that (25) admits a C ∞ -
ε
extension to e s = 1 and all of its derivatives on that line are bounded by O(|t| )
for each ε > 0. This already yields that the function G(s) given by (16), no matter
the value of the constant a, has also a C ∞ -extension to e s = 1 except possibly
at s = 1. Moreover, since T1 (x) = O(ex /x), we even get from Proposition 1 that
ε
G1 (t) = o(log |t|) for |t| ! 1, or, which amounts to the same, ζ(1+it) = O(|t| ), for
each ε > 0. Thus, by this bound and the bounds on the derivatives of log ζ(1 + it),
ε
we have that ζ (n) (1 + it) = O(|t| ), as can easily be deduced by induction with the
aid of the Leibniz formula.
Summarizing, we only need to show that there exists a > 0 for which ζ(s) −
a/(s − 1) has a C ∞ -extension on the whole line e s = 1. The function log((s −
1)ζ(s)) however admits a C ∞ -extension to this line, and its value at s = 1 coincides
with that of the function G1 , as shown by the expression (25). Therefore, (s−1)ζ(s)
also extends to e s ≥ 1 as a C ∞ -function, and its value at s = 1 can be calculated
as a = limσ→1+ eG1 (σ) = eG1 (1) > 0, because G1 (σ) is real-valued when σ is real.
Hence ζ(s) − a/(s − 1) has also a C ∞ -extension to e s ≥ 1. (This follows from
the general fact that t−1 (f (t) − f (0)) is C k−1 for a C k -function f .) This concludes
the proof of the theorem.
5. A Cesàro version of the PNT with remainder

In this last section we obtain an average version of Theorem 3 where the remainders
in (15) and (19) are taken in the Cesàro sense. The motivation of this new PNT
comes from a natural replacement of (ii), or equivalently (iii), by a certain weaker
growth requirement on ζ.
Let us introduce some function and distribution spaces. The space OC (R)
consists of all g ∈ C ∞ (R) such that there is some β ∈ R with g (n) (t) = On (|t|β ),
for each n ∈ N. This space is well known in distribution theory. When topologized

in a canonical way, its dual space OC (R) corresponds to the space of convolutors of
the tempered distributions [5, 13]. Another well-known space is that of multipliers
of S (R), denoted as OM (R) and consisting of all g ∈ C ∞ (R) such that for each
n ∈ R there is βn ∈ R such that g (n) (t) = On (|t|βn ). Of course, we have the
inclusion relation OC (R) OM (R).
Observe that condition (ii) from Theorem 3 precisely tells that for some a > 0
the analytic function G(s) = ζ(s) − a/(s − 1) has boundary values on e s = 1 in
the space OC (R), that is, G(1 + it) ∈ OC (R). We now weaken this membership
relation to G(1 + it) ∈ OM (R). To investigate the connection between the latter
condition and the asymptotic behavior of N and Π, we need to use asymptotics in
the Cesàro sense. For a locally integrable function E, with support in [0, ∞), and
α ∈ R, we write
x
E(x) = O (C) (x → ∞) (26)
logα x
if there is some (possibly large) m ∈ N such that the following average growth
estimate holds: x
E(u) u m x
1− du = O . (27)
0 u x logα x
The order m of the Cesàro–Riesz mean to be taken in (27) is totally irrelevant
for our arguments below and we therefore choose to omit it from the notation
in (26). The meaning of an expression f (x) = g(x) + O (x/ logα x) in the Cesàro
sense should be clear. We remark that Cesàro asymptotics can also be defined
for distributions, see [5, 13]. The notion of Cesàro summability of integrals is well
known, see, e.g., [5].
We have the following PNT with remainder in the Cesàro sense:
Theorem 4. For a generalized number system the following four statements are
equivalent:
(i) For some a > 0, the generalized integer distribution function N satisfies
N (x) = ax + O ( x/ (logn x)) (C), for all n ∈ N. (28)
(ii) For some a > 0, the function
a
G(s) = ζ(s) − (29)
s−1
has a C ∞ -extension to e s ≥ 1 and G(1 + it) ∈ OM (R).
(iii) For some a > 0, there is a positive sequence {βn }∞

n=0 such that the function
(29) satisfies
G(n) (s) = O((1 + |s|)βn ), for all n ∈ N, (30)
on e s > 1 with global On -constants.
(iv) The Riemann prime distribution function Π satisfies
x
Π(x) = Li(x) + O (C), for all n ∈ N. (31)
logn x
We indicate that, as in Remark 1, the bounds (30) also imply that G has a
C ∞ -extension to e s ≥ 1 and that (30) remains valid on e s ≥ 1. Naturally, the
PNT (31) delivered by Theorem 4 is much weaker than (19). Before discussing the
proof of Theorem 4, we give a family of examples of generalized number systems
which satisfy condition (ii) from Theorem 4 but not those from Theorem 3.
Example 4. The family of continuous generalized number systems whose Riemann
prime distribution functions are given by
x
1 − cos(logα u)
Πα (x) = du , for α > 1,
1 log u
was studied in [3]. It follows from [3, Thm. 3.1] that there are constants aα such
that their zeta functions have the property that Gα (s) = ζα (s) − aα /(s − 1) are
entire. In this case, [3, Thm. 3.1] also implies that Gα (1 + it) ∈ OM (R), but it
does not belong to OC (R).
We need some auxiliary results in order to establish Theorem 4. The next
theorem is of Tauberian character. Part of its proof is essentially the same as that
of [3, Lemma 2.1], but we include it for the sake of completeness.
Theorem 5. Let E be of locally bounded variation with support on [1, ∞) and
suppose that E(x) = O(x) (C). Set
∞
F (s) = x−s dE(x) (C), e s > 1.
1−
Then, E satisfies (26) for every α > 0 if and only if F has a C ∞ -extension to
e s ≥ 1 that satisfies F (1 + it) ∈ OM (Rt ). If this is the case, then there is a
sequence {βn }∞
n=0 such that for each n
F (n) (s) = O((1 + |s|)βn ), on e s ≥ 1. (32)

Furthermore, assume additionally that
x
x
V (E, [1, x]) = |dE|(u) = O , (33)
1− log x
where |dE| stands for the total variation measure of dE. Then,
F (s) = O(log(1 + |$m s|)), on e s ≥ 1. (34)
Proof. Note that the Cesàro growth assumption implies that F (s) is Cesàro sum-
mable for e s > 1 and therefore analytic there. Let F1 (s) = F (s)/s and R(u) =
e−u E(eu ). It is clear that F1 (s) has a C ∞ -extension to e s = 1 that satisfies
F (1 + it) ∈ OM (R) if and only if F1 has the same property. The latter property

holds if and only if R ∈ OC (R). Indeed, since R ∈ S (R) and F1 (s + 1) = L{R; s},
we obtain that R̂(t) = F1 (1 + it), whence our claim follows because the spaces

OC (R) and OM (R) are in one-to-one correspondence via the Fourier transform [5].

Now, by definition of the convolutor space, R ∈ OC (R) if and only if
∞
R(u + h)φ(u)du = O(h−α ),
−∞
for each α > 0 and φ ∈ D(R) [13]. Writing h = log λ and φ(x) = ex ϕ(ex ), we

obtain that R ∈ OC (R) if and only if E(x)/x has the quasiasymptotic behavior
[5, 13]
E(λx) 1
=O , λ → ∞ , in D(0, ∞) , (35)
λx logα λ
which explicitly means that
∞
E(λx) 1
ϕ(x)dx = O , λ → ∞,
1 λx logα λ
for every test function ϕ ∈ D(0, ∞). Using [13, Thm. 2.37, p. 154], we obtain
that the quasiasymptotic behavior (35) in the space D(0, ∞) is equivalent to the
same quasiasymptotic behavior in the space D(R), and, because of the structural
theorem for quasiasymptotic boundedness [13, Thm. 2.42, p. 163] (see also [16, 17]),

we obtain that R ∈ OC (R) is equivalent to the Cesàro behavior (26) for every α.
Note that we have E(x) logn x = O(x/ logα x) (C) for every α > 0 as well.
So the bounds (32) can be obtained from these Cesàro asymptotic estimates by
integration by parts. The bound (34) under the assumption (33) can be shown via
a similar argument to the one used in the proof of Proposition 1. It is enough to
show the bound for σ = e s > 1. Consider the splitting
eX ∞
−s
F (s) = x dE(x) + x−s dE(x),
1− eX
with X ! 1. We can actually assume that 1 < σ < 2 and |t| ! 1 because
otherwise F is already bounded in view of (33). The first term in this formula is
clearly O(log X) because of (33). We handle the second term via integration by
parts. Let Em be an m-primitive of E(x)/x such that Em (x) = O(xm / log2 x).
The absolute value of the term in question is then
∞ ∞
Em (x) dx |t|m+1

≤ |s| · · · |s + m| C + dx ≤ C |t| m+1
= C ,
X xs+m m 2
X x log x
m
X
e e
and we obtain F (s) = O(log |t|) by taking X = |t| m+1

.
With the same technique as the one employed in Lemma 1, one shows the
following bound on the inverse of ζ:
Lemma 2. Suppose that condition (iii) from Theorem 4 is satisfied. Then, (s−1)ζ(s)
has no zeros on es ≥ 1 and, in particular, 1/ζ(s) has a C ∞ -extension to es ≥ 1
as well. Furthermore, there is some β > 0 such that
1
= O((1 + |s|)β ), on e s ≥ 1.
ζ(s)
Let us point out that the Cesàro asymptotics (28) always leads to N (x) ∼ ax,
while (31) leads to Π(x) ∼ x/ log x, which can be shown by standard Tauberian
arguments. This comment allows us the application of Theorem 5 to the functions
E1 (x) = N (x) − ax and E2 (x) = Π(x) − Li(x).
The rest of the proof goes exactly along the same lines as that of Theorem 3
(using Theorem 5 instead of Theorem 2 and Proposition 1), and we thus omit the
repetition of details. So, Theorem 4 has been established.
Acknowledgement
G. Debruyne gratefully acknowledges support by Ghent University, through a BOF
Ph.D. grant. The work of J. Vindas was supported by the Research Foundation –
Flanders, through the FWO-grant number 1520515N.
References
[1] Bateman, P.T., Diamond, H.G.: Asymptotic distribution of Beurling’s generalized
prime numbers. In: Studies in Number Theory, pp. 152–210. Math. Assoc. Amer.,
Prentice-Hall, Englewood Cliffs, N.J. (1969)
[2] Beurling, A.: Analyse de la loi asymptotique de la distribution des nombres premiers
généralisés. Acta Math. 68, 255–291 (1937)
[3] Debruyne, G., Schlage-Puchta, J.-C., Vindas, J.: Some examples in the theory of
Beurling’s generalized prime numbers. Acta Arith. 176, 101–129 (2016).
[4] Diamond, H.G.: Asymptotic distribution of Beurling’s generalized integers. Illinois
J. Math. 14, 12–28 (1970)
[5] Estrada, R., Kanwal, R.P.: A distributional approach to asymptotics. Theory and
applications. Birkhäuser, Boston (2002)
[6] Hilberdink, T.: Generalized prime systems with periodic integer counting function.
Acta Arith. 152, 217–241 (2012)
[7] Ingham, A.E.: A Tauberian theorem for partitions. Ann. of Math. 42, 1075–1090
(1941)
[8] Ingham, A.E.: Review #MR0032693 (11,332b). Mathematical Reviews, American
Mathematical Society
[9] Ingham, A.E.: The distribution of prime numbers. Cambridge University Press, Cam-
bridge (1990)
[10] Kahane, J.-P.: Sur les nombres premiers généralisés de Beurling. Preuve d’une con-
jecture de Bateman et Diamond. J. Théor. Nombres Bordeaux 9, 251–266 (1997)
[11] Matula, D.W.: A natural rooted tree enumeration by prime factorization. SIAM Rev.
10, 273 (1968)
[12] Nyman, B.: A general prime number theorem. Acta Math. 81, 299–307 (1949)
[13] Pilipović, S., Stanković, B., Vindas, J.: Asymptotic behavior of generalized functions.
Series on Analysis, Applications and Computation, 5. World Scientific Publishing Co.
Pte. Ltd., Hackensack, NJ (2012)
[14] Schlage-Puchta, J.-C., Vindas, J.: The prime number theorem for Beurling’s gener-
alized numbers. New cases. Acta Arith. 153, 299–324 (2012)
[15] Vernaeve, H., Vindas, J., Weiermann, A.: Asymptotic distribution of integers with
certain prime factorizations. J. Number Theory 136, 87–99 (2014)
[16] Vindas, J.: Structural theorems for quasiasymptotics of distributions at infinity.
Publ. Inst. Math. (Beograd) (N.S.) 84, 159–174 (2008)
[17] Vindas, J.: The structure of quasiasymptotics of Schwartz distributions. In: Linear
and non-linear theory of generalized functions and its applications, pp. 297–314.
Banach Center Publ. 88, Polish Acad. Sc. Inst. Math., Warsaw (2010)
[18] Vladimirov, V.S.: Methods of the theory of generalized functions. Analytical Methods
and Special Functions, 6. Taylor & Francis, London (2002)
[19] Zhang, W.-B.: Extensions of Beurling’s prime number theorem. Int. J. Number The-
ory 11, 1589–1616 (2015)
Gregory Debruyne and Jasson Vindas

Ghent University
[email protected]
Operator Theory:

Inverse Function Theorems for

Generalized Smooth Functions
Paolo Giordano and Michael Kunzinger
Abstract. Generalized smooth functions are a possible formalization of the

original historical approach followed by Cauchy, Poisson, Kirchhoff, Helm-
holtz, Kelvin, Heaviside, and Dirac to deal with generalized functions. They
are set-theoretical functions defined on a natural non-Archimedean ring, and
include Colombeau generalized functions (and hence also Schwartz distribu-
tions) as a particular case. One of their key property is the closure with
respect to composition. We review the theory of generalized smooth functions
and prove both the local and some global inverse function theorems.
1. Introduction
Since its inception, category theory has underscored the importance of unrestricted
composition of morphisms for many parts of mathematics. The closure of a given
space of “arrows” with respect to composition proved to be a key foundational
property. It is therefore clear that the lack of this feature for Schwartz distributions
has considerable consequences in the study of differential equations [28, 14], in
mathematical physics [4, 6, 8, 10, 18, 21, 25, 41, 42, 43, 46], and in the calculus of
variations [30], to name but a few.
On the other hand, Schwartz distributions are so deeply rooted in the linear
framework that one can even isomorphically approach them focusing only on this
aspect, opting for a completely formal/syntactic viewpoint and without requiring
any functional analysis, see [49]. So, Schwartz distributions do not have a notion
of pointwise evaluation in general, and do not form a category, although it is well
known that certain subclasses of distributions have meaningful notions of pointwise
evaluation, see, e.g., [35, 36, 47, 45, 16, 15, 51].
This is even more surprising if one takes into account the earlier histori-
cal genesis of generalized functions dating back to authors like Cauchy, Poisson,
Kirchhoff, Helmholtz, Kelvin, Heaviside, and Dirac, see [29, 33, 34, 50]. For them,
this “generalization” is simply accomplished by fixing an infinitesimal or infinite
96 P. Giordano and M. Kunzinger
parameter in an ordinary smooth function, e.g., an infinitesimal and invertible

standard deviation in a Gaussian probability density. Therefore, generalized func-
tions are thought of as some kind of smooth set-theoretical functions defined and
valued in a suitable non-Archimedean ring of scalars. From this intuitive point of
view, they clearly have point values and form a category.
This aspect also bears upon the concept of (a generalized) solution of a
differential equation. In fact, any theory of generalized functions must have a
link with the classical notion of (smooth) solution. However, this classical notion
is deeply grounded on the concept of composition of functions and, at the same
time, it is often too narrow, as is amply demonstrated, e.g., in the study of PDE in
the presence of singularities. In our opinion, it is at least not surprising that also the
notion of distributional solution did not lead to a satisfying theory of nonlinear
PDE (not even of singular ODE). We have hence a wild garden of flourishing
equation-dependent techniques and a zoo of counter-examples. The well-known
detaching between these techniques and numerical solutions of PDE is another
side of the same question.
One can say that this situation presents several analogies with the classical
compass-and-straightedge solution of geometrical problems, or with the solution of
polynomial equations by radicals. The distinction between algebraic and irrational
numbers and the advent of Galois theory were essential steps for mathematics
to start focusing on a different concept of solution, frequently nearer to applied
problems. In the end, these classical problems stimulated more general notions of
geometrical transformation and numerical solution, which nowadays have super-
seded their origins. The analogies are even greater when observing that first steps
toward a Galois theory of nonlinear PDE are arising, see [5, 7, 38, 39].
Generalized smooth functions (GSF) are a possible formalization of the orig-
inal historical approach of the aforementioned classical authors. We extend the
field of real numbers into a natural non Archimedean ring ρ R and we consider the
simplest notion of smooth function on the extended ring of scalars ρ R. To define
a GSF f : X −→ Y , X ⊆ ρ Rn , Y ⊆ ρ Rd , we simply require the minimal logical
conditions so that a net of ordinary smooth functions fε ∈ C ∞ (Ωε , Rd ), Ωε ⊆ Rn ,
defines a set-theoretical map X −→ Y which is infinitely differentiable; see below
for the details. This freedom in the choice of domains and codomains is a key
property to prove that GSF are closed with respect to composition. As a result,
GSF share so many properties with ordinary smooth functions that frequently we
only have to formally generalize classical proofs to the new context. This allows
an easier approach to this new theory of generalized functions.
It is important to note that the new framework is richer than the classical
one because of the possibility to express non-Archimedean properties. So, e.g., two
different infinitesimal standard deviations in a Gaussian result in infinitely close
Dirac-delta-like functionals but, generally speaking, these two GSF could have
different infinite values at infinitesimal points h ∈ ρ R. For this reason, Schwartz
distributions are embedded as GSF, but this embedding is not intrinsic and it has
Inverse Function Theorems for Generalized Smooth Functions 97
to be chosen depending on the physical problem or on the particular differential

equation we aim to solve.
In the present work, we establish several inverse function theorems for GSF.
We prove both the classical local and also some global versions of this theorem.
It is remarkable to note that the local version is formally very similar to the
classical one, but with the sharp topology instead of the standard Euclidean one.
We also show the relations between our results and the inverse function theorem
for Colombeau functions established by using the discontinuous calculus of [2, 3].
The paper is self-contained in the sense that it contains all the statements of
results required for the proofs of the new inverse function theorems. If proofs of
preliminaries are omitted, we give references to where they can be found.
1.1. Basic notions
The ring of generalized scalars. In this work, I denotes the interval (0, 1] ⊆ R and
we will always use the variable ε for elements of I; we also denote ε-dependent nets
x ∈ RI simply by (xε ). By N we denote the set of natural numbers, including zero.
We start by defining the non-Archimedean ring of scalars that extends the
real field R. For all the proofs of results in this section, see [19, 18].
Definition 1. Let ρ = (ρε ) ∈ RI be a net such that limε→0+ ρε = 0+ , then
(i) I(ρ) := {(ρ−a
ε ) | a ∈ R>0 } is called the asymptotic gauge generated by ρ.
(ii) If P(ε) is a property of ε ∈ I, we use the notation ∀0 ε : P(ε) to denote
∃ε0 ∈ I ∀ε ∈ (0, ε0 ] : P(ε). We can read ∀0 ε as for ε small.
(iii) We say that a net (xε ) ∈ RI is ρ-moderate, and we write (xε ) ∈ Rρ if
∃(Jε ) ∈ I(ρ) : xε = O(Jε ) as ε → 0+ .
(iv) Let (xε ), (yε ) ∈ RI , then we say that (xε ) ∼ρ (yε ) if ∀(Jε ) ∈ I(ρ) : xε =
yε + O(Jε−1 ) as ε → 0+ . This is an equivalence relation on the ring Rρ of
moderate nets with respect to pointwise operations, and we can hence define
ρ
R := Rρ / ∼ρ ,
which we call Robinson–Colombeau ring of generalized numbers, [48, 8]. We
denote the equivalence class x ∈ ρ R simply by x =: [xε ] := [(xε )]∼ ∈ ρ R.
In the following, ρ will always denote a net as in Definition 1, and we will
use the simpler notation R for the case ρε = ε. The infinitesimal ρ can be chosen
depending on the class of differential equations we need to solve for the gener-
alized functions we are going to introduce, see [20]. For motivations concerning
the naturality of ρ R, see [18]. We also use the notation dρ := [ρε ] ∈ ρ R and
dε := [ε] ∈ (ε) R.
We can also define an order relation on ρ R by saying [xε ] ≤ [yε ] if there
exists (zε ) ∈ RI such that (zε ) ∼ρ 0 (we then say that (zε ) is ρ-negligible) and
xε ≤ yε + zε for ε small. Equivalently, we have that x ≤ y if and only if there
exist representatives (xε ), (yε ) of x, y such that xε ≤ yε for all ε. Clearly, ρ R
is a partially ordered ring. The usual real numbers r ∈ R are embedded in ρ R
considering constant nets [r] ∈ ρ R.
Even if the order ≤ is not total, we still have the possibility to define the
infimum [xε ] ∧ [yε ] := [min(xε , yε )], and analogously the supremum function [xε ] ∨
[yε ] := [max(xε , yε )] and the absolute value |[xε ]| := [|xε |] ∈ ρ R. Our notations for
intervals are: [a, b] := {x ∈ ρ R | a ≤ x ≤ b}, [a, b]R := [a, b] ∩ R, and analogously
for segments [x, y] := {x + r · (y − x) | r ∈ [0, 1]} ⊆ ρ Rn and [x, y]Rn = [x, y] ∩
Rn . Finally, we write x ≈ y to denote that |x − y| is an infinitesimal number,
i.e., |x − y| ≤ r for all r ∈ R>0 . This is equivalent to limε→0+ |xε − yε | = 0 for all
representatives (xε ), (yε ) of x, y.
Topologies on ρ Rn . On the ρ R-module ρ Rn , we can consider the natural extension

of the Euclidean norm, i.e., |[xε ]| := [|xε |] ∈ ρ R, where [xε ] ∈ ρ Rn . Even if this
generalized norm takes values in ρ R, it shares several properties with usual norms,
like the triangular inequality or the property |y · x| = |y| · |x|. It is therefore natural
to consider on ρ Rn topologies generated by balls defined by this generalized norm
and suitable notions of being “strictly less than a given radius”:
Definition 2. Let c ∈ ρ Rn and x, y ∈ ρ R, then:

(i) We write x < y if ∃r ∈ ρ R≥0 : r is invertible, and r ≤ y − x
(ii) We write x <R y if ∃r ∈ R>0 : r ≤ y − x.

(iii) Br (c) := x ∈ ρ Rn | |x − c| < r for each r ∈ ρ R>0 .

(iv) BrF (c) := x ∈ ρ Rn | |x − c| <R r for each r ∈ R>0 .
(v) BrE (c) := {x ∈ Rn | |x − c| < r}, for each r ∈ R>0 , denotes an ordinary
Euclidean ball in Rn .
The relations <, <R have better topological properties as compared to the usual
0 relation a ≤ b and a = b 1
strict order (that we
0 will never use) because both
1 the sets
of balls Br (c) | r ∈ R>0 , c ∈ R
ρ ρ n
and Br (c) | r ∈ R>0 , c ∈ R
F ρ n
are bases
for two topologies on ρ Rn . The former is called sharp topology, whereas the latter
is called Fermat topology. We will call sharply open set any open set in the sharp
topology, and large open set any open set in the Fermat topology; clearly, the
latter is coarser than the former. The existence of infinitesimal neighborhoods
implies that the sharp topology induces the discrete topology on R. This is a
necessary result when one has to deal with continuous generalized functions which
have infinite derivatives. In fact, if f (x0 ) is infinite, only for x ≈ x0 we can have
f (x) ≈ f (x0 ).
The following result is useful to deal with positive and invertible generalized
numbers (cf. [24, 40]).
Lemma 1. Let x ∈ ρ R. Then the following are equivalent:

(i) x is invertible and x ≥ 0, i.e., x > 0.
(ii) For each representative (xε ) ∈ Rρ of x we have ∀0 ε : xε > 0.
(iii) For each representative (xε ) ∈ Rρ of x we have ∃m ∈ N ∀0 ε : xε > ρm
ε
Internal and strongly internal sets. A natural way to obtain sharply open, closed
and bounded sets in ρ Rn is by using a net (Aε ) of subsets Aε ⊆ Rn . We have
two ways of extending the membership relation xε ∈ Aε to generalized points
[xε ] ∈ ρ R:
Definition 3. Let (Aε ) be a net of subsets of Rn , then
0 1
(i) [Aε ] := [xε ] ∈ ρ Rn | ∀0 ε : xε ∈ Aε is called the internal set generated by
the net (Aε ). See [44] for the introduction and an in-depth study of this
notion.
(ii) Let (xε ) be a net of points of Rn , then we say that xε ∈ε Aε , and we read it

as (xε ) strongly belongs to (Aε ), if ∀0 ε : xε ∈ Aε and
0 if (xε ) ∼ρ (xε ), then
1
also xε ∈ Aε for ε small. Moreover, we set Aε := [xε ] ∈ ρ Rn | xε ∈ε Aε ,
and we call it the strongly internal set generated by the net (Aε ).
(iii) Finally, we say that the internal set K = [Aε ] is sharply bounded if there
exists r ∈ ρ R>0 such that K ⊆ Br (0). Analogously, a net (Aε ) is sharply
bounded if there exists r ∈ ρ R>0 such that [Aε ] ⊆ Br (0).
Therefore, x ∈ [Aε ] if there exists a representative (xε ) of x such that xε ∈ Aε
for ε small, whereas this membership is independent from the chosen representative
in the case of strongly internal sets. Note explicitly that an internal set generated
by a constant net Aε = A ⊆ Rn is simply denoted by [A].
The following theorem shows that internal and strongly internal sets have
dual topological properties:
Theorem 1. For ε ∈ I, let Aε ⊆ Rn and let xε ∈ Rn . Then we have
(i) [xε ] ∈ [Aε ] if and only if ∀q ∈ R>0 ∀0 ε : d(xε , Aε ) ≤ ρqε . Therefore [xε ] ∈ [Aε ]
if and only if [d(xε , Aε )] = 0 ∈ ρ R.
(ii) [xε ] ∈ Aε if and only if ∃q ∈ R>0 ∀0 ε : d(xε , Acε ) > ρqε , where Acε := Rn \Aε .
Therefore, if (d(xε , Acε )) ∈ Rρ , then [xε ] ∈ Aε if and only if [d(xε , Acε )] > 0.
(iii) [Aε ] is sharply closed and Aε is sharply open.
(iv) [Aε ] = [cl (Aε )], where cl (S) is the closure of S ⊆ Rn . On the other hand
Aε = int(Aε ), where int (S) is the interior of S ⊆ Rn .
We will also use the following:
Lemma 2. Let (Ωε ) be a net of subsets in Rn for all ε, and (Bε ) a sharply bounded
net such that [Bε ] ⊆ Ωε , then
∀0 ε : Bε ⊆ Ωε .
Sharply bounded internal sets (which are always sharply closed by Theo-
rem 1 (iii)) serve as compact sets for our generalized functions. For a deeper study
of this type of sets in the case ρ = (ε) see [44, 17]; in the same particular setting,
see [19] and references therein for (strongly) internal sets.
Generalized smooth functions. For the ideas presented in this section, see also,
e.g., [19, 18].
Using the ring ρ R, it is easy to consider a Gaussian with an infinitesimal
standard deviation. If we denote this probability density by f (x, σ), and if we set
σ = [σε ] ∈ ρ R>0 , where σ ≈ 0, we obtain the net of smooth functions (f (−, σε ))ε∈I .
This is the basic idea we develop in the following
Definition 4. Let X ⊆ ρ Rn and Y ⊆ ρ Rd be arbitrary subsets of generalized points.

Then we say that
f : X −→ Y is a generalized smooth function
if there exists a net of functions fε ∈ C ∞ (Ωε , Rd ) defining f in the sense that
X ⊆ Ωε , f ([xε ]) = [fε (xε )] ∈ Y and (∂ α fε (xε )) ∈ Rdρ for all x = [xε ] ∈ X and all
α ∈ Nn . The space of GSF from X to Y is denoted by ρ GC ∞ (X, Y ).
Let us note explicitly that this definition states minimal logical conditions
to obtain a set-theoretical map from X into Y and defined by a net of smooth
functions. In particular, the following Theorem 2 states that the equality f ([xε ]) =
[fε (xε )] is meaningful, i.e., that we have independence from the representatives for
all derivatives [xε ] ∈ X → [∂ α fε (xε )] ∈ ρ Rd , α ∈ Nn .
Theorem 2. Let X ⊆ ρ Rn and Y ⊆ ρ Rd be arbitrary subsets of generalized points.

Let fε ∈ C ∞ (Ωε , Rd ) be a net of smooth functions that defines a generalized smooth
map of the type X −→ Y , then
(i) ∀α ∈ Nn ∀(xε ), (xε ) ∈ Rnρ : [xε ] = [xε ] ∈ X ⇒ (∂ α uε (xε )) ∼ρ (∂ α uε (xε )).
(ii) ∀[xε ] ∈ X ∀α ∈ Nn ∃q ∈ R>0 ∀0 ε : supy∈B Eq (xε ) |∂ α uε (y)| ≤ ε−q .
ε
(iii) For all α ∈ Nn , the GSF g : [xε ] ∈ X → [∂ α fε (xε )] ∈ Rd is locally Lipschitz

in the sharp topology, i.e., each x ∈ X possesses a sharp neighborhood U such
that |g(x) − g(y)| ≤ L|x − y| for all x, y ∈ U and some L ∈ ρ R.
(iv) Each f ∈ ρ GC ∞ (X, Y ) is continuous with respect to the sharp topologies
induced on X, Y .
(v) Assume that the GSF f is locally Lipschitz in the Fermat topology and that
its Lipschitz constants are always finite: L ∈ R. Then f is continuous in the
Fermat topology.
(vi) f : X −→ Y is a GSF if and only if there exists a net vε ∈ C ∞ (Rn , Rd )
defining a generalized smooth map of type X −→ Y such that f = [vε (−)]|X .
(vii) Subsets S ⊆ ρ Rs with the trace of the sharp topology, and generalized smooth
maps as arrows form a subcategory of the category of topological spaces. We
will call this category ρ GC ∞ , the category of GSF.
The differential calculus for GSF can be introduced showing existence and
uniqueness of another GSF serving as incremental ratio. For its statement, if P(h)
is a property of h ∈ ρ R, then we write ∀s h : P(h) to denote ∃r ∈ ρ R>0 ∀h ∈
Br (0) : P(h) and ∀F h : P(h) for ∃r ∈ R>0 ∀h ∈ BrF (c) : P(h).
Theorem 3. Let U ⊆ ρ Rn be a sharply open set, let v = [vε ] ∈ ρ Rn , and let

f ∈ ρ GC ∞ (U, ρ R) be a generalized smooth map generated by the net of smooth
functions fε ∈ C ∞ (Ωε , R). Then
(i) There exists a sharp neighborhood T of U × {0} and a generalized smooth
map r ∈ ρ GC ∞ (T, ρ R), called the generalized incremental ratio of f along v,
such that
∀x ∈ U ∀s h : f (x + hv) = f (x) + h · r(x, h).
(ii) If r̄ ∈ ρ GC ∞ (S, ρ R) is another generalized incremental ratio of f along v
defined on a sharp neighborhood S of U × {0}, then
∀x ∈ U ∀s h : r(x, h) = r̄(x, h).
% &
(iii) We have r(x, 0) = ∂fε
∂vε (xε ) for every x ∈ U and we can thus define ∂f
∂v (x) :=
∞
r(x, 0), so that ∂f
∂v∈ ρ GC (U, ρ R).
If U is a large open set, then an analogous statement holds replacing ∀s h by ∀F h
and sharp neighborhoods by large neighborhoods.
Note that this result permits to consider the partial derivative of f with
respect to an arbitrary generalized vector v ∈ ρ Rn which can be, e.g., infinitesimal
or infinite.
Using this result we obtain the usual rules of differential calculus, including
the chain rule. Finally, we note that for each x ∈ U , the map Df (x).v := ∂f ∂v (x) ∈
ρ
Rd is ρ R-linear in v ∈ ρ Rn . The set of all the ρ R-linear maps ρ Rn −→ ρ Rd will be
denoted by L(ρ Rn , ρ Rd ). For A = [Aε (−)] ∈ L(ρ Rn , ρ Rd ), we set |A| := [|Aε |], the
generalized number defined by the operator norms of the matrices Aε ∈ L(Rn , Rd ).
Embedding of Schwartz distributions and Colombeau functions. We finally recall
two results that give a certain flexibility in constructing embeddings of Schwartz
distributions. Note that both the infinitesimal ρ and the embedding of Schwartz
distributions have to be chosen depending on the problem we aim to solve. A
trivial example in this direction is the ODE y = y/ dε, which cannot be solved
for ρ = (ε), but it has a solution for ρ = (e−1/ε ). As another simple example, if
we need the property H(0) = 1/2, where H is the Heaviside function, then we
have to choose the embedding of distributions accordingly. This corresponds to
the philosophy followed in [26]. See also [20] for further details.
If ϕ ∈ D(Rn ), r ∈ R>0 and x ∈ Rn , we use the notations r ( ϕ for the function
x ∈ Rn → r1n · ϕ xr ∈ R and x ⊕ ϕ for the function y ∈ Rn → ϕ(y − x) ∈ R.
These notations permit to highlight that ( is a free action of the multiplicative
group (R>0 , ·, 1) on D(Rn ) and ⊕ is a free action of the additive group (R>0 , +, 0)
on D(Rn ). We also have the distributive property r ( (x ⊕ ϕ) = rx ⊕ r ( ϕ.
Lemma 3. Let b ∈ Rρ be a net such that limε→0+ bε = +∞. Let d ∈ (0, 1). There
exists a net (ψε )ε∈I of D(Rn ) with the properties:
(i) supp(ψε ) ⊆ B1 (0) for all ε ∈ I.

(ii) ψε = 1 for all ε ∈ I.
(iii) ∀α ∈ Nn ∃p ∈ N : supx∈Rn |∂α ψε (x)| = O(bpε ) as ε → 0+ .
(iv) ∀j ∈ N ∀0 ε : 1 ≤ |α| ≤ j ⇒ xα · ψε (x) dx = 0.
(v) ∀η ∈ R>0 ∀0 ε : |ψε | ≤ 1 + η.
0
(vi) If n = 1, then the net (ψε )ε∈I can be chosen so that −∞ ψε = d.
If ψε satisfies (i)–(vi) then in particular ψεb := b−1
ε ( ψε satisfies (ii)–(v).
Concerning embeddings of Schwartz distributions, we have the following result,

where ρ Ωc := {[xε ] ∈ [Ω] | ∃K Ω ∀0 ε : xε ∈ K} is called the set of compactly
supported points in Ω ⊆ Rn .
Theorem 4. Under the assumptions of Lemma 3, let Ω ⊆ Rn be an open set and

let (ψεb ) be the net defined in Lemma 3. Then the mapping
) *
ιbΩ : T ∈ E (Ω) → T ∗ ψεb (−) ∈ ρ GC ∞ (ρ Ωc , ρ R)
uniquely extends to a sheaf morphism of real vector spaces
2c , ρ R),
ιb : D −→ ρ GC ∞ (ρ (−)
and satisfies the following properties:
2c , ρ R)
(i) If b ≥ dρ−a for some a ∈ R>0 , then ιb |C ∞ (−) : C ∞ (−) −→ ρ GC ∞ (ρ (−)
is a sheaf morphism of algebras.
(ii) If T ∈ E (Ω) then supp(T ) = supp(ιbΩ (T )).
(iii) limε→0+ Ω ιbΩ (T )ε · ϕ = T, ϕ for all ϕ ∈ D(Ω) and all T ∈ D (Ω).

(iv) ιb commutes with partial derivatives, i.e., ∂ α ιbΩ (T ) = ιbΩ (∂ α T ) for each
T ∈ D (Ω) and α ∈ N.
Concerning the embedding of Colombeau generalized functions, we recall

that the special Colombeau algebra on Ω is defined as the quotient G s (Ω) :=
EM (Ω)/N s (Ω) of moderate nets over negligible nets, where the former is
EM (Ω) := {(uε ) ∈ C ∞ (Ω)I | ∀K Ω ∀α ∈ Nn ∃N ∈ N : sup |∂ α uε (x)| = O(ε−N )}
x∈K
and the latter is
N s (Ω) := {(uε ) ∈ C ∞ (Ω)I | ∀K Ω ∀α ∈ Nn ∀m ∈ N : sup |∂ α uε (x)| = O(εm )}.
x∈K
Using ρ = (ε), we have the following compatibility result:
Theorem 5. A Colombeau generalized function u = (uε ) + N s (Ω)d ∈ G s (Ω)d de-

fines a generalized smooth map u : [xε ] ∈ ρ Ωc −→ [uε (xε )] ∈ Rd which is locally
Lipschitz on the same neighborhood of the Fermat topology for all derivatives. This
assignment provides a bijection of G s (Ω)d onto ρ GC ∞ (ρ Ωc , ρ Rd ) for every open set
Ω ⊆ Rn .
For GSF, suitable generalizations of many classical theorems of differential

and integral calculus hold: intermediate value theorem, mean value theorems, Tay-
lor formulas in different forms, a sheaf property for the Fermat topology, and the
extreme value theorem on internal sharply bounded sets (see [18]). The latter are
called functionally compact subsets of ρ Rn and serve as compact sets for GSF.
A theory of compactly supported GSF has been developed in [17], and it closely
resembles the classical theory of LF-spaces of compactly supported smooth func-
tions. It results that for suitable functionally compact subsets, the corresponding
space of compactly supported GSF contains extensions of all Colombeau gener-
alized functions, and hence also of all Schwartz distributions. Finally, in these
spaces it is possible to prove the Banach fixed point theorem and a corresponding
Picard–Lindelöf theorem, see [37].
2. Local inverse function theorems

As in the case of classical smooth functions, any infinitesimal criterion for the
invertibility of generalized smooth functions will rely on the invertibility of the
corresponding differential. We therefore note the following analogue of [24, Lemma
1.2.41] (whose proof transfers literally to the present situation):
Lemma 4. Let A ∈ ρ Rn×n be a square matrix. The following are equivalent:
(i) A is nondegenerate, i.e., ξ ∈ ρ Rn , ξ t Aη = 0 ∀η ∈ ρ Rn implies ξ = 0.
(ii) A : ρ Rn → ρ Rn is injective.
(iii) A : ρ Rn → ρ Rn is surjective.
(iv) det(A) is invertible.
Theorem 6. Let X ⊆ ρ Rn , let f ∈ ρ GC ∞ (X, ρ Rn ) and suppose that for some x0 in
the sharp interior of X, Df (x0 ) is invertible in L(ρ Rn , ρ Rn ). Then there exists a
sharp neighborhood U ⊆ X of x0 and a sharp neighborhood V of f (x0 ) such that
f : U → V is invertible and f −1 ∈ ρ GC ∞ (V, U ).
Proof. Theorem 2.(vi) entails that f can be defined by a globally defined net fε ∈
∞
C3 (Rn , Rn ). Hadamard’s inequality (cf. [11, Prop. 3.43]) implies |Df (x0 )−1 | ≥
n 1
C|det (Df (x0 )−1 )|, where C ∈ R>0 is a universal constant that only depends
on the dimension n. Thus, by Lemma 4 and Lemma 1, det Df (x0 ) and consequently
also a := |Df (x0 )−1 | is invertible. Next, pick positive invertible numbers b, r ∈ ρ R
such that ab < 1, B2r (x0 ) ⊆ X and
|Df (x0 ) − Df (x)| < b
for all x ∈ B2r (x0 ). Such a choice of r is possible since every derivative of f is
continuous with respect to the sharp topology (see Thm. 2.(iv) and Thm. 3.(iii)).
Pick representatives (aε ), (bε ) and (rε ) of a, b and r such that for all ε ∈ I we
have bε > 0, aε bε < 1, and rε > 0. Let (x0ε ) be a representative of x0 . Since
[Brε (x0ε )] ⊆ B2r (x0 ), by Lemma 2 we can also assume that Brε (x0ε ) ⊆ Ωε , and
|Dfε (x0ε ) − Dfε (x)| < bε for all x ∈ Uε := Brε (x0ε ). Now let cε := 1−a aε
ε bε
. Then
c := [cε ] > 0 and by [13, Th. 6.4] we obtain for each ε ∈ I:
(a) For all x ∈ Uε := Brε (x0ε ), Dfε (x) is invertible and |Dfε (x)−1 | ≤ cε .
(b) Vε := fε (Brε (x0ε )) is open in Rn .
(c) fε |Uε : Uε −→ Vε is a diffeomorphism, and
(d) setting y0ε := fε (x0ε ), we have Brε /cε (y0ε ) ⊆ fε (Brε (x0ε )).
The sets U := Uε = Br (x0 ) ⊆ X and V := Vε are sharp neighborhoods of
x0 and f (x0 ), respectively, by (d), and so it remains to prove that [fε |−1 Uε (−)] ∈
ρ
GC ∞ (V, U ).
We first note that by (a), |Dfε (x)−1 | ≤ cε for all x ∈ Brε (x0ε ), which by
Hadamard’s inequality implies
1
| det(Dfε (x))| ≥ (x ∈ Brε (x0ε )). (1)
C · cnε
Now for [yε ] ∈ V and 1 ≤ i, j ≤ n we have (see, e.g., [11, (3.15)])
1
∂j (fε−1 )i (yε ) = · Pij ((∂s fεr (fε−1 (yε )))r,s ), (2)
det(Dfε (fε−1 (yε )))
where Pij is a polynomial in the entries of the matrix in its argument. Since
[fε−1 (yε )] ∈ U ⊆ X, it follows from (1) and the fact that f |U ∈ ρ GC ∞ (U, ρ Rn ) that
(∂j (fε−1 )i (yε )) ∈ Rnρ .
Higher-order derivatives can be treated analogously, thereby establishing that ev-
ery derivative of gε := fε |−1Uε is moderate. To prove the claim, it remains to show
that [gε (yε )] ∈ U = Uε for all [yε ] ∈ V = Vε . Since gε : Vε −→ Uε , we only prove
that if (xε ) ∼ρ (gε (yε )), then also xε ∈ Uε for ε small. We can set yε := fε (xε )
because fε is defined on the entire Rn . By the mean value theorem applied to fε
and the moderateness of f , we get
|yε − yε | = |fε (xε ) − fε (gε (yε ))| ≤ ρN
ε · |xε − gε (yε )|.
Therefore (yε ) ∼ρ (yε ) and hence yε ∈ Vε and gε (yε ) = xε ∈ Uε for ε small.
From Theorem 2.(iv), we know that any generalized smooth function is
sharply continuous. Thus we obtain:
Corollary 1. Let X ⊆ ρ Rn be a sharply open set, and let f ∈ ρ GC ∞ (X, ρ Rn ) be
such that Df (x) is invertible for each x ∈ X. Then f is a local homeomorphism
with respect to the sharp topology. In particular, it is an open map.
Any such map f will therefore be called a local generalized diffeomorphism.
If f ∈ ρ GC ∞ (X, Y ) possesses an inverse in ρ GC ∞ (Y, X), then it is called a global
generalized diffeomorphism.
Following the same idea we used in the proof of Theorem 6, we can prove a
sufficient condition to have a local generalized diffeomorphism which is defined in
a large neighborhood of x0 :
Theorem 7. Let X ⊆ ρ Rn , let f ∈ ρ GC ∞ (X, ρ Rn ) and suppose that for some x0

in the Fermat interior of X, Df (x0 ) is invertible in L(ρ Rn , ρ Rn ). Assume that
|Df (x0 )−1 | is finite, i.e., |Df (x0 )−1 | ≤ k for some k ∈ R>0 , and Df is Fermat
continuous. Then there exists a large neighborhood U ⊆ X of x0 and a large
neighborhood V of f (x0 ) such that f : U → V is invertible and f −1 ∈ ρ GC ∞ (V, U ).
Proof. We proceed as above, but now we have rε = r ∈ R>0 , bε = b ∈ R>0

because of our assumptions. Setting cε := 1−a aε
εb
, we have that c := [cε ] ∈ ρ R>0 is
finite. Therefore, there exists s ∈ R>0 such that s < rc . We can continue as above,
noting that now BrF (x0 ) ⊆ U = Br (x0 ) ⊆ X and BsF (y0 ) ⊆ Br/c (y0 ) ⊆ V are large
neighborhoods of x0 and f (x0 ) respectively.
Example 1.
(i) Theorem 4, for n = 1, shows that δ(x) = [bε ψε (bε x)] is, up to sheaf isomor-
phism, the Dirac delta. This also shows directly that δ ∈ ρ GC ∞ (ρ R, ρ R). We
can take the net (ψε ) so that ψε (0) = 1 for all ε. In this way, H (0) = δ(0) = b
is an infinite number. We can thus apply the local inverse function theorem
6 to the Heaviside function H obtaining that H is a generalized diffeomor-
phism in an infinitesimal neighborhood of 0. This neighborhood cannot be
finite because H (r) = 0 for all r ∈ R=0 .
(ii) By the intermediate value theorem for GSF (see [18, Cor. 42]), in the interval
[0, 1/2] the Dirac delta takes any value in [0, δ(0)]. So, let k ∈ [0, 1/2] such
that δ(k) = 1. Then by the mean value theorem for GSF (see [18, Thm. 43])
δ(δ(1)) − δ(δ(k)) = δ(0) − δ(1) = b − 0 = (δ ◦ δ) (c) · (1 − k) for some c ∈ [k, 1].
Therefore (δ ◦ δ) (c) = 1−kb
∈ ρ R>0 , and around c the composition δ ◦ δ is
invertible. Note that (δ ◦ δ)(r) = b for all r ∈ R=0 , and (δ ◦ δ)(h) = 0 for all
h ∈ ρ R such that δ(h) is not infinitesimal.
Now, let r ∈ ρ R>0 be an infinitesimal generalized number, i.e., r ≈ 0.
(iii) Let f (x) := r · x for x ∈ ρ Rc . Then f (x0 ) = r ≈ 0 and Theorem 6 yields
f −1 : y ∈ Bs (rx0 ) → y/r ∈ ρ Rc for some s ∈ ρ R>0 . But y/r is finite only if y
is infinitesimal, so that s ≈ 0. This shows that the assumption in Theorem 7
on |Df (x0 )−1 | being finite is necessary.
(iv) Let f (x) := sin xr . We have f ∈ ρ GC ∞ (ρ R, ρ R) and f (x) = 1r cos xr , which is
always an infinite number, e.g., if ∃ limε→0+ xε = r(2k + 1) π2 ≈ 0, k ∈ Z. By
Theorem 6, we know that f is invertible, e.g., around x = 0. It is easy to
recognize that f is injective in the infinitesimal interval − π2 r, + π2 r . In [11,
Exa. 3.9], it is proved that f is not injective in any large neighborhood of
−1
x = 0. Therefore, f |(− π r,+ π r) is a GSF that cannot be extended to a
2 2
Colombeau generalized function.
(v) Similarly, f (x) := r sin x, x ∈ ρ R, has an inverse function which cannot be
extended outside the infinitesimal neighborhood (−r, r).
(vi) Theorem 6 cannot be applied to f (x) := x3 at x0 = 0. However, if we restrict

to x ∈ (−∞, −r)∪(r, +∞), then the inverse function f −1 (y) = y 1/3 is defined
in y ∈ (−∞, −r3 ) ∪ (r3 , +∞) and has infinite derivative at each infinitesimal
point in its domain.
In [2], Aragona, Fernandez and Juriaans introduced a differential calculus on
spaces of Colombeau generalized points based on a specific form of convergence of
difference quotients. Moreover, in [3], an inverse function theorem for Colombeau
generalized functions in this calculus was established. In the one-dimensional case
it was shown in [19] that any GSF is differentiable in the sense of [2, 3], with
the same derivative. Below we will show that this compatibility is in fact true in
arbitrary dimensions and that Theorem 6 implies the corresponding result from
[3]. In the remaining part of the present section, we therefore restrict our attention
to the case ρε = ε, the gauge that is used in standard Colombeau theory (as well
as in [2, 3]), and hence ρ R = R and ρ Ωc = Ωc .
First, we recall the definition from [2]:
Definition 5. A map f from some sharply open subset U of Rn to Rm is called
differentiable in x0 ∈ U in the sense of [2] with derivative A ∈ L(Rn , Rm ) if
|f (x) − f (x0 ) − A(x − x0 )|e
lim = 0, (3)
x→x0 |x − x0 |e
where
v : (xε ) ∈ Rn(ε) → sup{b ∈ R | |xε | = O(εb )} ∈ (−∞, ∞]
| − |e : x ∈ Rn → exp(−v(x)) ∈ [0, ∞).
The following result shows compatibility of this notion with the derivative in
the sense of GSF.
Lemma 5. Let U be sharply open in Rn , let x0 ∈ U and suppose that f ∈
GC ∞ (U, Rm ). Then f is differentiable in the sense of [2] in x0 with derivative
Df (x0 ).
Proof. Without loss of generality we may suppose that m = 1. Let f be defined
by the net fε ∈ C ∞ (Rn , R) for all ε. Since (D2 fε (xε )) is moderate, it follows from
Theorem 2.(ii) that there exists some q > 0 such that supy∈B Eq (xε ) |D2 fε (y)| ≤ ε−q
ε
for ε small. Then by Taylor’s theorem we have
fε (xε ) − fε (x0ε ) − Dfε (x0ε )(xε − x0ε )
|α| 1
= (1 − t)|α|−1 ∂ α fε (x0ε + t(xε − x0ε )) dt · (xε − x0ε )α .
α! 0
|α|=2
For [xε ] ∈ Bdεq (x0 ) this implies that

|f (x) − f (x0 ) − Df (x0 )(x − x0 )|e ≤ eq |x − x0 |2e ,
thereby establishing (3) with A = Df (x0 ), as claimed.
It follows that any f ∈ GC ∞ (U, Rm ) is in fact even infinitely often differen-

tiable in the sense of [2].
Based on these observations we may now give an alternative proof for
[3, Thm. 3]:
Theorem 8. Let Ω ⊆ Rn be open, f ∈ G s (Ω)n , and x0 ∈ Ωc such that det Df (x0 )
is invertible in R. Then there are sharply open neighborhoods U of x0 and V of
f (x0 ) such that f : U → V is a diffeomorphism in the sense of [2].
Proof. By Theorem 5, f can be viewed as an element of GC ∞ (Ωc , Rn ). Moreover,
Ωc is sharply open, which together with Lemma 4 shows that all the assumptions of
Theorem 6 are satisfied. We conclude that f possesses an inverse f −1 in GC ∞ (V, U )
for a suitable sharp neighborhood V of f (x0 ). Finally, by Lemma 5, both f and
f −1 are infinitely differentiable in the sense of [2].
3. Global inverse function theorems

The aim of the present section is to obtain statements on the global invertibility of
generalized smooth functions. For classical smooth functions, a number of criteria
for global invertibility are known, and we refer to [9, 31] for an overview.
The following auxiliary result will repeatedly be needed below:
Lemma 6. Let f ∈ ρ GC ∞ (X, Y ) be defined by (fε ), where X ⊆ ρ Rn and Y ⊆ ρ Rd .
Assume that ∅ = [Aε ] ⊆ X. Let b : Rd −→ R be a set-theoretical map such that
b̄ : [yε ] ∈ Y → [b(yε )] ∈ ρ R is well defined (e.g., b(x) = |x|). If f satisfies
∀x ∈ X : b̄ [f (x)] > 0, (4)
then
(i) ∃q ∈ R>0 ∀0 ε ∀x ∈ Aε : b (fε (x)) > ρqε .
(ii) For all K Rn , if [K] ⊆ X then ∀0 ε ∀x ∈ K : b (fε (x)) > 0.
Proof. In fact, suppose to the contrary that there was a sequence (εk )k ↓ 0 and a
sequence xk ∈ Aεk such that b(fεk (xk )) ≤ ρkεk . Let Aε = ∅ for ε ≤ ε0 , and pick
aε ∈ Aε . Set
xk for ε = εk
xε := (5)
aε otherwise.
It follows that x := [xε ] ∈ [Aε ] ⊆ X, and hence b̄ [f (x)] > 0 by (4). Therefore,
b (fεk (xk )) > ρpεk for some p ∈ R>0 by Lemma 1, and this yields a contradiction.
The second part follows by setting Aε = K in the first one and by noting that
ρε > 0.
After these preparations, we now turn to generalizing global inverse function
theorems from the smooth setting to GSF. We start with the one-dimensional case.
Here it is well known that a smooth function f : R → R is a diffeomorphism onto
its image if and only if |f (x)| > 0 for all x ∈ R. It is a diffeomorphism onto R if
in addition there exists some r > 0 with |f (x)| > r for all x ∈ R. Despite the fact
that ρ Rc is non-Archimedean, there is a close counterpart of this result in GSF.
Theorem 9. Let f ∈ ρ GC ∞ (ρ Rc , ρ Rc ) and suppose that there exists some r ∈ R≥0

such that |f (x)| > r for all x ∈ ρ Rc . Then
(i) f has a defining net (f¯ε ) consisting of diffeomorphisms f¯ε : R → R.
(ii) f is a global generalized diffeomorphism in ρ GC ∞ (ρ Rc , f (ρ Rc )).
(iii) If r > 0, then f (ρ Rc ) = ρ Rc , so f is a global generalized diffeomorphism in
ρ
GC ∞ (ρ Rc , ρ Rc ).
Proof. Let (fε ) be a defining net for f such that fε ∈ C ∞ (R, R) for each ε (cf.
Thm. 2 (vi)). Since |f (x)| > 0 for every x ∈ ρ Rc , Lemma 6 implies that for each
n ∈ N there exists some εn > 0 and some qn > 0 such that for each ε ∈ (0, εn ]
and each x ∈ [−n, n] we have |fε (x)| > ρqεn . Clearly we may suppose that εn ↓ 0,
qn+1 > qn for all n and that ρqεn < 1. Now for any n ∈ N>0 let ϕn : R → [0, 1] be
a smooth cut-off function with ϕn ≡ 1 on [−(n − 1), n − 1] and suppϕn ⊆ [−n, n].
Supposing that fε (x) > 0 on [−n, n] (the case fε (x) < 0 on [−n, n] can be handled
analogously), we set
vnε (x) := fε (x)ϕn (x) + 1 − ϕn (x) (x ∈ R)
x
¯
fε (x) := fε (0) + vnε (t) dt (x ∈ R, εn+1 < ε ≤ εn ),
0
and f¯ε := fε for ε ∈ (ε0 , 1]. Then f¯ε ∈ C ∞ (R, R) for each ε, and for each x ∈ R
and each ε ∈ (εn+1 , εn ], we have f¯ε (x) = fε (x)ϕn (x) + 1 − ϕn (x) > ρqεn if and
only if ϕn (x) · [1 − fε (x)] < 1 − ρqεn . The latter inequality holds if x ∈
/ [−n, n] or if
1−ρqn
fε (x) ≥ 1. Otherwise, ϕn (x) ≤ 1 < 1−f ε(x) because 1 > fε (x) > ρqεn . Any such f¯ε
ε
therefore is a diffeomorphism from R onto R. Also, f¯ε (x) = fε (x) for all x ∈ [−n, n]
as soon as ε ≤ εn+1 . Hence also (f¯ε ) is a defining net for f . This proves (i).
For each ε ≤ ε0 , let gε be the global inverse of f¯ε . We claim that g := [gε ] is
a GSF from f (ρ Rc ) onto ρ Rc that is inverse to f . For this it suffices to show that
(k)
whenever y = [(yε )] ∈ f (ρ Rc ), then for each k ∈ N, the net (gε (yε )) is ρ-moderate.
To see this, it suffices to observe that for y = f (x), f satisfies the assumptions of
the local inverse function theorem (Thm. 6) at x, and so the proof of that result
shows that g is a GSF when restricted to a suitable sharp neighborhood of y. But
this in particular entails the desired moderateness property at y, establishing (ii).
Finally, assume that r > 0. The same reasoning as in the proof of (i) now
produces a defining net (f¯ε ) with the property that |f¯ε (x)| > r for all ε ≤ ε0 and
all x ∈ R. Again, each f¯ε is a diffeomorphism from R onto R, and we denote its
inverse by gε : R → R. Due to (ii) it remains to show that f : ρ Rc → ρ Rc is onto.
To this end, note first that |gε (y)| < 1/r for all ε ≤ ε0 and all y ∈ Rn .
Also, since f ∈ ρ GC ∞ (ρ Rc , ρ Rc ), there exists some real number C > 0 such that
|fε (0)| ≤ C for ε small. For such ε and any [yε ] ∈ ρ Rc we obtain by the mean value
theorem
1 1
|yε − fε (0)| ≤ (|yε | + C),
|gε (yε )| = |gε (yε ) − gε (fε (0))| ≤ (6)
r r
so that gε (yε ) remains in a compact set for ε small. Based on this observation,
the same argument as in (2) shows that, for any y = [yε ] ∈ ρ Rc and any k ≥ 1,
(k)
(gε (yε )) is moderate, so (gε ) defines a GSF ρ Rc −→ ρ Rc . Hence given y ∈ ρ Rc it
suffices to set x := g(y) to obtain f (x) = y.
Turning now to the multi-dimensional case, we first consider Hadamard’s
global inverse function theorem. For its formulation, recall that a map between
topological spaces is called proper if the inverse image of any compact subset is
again compact. As is easily verified, a continuous map α : Rn → Rm is proper if
and only if
|α(x)| → ∞ as |x| → ∞. (7)
Theorem 10 (Hadamard). A smooth map f : Rn → Rn is a global diffeomorphism
if and only if it is proper and its Jacobian determinant never vanishes.
For a proof of this result we refer to [22].
The following theorem provides an extension of Theorem 10 to the setting
of GSF.
Theorem 11. Suppose that f ∈ ρ GC ∞ (ρ Rnc , ρ Rnc ) possesses a defining net fε :
Rn −→ Rn such that:
(i) ∀x ∈ Rn ∀ε ∈ I : Dfε (x) is invertible in L(Rn , Rn ), and for each x ∈ ρ Rnc ,
Df (x) is invertible in L(ρ Rn , ρ Rn ).
(ii) There exists some ε ∈ I such that inf ε∈(0,ε ] |fε (x)| → +∞ as |x| → ∞.
Then f is a global generalized diffeomorphism in ρ GC ∞ (ρ Rnc , ρ Rnc ).
Proof. By Theorem 10, each fε is a global diffeomorphism Rn → Rn for each
ε ≤ ε and we denote by gε : Rn → Rn the global inverse of fε . In order to prove
that the net (gε )ε≤ε defines a GSF, we first note that, by (ii), the net (fε )ε≤ε is
‘uniformly proper’ in the following sense: Given any M ∈ R≥0 there exists some
M ∈ R≥0 such that when |x| ≥ M then ∀ε ≤ ε : |fε (x)| ≥ M .
Hence, for any K Rn , picking M > 0 with K ⊆ BM (0) it follows that
gε (K) ⊆ BM (0) =: K Rn for all ε ≤ ε . Thereby, the net (gε )ε≤ε maps ρ Rnc
into itself, i.e.,
∀[yε ] ∈ ρ Rnc : [gε (yε )] ∈ ρ Rnc . (8)
Moreover, for each K Rn , assumption (i), Lemma 4, Lemma 1 and Lemma 6
yield
∃q ∈ R>0 ∀0 ε ∀x ∈ K : | det Dfε (x)| > ρqε . (9)
From (8) and (9) it follows as in (2) that, for any y = [yε ] ∈ ρ Rc and any |β| ≥ 1,
(∂ β gε (yε )) is moderate, so g := [yε ] → [gε (yε )] ∈ ρ GC ∞ (ρ Rnc , ρ Rnc ). Finally, that g
is the inverse of f on ρ Rnc follows as in Theorem 9.
The next classical inversion theorem we want to adapt to the setting of gen-
eralized smooth functions is the following one:
Theorem 12 (Hadamard–Lévy). Let f : X → Y be a local diffeomorphism be-
tween Banach spaces. Then f is a diffeomorphism if there exists a continuous
non-decreasing function β : R≥0 → R>0 such that
∞
1
ds = +∞, |Df (x)−1 | ≤ β(|x|) ∀x ∈ X.
0 β(s)
This holds, in particular, if there exist a, b ∈ R>0 with |Df (x)−1 | ≤ a + b|x| for
all x ∈ X.
For a proof, see [9].
We can employ this result to establish the following global inverse function
theorem for GSF.
Theorem 13. Suppose that f ∈ ρ GC ∞ (ρ Rnc , ρ Rnc ) satisfies:
(i) f possesses a defining net fε : Rn −→ Rn such that ∀x ∈ Rn ∀ε ∈ I :
Dfε (x) is invertible in L(Rn , Rn ), and for each x ∈ ρ Rnc , Df (x) is invertible
in L(ρ Rn , ρ Rn ).
(ii) There exists a net of continuous non-decreasing functions βε : R≥0 −→ R>0
such that ∀0 ε ∀x ∈ Rn : |Dfε (x)−1 | ≤ βε (|x|) and
∞
1
ds = +∞.
0 βε (s)
Then f is a global generalized diffeomorphism in ρ GC ∞ (ρ Rnc , f (ρ Rnc )).
If instead of (ii) we make the stronger assumption
(iii) ∃C ∈ R>0 : ∀0 ε ∀x ∈ Rn : |Dfε (x)−1 | ≤ C,
then f is a global generalized diffeomorphism in ρ GC ∞ (ρ Rnc , ρ Rnc ).
In particular, (ii) applies if there exist a, b ∈ ρ R>0 that are finite (i.e., aε , bε < R
for some R ∈ R and ε small) with |Dfε (x)−1 | ≤ aε + bε |x| for ε small and all
x ∈ ρ Rnc .
Proof. From (ii) it follows by an ε-wise application of Theorem 12 that there exists
some ε0 > 0 such that each fε with ε < ε0 is a diffeomorphism: Rn → Rn . We
denote by gε its inverse. Using assumption (i), it follows exactly as in the proof of
Theorem 9 (ii) that g := [gε ] is an element of ρ GC ∞ (f (ρ Rnc ), ρ Rnc ) that is inverse
to f .
Assuming (iii), for any [yε ] ∈ ρ Rnc and ε small, we have
|Dgε (yε )| = |(Dfε (gε (yε )))−1 | ≤ C,
so the mean value theorem yields
|gε (yε )| = |gε (yε ) − gε (fε (0))| ≤ C|yε − fε (0)|, (10)
which is uniformly bounded for ε small since f ∈ ρ GC ∞ (ρ Rnc , ρ Rnc ). We conclude

that (gε ) satisfies (8). From here, the proof can be concluded literally as in Theo-
rem 11.
Remark 1. By Theorem 5, for ρ = (ε), the space ρ GC ∞ (ρ Rnc , ρ Rn ) can be identi-
fied with the special Colombeau algebra G s (Rn )n . In this picture, ρ GC ∞ (ρ Rnc , ρ Rnc )
corresponds to the space of c-bounded Colombeau generalized functions on Rn
(cf. [32, 24]). Therefore, under the further assumption that f (ρ Rnc ) = ρ Rnc , theo-
rems 11 and 13 can alternatively be viewed as global inverse function theorems for
c-bounded Colombeau generalized functions.
4. Conclusions
Once again, we want to underscore that the statement of the local inverse function
theorem 6 is the natural generalization to GSF of the classical result. Its simplicity
relies on the fact that the sharp topology is the natural one for GSF, as explained
above. This natural setting permits to include examples in our theory that cannot
be incorporated in an approach based purely on Colombeau generalized functions
on classical domains (cf. Example 1 and [11]).
Moreover, as Theorem 7 shows, the concept of Fermat topology leads, with
comparable simplicity, to sufficient conditions that guarantee solutions defined on
large (non-infinitesimal) neighborhoods.
Acknowledgement
We are indebted to the referee for several helpful comments that have substantially
improved the results of Section 3. P. Giordano has been supported by grants
P25116 and P25311 of the Austrian Science Fund FWF. M. Kunzinger has been
supported by grants P23714 and P25326 of the Austrian Science Fund FWF.
References
[1] Abraham, R., Marsden, J.E., Ratiu, T.: Manifolds, Tensors, Analysis and Applica-
tions, second ed., Springer-Verlag, New York (1988)
[2] Aragona, J., Fernandez, R., Juriaans, S.O.: A discontinuous Colombeau differential
calculus. Monatsh. Math. 144, 13–29 (2005)
[3] Aragona, J., Fernandez, R., Juriaans, S.O., Oberguggenberger, M.: Differential cal-
culus and integration of generalized functions over membranes. Monatsh. Math. 166,
1–18 (2012)
[4] Bampi, F., Zordan, C.: Higher order shock waves. Journal of Applied Mathematics
and Physics, 41, 12–19 (1990)
[5] Bertrand, D.: Review of ‘Lectures on differential Galois theory’, by A. Magid.
Bull. Amer. Math. Soc. 33, 289–294 (1996)
[6] Bogoliubov, N.N., Shirkov, D.V.: The Theory of Quantized Fields. New York, Inter-
science (1959)
[7] Casale, G.: Morales-Ramis theorems via Malgrange pseudogroup. Ann. Inst. Fourier
59, no. 7, 2593–2610 (2009)
[8] Colombeau, J.F.: Multiplication of distributions – A tool in mathematics, numerical
engineering and theoretical Physics. Springer-Verlag, Berlin Heidelberg (1992)
[9] De Marco, G., Gorni, G., Zampieri, G.: Global inversion of functions: an introduction.
NoDEA Nonlinear Differential Equations Appl. 1 no. 3, 229–248 (1994)
[10] Dirac, P.A.M.: The physical interpretation of the quantum dynamics. Proc. R. Soc.
Lond. A, 113 27, 621–641 (1926–27)
[11] Erlacher, E.: Local existence results in algebras of generalized functions. Doctoral
thesis, University of Vienna (2007)
http://www.mat.univie.ac.at/~diana/uploads/publication47.pdf
[12] Erlacher, E., Grosser, M.: Inversion of a ‘discontinuous coordinate transformation’
in General Relativity. Appl. Anal. 90, 1707–1728 (2011)
[13] Erlacher, E., Inversion of Colombeau generalized functions. Proc. Edinb. Math. Soc.
56, 469–500 (2013)
[14] Erlacher, E., Grosser, M.: Ordinary Differential Equations in Algebras of Generalized
Functions. In: Pseudo-Differential Operators, Generalized Functions and Asymp-
totics, S. Molahajloo, S. Pilipović, J. Toft, M.W. Wong eds., Operator Theory: Ad-
vances and Applications Volume 231, 253–270 (2013)
[15] Estrada, R., Vindas, J.: A general integral. Dissertationes Math. 483, 1–49 (2012)
[16] Campos Ferreira, J.: On some general notions of superior limit, inferior limit and
value of a distribution at a point. Portugaliae Math. 28, 139–158 (2001)
[17] Giordano P., Kunzinger M.: A convenient notion of compact set for generalized
functions. Proc. Edinb. Math. Soc., in press. See arXiv 1411.7292.
[18] Giordano P., Kunzinger M., Steinbauer R.: A new approach to generalized functions
for mathematical physics.
See http://www.mat.univie.ac.- at/˜giordap7/GenFunMaps.pdf.
[19] Giordano, P., Kunzinger, M., Vernaeve, H.: Strongly internal sets and generalized
smooth functions. Journal of Mathematical Analysis and Applications 422, issue 1,
56–71 (2015)
[20] Giordano, P., Luperi Baglini, L.: Asymptotic gauges: Generalization of Colombeau
type algebras. Math. Nachr. 289, 2-3, 1–28 (2015)
[21] Gonzales Dominguez, A., Scarfiello, R.: Nota sobre la formula v.p. 1
x
· δ = − 12 δ .
Rev. Un. Mat,. Argentina 1, 53–67 (1956)
[22] Gordon, W.B.: On the diffeomorphisms of Euclidean space. Amer. Math. Monthly
79, 755–759 (1972)
[23] Grant, K.D.E., Mayerhofer, E., Steinbauer, R.: The wave equation on singular space-
times. Commun. Math. Phys. 285, 399–420 (2009)
[24] Grosser, M., Kunzinger, M., Oberguggenberger, M., Steinbauer, R.: Geometric the-
ory of generalized functions. Kluwer, Dordrecht (2001)
[25] Gsponer, A.: A concise introduction to Colombeau generalized functions and their
applications in classical electrodynamics. European J. Phys. 30, no. 1, 109–126 (2009)
[26] Hairer, M.: A theory of regularity structures. Invent. Math. 198, 269–504 (2014)
[27] Hörmann, G., Kunzinger, M., Steinbauer, R.: Wave equations on non-smooth space-
times. In: Asymptotic Properties of Solutions to Hyperbolic Equations, M. Ruzhan-
sky and J. Wirth (eds.) Progress in Mathematics 301. 162–186, Birkhäuser (2012)
[28] Kanwal, R.P.: Generalized functions. Theory and technique. 2nd edition, Birkhäuser
(1998)
[29] Katz, M.G., Tall, D.: A Cauchy–Dirac delta function. Foundations of Science
(2012). See http://dx.doi.org/10.1007/s10699-012-9289-4 and http://arxiv.org/abs/
1206.0119.
[30] Konjik, S., Kunzinger, M., Oberguggenberger, M.: Foundations of the Calculus of
Variations in Generalized Function Algebras. Acta Applicandae Mathematicae 103
n. 2, 169–199 (2008)
[31] Krantz, S.G., Parks, H.R.: The implicit function theorem. History, theory, and
applications. Reprint of the 2003 edition. Modern Birkhäuser Classics. Birkhäu-
ser/Springer, New York (2013)
[32] Kunzinger: M. Generalized functions valued in a smooth manifold. Monatsh. Math.,
137 31–49 (2002)
[33] Laugwitz, D.: Definite values of infinite sums: aspects of the foundations of infini-
tesimal analysis around 1820. Arch. Hist. Exact Sci. 39, no. 3, 195–245 (1989)
[34] Laugwitz, D.: Early delta functions and the use of infinitesimals in research. Revue
d’histoire des sciences, tome 45, 1, 115–128 (1992)
[35] L
ojasiewicz, S.: Sur la valeur et la limite d’une distribution en un point. Studia Math.
16, 1–36 (1957)
[36] L
ojasiewicz, S.: Sur la fixation des variables dans une distribution. Studia Math. 17,
1–64 (1958)
[37] Luperi Baglini, L., Giordano, P.: A fixed point iteration method for arbitrary gen-
eralized ODE. In preparation.
[38] Malgrange, B.: On nonlinear differential Galois theory. Dedicated to the memory of
Jacques-Louis Lions. Chinese Ann. Math. Ser. B 23, no. 2, pp. 219–226 (2002).
[39] Malgrange, B.: Pseudogroupes de Galois et théorie de Galois différentielle. Prépu-
blications IHES M/10/11 (2010).
[40] Mayerhofer, E.: On Lorentz geometry in algebras of generalized functions. Proc. Roy.
Soc. Edinburgh Sect. A 138, no. 4, 843–871 (2008)
[41] Marsden, J.E.: Generalized Hamiltonian mechanics a mathematical exposition of
non-smooth dynamical systems and classical Hamiltonian mechanics. Archive for
Rational Mechanics and Analysis 28, n. 5, 323–361 (1968).
[42] Marsden, J.E.: Non smooth geodesic flows and classical mechanics. Canad. Math.
Bull. 12, 209–212 (1969)
[43] Mikusinski, J.: On the square of the Dirac delta distribution. Bull. Acad. Polon. Sci.
14, 511–513 (1966)
[44] Oberguggenberger, M., Vernaeve, H.: Internal sets and internal functions in Colom-
beau theory. J. Math. Anal. Appl. 341, 649–659 (2008)
[45] Peetre, J.: On the value of a distribution at a point. Portugaliae Math. 27, 149–159
(1968)
[46] Raju, C.K.: Products and compositions with the Dirac delta function. J. Phys. A:
Math. Gen. 15, 381–396 (1982)
[47] Robinson, A., Non-Standard Analysis. North-Holland, Amsterdam, (1966)
[48] Robinson, A.: Function theory on some nonarchimedean fields. Amer. Math. Monthly
80 (6) 87–109; Part II: Papers in the Foundations of Mathematics (1973)
[49] Sebastio e Silva, J.: Sur une construction axiomatique de la théorie des distributions.
Rev. Fac. Cincias Lisboa, 2a Serie A, 4, pp. 79–186, (1954/55)
[50] van der Pol, B., Bremmer, H.: Operational calculus (3rd ed.). New York: Chelsea
Publishing Co. (1987)
[51] Vernaeve, H., Vindas, J.: Characterization of distributions having a value at a point
in the sense of Robinson. J. Math. Anal. Appl. 396, 371–374 (2012)
Paolo Giordano and Michael Kunzinger

University of Vienna
Oskar-Morgenstern-Platz 1
A-1090 Wien
[email protected]
Operator Theory:

The Stochastic LQR Optimal Control

with Fractional Brownian Motion
Tijana Levajković, Hermann Mena and Amjad Tuffaha
Abstract. We consider the stochastic linear quadratic optimal control prob-

lem where the state equation is given by a stochastic differential equation
of the Itô–Skorokhod type with respect to fractional Brownian motion. The
dynamics are driven by strongly continuous semigroups and the cost func-
tional is quadratic. We use the fractional isometry mapping defined between
the space of square integrable stochastic processes with respect to fractional
Gaussian white noise measure and the space of integrable stochastic processes
with respect to the classical Gaussian white noise measure. By this mapping
we transform the fractional state equation to a state equation with Brownian
motion. Applying the chaos expansion approach, we can solve the optimal
control problem with respect to a state equation with the standard Brown-
ian motion. We recover the solution of the original problem by the inverse
of the fractional isometry mapping. Finally, we consider a general form of
the state equation related to the Gaussian colored noise, we study the con-
trol problem, a system with an algebraic constraint and a particular example
involving generalized operators from the Malliavin calculus.
1. Introduction
The linear quadratic Gaussian control problem for the control of finite-dimensional
linear stochastic systems with Brownian motion is well understood, see, e.g.,
[15]. The case for fractional Brownian motion [10, 11, 12] as well as the infinite-
dimensional case have been studied recently [9]. A more general problem arises
if the noise depends on the state variable, this is the so-called stochastic linear
quadratic regulator (SLQR) problem. The SLQR problem in infinite dimensions
was solved by Ichikawa in [22] using a dynamic programming approach. Da Prato
[8] and Flandoli [14] later considered the SLQR for systems driven by analytic
semigroups with Dirichlet or Neumann boundary controls, but with disturbance
in the state only. The infinite-dimensional SLQR with random coefficients has
been investigated in [16, 17] along with the associated backward stochastic Riccati
116 T. Levajković, H. Mena and A. Tuffaha
equation. Recently, a theoretical framework for the SLQR has been laid for sin-
gular estimates control systems in the presence of noise in the control and in the
case of finite time penalization in the performance index [18]. Considering the gen-
eral setting described in [18, 26], an approximation scheme for solving the control
problem and the associated Riccati equation has been proposed in [28]. In [27],
a novel approach for solving the SLQR based on the concept of chaos expansion
from white noise analysis is proposed. In this paper we extend the results from
[27] to the SLQR problem with fractional Brownian motion.
Fractional Brownian motion B (H) is a one-parameter extension of a stan-
dard Brownian motion and the main properties of such a Gaussian process de-
pend on values of the Hurst parameter H ∈ (0, 1). Fractional Brownian motion,
as a process with independent increments which have a long-range dependence
and self-similarity properties found many applications when modeling wide range
of problems in hydrology, telecommunications, queueing theory and mathematical
finance [5]. A specific construction of a stochastic integral with respect to a frac-
tional Brownian motion defined for all possible values H ∈ (0, 1), was introduced
by Elliot and van der Hoek in [13]. Several different definitions of stochastic inte-
gration for fractional Brownian motion appear in literature [5, 13, 39, 42]. In this
paper we follow [13] and use the definition of the fractional white noise spaces by
use of the fractional transform mapping for all values of H ∈ (0, 1) and the ex-
tension of the action of the fractional transform operator to a class of generalized
stochastic processes. The main properties of the fractional transform operator and
the connection of a fractional Brownian motion with a classical Brownian motion
on the classical white noise space were presented in [5, 33].
We consider the infinite-dimensional SLQR problem, which consists of the
state equation
dy (t) = (A y(t) + B u(t)) dt + C y(t) dB (H) (t), y (0) = y 0 , t ∈ [0, T ], (1)
defined on Hilbert state space H, where A and C are operators on H and B

acts from the control space U to the state space H and y 0 is a random variable.
Spaces H and U are Hilbert spaces. The operators B and C are considered to be
linear and bounded, while A could be unbounded. The objective is to minimize
the functional
' (
T
J (u) = E
(H)
R y W + u U dt + G yT Z
2 2 2
(2)
0
over all possible controls u and subject to the condition that y satisfies the state
equation (1). The operators R and G are bounded observation operators taking
values in Hilbert spaces W and Z respectively, E denotes the expectation and
yT = y(T ). A control process u∗ is called optimal if it minimizes the cost (2) over
all control processes, i.e.,
min J(H) (u) = J(H) (u∗ ).
u
The SLQR with Fractional Brownian motion 117
The corresponding optimal trajectory is denoted by y ∗ . Thus, the pair (y ∗ , u∗ ) is

the optimal solution of the problem (1) and (2) and is called the optimal pair.
Following [13] and [33] we construct a fractional isometry in order to trans-
form optimal control problem (1)–(2) from a fractional space to the corresponding
optimal control problem with the state equation given with respect to Brownian
motion
dy(t) = (A y(t) + B u(t)) dt + C y(t) dB(t), y(0) = y 0 , t ∈ [0, T ],
and the performance index
' (
T
J(u) = E Ry 2
W + u 2
U dt + G yT 2
Z .
0
We combine the chaos expansion method with deterministic theory of optimal

control to solve the above optimal control problem. The solution of the initial
problem is thus obtained through the inverse fractional map.
Moreover, we also consider a general state equation of the form
ẏ = Ay + T♦y + Bu, y(0) = y0 , (3)
where A is an operator which generates a strongly continuous semigroup, and
T is a linear bounded operator which combined with Wick product ♦ introduces
convolution-type perturbations into the equation. Equation (3) is related to Gauss-
ian colored noise. The existence and uniqueness of its generalized solution was
proven in [34]. Examples of this type of equations are: the heat equation with
random potential, the heat equation in random (inhomogeneous and anisotropic)
media, the Langevin equation, etc. The related control problem for (3) will lead
to an optimal control defined in a space of generalized processes. A particular case
of (3) together with an algebraic constraint arises in fluid dynamics, e.g., Stokes
equations. The resulting system is known as semi-explicit operator differential al-
gebraic equation (ODAE) and it has the form
ẏ = Ay + B u + T♦y + f, B y = g.
We conclude the paper with the study of an ODAE involving generalized opera-
tors of Malliavin calculus. Particularly, we set the operator B to be the Skorohod
integral δ and B the Malliavin derivative D. Equations involving generalized ope-
rators of Malliavin calculus were studied in [29, 30, 31, 34, 35].
The paper is organized as follows. In Section 2 we briefly state the theoretical
background needed, then in Section 3 we define the fractional isometry operator
M, prove its properties and study the optimal control problem with state equation
given in the form of fractional Itô–Skorokhod integral in fractional space. By using
the fractional isometry we study the control problem in the standard space, prove
the existence and uniqueness of the control and characterize the optimality of
our approach. Finally, we extend our results and solve an ODAE involving the
operators of Malliavin calculus.
2. Theoretical background
Let U and H be separable Hilbert spaces of controls and states, respectively,
with norms · U and · H , generated by the corresponding scalar products.
Let (Ω, F , P) be a complete probability space and let (bt )t≥0 be a real-valued
one-dimensional Brownian motion defined on (Ω, F , P). Let (Ft )t≥0 be the com-
plete right continuous σ-algebra generated by (bt )t≥0 . We assume that all function
spaces are adapted to the filtration (Ft )t≥0 . Let L2 (Ω, P) = L2 (Ω, F , P) be the
Hilbert space of square integrable real-valued random variables endowed with the
norm F 2L2 (Ω,P) = EP (F 2 ), for F ∈ L2 (Ω, P), induced by the scalar product
(F, G)L2 (Ω,P) = EP (F G), for F, G ∈ L2 (Ω, P), and EP denotes the expectation
with respect to the measure P. Throughout the paper, when it is clear which mea-
sure P is used, we will write E for the expectation and L2 (Ω) for L2 (Ω, P) omitting
P. We denote by L2 (Ω, U) the Hilbert space of U-valued square integrable random
variables and by L2 ([0, T ] × Ω, U) we denote the Hilbert space of square integrable
FT -predictable U-valued stochastic processes u endowed with the norm
T
2
u L2 ([0,T ]×Ω,U ) = E ( u(t) 2U ) dt.
0
Let C([0, T ], L (Ω, H)) be the Hilbert space of FT -predictable continuous H-valued
2
stochastic processes y endowed with the norm

y 2
C([0,T ],L2 (Ω,H)) = sup E ( y(t) 2
H ).
t∈[0,T ]
2.1. The SLQR problem: existence of solution

The infinite-dimensional linear quadratic regulator (LQR) stochastic optimal con-
trol problem on Hilbert spaces with respect to Brownian motion is given by the
state equation
dy(t) = (Ay(t) + Bu(t)) dt + Cy(t) dB(t), y(0) = y 0 , t ∈ [0, T ], (4)
subject to the quadratic cost functional
' (
T
J(u) = E Ry 2W + u 2
U dt + GyT 2Z . (5)
0
The dynamics of the problem, the operator A, is deterministic and represents an

infinitesimal generator of a strongly continuous semigroup (eAt )t≥0 on the state
space H. The operators A and C are operators on H, while B is the operator
acting from the control space U to the state space H. We take the operator C
to be linear and bounded. We assume the operators R and G to be linear and
bounded operators acting on the state space H into Hilbert spaces W and Z
respectively. For simplicity, we shall assume that W = Z = H from here onwards.
We denote by D(S) the domain of a certain operator S, and by S the adjoint
operator of S.
The aim of the stochastic linear quadratic problem is to minimize the cost
functional J(u) over a set of square integrable controls u ∈ L2 ([0, T ], L2 (Ω, U)),
which are adapted to the filtration (Ft )t≥0 .
The following theorem gives conditions for the existence of the optimal control
in the feedback form using the associated Riccati equation. For more details on
existence of mild solutions to the SDE (4) we refer to [7] and for the optimal
control and Riccati feedback synthesis we refer the reader to [22].
Theorem 1 ([7, 22]). Let the following assumptions hold:
(a1) The linear operator A is the infinitesimal generator of a C0 -semigroup
(eAt )t≥0 on the space H.
(a2) The linear control operator B is bounded U → H.
(a3) The operators R, G, C are bounded linear operators.
Then the optimal control u∗ of the linear quadratic problem (4)–(5) satisfies the
feedback characterization in terms of the optimal state y ∗
u∗ (t) = −B P(t) y ∗ (t),
where P(t) is a positive self-adjoint operator solving the Riccati equation
Ṗ(t) + P(t)A + A P(t) + C P(t)C + R R − P(t)BB P(t) = 0,
(6)
P(T ) = G G.
2.1.1. Inhomogeneous deterministic LQR problem. Here we invoke the solution to
the inhomogeneous deterministic control problem of minimizing the performance
index T
2 2 2
J(u) = ( Rx H + u U ) dt + Gx(T ) H (7)
0
subject to the inhomogeneous differential equation
x (t) = Ax(t) + Bu(t) + f (t), x(0) = x0 . (8)
Besides the assumptions (a1) and (a2) from Theorem 1 made on A and B, it
is enough to assume that f ∈ L2 ((0, T ), H) to obtain the optimal solution for
the state and control (x∗ , u∗ ). The feedback form of the optimal control for the
inhomogeneous problem (7)–(8) is given by
u∗ (t) = −B Pd (t)x∗ (t) − B k(t),
where Pd (t) solves the Riccati equation
(P˙d + Pd A + A Pd + R R − Pd BB Pd ) v, w = 0,
(9)
Pd (T )v = G Gv,
for all v, w in D(A), while k(t) is a solution to the auxiliary differential equation
k (t) + (A − Pd (t)BB )k(t) + Pd (t)f (t) = 0
with the boundary conditions
Pd (T ) = G G and k(T ) = 0.
For the homogeneous problem we refer to [24], and for the inhomogeneous optimal
control problem for singular estimate type systems we refer to [25].
2.1.2. Strong and mild solutions. Let g(t) be an FT -predictable Bochner integrable
H-valued function. An H-valued adapted process y(t) is a strong solution to the
state equation (4) over [0, T ] if:
(1) y(t) takes values in D(A) ∩ D(C) for almost all t and ω;
T T
(2) P ( 0 y(s) H + Ay(s) H ds < ∞) = 1 and P ( 0 Cy(s) 2H ds < ∞) = 1;
(3) for arbitrary t ∈ [0, T ] and P-almost surely, it satisfies the integral equation
t t t
y(t) = y 0 + Ay(s) ds + g(s)ds + Cy(s) dBs .
0 0 0
An H-valued adapted process y(t) is a mild solution to the state equation

dy(t) = (Ay(t) + g(t)) dt + Cy(t) dB(t), y(0) = y 0 , t ∈ [0, T ],
over [0, T ] if:
(1) y(t) takes values in D(C);
T T
(2) P ( 0 y(s) H ds < ∞) = 1 and P ( 0 Cy(s) 2H ds < ∞) = 1;
(3) for arbitrary t ∈ [0, T ] and P-almost surely, it satisfies the integral equation
t t
y(t) = eAt y 0 + eA(t−s) g(s) ds + eA(t−s) Cy(s) dBs .
0 0
Note that, under the assumptions of Theorem 1, and given a control u from
L2 ([0, T ]; L2(Ω, U)), i.e., g(t) = Bu(t), and the deterministic initial data y 0 ∈ H,
there exits a unique mild solution y ∈ L2 ([0, T ]; L2 (Ω, H)) to the controlled state
equation (4), cf. [7].
2.2. Fractional Brownian motion

Fractional Brownian motion is one-parameter extension of a Brownian motion.
It depends on the Hurst index H which takes values in (0, 1). The name is due
to the climatologist Hurst, who developed statistical analysis of the early water
run-offs of the river Nile. In the framework of Hilbert spaces, fractional Brownian
motion was first introduced by Kolmogorov in 1940, where it was called the Wiener
Spirals. The name fractional Brownian motion is due to Mandelbrot and Van Ness,
who gave a stochastic integral representation of this process in terms of Brownian
motion on an infinite interval [38].
Fractional Brownian motion is a process with dependent increments which
have long-range dependence and self-similarity properties. For H > 12 fractional
Brownian motion has a certain memory feature, which is suitable for modeling
weather derivatives, temperature at a specific place as a function of time, water
level in a river as a function of time or for describing the values of the log returns
of a stock. On the other hand, for H < 12 fractional Brownian motion has a certain
turbulence feature, which is applicable in mathematical finance in the modeling of
financial turbulence, i.e., empirical volatility of a stock or in modeling the prices

of electricity in a liberated Nordic electricity market [5, 13, 38, 39, 40].
Definition 1. A one-dimensional real-valued fractional Brownian motion with the

Hurst index H ∈ (0, 1) on a probability space (Ω, F , P) is a Gaussian process
b(H) = (b(H) (·))t∈R satisfying:
(H)
(a) b0 = 0 a.s.,
(H)
(b) zero expectation, i.e., E[bt ] = 0 for all t ∈ R, and
(c) the covariance function is of the form
(H) 1 2H
E b(H)
s bt = |t| + |s|2H − |t − s|2H , s, t ∈ R. (10)
2
Fractional Brownian motion is a centered Gaussian process with non-inde-
pendent stationary increments and its dependence structure is modified by the
Hurst parameter H ∈ (0, 1). For H = 12 the covariance function can be written in
(1) (1) (1)
the form E(bt 2 bs 2 ) = min{s, t} and the process bt 2 becomes a Brownian motion
bt , which has independent increments. Moreover, for H = 12 fractional Brownian
motion is neither a semimartingale nor a Markov process. From (10) it follows that
(H)
E(bt − b(H)
s ) = |t − s|
2 2H
.
According to the Kolmogorov continuity criterion fractional Brownian motion b(H)

has a continuous modification [39]. The parameter H controls the regularity of tra-
jectories. The covariance function (10) is homogeneous of order 2H, thus fractional
(H) (H)
Brownian motion b(H) is an H self-similar process, i.e., bkt = k H bt , k > 0.
For any n ∈ Z, n = 0 it holds
1 n+1

(H) (H)
r(n) = E[b1 (bn+1 − b(H)
n )] = H(2H − 1) (u − v)2H−2 dudv
0 n
∼ H(2H − 1)|n|2H−1 , as |n| → ∞.
Therefore, the increments are positively correlated for H ∈ ( 12 , 1) and negatively

correlated for H ∈ (0, 12 ). More precisely, for H$∈ ( 12 , 1) fractional Brownian
∞
motion has the long-range dependence
$ property n=1 r(n) = ∞ and for H ∈
∞
(0, 12 ) the short-range property n=1 |r(n)| < ∞. For more details we refer to
[5, 20, 39, 41, 46].
2.3. White noise analysis and chaos expansions

In this section, we briefly recall some basic facts from white noise analysis. De-
x2 d n 2
− x2
note by hn (x) = (−1)n e 2 dx n (e ), n ∈ N0 , N0 = N ∪ {0}, the family of
x2 √
Hermite polynomials and ξn (x) = √
4
√1 e− 2 hn−1 ( 2x), n ∈ N, the fam-
π (n−1)!
ily of Hermite functions. The family of Hermite functions forms a complete or-
thonormal system in L2 (R). These functions are the eigenfunctions for the har-
monic oscillator in quantum mechanics. Clearly, the elements of {ξn }n∈N be-
long to the Schwartz space of rapidly decreasing functions S(R), i.e., they de-
cay faster than polynomials of any degree. The Schwartz spaces can be char-
acterized in terms of the Hermite basis in the following manner: The space of
rapidly decreasing functions as a projective limit space S(R) = l∈N0 Sl (R) where
$∞ $∞
Sl (R) = {f = k=1 ak ξk ∈ L2 (R) : f 2l = k=1 a2k (2k)l < ∞}, l ∈ N0 and the
space of tempered distributions as an inductive limit space S (R) = l∈N0 S−l (R)
$ $∞ 2
where S−l (R) = {f = ∞ 2
k=1 ak ξk : f l = k=1 ak (2k)
−l
< ∞}, l ∈ N0 . Also, we

have a Gel’fand triple S(R) ⊆ L (R) ⊆ S (R) with continuous inclusions.
2
2.3.1. Gaussian white noise space. Throughout the paper all analysis is provided
on two white noise spaces. Here we introduce the (classical) Gaussian white noise
space (S (R), B, μ) and later in Section 2.3.6 we will introduce the fractional Gauss-
ian white noise space (S (R), B, μH ). In both cases, we follow the ideas of Hida
from [19]. The underlying space is the space of tempered distributions S (R). By
B we denote the Borel sigma-algebra generated by the weak topology on S (R)
and μ is the Gaussian white noise measure given by the Bochner–Minlos theorem

− 1 φ2
eiω,φ dμ(ω) = e 2 L2 (R) , φ ∈ S(R),
S (R)
where ω, φ denotes the dual pairing between a tempered distribution ω ∈ S (R)
and a test function φ ∈ S(R).
Denote by I = (NN 0 )c the set of sequences of non-negative integers which
have only finitely many nonzero components. All multi-indices α ∈ I are of the
form α = (α1 , α2 , . . . , αm , 0, 0, . . .), αi ∈ N0 , i = 1, 2, . . . , m, m ∈ N. Particularly,
0 = (0, 0, . . . ) is the zeroth vector and ε(k) = (0, . . . , 0, 1, 0, . . . ), k ∈$N is the kth
∞
unit vector.
-∞The length of a multi-index α$ ∈ I is defined by |α| = k=1 αk . Let
−pα
α αk
(2N) = k=1 (2k) . It was proven that α∈I (2N) < ∞ for p > 1, cf [21].
We say α ≥ β if αk ≥ βk for all k ∈ N. In this case α − β = (α1 − β1 , α2 − β2 , . . . ).
For α < β the difference α − β is not defined.
The space L2 (μ) = L2 (Ω, μ) = L2 (S (R), B, μ) is the Hilbert space of square
integrable random variables with respect to the Gaussian measure μ, i.e., the space
of random variables with finite second moments.
Definition 2. The Fourier–Hermite polynomials on L2 (μ) are defined by
∞

Hα (ω) = hαk (ω, ξk ), α ∈ I. (11)
k=1
Particularly, H0 (ω) = 1 and Hε(k) (ω) = ω, ξk , k ∈ N. The family {Hα }α∈I
forms an orthogonal basis of L2 (μ) with Hα 2L2 (μ) = α!, see [21].
Theorem 2 (Wiener–Itô chaos expansion theorem). Each element F ∈ L2 (μ) has

a unique representation of the form

F (ω) = aα Hα (ω),
α∈I
$
with real coefficients aα , α ∈ I, ω ∈ Ω, such that F 2
L2 (μ) = α∈I a2α α! < ∞.
The space spanned by {Hα : |α| = k} is called the Wiener chaos of order
k and is denoted by Hk , k ∈ N0 . Each Hk is an infinite-dimensional subspace of
L2 (μ) and
∞
4
2
L (μ) = Hk ,
k=0
where the sum is an orthogonal sum [21].

Let H be a real separable Hilbert space. Then each element5F of the space of
∞
Hilbert-valued square integrable random
$ variables L2 (Ω, H) = k=0 Hk (H), can
be represented in the form F (ω) = α∈I fα Hα (ω), for fα ∈ H, α ∈ I, such that

F 2L2 (Ω,H) = fα 2H α! < ∞.
α∈I
One of the typical complications that arise in solving stochastic differen-

tial equations is the blowup of L2 -norms of F , i.e., infinite variance. Therefore,
weighted spaces of random variables in which the considered equation has a solu-
tion have to be introduced. For example, such spaces are the Kondratiev spaces
(S)−ρ , ρ ∈ [0, 1] of generalized random variables, which represent the stochastic
analogue of the Schwartz spaces as generalized function spaces. The largest space
of Kondratiev stochastic distributions is (S)−1 , obtained for ρ = 1.
The space of the Kondratiev test random variables (S)1 can be constructed
as the projective limit of the family of spaces

(S)1,p = f (ω) = aα Hα (ω) ∈ L (μ) : f 1,p =
2 2
aα (α!) (2N) < ∞ ,
2 2 pα
α∈I α∈I
p ∈ N0 . The space of the Kondratiev generalized random variables (S)−1 can be

constructed as the inductive limit of the family of spaces

(S)−1,−p = F (ω) = bα Hα (ω) : f 2−1,−p = b2α (2N)−pα < ∞ , p ∈ N0 .
α∈I α∈I

It holds (S)1 = p∈N0 (S)1,p and (S)−1 = p∈N0 (S)−1,p . The action of a gene-
$
$ random variable F = α∈I bα Hα (ω) ∈ $
ralized (S)−1 on a test random variable
f = α∈I aα Hα (ω) ∈ (S)1 is given by F, f = α∈I α! aα bα . It holds that (S)1
is a nuclear space with the Gel’fand triple (S)1 ⊆ L2 (μ) ⊆ (S)−1 with continuous
inclusions [21].
$ $
Definition 3. For F (ω) = α∈I fα Hα (ω) and G(ω) = β∈I gβ Hβ (ω) the element
F ♦G is called the Wick product of F and G and is given in the form

F ♦G(ω) = fα gβ Hα+β (ω) = fα gγ−α Hγ (ω). (12)
α∈I β∈I γ∈I α≤γ
The Kondratiev spaces (S)1 and (S)−1 are closed under the Wick multiplica-
tion. The Wick product is a commutative, associative operation, and is distributive
with respect to addition. In particular, for the orthogonal polynomial basis of L2 (μ)
we have Hα ♦Hβ = Hα+β , for all α, β ∈ I. Whenever F , G and F ♦G are integrable
it holds that E(F ♦G) = E(F ) · E(G), without independence requirement [21, 31].
2.3.2. Stochastic processes. A square integrable real-valued stochastic process is

defined as a measurable mapping [0, T ] → L2 (μ). A generalized stochastic process
is considered to be a measurable mapping from [0, T ] into a Kondratiev space
(S)−1 . The chaos expansion representation of generalized stochastic process F
follows from Theorem 2. A generalized process F can be represented in the form

Ft (ω) = fα (t) Hα (ω), t ∈ [0, T ],
α∈I
where fα , α ∈ I are measurable real functions and there exists p ∈ N0 such that
for all t ∈ [0, T ]

|fα (t)|2 (2N)−pα < ∞.
α∈I
If we assume H to be a real separable Hilbert space, then Theorem 2 can be
extended also for H-valued stochastic processes. Particularly, a square integrable
H-valued stochastic processes v is an element of L2 ([0, T ]×Ω, H) ∼
= L2 ([0, T ], H)⊗
2
L (Ω, μ) and can be represented in the chaos expansion form

v(t, ω) = vα (t)Hα (ω)
α∈I
(13)
= v0 (t) + vε(k) (t) Hε(k) (ω) + vα (t) Hα (ω), t ∈ [0, T ],
k∈N |α|>1
where vα ∈ L ([0, T ], H) such that it holds

2

vα 2L2 ([0,T ],H) α! < ∞. (14)
α∈I
A process v with the chaos expansion representation (13) that instead of (14)
satisfies the condition

vα 2L2 ([0,T ],H) (2N)−pα < ∞ (15)
α∈I
belongs to L ([0, T ], H) ⊗ (S)−1 and is considered to be a generalized stochastic

2
process. The coefficient v0 (t) is the deterministic part of v in (13) and represents
the (generalized) expectation of the process v.
Denote by {en (t)}n∈N the orthonormal basis of L2 ([0, T ], H), i.e., the basis
obtained by diagonalizing the orthonormal basis {bi (t)sj }i,j∈N , where {bi (t)}i∈N
is the orthonormal basis of L2 ([0, T ]) and {sj }j∈N is the orthonormal basis of H.
The coefficients vα (t) ∈ L2 ([0, T ], H), α ∈ I can be represented in the form

vα (t) = vα,j (t) sj = vα,j,i bi (t) sj , α ∈ I
j∈N j∈N i∈N
with vα,j ∈ L ([0, T ]) and vα,j,i ∈ R. Then the chaos expansion (13) of a stochastic
2
process v ∈ L2 ([0, T ], H) ⊗ L2 (Ω, μ) can be written as

v(t, ω) = vα (t)Hα (ω) = vα,j,i sj bi (t)Hα (ω).
α∈I α∈I j∈N i∈N
After a diagonalization of N × N → N it can be rearranged to

v(t, ω) = vα,n en (t) Hα (ω), vα,n ∈ R, ω ∈ Ω, t ∈ [0, T ].
α∈I n∈N
Example 1. (a) A one-dimensional real-valued Brownianmotion can be represented

$∞ t
in the chaos expansion form bt (ω) = k=1 0 ξk (s)ds Hε(k) (ω), t ≥ 0. For each
t it is an element of L2 (μ). A singular real-valued white noise is defined by the
formal chaos expansion
∞
wt (ω) = ξk (t)Hε(k) (ω). (16)
k=1
$∞ $∞ $∞
Since k=1 |ξk (t)| >
2 1
k=1 k= ∞ and k=1 |ξk (t)|2 (2k)−p < ∞ holds for
p > 1, it follows that the singular white noise is an element of the space (S)−1 ,
for all t ≥ 0, see [21]. It is integrable and the relation dt
d
bt = wt holds in the
distributional sense. Both Brownian motion and singular white noise are Gaussian
processes and belong to the Wiener chaos space of order one.
(b) An H-valued white noise process is given in the chaos expansion form
∞

Wt (ω) = ek (t) Hε(k) (ω). (17)
k=1
$
Note that the H-valued white noise can be also defined as n∈N wtn (ω) sn , where
(n)
wt (ω) are independent copies of one-dimensional white noise (16) and {sn }n∈N
is the orthonormal basis of H. This definition can be reduced to (17) since
(n)
wt (ω)sn = ξk (t) Hε(k) (ω)sn
n∈N n∈N k∈N
∞

= ξi (t) si Hε(i) (ω) sn = ei (t)Hε(i) (ω),
i∈N i=1
where {ei }i∈N is the orthogonal basis of L2 (R, H) obtained by diagonalizing the
basis {ξk (t)sn }k,n∈N .
(c) In general, the chaos expansion representation of an H-valued Gaussian

process that belongs to the Wiener chaos space of order one is given in the form

Gt (ω) = gk (t) Hε(k) (ω) = gki ei (t) Hε(k) (ω), (18)
k∈N k∈N i∈N
with gk ∈ L2 ([0, T ], H) and gki = (gk , ei )L2 ([0,T ],H) is a real constant. If the condi-
tion

gk 2L2 ([0,T ],H) < ∞ (19)
k∈N
is fulfilled, then Gt belongs to the space L2 ([0, T ]×Ω, H) ∼ = L2 ([0, T ], H)⊗L2(Ω, μ).
If the sum in (19) is infinite then the representation (18) is formal, and if addi-
tionally
(k)
gk 2L2 ([0,T ],H) (2N)−pε = gk 2L2 ([0,T ],H) (2k)−p < ∞,
k∈N k∈N
holds for some p ∈ N0 , the process Gt , for each t, belongs to the Kondratiev space
of stochastic distributions (S)−1 , i.e., G ∈ L2 ([0, T ], H) ⊗ (S)−1 , see [33, 36, 44].
Note that a Gaussian noise represented in (18) can be interpreted as a colored
noise with the representation operator N and the correlation function C = N N ,
such that

N fk (t) Hε(k) (ω) = N fki ei (t) Hε(k) (ω)
k∈N k∈N i∈N

= λi fki ei (t)Hε(k) (ω),
k∈N i∈N
with N ei (t) = λi ei (t), i ∈ N, [37]. Particularly, we will consider the color noise
to be a Gaussian process of the form

Lt (ω) = lk ek (t) Hε(k) (ω), (20)
k∈N
with a sequence of real coefficients {lk }k∈N such that for some p ∈ N it holds

lk2 (2k)−p < ∞ (21)
k∈N
The Wick product of two stochastic processes is defined in an analogous way

as it was defined for random variables and generalized random variables (12), for
more details see [30].
2.3.3. Operators. Following [34], we define two classes of operators on spaces of

stochastic processes, namely coordinatewise and simple coordinatewise operators,
that we are going to deal with in the paper.
Definition 4. An operator O is called a coordinatewise operator

$ if there exists a
family of operators {Oα }α∈I , such that for a process v = vα Hα it holds
α∈I

Ov = Oα (vα ) Hα . (22)
α∈I
Moreover, operator O is a simple coordinatewise operator if Oα = O for all α ∈ I,

i.e., if it holds that

Ov = O(vα ) Hα = O(v0 ) + O(vα ) Hα .
α∈I |α|>0
Lemma 1. Let O : L2 ([0, T ], H) ⊗ L2 (Ω, μ) → L2 ([0, T ], H) ⊗ L2 (Ω, μ) be a

coordinatewise operator that corresponds to a deterministic family of operators
Oα : L2 ([0, T ], H) → L2 ([0, T ], H), α ∈ I. If the operators Oα , α ∈ I are uni-
formly bounded by c > 0 then O is a bounded operator on L2 ([0, T ], H) ⊗ L2 (Ω, μ).
$
Proof. Let Oα op ≤ c for all α ∈ I. Then, for v = α∈I vα Hα in L2 ([0, T ], H) ⊗
L2 (Ω, μ) it holds
Ov 2
L2 ([0,T ],H)⊗L2 (Ω,μ)

= Oα vα 2
L2 ([0,T ],H) α! ≤ Oα 2
op vα 2
L2 ([0,T ],H) α!
α∈I α∈I

≤c 2
vα 2L2 ([0,T ],H) α! = c2 v 2
L2 ([0,T ],H)⊗L2 (Ω,μ) .
α∈I
2.3.4. Stochastic integration and Wick multiplication. For a square integrable pro-
cess v that is adapted in the filtration (Ft )t≥0 generated by an H-valued Brownian
T
motion (Bt )t≥0 , the corresponding stochastic integral 0 vt dBt is considered to
be the Itô integral I(v). When v is not adapted to the filtration, then the stochas-
tic integral is interpreted as the Itô–Skorokhod integral. From the fundamental
theorem of stochastic calculus it follows that the Itô–Skorokhod integral of an H-
valued stochastic process v = vt (ω) can be represented as a Riemann integral of
the Wick product of vt with a singular white noise
T T
δ(v) = v dBt (ω) = v ♦Wt (ω) dt, (23)
0 0
d
where the derivative Wt = Bt is taken in sense of distributions [21].
dt
Thus, for an H-valued adapted processes v the Itô integral and the Skorokhod
integral coincide, i.e., I(v) = δ(v). Note that the Itô integral is an H-valued random
variable. From the Wiener–Itô chaos expansion theorem, Theorem 2, it follows that
there exists a unique family aα , α ∈ I such that the Itô integral can be represented
in the chaos expansion form

I(v) = a α Hα . (24)
α∈I
On the other hand, by (12), (17) and (23) we $ obtain a chaos expansion represen-
tation of the Skorokhod integral, i.e., for v = vα (t)Hα we have
α∈I

v ♦ Wt (ω) = vα (t) Hα (ω) ♦ ek (t) Hε(k) (ω)
α∈I k∈N
(25)
= vα (t) ek (t) Hα+ε(k) (ω).
α∈I k∈N
Thus, from (23) and (25) we obtain

δ(v) = vα,k Hα+ε(k) (ω), (26)
α∈I k∈N
with real coefficients vα,k = (vα , ek )L2 ([0,T ],H) and ω ∈ Ω. Combining (26) and
(24) we obtain the coefficients aα , for all α ∈ I and α > 0 in the form

aα = vα−ε(k) ,k . (27)
k∈N
We use the following convention: vα−ε(k) is not defined if the kth component
of α, i.e., αk equals zero. For example, for α = (0, 3, 0, 2, 0, . . . ) the coefficient
a(0,3,0,2,0,... ) is expressed as the sum of two coefficients of the process v, i.e., from
(27) we have a(0,3,0,2,0,... ) = v(0,2,0,2,0,... ),2 + v(0,3,0,1,0,... ),4 . The obtained chaos
expansion representation form of the Itô–Skorokhod integral (26) will be used in
Section 3, where we will be able to represent explicitly the stochastic perturbation
in the optimal control problem (4). Note also that δ(v) belongs to the Wiener
chaos space of higher order than v, see also [21, 35].
$ 5. A square integrable H-valued stochastic process

Definition v given in the form
v = $ v
α∈I α (t) H α (ω), with the coefficients vα ∈ L 2
([0, T ], H) such that
vα (t) = v
k∈N α,k ke (t), vα,k ∈ R for all α ∈ I is integrable in Itô–Skorokhod
sense if the condition
2 √ 2
α! vα−ε(k) ,k = α! vα,k αk + 1 <∞ (28)
α∈I,|α|>0 k∈N α∈I k∈N
holds. Then the chaos expansion form of the Itô–Skorokhod integral of v is given
by (26) and we write v ∈ Dom(δ).
Theorem 3. The Skorokhod integral δ of an H-valued square integrable stochastic

process is a linear and continuous mapping
δ: Dom(δ) → L2 (Ω).
Proof. Let u, v ∈ L2 ([0, T ], H)⊗L2 (Ω) be integrable in Itô–Skorokhod sense. Then,

for a, b ∈ R it holds

δ(au + bv) = δ (auα,k + bvα,k )ek Hα = (auα,k + bvα,k )Hα+ε(k)
α∈I k∈N α∈I k∈N

=a aα,k Hα+ε(k) + b vα,k Hα+ε(k) = aδ(u) + bδ(v).
Moreover, from (28) and (α + ε(k) )! = (αk + 1) α! for α ∈ I, k ∈ N we obtain

6 62
6 6
2 6
δ(v) L2 (Ω) = 6 vα−ε(k) ,k Hα 6
6
α∈I,|α|>0 k∈N L2 (Ω)
2
= vα−ε(k) ,k α! < ∞.
|α|>0 k∈N
From the estimates

2
vα 2
L2 ([0,T ],H) α! = α! 2
vα,k ≤ α! vα,k
α∈I α∈I k∈N α∈I k∈N
√ 2
≤ α! vα,k αk + 1 <∞
α∈I k∈N
we conclude that if v ∈ Dom(δ) then v ∈ L2 ([0, T ], H) ⊗ L2 (Ω). Moreover, if the

condition

|α| vα 2L2 ([0,T ],H) α! < ∞ (29)
α∈I
is fulfilled then v ∈ Dom(δ). This follows from

√ 2
α! uα,k αk + 1 ≤c α! |α| u2α,k < ∞.
A detailed analysis of domain and range of operators of the Malliavin calculus in

spaces of generalized stochastic processes can be found in [31, 35].
Lemma 2. Let O : L2 ([0, T ], H)⊗L2(Ω) → L2 ([0, T ], H)⊗L2(Ω) be a coordinatewise

operator that corresponds to a uniformly bounded family of linear $ operators Oα :
L2 ([0, T ], H) → L2 ([0, T ], H), α ∈ I. If a stochastic process v = α∈I vα Hα ∈
L2 ([0, T ], H) ⊗ L2 (Ω) satisfies the condition (29) then Ov ∈ Dom(δ).
Proof. Since v ∈ L2 ([0, T ], H) ⊗ L2 (Ω) satisfies (29) then v ∈ Dom(δ), i.e., (28)
holds. Let O corresponds to the family Oα : L2 ([0, T ], H) → L2 ([0, T ], H), α ∈ I
such that Oα L(H) ≤ c, α ∈ I, where L(H) denotes the set of linear bounded
operators on L2 ([0, T ], H). From

Oα vα 2L2 ([0,T ],H) |α| α! ≤ Oα 2
L(H) vα L2 ([0,T ],H) |α| α!
2
α∈I α∈I

≤ c2 vα L2 ([0,T ],H) |α| α!
2
<∞
α∈I
it follows that Ov ∈ Dom(δ).
2.3.5. The fractional transform operator M (H) . In [13] the authors developed the
fractional white noise theory for a Hurst parameter H ∈ (0, 1). They introduced
the fractional transform operator M (H) , which connects the fractional Brownian
(H)
motion bt and the standard Brownian motion bt on the white noise probability
space (S (R), B, μ). We extend these results for H-valued Brownian motion Bt
(H)
and H-valued white noise Wt and their corresponding fractional versions Bt
(H)
and Wt .
Definition 6 ([13]). Let H ∈ (0, 1). The fractional transform operator M (H) :
S(R) → L2 (R) ∩ C ∞ (R) is defined by

M (H) f (y) = |y| 2 −H f(y),
1
y ∈ R, f ∈ S(R), (30)

where f(y) := R e−ixy f (x)dx denotes the Fourier transform of f .
Equivalently, the operator M (H) for all H ∈ (0, 1) can be defined as a constant
multiple of

d 3
− (t − x) |t − x|H− 2 f (t) dt, (31)
dx R
such that the constant is chosen so that (30) holds. The operator M (H) has the
structure of a convolution operator. Particularly, from (31) it follows that for

H ∈ (0, 12 ) the fractional operator is of the form M (H) f (x) = CH R f (x−t)−f3
(x)
dt,
|t| 2 −H
f (t)
then for H ∈ ( 12 , 1) it is of the form M (H) f (x) = CH R 3 −H dt and for
|t−x| 2
1
H = 12 it reduces to the identity operator, i.e., M ( 2 ) f (x) = f (x). The normalizing
constant is CH = (2Γ(H − 12 ) cos( π2 (H − 12 )))−1 and Γ is the Gamma function.
From (30) we have that the inverse fractional transform operator of the op-
erator M (H) is the operator M (1−H) , which is defined by
M
(1−H) f (y) = |y|H− 2 f(y),
1
y ∈ R, f ∈ S(R).
Denote by L2H (R) = {f : R → R : M (H) f (x) ∈ L2 (R)} the closure of S(R)

with respect to the norm f L2H (R) = M (H) f L2 (R) , for f ∈ S(R), induced by the
inner product
(f, g)L2H (R) = (M (H) f, M (H) g)L2 (R) .
The operator M (H) is a self-adjoint operator and for f, g ∈ L2 (R) ∩ L2H (R) we
have

(f, M (H) g)L2H (R) = (f, M (H) g) 2 = |y| 2 −H f(y)
1
g(y)dy
L (R)
R

= (M (H) f ,
g)L2 (R) = (M (H) f, g)L2H (R) .
Remark 1. For fixed H ∈ ( 12 , 1), define φ(s, t) = H(2H − 1)|s − t|2H−2 , s, t ∈ R.
Then,
(M (H) f (x))2 dx = cH f (s)f (t)φ(s, t)dsdt, (32)
R R R
with cH constant. The property (32) was used in [13, 20, 32] and [38] in order to
adapt the classical white noise calculus to the fractional one.
Theorem 4 ([6, 13]). Let M (H) : L2H (R) → L2 (R) defined by (30) be the extension
of the operator M from Definition 6. Then, M (H) is an isometry between the two
Hilbert spaces L2 (R) and L2H (R). The functions
e(H)
n (x) = M
(1−H)
ξn (x), n ∈ N, (33)
belong to S(R) and form an orthonormal basis in L2H (R).
(1−H)
From (33) it also follows en = M (H) ξn , n ∈ N, where we used the fact
(1−H)
that M is the inverse operator of the operator M (H) . Following [6] and [13]
we extend M (H)
onto S (R) and define the fractional operator M (H) : S (R) →

S (R) by
M (H) ω, f = ω, M (H) f , f ∈ S(R), ω ∈ S (R).
2.3.6. Fractional Gaussian white noise space. Following [5], for H ∈ (0, 1) we
denote by
L2 (μH ) = L2 (μ ◦ M (1−H) ) = {G : Ω → R ; G ◦ M (H) ∈ L2 (μ)}.
the stochastic analogue of L2H (R). It is the space of square integrable functions
on S (R) with respect to fractional Gaussian white noise measure μH . Thus, the
space (S (R), B, μH ) denotes the fractional Gaussian white noise space.
Since G ∈ L2 (μH ) if and only if G ◦ M (H) ∈ L2 (μ), it follows that G has an
expansion of the form
∞
(H)
G(M ω) = cα Hα (ω) = cα hαi (ω, ξi )
α∈I α∈I i=1
∞
∞

= cα hαi (ω, M (H) ei ) = cα hαi (M (H) ω, ei ).
α∈I i=1 α∈I i=1
Definition 7. The family of fractional Fourier–Hermite polynomials is defined by

∞

Hα (ω) = hαk (ω, ek ), α ∈ I. (34)
k=1
The family {Hα }α∈I forms an orthogonal basis of L2 (μH ) and for all α ∈ I
it holds Hα 2L2 (μH ) = α!. Therefore, Theorem 2 can be formulated for fractional
square integrable random variables.
Theorem 5. Each G ∈ L2 (μH ) can be uniquely represented in the form

G(ω) = cα Hα (ω), cα ∈ R, α ∈ I, ω ∈ Ω
α∈I
$
such that G 2
L2 (μH ) = c2α α! is finite and G L2 (μH ) = G ◦ M (H) L2 (μ) .
α∈I
(H) (H)
The fractional Kondratiev spaces (S)1 and (S)−1 are defined in an anal-
ogous way as it was done in Section 2.3.1 for stochastic random variables in the
Gaussian white noise case. An H-valued fractional stochastic process v as element
of L2 ([0, T ], H) ⊗ L2 (Ω, μH ) is uniquely defined by

vt (ω) = vα (t) Hα (ω), (35)
α∈I
where vα ∈ L ([0, T ], H), α ∈ I such that (14) holds. Moreover, (35) can be written
2
in the form

vt (ω) = vα,n en (t) Hα (ω), vα,n ∈ R, ω ∈ Ω, t ∈ [0, T ].
α∈I n∈N
(H)
The fractional generalized process v from L2 ([0, T ], H) ⊗ (S)−1 has a chaos ex-
pansion representation of the form (35) such that (15) holds.
The definitions of coordinatewise and simple coordinatewise operators, Sec-
tion 2.3.3, hold for processes defined on both classical white noise space and frac-
tional white noise space.
3. The Stochastic LQR problem with fractional Brownian motion

In order to study the stochastic LQR problem on fractional spaces we introduce
an isometry M between the space of square integrable fractional random variables
L2 (μH ) and the space of integrable random variables L2 (μ). Extending this map-
ping to stochastic processes we can transform the state equation with fractional
Brownian motion to an equation with standard Brownian motion. Therefore, we
can solve the optimal control problem with respect to an equation with standard
Brownian motion and find the solution of the original problem by applying M−1 .
3.1. The fractional operator M
Since M (H) is self-adjoint we can connect (11) and (34) for all α ∈ I
∞
∞
∞

Hα (ω) = hαk (ω, ξk ) = hαk (ω, M (H) ek ) = hαk (M (H) ω, ek )
k=1 k=1 k=1
(H)
= Hα (M ω)
and similarly
Hα (ω) = Hα (M (1−H) ω).
Therefore we define a new (fractional) operator M which maps the orthogonal
basis of L2 (μH ) into the orthogonal basis of L2 (μ).
Definition 8 ([33]). Let M : L2 (μH ) → L2 (μ) be defined by
M(Hα (ω)) = Hα (ω), α ∈ I, ω ∈ Ω.
The operator M and the fractional operator M (1−H) correspond to each
$
other. For G = α∈I cα Hα (ω) ∈ L2 (μH ), by linearity and continuity we extend
M to

M cα Hα (ω) = cα Hα (ω). (36)
α∈I α∈I
Theorem 6 ([33]). The operator M is an isometry between spaces of classical

Gaussian and fractional Gaussian random variables.
Proof. The operator M is the isometry between L2 (μH ) and L2 (μ) because it
holds M(Hα ) L2 (μ) = Hα L2 (μ) = α! = Hα L2 (μH ) .
The action of M can be seen as a transformation of the corresponding ele-
ments of the orthogonal basis {Hα }α∈I into {Hα }α∈I , see [33]. For every element
F ∈ L2 (μ) there exists a unique F ∈ L2 (μH ) so F = MF and also for each
F ∈ L2 (μH ) there exists a unique F ∈ L2 (μ) so F = M−1 F . Further on, such
pairs of elements F and F , that are connected via M, will be called the associ-
ated pairs. The coefficients of the chaos expansion representations of associated
elements F and F coincide.
$ $
Lemma 3. Let F = α∈I fα Hα ∈ L2 (μ) and F = α∈I fα Hα ∈ L2 (μH ). Then
F and F are associated if and only if fα = fα for all α ∈ I.
Proof. Let F and F be associated. Then it holds

fα Hα = F = M (F ) = M f α Hα = f α Hα .
α∈I α∈I α∈I
Since the chaos expansion representation in the orthogonal basis {Hα }α∈I is
unique, it follows that fα = fα for all α ∈ I.
The action of the operator M can be extended to a Kondratiev space of
(H)
stochastic distributions M : (S)−1 → (S)−1 by

M aα Hα (ω) = aα Hα (ω), aα ∈ R.
α∈I α∈I
$
The extension is well defined since there exists p ∈ N so α∈I a2α (2N)−pα < ∞.
In an analogous way the action of the operator M can be extended to stochastic
processes and H-valued (generalized) stochastic processes.
(H)
Example 2. (a) A real-valued fractional Brownian motion bt (ω), H ∈ (0, 1) as
an element of the fractional Gaussian space L2 (μ(1−H) ) = L2 (μ◦M (H) ) is given by
∞ t
(H)
bt (ω) = ξk (s)ds Hε(k) (ω),
k=1 0
(1−H)
with help of the property M (H) ξk = ek , see [33].
(H)
(b) A one-dimensional real-valued fractional singular white noise wt as an
(1−H)
element of the fractional Kondratiev space (S)−1 is defined by the chaos ex-
(H) $ ∞ d (H)
pansion wt (ω) = k=1 ξk (t) Hε(k) (ω). It is integrable and the relation dt bt =
(H)
wt holds in the sense of distributions.
Moreover, combining (16) and (36) we obtain
∞ ∞
M−1 (wt ) = M−1
(H)
ξk Hε(k) = ξk Hε(k ) (ω) = wt .
k=1 k=1
(c) An H-valued fractional white noise in the fractional space is given by
∞
(H)
Wt (ω) = ek (t) Hε(k) (ω), (37)
k=1
where {ek }k∈N is an orthonormal basis in L2 ([0, T ], H). By (17) and (37) the
relations M(Wt ) = Wt and M−1 (Wt ) = Wt
(H) (H)
follow.
From here onwards we will keep the following notation: all processes denoted
with tilde in subscript will be considered $ as elements of a fractional space. There-
fore, due to Lemma 3, each process v = α∈I vα Hα from an H-valued classical
space (particularly L2 ([0, T ], H) ⊗ L2 (Ω, μ) or L2 ([0, T ], H) ⊗ (S)−1 ) will be asso-
$
ciated to a process v = α∈I vα Hα from the corresponding H-valued fractional
(H)
space (particularly L2 ([0, T ], H)⊗L2 (Ω, μH ) or L2 ([0, T ], H)⊗(S)−1 ) via the frac-
tional mapping M, i.e., M(v) = v. Since the coefficients of processes v and v are
equal, it also follows

v 2L2 ([0,T ],H)⊗L2 (Ω,μH ) = α! vα 2L2 ([0,T ],H) = v 2L2 ([0,T ],H)⊗L2 (Ω,μ) . (38)
α∈I
Theorem 7. The fractional mapping M satisfies the following properties:

(1) Let the operators O : L2 ([0, T ], H) ⊗ L2 (Ω, μH ) → L2 ([0, T ], H) ⊗ L2 (Ω, μH )
and O : L2 ([0, T ], H) ⊗ L2 (Ω, μ) → L2 ([0, T ], H) ⊗ L2 (Ω, μ) be coordinatewise
operators that correspond to the same family of operators Oα : L2 ([0, T ], H) →
L2 ([0, T ], H), α ∈ I. Then it holds
M(Ov) = O(Mv),
(2) M is linear and it also holds M(u♦y) = M(u)♦M(y) and
(3) M(EμH v) = Eμ (Mv),
(H)
for v ∈ L2 ([0, T ], H) ⊗ L2 (Ω, μH ) and u, y ∈ L2 ([0, T ], H) ⊗ (S)−1 .
Proof. Since M acts on the orthogonal basis of L2 (Ω, μH ) the following is valid:
(1) Let v ∈ L2 ([0, T ], H) ⊗ L2 (Ω, μH ). From (22) and (36) we obtain

M(Ov) = M Oα vα Hα = Oα vα Hα = O vα Hα = O(Mv).
(2) By definition, the fractional operator M is linear. It is also homogeneous

with respect to the Wick multiplication, i.e., it holds

M(u♦y) = M uα yβ Hα+β = uα yβ Hα+β
α∈I β∈I α∈I β∈I

=M uα Hα ♦M y β Hβ = M(u)♦M(y).
α∈I β∈I
(3) For v ∈ L2 ([0, T ], H)⊗L2 (Ω, μH ) an element EμH v is the zeroth coefficient
of fractional expansion of v, i.e., EμH v = v0 . Thus, M(EμH v) = v0 . On the other
side, Eμ (Mv) is the zeroth coefficient of the expansion of Mv, which is also equal
to v0 . Thus, M(EμH v) = Eμ (Mv).
Theorem 8. For a differentiable H-valued process z from the fractional space the
following holds
d d
M z = Mz .
dt dt
Proof. Differentiation of a stochastic process is a simple coordinatewise operator,
i.e., a process is considered to be differentiable if and only if its coordinates are
differentiable deterministic functions [34]. The assertion follows by applying M to
d $ d $
dt z = α∈I dt zα (t) Hα (ω) = α∈I zα (t) Hα (ω). We obtain
d
M( z) = M zα (t)Hα = zα (t) Hα
dt
α∈I α∈I
d d
= zα (t)Hα = Mz .
dt dt
α∈I
3.1.1. Fractional integral. The fractional Itô–Skorokhod integral δ (H) of an H-

valued process u that belongs to Dom(δ (H) ) in the fractional space is defined in
an analogous way as the Itô–Skorokgod integral (23) in classical space, see Section
$
2.3.4. Clearly, we say that u = α∈I uα Hα ∈ Dom(δ (H) ) if (28) holds. Then
$
the fractional Itô–Skorokhod integral of a process u = α∈I uα Hα in fractional
space
T T
(H) (H)
δ (H) (u) = u dBt = u ♦ Wt dt
0 0
has the chaos expansion representation of the form

δ (H) (u) = uα,k Hα+ε(k) . (39)
α∈I k∈N
The coefficients uα,k are the coefficients of the$expansion of the corresponding

δ(u), where u is the associated process to u = α∈I uα Hα , uα ∈ L2 ([0, T ], H),
α ∈ I.
Theorem 9. For u ∈ Dom(δ) it holds
M(δ (H) (u)) = δ(M(u)). (40)
Proof. From (26), (39) and the definition of operator M the property (40) follows,
M(δ (H) (u)) = δ(u) = δ(M(u)) holds for all associated pairs of processes u and
u = Mu. Since M is an isometry it holds
δ(u) 2
L2 (Ω,μ) = M(δ (H) (u)) 2
L2 (Ω,μ) = δ (H) (u)) 2
L2 (Ω,μH )
√ 2
= α! uα,k αk + 1 < ∞.
α∈I k∈N
Remark 2. The definition of the fractional Itô–Skorokhod integral in the classical

Gaussian space is given in [5, 6, 39]. In [33] the authors provided a detailed analysis
on generalized classical and fractional operators of Malliavin calculus on white
noise spaces.
3.2. The optimal control problem
We consider the state equation
(H)
dy(t) = [A y(t) + B u(t)] dt + C y(t) dBt , y(0) = y 0 , t ∈ [0, T ], (41)
with respect to an H-valued fractional Brownian motion in the fractional Gaussian
white noise space. The objective is to minimize the functional
' (
T
J (u) = EμH
(H)
Ry H + u U dt + GyT H
2 2 2
(42)
0
over all u ∈ L2 ([0, T ] × Ω, U).

Due to the fundamental theorem of stochastic calculus, for admissible square
integrable processes, the fractional state equation (41) is equivalent to its Wick
version
˙
y(t) = Ay(t) + Bu(t) + C y(t) ♦ W (H) (t), y(0) = y 0 , t ∈ [0, T ]. (43)
By using the fractional mapping M one can transfer the optimal control
problem (41)–(42) from the fractional space to the corresponding optimal control
problem with the state equation
dy(t) = [Ay(t) + Bu(t)] dt + Cy(t) dBt , y(0) = y 0 , t ∈ [0, T ]. (44)
with respect to Brownian motion subject to
' (
T
J(u) = Eμ Ry 2H + u 2
U dt + GyT 2H , (45)
0
in the classical Gaussian white noise space. Instead of the state equation (44),
on a set of square integrable processes, one can consider its equivalent Wick-type
equation
ẏ(t) = Ay(t) + Bu(t) + Cy(t) ♦ Wt , y(0) = y 0 , t ∈ [0, T ]. (46)
Once the solution of the optimal control problem (44)–(45) is obtained, then
using the fractional isometry M one can also obtain the solution to the initial
optimal control problem (41)–(42). That is the statement of the following theorem.
Theorem 10. Let the fractional operators A, B, C, R and G defined on fractional

space be coordinatewise operators that correspond to the families of deterministic
operators {Aα }α∈I , {Bα }α∈I , {Cα }α∈I , {Rα }α∈I and {Gα }α∈I respectively. Let
the pair (u∗ , y ∗ ) be the optimal solution of the fractional stochastic optimal control
problem (41)–(42). Then, the pair (M u∗ , M y ∗ ) is the optimal solution (u∗ , y ∗ )
of the associated optimal control problem (44)–(45), where the operators A, B, C,
R and G defined on classical space, are coordinatewise operators that correspond
respectively to the same families of deterministic operators {Aα }α∈I , {Bα }α∈I ,
{Cα }α∈I , {Rα }α∈I and {Gα }α∈I . Moreover, if (u∗ , y ∗ ) is the optimal solution of
the stochastic optimal control problem (44)–(45), then the pair (M−1 u∗ , M−1 y ∗ )
is the optimal solution (u∗ , y ∗ ) to the corresponding fractional optimal control prob-
lem (41)–(42).
Proof. Let (u∗ , y ∗ ) be the optimal pair of the problem (41)–(42), i.e., its equivalent
problem (42)–(43). Then minu J(u) = J(u∗ ), while y ∗ solves (41) and also (43).
Let all operators appearing in the control problem be coordinatewise operators. By
applying the chaos expansion method and the properties of the fractional operator
M stated in Theorem 7 and Theorem 8, we transform (43) in fractional space to
the corresponding state equation in classical space, i.e.,
ẏ(t) = M(Ay(t) + Bu(t) + Cy(t) ♦ W (H) (t))
(H)
= M(Ay) + M(Bu) + M(Cy) ♦ M(Wt )
= Ay + Bu + Cy ♦ Wt ,
where y and u are the associated processes to y and u respectively. Moreover,
by Theorem 7 part (3) and (38) the operator M transforms the cost functional
J(H) to
M(J(H) (u)) = M(EμH (v)) = Eμ (Mv) = Eμ (v) = J(u),
T
where v and v are associated elements v = 0 Ry 2H + u 2U dt + GyT 2H and
T
v= 0 Ry 2H + u 2U dt + GyT 2H .
We will solve the control problem in the classical space (we will generalize
the results from [27]) and then, by use of Theorem 10 via the inverse fractional
mapping M−1 , we obtain the optimal solution for the corresponding fractional
problem.
Theorem 11. Let the following assumptions hold:

(A1) The operator A : L2 ([0, T ], D) ⊗ L2 (Ω, μ) → L2 ([0, T ], D) ⊗ L2 (Ω, μ) is a
coordinatewise linear operator that corresponds to the family of deterministic
operators Aα : L2 ([0, T ], D) → L2 ([0, T ], H), α ∈ I, where Aα are infinites-
imal generators of strongly continuous semigroups (eAα t )α∈I , t ≥ 0, defined
on a common domain D that is dense in H, such that for some m, θ > 0 and
all α ∈ I we have
(eAα t )α L(H) ≤ m eθt , t ≥ 0.
(A2) The operator C : L2 ([0, T ], H) ⊗ L2 (Ω, μ) → L2 ([0, T ], H) ⊗ L2 (Ω, μ) is a

coordinatewise operator corresponding to a family of uniformly bounded deter-
ministic operators Cα : L2 ([0, T ], H) → L2 ([0, T ], H), α ∈ I.
(A3) The control operator B is a simple coordinatewise operator B : L2 ([0, T ], U)⊗
L2 (Ω, μ) → L2 ([0, T ], H) ⊗ L2 (Ω, μ) that is defined by a family of uniformly
bounded deterministic operators Bα : L2 ([0, T ], U) → L2 ([0, T ], H), α ∈ I.
(A4) The operators R and G are bounded coordinatewise operators corresponding
to the families of deterministic operators {Rα }α∈I and {G}α∈I respectively.
(A5) Eμ y 0 2H < ∞.
Then, the optimal control problem (45)–(46) has a unique optimal control u∗ given
in the chaos expansion form

u∗ = − Bα Pd,α (t) yα∗ (t) Hα − Bα kα (t) Hα ,
α∈I |α|>0
where Pd,α (t) for every α ∈ I solves the Riccati equation
Ṗd,α (t) + Pd,α (t)Aα + Aα Pd,α (t) + Rα Rα

− Pd,α (t)Bα Bα Pd,α (t) = 0
(47)
Pd,α (T ) = Gα Gα
and kα (t) is for each α ∈ I a solution to the auxiliary differential equation

kα (t) + (Aα − Pd,α (t)Bα Bα ) kα (t) + Pd,α (t) Cα−ε(i) yα−ε(i) (t) · ei (t) = 0,
i∈N
$ (48)
with the terminal condition kα (T ) = 0 and y ∗ = y
α∈I α
∗
H α is the optimal state.
$ A, B and C are coordinatewise,

Proof. Since all the operators $ by (22) the actions
are given by$Ay(t, ω) = α∈I Ayα (t) Hα (ω), Bu(t) = α∈I Buα (t) Hα (ω) and
Cy(t, ω) = α∈I Cyα (t) Hα (ω), for

y(t, ω) = yα (t)Hα (ω), u(t, ω) = uα (t)Hα (ω) (49)
α∈I α∈I
such that for all α ∈ I the coefficients yα ∈ L ([0, T ], H) and uα ∈ L2 ([0, T ], U).
2
From (A2) and (A3) we conclude that the operators C and B are bounded and
by Lemma 1 it holds

Bu 2
L2 ([0,T ],H)⊗L2 (Ω,μ) = α! Bα uα 2
L2 ([0,T ],H)
α∈I

≤ c2 α! uα 2
L2 ([0,T ],U ) = c2 u 2
L2 ([0,T ],U )⊗L2 (Ω,μ) ,
α∈I
where Bα ≤ c for all α ∈ I.

We divide the proof into several steps. First, we consider the Wick version
(46) of the state equation (44), we apply the chaos expansion method and obtain a
system of deterministic equations. By representing y and y 0 in their chaos expan-
sion forms, the initial condition y(0) = y 0 , for a given H-valued random variable
y 0 , is reduced to a family of initial conditions for the coefficients of the state
yα (0) = yα0 , for all α ∈ I, where yα0 ∈ H, α ∈ I.
With the chaos expansion method the state equation (46) transforms to the
system of infinitely many deterministic initial value problems:
1◦ for α = 0:
y0 (t) = A0 y0 (t) + B0 u0 (t), y0 (0) = y00 , (50)

◦
2 for |α| > 0:

yα (t) = Aα yα (t) + Bα uα (t) + Cα−ε(i) yα−ε(i) (t) · ei (t),
i∈N (51)
yα (0) = yα0 ,
where the unknowns correspond to the coefficients of the control and the state
variables. It describes how the stochastic state equation propagates chaos through
different levels. Note that for α = 0, the equation (50) corresponds to the determi-
nistic version of the problem and the state y0 is the expected value of y. The terms
yα−ε(i) (t) are obtained recursively with respect to the length of α. The sum in (51)
goes through all possible decompositions of α, i.e., for all j for which α − ε(j) is
defined. Therefore, the sum has as many terms as multi-index α has non-zero
components. Existence and uniqueness of solutions of (50), (51) follow from the
assumptions (A1), (A2) and (A3) for the operators Aα , Bα and Cα , α ∈ I.
In the second step, we set up optimal control problems for each α-level.
We seek for the optimal control u and the corresponding optimal state y in the
chaos expansion representation form (49), i.e., the goal is to obtain the unknown
coefficients uα and yα for all α ∈ I.
The problems are defined in the following way:
1◦ for α = 0 the control problem
T
min J(u0 ) = ( R0 y0 (t) 2H + u0 (t) 2
U) dt + G0 y0 (T ) 2
H (52)
u0 0
subject to
y0 (t) = A0 y0 (t) + B0 u0 (t), y0 (0) = y00 , and
◦
2 for |α| > 0 the control problem
T
J(uα ) = ( Rα yα (t) 2H + uα (t) 2
U) dt + Gα yα (T ) 2
H, (53)
0
subject to

yα (t) = Aα yα (t) + Bα uα (t) + Cα−ε(i) yα−ε(i) (t) · ei (t), yα (0) = yα0 ,
i∈N
and can be solved by the induction on the length of multi-index α ∈ I. Next we

solve the family of deterministic control problems, i.e., we discuss the solution of
the deterministic system of control problems (52) and (53):
1◦ For α = 0 the state equation (50) is homogeneous, thus the optimal control
for (50)-(52) is given in the feedback form
u∗0 (t) = −B0 Pd,0 (t) y0∗ (t), (54)
where Pd,0 (t) solves the Riccati equation (9).
2◦ For each |α| > 0 the state equation (51) is inhomogeneous and the optimal
control for (53) is given by
u∗α (t) = −Bα Pd,α (t) yα∗ (t) − Bα kα (t), (55)
where Pd,α (t) solves the Riccati equation (47), while kα (t) is a solution to the
auxiliary differential equation (48) with the terminal condition kα (T ) = 0, as
discussed in Section 2.1.1.
Summing up all the coefficients we obtain the optimal solution (u∗ , y ∗ ) re-
presented in terms of chaos expansions. Thus, the optimal state is given in the
form
y∗ = yα∗ (t) Hα = y0∗ + yα∗ (t) Hα
α∈I |α|>0
and the corresponding optimal control

u∗ = u∗α (t) Hα = u∗0 + u∗α (t) Hα
α∈I |α|>0

= −B0 Pd,0 (t) y0∗ − Bα Pd,α (t) yα∗ (t) Hα − Bα kα (t) Hα
|α|>0 |α|>0 (56)

=− Bα Pd,α (t) yα (t) Hα − Bα kα (t) Hα
α∈I α∈I
∗
= −B Pd y (t) − B K,

where Pd (t) is a coordinatewise operator corresponding to the deterministic family

of operators {Pd,α }α∈I and
$ K is a stochastic process with coefficients kα (t), i.e.,
process of the form K = α∈I kα (t) Hα , with k0 = 0.
In the following step we prove the optimality of the obtained solution. As-
suming (A1)–(A4) it follows that the assumptions of Theorem 1 are fulfilled and
thus the optimal control of the problem (4)–(5) is given in the feedback form by
u∗ (t) = −B P(t) y ∗ (t), (57)
with a positive self-adjoint operator P(t) solving the stochastic Riccati equation
(6). Since the state equations (4) and (46) are equivalent, we are going to interpret
the optimal solution (57), involving the Riccati operator P(t) in terms of chaos
expansions. It holds J(u∗ ) = min J(u), for u∗ of the form (57).
u
On the other hand, the stochastic cost function J is related with the deter-
ministic cost function J by
' (
T
J(u) = E Ry W + u U dt + GyT Z
2 2 2
0
T T
= E Ry 2W dt + E u 2U dt +E GyT 2Z
0 0

= α! Rα yα 2L2 ([0,T ],W) + α! uα 2
L2 ([0,T ],U ) + α! Gα yα (T ) 2
Z

= α! Rα yα 2L2 ([0,T ],W) + uα 2L2 ([0,T ],U ) + Gα yα (T ) 2
Z
α∈I

= α! J(uα ).
α∈I
Thus,

J(u∗ ) = min J(u) = min α! J(uα ) = α! min J(uα ) = α! J(u∗α )
u u uα
and therefore
u∗ (t, ω) = u∗α (t) Hα (ω), (58)
α∈I
i.e., the optimal control obtained via direct Riccati approach$u∗ coincides with
∗
the optimal control obtained via chaos expansion approach α∈I uα (t)Hα (ω).
Moreover, the optimal states are the same and the existence and uniqueness of the
solution of the optimal state equation via chaos expansion approach follows from
the direct Riccati approach.
Finally, we prove the convergence of the chaos expansions of the optimal
state. We include the feedback forms (54) and (55) of the optimal controls u∗α ,
α ∈ I in the state equations (50) and (51) and obtain the system
y0 (t) = (A0 − B0 B0 Pd,0 (t)) y0 (t)

yα (t) = (Aα − Bα Bα Pd,α (t)) yα (t) − Bα Bα kα (t) + Cyα−ε(i) (t) ei (t), (59)
i∈N
for |α| ≥ 1, with the initial conditions yα (0) = yα0 , α ∈ I.

From the assumption (A1) it follows that Aα , α ∈ I are infinitesimal genera-

tors of strongly continuous semigroups (Tt )α = (eAα t )α , t ≥ 0 which are uniformly
bounded, i.e., eAα t L(H) ≤ meθt , α ∈ I holds for some positive constants m
and θ, where L(H) denotes the set of linear bounded mappings on L2 ([0, T ], H).
Moreover, the family (Tt )α = (eAα t )α , t ≥ 0 is a family of strongly continuous
semigroups whose infinitesimal generators are Aα , α ∈ I, the adjoint operators of
Aα , α ∈ I. This follows from the fact that each Hilbert space is a reflexive Banach
space, see [43].
We denote by Sα (t) = Aα − Bα Bα Pd,α (t), α ∈ I and rewrite (59) in simpler
form
y0 (t) = S0 (t) y0 (t), y0 (0) = y00 ,
(60)
yα (t) = Sα (t) yα (t) + fα (t), yα (0) = yα0 , |α| > 1
$
where fα (t) = − Bα Bα kα (t) + i∈N Cyα−ε(i) (t) ei (t), α ∈ I.
The operators Sα (t), α ∈ I can be understood as time dependent continuous
perturbations of the operators Aα . From Theorem 1 it follows that Pd,α (t), α ∈ I
are self adjoint and uniformly bounded operators, i.e., Pd,α (t) ≤ p, α ∈ I,
t ∈ [0, T ]. The operators Bα and thus Bα are uniformly bounded, i.e., for all
α ∈ I we have Bα ≤ b and Bα∗ ≤ b, b > 0. Therefore, Bα Bα Pd,α (t), α ∈ I are
uniformly bounded. Hence, we can associate a family of evolution systems Uα (t, s),
α ∈ I, 0 ≤ s ≤ t ≤ T to the initial value problems (60) such that
Uα (t, s) L(H) ≤ eθ1 t , for all 0 ≤ s ≤ t ≤ T.
The family of solution maps α ∈ I to the non-autonomous system
Uα (t, s)yα0 ,
(60) is a family of evolutions which are in C([0, T ], H) since Bα Bα Pd,α , α ∈ I are
bounded for every t, and are for all α ∈ I continuous in time, i.e., elements of
C([0, T ], L(H)), [43]. The adjoint operators (S(t))α = Aα + Pd,α (t) Bα Bα , α ∈ I
are associated to the corresponding adjoint evolution systems Uα (t, s), α ∈ I,
0 ≤ s ≤ t ≤ T , see [43].
The operators Cα , α ∈ I are uniformly bounded and for all α ∈ I it holds
Cα ≤ d, d > 0. For a fixed control u it also holds Cy ∈ Dom(δ), i.e., (28) holds
for Cy.
Consider a small interval [0, T0 ], for fixed T0 ∈ (0, T ]. Denote by M1 (t) = eθ1 t
and M2 (t) = 2θ11 (e2θ1 t − 1)2 for t ∈ (0, T0 ].
For every yα0 ∈ Dom(S(t))α the mild solution of (60) is given in the form
y0 (t) = U0 (t, 0) y00
t
yα (t) = Uα (t, 0) yα0 + Uα (t, s) Cα−ε(i) yα−ε(i) (s) ei (s) − Bα Bα kα (s) ds,
0 i∈N
for |α| ≥ 1 and 0 ≤ s ≤ t ≤ T and yα are continuous functions for all α ∈ I.

The operators Cα , Bα and Bα , α ∈ I are uniformly bounded and therefore the
inhomogeneity part of (59) belongs to the space L2 ([0, T ], H), where the functions
kα , α ∈ I are given in (48). Denote by X = L2 ([0, T0 ], H) and X = L2 ([0, T0 ], H)⊗
L2 (μ). Thus it holds

y 2X = α! yα 2X = y0 2
X + α! yα 2
X ≤2 α! Uα (t, 0) yα0 2
X
α∈I |α|≥1 α∈I
t
+2 α! (Uα (t, s) Cα−ε(i) yα−ε(i) (s)ei (s) − Bα Bα kα (s) ds 2
X
|α|≥1 0 i∈N

≤ 2M12 (T0 ) α! yα0 2X
α∈I

+ 8M2 (T0 )d2 α! |α| yα 2
X + 4M2 (T0 ) b4 α! kα (s) 2
X
|α|≥1 |α|≥1
≤ 2M12 (T0 ) y 0 2X + 4M2 (T0 ) d 2

y 2
Dom(δ) + 4M2 (T0 ) b4 K 2
X, (61)
$
where K 2X = α∈I kα 2X α!. The coefficients kα are the solutions of (48) and
are expressed in terms of the adjoint evolution system Uα (t, s), α ∈ I. Clearly, the
coefficients are of the form
T

kα (t) = Uα (T, t)kα (T ) + Uα (s, t)Pd,α (s) Cα−ε(i) yα−ε(i) ei (s) ds, t < T
t i∈N
for α ∈ I. Denote by X1 = L ([T0 , T ]) and Uα (T, t) ≤ eθ̃t = M3 (t), for θ̃ > 0,
2
α ∈ I and M4 (t) = 21θ̃ (e2θ̃(T −t) − 1)2 . Since kα (T ) = 0 we obtain

T
K 2X1 = α! Uα (s, t) Pd,α (t) Cα−ε(i) yα−ε(i) ei (s) ds 2X
α∈I t i∈N

≤ 2M4 (T0 ) p d 2 2
α! |α| uα 2X ≤ M4 (T0 ) p2 d2 y 2
Dom(δ) < ∞.
α∈I
Thus, K 2X < ∞. With this bound we return to (61) and conclude that y 2X < ∞.
The interval (0, T ] can be covered by the intervals of the form [kT0 , (k + 1)T0 ]
in finitely many steps. Thus, y ∈ L2 ([0, T ], H) ⊗ L2 (μ).
Theorem 11 is an extension of results from [27], where the case with simple
coordinatewise operators was considered. The importance of the convergence re-
sult can be seen in the error analysis that arises in the actual truncation when
implementing the algorithm numerically.
Remark 3. The previous results might be extended for optimal control prob-
lems with state equations of the form (3), in spaces of stochastic distributions.
By replacing the uniform boundedness conditions on the operators Bα and Cα ,
$ ∈ I in (A2)
α and (A3) with the polynomial growth conditions of the type
−sα
α∈I Cα
2
(2N) < ∞, for some s > 0 one can prove that for fixed admissible
control, the state equation has a unique solution in the space L2 ([0, T ], H)⊗(S)−1 .
A similar theorem to Theorem 11 for the optimal control can be proven. Moreover,
the corresponding optimal control problem with fractional noise can be solved.
The following theorem gives the characterization of the optimal solution (58)
in terms of the solution of the stochastic Riccati equation (6).
Theorem 12. Let the conditions (A1)–(A5) from Theorem 11 hold and let P be a
coordinatewise operator that corresponds to the family of operators {Pα }α∈I . Then,
the solution of the optimal control problem (4)–(5) obtained via chaos expansion
(56) is equal to the one obtained via Riccati approach (57) if and only if

Cα Pα (t) Cα yα∗ (t) = Pα (t) ∗
Cα−ε(i) yα−ε(i) (t) · ei (t) , |α| > 0, k ∈ N (62)
i∈N
hold for all t ∈ [0, T ].

Proof. Let us assume first that (56) is equal to (57), then
−B P y ∗ (t) = −B Pd y ∗ (t) − B K;
we obtain
(P(t) − Pd ) y ∗ (t) = K.
The difference between P(t) and Pd (t) is expressed through the stochastic pro-
cess K, which comes from the influence of inhomogeneities. Assuming that P is a
coordinatewise operator that corresponds to the family of operators {P }α∈I , we
will be able to see the action of stochastic operator P on the deterministic level,
i.e., level of
$coefficients. Thus, for y given in the chaos expansion form (49) and
P(t) y ∗ = α∈I Pα (t) yα∗ (t) Hα it holds

(Pα (t) − Pd,α (t)) yα∗ (t) Hα = kα (t) Hα . (63)
α∈I α∈I,|α|>0
Since k0 (t) = 0 it follows P0 (t) = Pd,0 (t), for t ∈ [0, T ] and for |α| > 0
(Pα (t) − Pd,α (t)) yα∗ (t) = kα (t),
such that (48) with the condition kα (T ) = 0 holds. We differentiate (63) and
substitute (48), together with (6), (9) and (51). Thus, after all calculations we
obtain for |α| = 0
(P0 (t) − Pd,0 (t)) y0∗ (t) = 0
and for |α| > 0

Cα Pα (t) Cα yα∗ (t) = Pα (t) Cα−ε(i) yα−ε ∗
(i) (t) · ei (t) , k ∈ N.
i∈N
Note that assuming (62) and P is a coordinatewise operator that corresponds to

operators Pα , α ∈ I we can go backwards in the analysis and prove that the
optimal controls (57) and (56) are the same.
Remark 4. The condition that characterizes the optimality (62) represents the
action of the stochastic Riccati operator in each level of the noise. Note that the
stochastic Riccati equation (6) and the deterministic one (9) differ only in the
term Cα Pα (t) Cα , i.e., the operator Cα Pα (t) Cα , α ∈ I captures the stochasticity
of the equation. Polynomial chaos projects the stochastic part in different levels
of singularity, the way that Riccati operator acts in each level is given by (62).
Remark 5. Following our approach the numerical treatment of the SLQR problem
relies on solving efficiently Riccati equations arising in the associated determin-
istic problems. In recent years, numerical methods for solving differential Riccati
equations have been proposed, e.g., [2, 3, 4, 23],
3.3. Further extensions
We consider now more general form of the state equation
ẏ = Ay + Bu + T♦y, y(0) = y 0 , (64)
$ coordinatewise operators A and B and T♦, where the operator T♦

for bounded
for y = α∈I yα Hα is defined by

T♦(y) = Tβ (yα−β ) Hα . (65)
α∈I β≤α
For more details about T♦ we refer to [34, 44]. We point out that in [34] the authors
proved that (64), for fixed u, has a unique solution in space of stochastic gene-
ralized processes. Here, we will show that the optimal control problem (45)–(64)
for a specific choice of the operator T can be reduced to the problem (45)–(46),
and thus its optimal control can be obtained from Theorem 11. Moreover, one
can also consider the corresponding fractional optimal control problem and thus
apply Theorem 10 and Theorem 11. This extension is connected to the form of a
Gaussian colored noise (20) with the condition (21). We denote X = L2 ([0, T ], H).
Theorem 13. Let Lt be of the form (20) such that (21) holds. Let N be a coor-
dinatewise operator which corresponds to a family of uniformly bounded operators
{Nα }α∈I and let the operators A, B and C satisfy the assumptions (A1)–(A4) of
Theorem 11. Let the operator T be a coordinatewise operator defined by a family
of operators {Tα }α∈I , Tα : X → X, α ∈ I, such that for |β| ≤ |α|
⎧
⎨ Nα (yα ) , |β| = 0
Tβ (yα−β ) = lk Nα−ε(k) (yα−ε(k) ) , |β| = 1, i.e., β = ε(k) , k ∈ N , (66)
⎩
0 , |β| > 1
for yα ∈ X, α ∈ I. Then the state equation (64) can be reduced to the state
equation (46). Thus, the optimal control problem (45)–(64) has a unique solution.
Proof. By the definition (65) and the chaos expansion method, the state equation
(64) reduces to the system:
1◦ for |α| = 0
ẏ0 = (A0 + T0 ) y0 + B0 u0 , y0 (0) = y00 , (67)
◦
2 for |α| ≥ 1

ẏα = (Aα + T0 ) yα + Bα uα + Tβ (yα−β ), yα (0) = yα0 . (68)
0<β≤α
From (66) it follows

T0 (yα ) = Nα (yα ), α ∈ I and Tε(k) (yα−ε(k) ) = lk Nα−ε(k) (yα−ε(k) ).
We define Âα = Aα + Nα , α ∈ I. Since the family {Nα } is uniformly bounded
and {Aα } are infinitesimal generators C0 -semigroups then the operators Âα are
also infinitesimal generators of C0 -semigroups and satisfy the condition (A1) of
Theorem 11, see [43]. Thus the system (67)–(68) transforms to:
1◦ for |α| = 0
ẏ0 = Â0 y0 + B0 u0 , y0 (0) = y00
2◦ for |α| ≥ 1

ẏα = Âα yα + Bα uα + lk Nα−ε(k) (yα−ε(k) ), yα (0) = yα0 .
k∈N
Define the operators Ĉ0 = N0 and Ĉα−ε(k) = lk Nα−ε(k) , for |α| ≥ 1, k ∈ N.

Therefore, the obtained system corresponds to the state equation of the form
ẏ = Â y + B u + Ĉ ♦ Wt , (69)
where Â and Ĉ are coordinatewise operators corresponding to the families {Âα }
and {Cˆα }, respectively. Moreover, the operators B and Ĉ satisfy the assumptions
(A2)–(A4) of Theorem 11. Therefore, it can be applied to the optimal control
problem (45)–(69).
4. An example involving operators from Malliavin calculus

In this section we focus on semi-explicit ODAEs, i.e., systems of a linear semi-
explicit equation subject to an algebraic constraint. These systems of equations are
motivated by applications, e.g., Stokes equations, linearized Navier–Stokes equa-
tions, etc. They are in most cases deterministic and finite-dimensional. However,
recently ODAEs with additive noise have been studied in [1]. Here, we consider an
ODAE of the form
ẏ = Ay + T♦y + B u + f, B y = g,
where the operator B is the Itô–Skorohod integral δ and B the Malliavin deriv-
ative D. The operator δ is the adjoint operator of D, i.e., the duality relationship
E (F · δ(y)) = E (DF, y) ,
holds for stochastic processes y and F belonging to appropriate spaces [41]. Thus,
we study the system
ẏ = Ay + δu + T♦y + f, D y = g. (70)
More details on properties of the generalized operators of the Malliavin calcu-
lus and the equations involving these operators can be found in [29, 31, 34]. Here we
$
assume that the space X is$ the Hilbert
$ space L2 ([0, T ], H). Let u = α∈I uα Hα ,
uα ∈ X, α ∈ I and F = α∈I k∈N fα,k ξk Hα , fα,k ∈ X, α ∈ I, k ∈ N and
{ξk }k∈N are the Hermite functions. The Malliavin derivative operator D represents
a stochastic gradient in the direction of white noise and is a linear and continuous
mapping D : X ⊗ (S)−1 → X ⊗ S (R) ⊗ (S)−1 given by

Du = αk uα ξk Hα−εk .
α∈I k∈N
A process u belongs to the domain Dom(D) if and only if for some p ∈ N0 it holds

|α|2 uα 2X (2N)−pα < ∞.
α∈I

$ mapping δ : X ⊗ S (R) ⊗
The Itô–Skorokhod integral δ is a linear and continuous
$
(S)−1 → X ⊗ (S)−1 and is defined by δ(F ) = α∈I k∈N fα,k Hα+εk . Note that
the domain Dom(δ) = X ⊗ S (R) ⊗ (S)−1 . In quantum theory D corresponds to
the annihilation operator and δ to the creation operator.
We reduce the system (70) to the following two problems: Dy = g, Ey = y 0
and δ(u) = v and then apply the results from [29] and [31].
Theorem 14. Let A : X ⊗ (S)−1 → X ⊗ (S)−1 be a coordinatewise operator
corresponding to a uniformly bounded family of deterministic operators Aα : X →
X, α ∈ I and T be a coordinatewise operator that corresponds
$ $to a polynomially
bounded family of operators Tα : X → X, α ∈ I. Let g = α∈I k∈N gα,k ξk Hα ∈
X ⊗ S (R) ⊗ (S)−1 and f ∈ X ⊗ (S)−1 , such that Ef = A0 y 0 + T0 y 0 . Then there
exists a unique solution y ∈ X ⊗ (S)−1 and u ∈ X ⊗ S (R) ⊗ (S)−1 of the system
(70) with the initial conditions Ey = y 0 ∈ X and Eẏ = y 1 ∈ X given by
1
y = y0 + gα−ε(k) ,k Hα (71)
|α|
α∈I,|α|>0 k∈N
and
vα+ε(k)
u= (αk + 1) ξk Hα , (72)
α∈I k∈N
|α + ε(k) |
where v = ẏ − Ay − T ♦y − f .
Proof. The initial value problem involving the Malliavin derivative operator
Dy = g, Ey = y 0 (73)
can be solved by applying the integral operator on both sides of the equation.
Given a process g ∈ X ⊗ S−p (R)
$ ⊗ (S)$−1,−q , p ∈ N0 , q > p + 1, represented in
its chaos expansion form g = α∈I k∈N gα,k ξk Hα , the equation (73) has a
unique solution in Dom(D) represented by (71). Additionally, it holds
y 2
X⊗(S)−1,−q ≤ u0 2
X +c g 2
X⊗S−l (R)⊗(S)−1,−q < ∞.
The operator A is a coordinatewise operator and it corresponds to an uni-
formly bounded family of operators {Aα }α∈I , i.e., it holds Aα ≤ M , α ∈ I. For

y ∈ X ⊗ (S)−1 Dom(D) it holds

Ay 2X⊗(S)−1,−q = Aα yα 2
X (2N)−qα ≤ M y 2
X⊗(S)−1,−q <∞
α∈I
and thus Ay ∈ X ⊗ (S)−1,−q . The operators {Tα }α∈I are polynomially bounded
and it holds T♦ : X ⊗ (S)−1,−q → X ⊗ (S)−1,−q . Since gα ∈ X ⊗ S−l (R) we can
use the formula for derivatives of the Hermite functions [21]. Thus,
7 7
d k k+1
ġα = gα,k ⊗ ξk = gα,k ⊗ ξk−1 − ξk+1
dt 2 2
k∈N k∈N
and ġα ∈ X ⊗ S−l−1 (R). We note that the problem Du̇ = ẏ with the initial
condition Eẏ = y 1 ∈ X can be solved as (73). Moreover,
ẏ 2
X⊗(S)−1,−q ≤ y1 2
X + c ġ 2
X⊗S−l−1 (R)⊗(S)−1,−q < ∞.
Let f ∈ X ⊗ (S)−1,−q and denote by v = ẏ − Ay − T♦y − f . From the given
assumptions it follows v ∈ $
X ⊗ (S)−1,−q such that Ev = 0. Then, v can be
represented in the form v = α∈I,|α|≥1 vα Hα and the integral equation
δ(u) = v ,
has a unique solution u in X ⊗ S−l−1 (R) ⊗ (S)−1,−q , for l > q, given in the form
(72), see [31, 35]. Moreover, the estimate

u 2X⊗(S)−1,−q ≤ c y 2X⊗(S)−1,−q + f 2X⊗(S)−1,−q + ẏ 2X⊗(S)−1,−q
also holds.
Acknowledgement
The authors would like to thank the referees for their valuable comments. They
greatly helped to improve this manuscript. This paper was partially supported
by the project Solution of large-scale Lyapunov Differential Equations (P 27926)
founded by the Austrian Science Foundation FWF.
References
[1] Altmann, R., Levajković, T., Mena, H.: Operator differential-algebraic equations
with noise arising in fluid dynamics. Monatsh. Math. (2016) doi: 10.1007/s00605-
016-0931-z.
[2] Benner, P., Mena, H.: Numerical solution of the infinite-dimensional LQR-problem
and the associated differential Riccati equations. Tech. Report MPIMD/12-13, MPI
Magdeburg Preprint (2012).
[3] Benner, P., Mena, H.: Rosenbrock methods for solving differential Riccati equations.
IEEE Trans. Automat. Control 58, 2950–2957 (2013).
[4] Benner, P., Ezzatti, P., Mena, H., Quintana-Ortı́, E.S., Remón, A.: Solving matrix
equations on multi-core and many-core architectures. Algorithms (Basel) 6(4), 857–
870 (2013).
[5] Biagini, F., Hu, Y., Øksendal, B., Zhang, T.: Stochastic calculus for fractional Brow-
nian motion and applications. Springer-Verlag London (2008).
[6] Biagini, F., Øksendal, B., Sulem, A., Wallner, N.: An introduction to white noise
theory and Malliavin calculus for fractional Brownian motion. Proc. Royal Soc. 460,
347–352 (2004).
[7] Da Prato, G., Zabczyk, J.: Stochastic equations in infinite dimensions. Encyclopedia
Math. Appl. 44, Cambridge Univ. Press, Cambridge (1992).
[8] Da Prato, G.: Direct solution of a Riccati equation arising in stochastic control
theory. Appl. Math. Optim. 11, 191–208 (1984).
[9] Duncan, T., Maslowski, B.: Pasik-Duncan, B.: Linear-quadratic control for stochastic
equations in a Hilbert space with fractional Brownian motions. SIAM J. Control
Optim. 50(1), 507–531 (2012).
[10] Duncan, T., Pasik-Duncan, B.: Control of some linear stochastic systems with a
fractional Brownian motion. Proceedings of the 48th IEEE Conference on Decision
and Control, Shanghai, 8518–8522 (2009).
[11] Duncan, T., Pasik-Duncan, B.: Stochastic linear-quadratic control for systems with
a fractional Brownian motion. Proceedings of the 49th IEEE Conference on Decision
and Control, Atlanta, GA, 6163–6168 (2010).
[12] Duncan, T., Pasik-Duncan, B.: Linear-quadratic fractional Gaussian control. SIAM
J. Control Optim. 51(6), 4504–4519 (2013).
[13] Elliot, R., van der Hoek, J.: A general fractional white noise theory and applications
to finance. Math. Finance 13(2), 301–330 (2003).
[14] Flandoli, F.: Direct solution of a Riccati equation arising in a stochastic control
problem with control and observation on the boundary. Appl. Math. Optim. 14,
107–129 (1986).
[15] Fleming, W., Rishel, R.: Deterministic and Stochastic Optimal Control. Springer
(1975).
[16] Guatteri, G., Tessitore, G.: On the bacward stochastic Riccati equation in infinite
dimensions. SIAM J. Control Optim. 44(1), 159–194 (2005).
[17] Guatteri, G., Tessitore, G.: Backward stochastic Riccati equations in infinite horizon
l-q optimal control with infinite dimensional state space and random coefficients.
Appl. Math. Optim. 57, 207–235 (2008).
[18] Hafizoglu, C., Lasiecka, I., Levajković, T., Mena, H., Tuffaha, A.: The stochastic
linear quadratic problem with singular estimates. SIAM J. Control Optim., accepted
(2016).
[19] Hida, T., Kuo, H., Potthoff, J., Streit, L.: White noise – An infinite dimensional
calculus. Kluwer Academic Publishers (1993).
[20] Hu, Y., Øksendal, B.: Fractional white noise calculus and applications to finance.
Infin. Dimens. Anal. Quantum Probab. Relat. Topics 6(1), 1–32 (2003).
[21] Holden, H., Øksendal, B., Ubøe, J., Zhang, T.: Stochastic partial differential equa-
tions. A modeling, White noise functional approach. 2nd Edition, Springer (2010).
[22] Ichikawa, A.: Dynamic programming approach to stochastic evolution equation.
SIAM J. Control. Optim., 17(1), 152–174 (1979).
[23] Lang, N., Mena, H., Saak J.: On the benefits of the LDL factorization for large-scale
differential matrix equation solvers. Linear Algebra and Appl. 480, 44–71 (2015).
[24] Lasiecka, I., Triggiani, R.: Control theory for partial differential equations: contin-
uous and approximation theories I. Abstract parabolic systems. Encyclopedia of
Mathematics and its Applications 74, Cambridge University Press (2000).
[25] Lasiecka, I., Triggiani, R.: Optimal control and differential Riccati equations under
singular estimates for eAt B in the absence of analyticity. Advances in Dynamics
and Control, Nonlinear Syst. Aviat. Aerosp. Aeronaut. Astronaut. 2, CRC Press,
270–307 (2004).
[26] Levajković, T., Mena, H.: On deterministic and stochastic linear quadratic control
problem. Current Trends in Analysis and Its Applications, Trends in Mathemat-
ics, Research Perspectives, Springer International Publishing Switzerland, 315–322
(2015).
[27] Levajković, T., Mena, H., Tuffaha, A.: The stochastic linear quadratic control prob-
lem: A chaos expansion approach. Evol. Equ. Control Theory 5(1), 105–134 (2016).
[28] Levajković, T., Mena, H., Tuffaha, A.: A numerical approximation framework for the
stochastic linear quadratic regulator problem on Hilbert spaces. Appl. Math. Optim.
(2016) doi: 10.1007/s00245-016-9339-3.
[29] Levajković, T., Pilipović, S., Seleši, D.: Chaos expansions: Applications to a gen-
eralized eigenvalue problem for the Malliavin derivative. Integral Transforms Spec.
Funct. 22(2), 97–105 (2011).
[30] Levajković, T., Pilipović, S., Seleši, D.: The stochastic Dirichlet problem driven by
the Ornstein-Uhlenbeck operator: Approach by the Fredholm alternative for chaos
expansions. Stochastic Anal. Appl. 29, 317–331 (2011).
[31] Levajković, T., Pilipović, S., Seleši, D.: Chaos expansion methods in Malliavin cal-
culus: A survey of recent results. Novi Sad J. Math. 45(1), 45–103 (2015).
[32] Levajković, T., Seleši, D.: Chaos expansion of generalized random processes on frac-
tional white noise space. Proceedings of 12th Serbian Mathematical Congress, Novi
Sad J. Math. 38(3), 137–146 (2009).
[33] Levajković, T., Seleši, D.: Chaos expansion methods for stochastic differential equa-
tions involving the Malliavin derivative Part I. Publ. Inst. Math., Nouv. Sér. 90(105),
65–85 (2011).
[34] Levajković, T., Pilipović, S., Seleši, D., Žigić, M.: Stochastic evolution equations with
multiplicative noise. Electron. J. Probab. 20(19), 1–23 (2015).
[35] Levajković, T., Pilipović, S., Seleši, D.: Fundamental equations with higher order
Malliavin operators. Stochastics 88(1), 106–127 (2016).
[36] Lototsky, S., Rozovsky, B.: Stochastic differential equations: A Wiener chaos ap-
proach. Book chapter in The Shiryaev Festschrift ”From Stochastic Calculus to
Mathematical Finance”, (ed: Yu. Kabanov et al.), 433–507, Springer Berlin (2006).
[37] Lototsky, S., Stemmann, K.: Solving SPDEs driven by colored noise: a chaos ap-
proach. Quart. Appl. Math. 66(3), 499–520 (2008).
[38] Mandelbrot, B.B., Van Nes, J.W.: Fractional Brownian motions, fractional noises
and applications. SIAM Rev. 10, 422–437 (1968).
[39] Mishura, Y.S.: Stochastic calculus for fractional Brownian motion and related pro-
cesses. Springer (2007).
[40] Nualart, D.: Fractional Brownian motion: stochastic calculus and applications. Pro-
ceedings of the International Congress of Mathematicians, Madrid (2006).
[41] Nualart, D.: The Malliavin calculus and related topics, Probability and its Applica-
tions. 2nd edition, Springer-Verlag, New York (2006).
[42] Nualart, D., Rozovsky, B.: Weighted stochastic Sobolev spaces and bilinear SPDEs
driven by space-time white noise. J. Funct. Anal. 149(1), 200–225 (1997).
[43] Pazy, A.: Semigroups of linear operators and applications to partial differential equa-
tions. Applied Mathematical Sciences 44, Springer-Verlag, New York (1983).
[44] Pilipović, S., Seleši, D.: Expansion theorems for generalized random processes, Wick
products and applications to stochastic differential equations. Infin. Dimens. Anal.
Quantum Probab. Relat. Top. 10(1), 79–110 (2007).
[45] Samko, S.G., Kilbas, A.A., Marichev, O.I.: Fractional integrals and derivatives. The-
ory and applications. Gordon and Breach Science Publishers, Yverdon (1993).
[46] Samorodnitsky, G., Taqqu, M.: Stable non-Gaussian random processes. Chapman
and Hall, New York (1994).
Tijana Levajković Hermann Mena

Department of Mathematics Department of Mathematics
Faculty of Mathematics, Faculty of Mathematics,
Computer Science and Physics Computer Science and Physics
University of Innsbruck University of Innsbruck
Technikerstraße 13 Technikerstraße 13
A-6020 Innsbruck, Austria A-6020 Innsbruck, Austria
and and
Faculty of Traffic and School of Mathematical Sciences
Transport Engineering and Information Technology
University of Belgrade Yachay Tech
Vojvode Stepe 305 San Miguel de Uricuquı́ Ecuador
11000 Belgrade, Serbia e-mails: [email protected]
e-mails: [email protected] [email protected]
[email protected]
Amjad Tuffaha
The American University of Sharjah
Sharjah, UAE
Operator Theory:

Multi-soliton Collision for

Essentially Nonintegrable Equations
George Omel’yanov
Abstract. We describe an approach to construct multi-soliton asymptotic so-

lutions for essentially nonintegrable equations. As for equations, we assume
the existence of an exact soliton type solution with an exponential decay rate,
however we do not suppose the smallness of the interacting waves. The gen-
eral idea is realized in the cases of two and three waves and for the gKdV-4
equation with small dispersion.
1. Introduction
Linear differential equations satisfy the superposition principle which states that
the sum of solutions is a solution. This principle allows us to simplify the analysis
by calculating particular elementary solutions instead of complete solutions. For
example, it is possible to analyze only one Fourier harmonic instead of the whole
wavetrain, since a general solution may then be constructed by superposition of
these elementary solutions in a Fourier integral or a Fourier series.
Obviously, the sum of solutions is not a solution for nonlinear equations. Thus,
to investigate a phenomenon we should calculate the whole wavetrain. Next let us
note that between all possible solutions of nonlinear equations there is a special
class of so-called “self-similar” solutions, which preserve in time their shape and
the basic characteristics. From the physical viewpoint such solutions are important
since most of them describe actual phenomena. For example, periodic wavetrains
or solitons describe wave phenomena in water and plasma physics, shock waves
– in gas dynamics and acoustics, heat waves – in biology and in the theory of
combustion, and so on. These solutions are of great interest from the mathematical
viewpoint also since they represent an almost unique example of exact solutions
for nonlinear partial differential equations. Moreover, stable solutions allow us to
construct approximations to describe more general phenomena. For example, the
154 G. Omel’yanov
nonlinear wave equation

∂2u ∂2u
ε2 − 2 + g(u) = 0, ε " 1, x ∈ R1 , t > 0, (1)
∂t2 ∂x
for g(u) = au2 + bu and constant a and b has the explicit solution u(x, t, ε) =
U (x − V t)/ε with U (τ ) = α + βcn(γτ, s), where α, β, γ, s are constants and
cn(ξ, s) is the elliptic Jacobi function. If a = a(x, t) and/or b = b(x, t) are functions,
then (1) has an almost periodic asymptotic solution

u(x, t, ε) = U0 S(x, t)/ε, x, t + εU1 S(x, t)/ε, x, t + · · · , (2)
where U0 (τ, x, t) is again the cnoidal wave, but with “slowly” varying parameters
(Whitham [20], 1965).
Another example is given by the generalized KdV equation
∂u ∂uμ ∂3u
+a + ε2 b 3 = 0, x ∈ R1 , t > 0, μ > 1, ε " 1. (3)
∂t ∂x ∂x
Again, if a and b are constants, then (3) has the explicit soliton solution

u(x, t, ε) = Aω β(x − V t)/ε , ω(η) = cosh−γ (η), (4)
where γ = 2/(μ − 1), V = bβ γ , A2 2 μ−1
= V (μ + 1)/2a. For variable coefficients
a = a(x, t) and/or b = b(x, t) the equation (3) has a soliton type asymptotic
solution

u(x, t, ε) = Aω β x − ϕ(t) /ε, t + εα U1 x, t, ε + · · · , (5)
where A = A(t), β = β(t), and ϕ(t) are nonlinear functions; α > 0; and the
correction U1 has a complicated structure. For the integrable case (μ = 2 and
μ = 3) such asymptotics have been elaborated by Karpman & E. Maslov [12]
(1977), Kaup & Newell [13] (1978), and others on the basis of formulas for the
explicit solution. For the non-integrable case the appropriate asymptotic theory
has been constructed by V. Maslov & Omel’yanov [15] (1981).
It seemed to be obvious that self-similar solutions of nonlinear equations
could not be used in the same way as “elementary” solutions of linear equations are
used to describe more complicated situations, since they do not preserve the basic
parameters after the collision. For example, interactions of shock waves generate
a shock wave but with a new velocity.
However, it turned out that there are some classes of nonlinear equations such
that self-similar solutions interact almost in linear manner. This result has been
obtained in the framework of the Inverse Scattering Transform Method (Gardner,
Greene, Kruskal, Miura (1967, 1974) and others, see the references in Ablowitz
[1]). In particular, the KdV equation demonstrates the elastic scenario of soliton
interaction: N solitons collide and form after that the sequence of N solitons again
with the same amplitudes and velocities. Some shifts of trajectories appear as the
unique result of the wave interaction. Another example of the elastic kink/kink and
kink/antikink interaction is given by the sine-Gordon equation, that is the equation
(1) with the nonlinearity g(u) = sin(u). It is important that explicit formulas have
Multi-soliton Collision for Nonintegrable Equations 155
been found for multi-soliton solutions in the framework of this method. At the same
time, the number of the integrable equations is not so large, thus there appears the
basic question: do the integrable equations form a compact cluster with a sharp
frontier or do there exist nonintegrable equations which provide an almost elastic
scenario of the wave interaction?
In the simple version the answer is well known: small perturbations of the
integrable equations preserve the elasticity of the interaction in the leading term,
however a small oscillating tail, the so-called “radiation”, appears on the left of
the soliton trajectories ([12, 13] and others, see also [1]).
Let us focus the attention on the essentially nonintegrable equations. We
assume the existence of a soliton-type solution, whereas explicit solutions for ar-
bitrary Cauchy data, in particular, for the problem of soliton interaction, remain
unknown. Let us analyze the simplest nonintegrable equation (3) with constant a
and b. We set a two-wave initial condition:
2

u|t=0 = Ai ω βi (x − x(i,0) )/ε , (6)
i=1
where we use the notation (4). Obviously, if x(1,0) > x(2,0) and A2 > A1 , then the
sum of the waves of the form (4) does not satisfy the gKdV equation in view of
the trajectories x = Vi t + x(i,0) intersection.
Let us consider shortly how it is possible to analyze the problem (3), (6).
There are some different cases:
1. Let A1 " A2 . Then one can construct an asymptotic solution in the form

u = W (x − ϕ2 (t))/ε, t, x, ε, μ1 , (7)
where μ1 = A1 /A2 " 1 and

W (x − ϕ2 (t))/ε, t, x, 0, 0 ≈ A2 ω β2 (x − V2 t − x(2,0) )/ε + O(μ1 )
before the interaction. Obviously, we obtain an ordinary differential equation for
the leading term and the main problem here is the analysis of the asymptotic
corrections. Moreover, it is possible now to set ε = 1 and to use the smallness of
μ1 only. Such scheme (with a slightly different viewpoint) has been realized by
Ostrovsky et al. (see, e.g., [10]).
2. Let A2 −A1 " 1. Then one can write the ansatz as a two-phase asymptotic
expansion:

u(x, t, ε) = W x − ϕ1 (t) /ε, x − ϕ2 (t) /ε, t, x, ε, μ2 , (8)
where μ2 = A2 − A1 " 1 and
2

W (x − ϕ1 (t))/ε, (x − ϕ2 (t))/ε, t, x, 0, 0 ≈ Ai ω βi (x − Vi t − x(i,0) )/ε
i=1
before the interaction. However, under the assumption μ2 /ε " 1,

W τ1 , τ2 , t, x, ε, μ2 = W τ2 , τ2 , t, x, ε, 0 + O(μ2 /ε), (9)
156 G. Omel’yanov
and the remainder is small for the soliton-type functions. Thus, we obtain again
the one phase problem for the leading asymptotic term. In fact, this case coincides
with the problem considered in [10].
3. The amplitudes A2 > A1 are arbitrary numbers. Then we write an ansatz
similar to (8) but without any additional parameter,

u(x, t, ε) = W x − ϕ1 (t) /ε, x − ϕ2 (t) /ε, t, x, ε . (10)

Substituting this ansatz into (3) we obtain for the leading term W0 (τ1 , τ2 , t :

2
∂ 2
∂
ÂW0 + aB̂W0ν + bB̂ 3 W0 = 0, Â = − ϕ̇i , B̂ = . (11)
i=1
∂τi i=1
∂τ i
Since ϕ̇1 = ϕ̇2 for interacting waves, we can pass to new variables,
η = (τ1 − τ2 )/(ϕ̇2 − ϕ̇1 ), ζ = (ϕ̇1 τ2 − ϕ̇2 τ1 )/(ϕ̇1 − ϕ̇2 ), (12)
and transform the equation (11) to the form
∂W0 ∂W0ν ∂ 3 W0
+a +b = 0. (13)
∂η ∂ζ ∂ζ 3
Thus, we come back to the gKdV equation again with the same question about
the existence of a 2-soliton type solution. Consequently, an attempt to construct
an asymptotic solution in a standard manner seems to be hopeless. However, it is
possible to consider the problem (3), (6) successfully changing the viewpoint. We
consider this approach in the next section.
Now let us discuss the scale in the equation (3). Obviously, in the sim-
plest one-soliton case the dispersion parameter can be easily removed from the
problem, however this situation is trivial. More interesting is the two-soliton (N -
soliton with N ≥ 2) case. We come back to the initial condition (6) and assume
x10 − x20 = c > 0 For small parameter ε the initial solitary waves (6) can be

treated as a sufficiently good approximation (with the accuracy O exp(−c/ε) ,
ε → 0) of the explicit solution. The rescaling x = x/ε, t = t/ε removes formally

$2
ε from the problem again. However, the combination i=1 Ai ω βi (x − xi0 ) ap-
proximates an explicit solution if and only if x10 − x20 ! 1. This implies the
following scattering problem: we set (asymptotically) at t → −∞ two waves with
infinitely large distance between the “supports” and should determine the solution
at t → +∞. Thus, we should explain the term “infinitely large distance” defining
the large parameter x10 − x20 ! 1. We do it using the small parameter ε defining
x10 − x20 = (x10 − x20 )/ε and considering the gKdV equation in the scale (3), that
is for ε " 1.
The contents of the paper is the following: in Section 2 we consider the main
ideas of the weak asymptotic method for the case of two solitons and apply it in
Section 3 to the case of three solitons. Next, in Section 4 we discuss non-uniqueness
which appears in the framework of the weak description of multi-soliton collision.
2. Weak asymptotics method: the main idea

The base of the “Weak Asymptotic Method” is very simple: rapidly varying solitary
wave solutions (soliton or kink type) tend to distributions as the small parameter
tends to zero. This allows us to treat the equation in the weak sense and, respec-
tively, look for singularities instead of regular functions. Non-integrability implies
that we cannot find neither classical nor weak exact solutions. However, we can
construct an asymptotic weak solution considering the smallness of the remainder
in the weak sense. The main advantage here is that in order to construct asymp-
totics it is enough to obtain and analyze ordinary differential equations instead
of partial differential equations. In a sense, the situation here is similar to that of
shock waves: various regularizations generate various profiles for the wave, but in
the limiting passage we obtain the same Rankine–Hugoniot conditions. For soli-
tons the passage to the weak representation results in the disappearance of the
shape, but preserves the soliton’s characteristics: amplitudes and phases. For the
problem of interaction these parameters vary in a neighborhood of the time instant
of the collision and stabilize themselves after that. Deriving uniform in time model
equations for the parameters we can describe the scenario of the wave interaction.
Originally, such idea had been suggested for shock wave type solutions
(Danilov & Shelkovich [5], 1997) and for soliton type solutions (Danilov & Omel’ya-
nov [3], 2003), and then generalized, developed, and adapted for many other prob-
lems, including the problem of kink/kink and ink/antikink interaction for the sine-
Gordon type equation (1) (V. Danilov, G. Omel’yanov, V. Shelkovich, D. Mitrovic
and others, [2, 4, 6, 7, 8, 9, 11, 14, 16, 17, 18, 19] and references therein). Let us
note finally that the treatment [16] of weak asymptotics as functions which satisfy
some conservation or balance laws takes us back to Whitham’s idea of constructing
one-phase asymptotic solution satisfying a Lagrangian, [20]. Now, for essentially
nonintegrable equations and multi-soliton solutions, we use the appropriate num-
ber of laws and require them to be satisfied in the weak sense.
Let us explain these ideas using as the example the problem of two soliton
interaction in the gKdV-4 model. Here and in what follows “gKdV-4” is the equa-
tion (3) with the nonlinearity u4 and a = b = 1. We start with the following
definition of the weak asymptotic solution [3]:
Definition 1. A sequence u(t, x, ε), belonging to C ∞ (0, T ; C ∞(Rx1 )) for ε > 0 and
belonging to C(0, T ; D (Rx1 )) for ε ≥ 0, is called a weak asymptotic mod OD (ε2 )
solution of the gKdV-4 equation if the relations
∞
d ∞ ∂ψ
uψdx − u4 dx = O(ε2 ), (14)
dt −∞ −∞ ∂x
∞
d ∞ 2 8 ∞ 5 ∂ψ
2
∂u ∂ψ
u ψdx − u dx + 3 ε dx = O(ε2 ) (15)
dt −∞ 5 −∞ ∂x −∞ ∂x ∂x
hold uniformly in t for any test function ψ = ψ(x) ∈ D(R1 ).

158 G. Omel’yanov
Here the right-hand sides are C ∞ -functions for ε > 0 and piecewise continuous
functions for ε ≥ 0. The estimates are understood in the C(0, T ) sense:
g(t, ε) = O(εk ) ↔ max |g(t, ε)| ≤ cεk .

t∈[0,T ]
Definition 2. A function v(t, x, ε) is said to be of the value OD (εk ) if the relation
∞
v(t, x, ε)ψ(x)dx = O(εk )
−∞
holds uniformly in t for any test function ψ ∈ D(Rx1 ).
The sense of the relation (14) is obvious: it is the adaptation of the standard
D -definition to asymptotic mod OD (ε2 ) solution which belongs to C(0, T ; D (Rx1 )).
Next we note that (14) is not the unique condition that has to be satisfied, since
here the difference between the gKdV-4 equation and the limiting first-order equa-
tion (with ε = 0) has been lost. To involve the dispersion term into the analysis,
we supplement (14) by the additional condition (15) which is the result of multipli-
cation of gKdV-4 equation by uψ and integration by parts. This condition can be
treated as a version of (14) but for special test functions u ψ(x), ψ ∈ D(Rx1 ), which
vary rapidly there where the solution varies rapidly. It is important also that (15)
duplicates the orthogonality condition which appears for single-phase asymptotics
(see [17]).
Next we present the ansatz as the sum of two distorted solitons, that is:

2

u= Gi ω βi (x − ϕi )/ε , (16)
i=1
where we use the notation (4) with μ = 4, a = b = 1,

Gi = Ai + Si (τ ), ϕi = ϕi0 (t) + εϕi1 (τ ), τ = β1 ϕ20 (t) − ϕ10 (t) /ε, (17)
and Ai are the original amplitudes. The functions ϕi and Si should be found;
however, to simplify formulas, we write ϕi0 = Vi t + xi0 with “unknown” Vi . Next
we assume that A1 < A2 and x10 − x20 > 0, therefore, the trajectories x = ϕ10
and x = ϕ20 intersect at a point (x∗ , t∗ ). Next we define the “fast time” τ to
characterize the distance between the trajectories ϕi0 and we assume that Si (τ ),
ϕi1 (τ ) are smooth functions such that
Si → 0 as τ → ±∞, (18)
ϕi1 → 0 as τ → −∞, ϕi1 → ϕ∞
i1 as τ → +∞. (19)
It is obvious that the existence of the weak asymptotics (16) with the properties
(18), (19) implies that the solitary waves (4) interact like the KdV solitons at least
in the leading term.
To construct the asymptotics we should calculate the weak expansions for uk

and (εux )2 . It is easy to check that

2
u = εa1 Gi βi−1 δ(x − ϕi ) + OD (ε3 ). (20)
i=1
Here and in what follows we use the notation
∞ ∞ l k
(l) dω
ak = ω k (η) dη, ak = l
dη, k ≥ 1, l ≥ 1. (21)
−∞ −∞ dη
Next we note
∞
2 ∞
G2
u2 (x, t)ψ(x)dx = ε i
ω 2 (η)ψ(ϕi + ε βηi )dη (22)
−∞ i=1
βi −∞
∞
x − ϕ1 x − ϕ2
+2G1 G2 ω β1 ω β2 ψ(x)dx. (23)
−∞ ε ε
We take into account that the integrand in (23) vanishes exponentially fast as
|ϕ1 − ϕ2 | grows, thus, the main contribution gives the point x∗ . We write
ϕi0 = x∗ + Vi (t − t∗ ) = x∗ + ετ Vi /ψ̇0 and ϕi = x∗ + εχi , (24)
where ψ̇0 = β1 (V2 − V1 ), χi = τ Vi /ψ̇0 + ϕi1 . Next we transform the integral in
(23) to the following form:
∞
ε
ω(θ12 η − σ12 )ω(η)ψ x∗ + εχ2 + ε βη2 dη, (25)
β2 −∞
where θ12 = β1 /β2 , σ12 = β1 (ϕ1 − ϕ2 )/ε. It remains to apply the formula
f (τ )δ(x − ϕi ) = f (τ )δ(x − x∗ ) − εχi f (τ )δ (x − x∗ ) + OD (ε2 ), (26)
which holds for each ϕi of the form (24) with slowly increasing χi and for f (τ )
from the Schwartz space. Moreover, the second term in (26) is OD (ε). Thus, under
the assumptions (18) we can modify (20) and (22), (23) to the final form:

2
Ai
2
Si
u=ε δ(x − ϕi ) + ε δ(x − x∗ ) + OD (ε3 ), (27)
i=1
βi i=1
βi

2
A2 02
1
u2 = εa2 i
δ(x − ϕi ) + εa2 (2Ai Si + Si2 )
i=1
βi i=1
β i
G1 G2 1
+2 λ(σ12 ) δ(x − x∗ ) + OD (ε2 ), (28)
β2
where the convolution λ(σ12 ) describes the product of two waves,
∞
1
λ(σ12 ) = ω(θ12 η − σ12 )ω(η)dη. (29)
a2 −∞
Calculating the weak expansions for others terms which are involved in Definition
1 and substituting them into (14), (15), we obtain linear combinations of δ (x−ϕi ),
160 G. Omel’yanov
i = 1, 2, δ(x − x∗ ), and δ (x − x∗ ). Therefore, we pass to the following system of 8

equations for 8 unknowns (see [3, 4] for the explicit formulas):
Hi,j (Ai , βi , Vi ) = 0, j = 1, 2, i = 1, 2, (30)
Qj (S1 , S2 , σ12 ) = 0, j = 1, 2, (31)
dϕj1
= Zj (S1 , S2 , σ12 ), j = 1, 2. (32)
dτ
The first four algebraic equations (30) imply again the relations after (4) between
Ai , βi , and Vi . Furthermore, functional equations (31) allow to define Si with the
property (18), whereas an analysis of the ODE (32) justifies the existence of the
required phase corrections ϕi1 with the property (19), [3, 4].
√
Theorem 1. Let θ∗ = 5 − 2 and let θ12 < θ∗ . Then there exist functions Gi (t, ε),
ϕi = ϕi (t, ε), i = 1, 2, such that (16) describes mod OD (ε2 ) the elastic scenario
of the solitary waves interaction for the gKdV-4 equation.
The next theorem allows us to treat the weak asymptotics (16) in the classical
sense [3, 4]:
Theorem 2. Let θ12 < θ∗ . Then the function u of the form (16) is a weak asymp-
totic mod OD (ε2 ) solution of the problem (3), (6) for the gKdV-4 equation if and
only if u satisfies the following conservation and balance laws:
∞
d ∞ d ∞ 2 d ∞
udx = 0, u dx = 0, x u dx −
2
u4 dx = 0,
dt −∞ dt −∞ dt −∞ −∞
∞
d ∞ 8 ∞ 5
2
∂u
x u2 dx − u dx + 3 ε dx = 0.
dt −∞ 5 −∞ −∞ ∂x
Results of numerical simulations of the problem (3), (6) for the gKdV-4 equa-
tion (Garcia-Alvarado & Omel’yanov [8, 9], 2012, 2014) confirm the traced asymp-
totic analysis uniformly in θ12 ∈ (0, 1). In particular, the profile of the solution
shown in Figure 1 demonstrates the elastic character of interaction for θ12 > θ∗ .
Note that the oscillating tail that appears after the collision is not a computational
error but a part of the solution (see [9]). Its local structure can be described by the
Airy function, and it is related to the “radiation” of the perturbed KdV equation.
3. Multi-soliton interaction
3.1. General construction
It turned out however that Definition 1 does not support asymptotics with three
or more phases since it implies the appearance of ill-posed model equations for
the parameters of the solutions (they are well posed for the case of two phases).
To overcome this obstacle it is necessary to change the viewpoint on the weak as-
ymptotic solution: an analysis (Omel’yanov [16, 17], 2015) shows that Definition 1
Figure 1. Evolution of two solitary waves for β1 = 0.7, β2 = 1.25, and

ε = 0.1
implies the fulfilment of two conservation laws in the weak sense. Moreover, the
one-phase asymptotic theory for perturbed equations implies the fulfilment of a
single “conservation law” again in the weak sense. Thus there appears the hypothe-
ses that to construct N -phase asymptotics it is necessary to use N conservation
laws.
To demonstrate this for the gKdV-4 equation let us consider a three-wave
initial condition:
3

u|t=0 = Ai ω βi (x − x(i,0) )/ε , (33)
i=1
where the notation (4) has been used again. We set A1 < A2 < A3 , x(i,0) > x(i+1,0) ,
i = 1, 2, and suppose the intersection of all trajectories x = ϕi0 (t) at the same
point (x∗ , t∗ ).
Let us write the conservation laws in the differential form:
∂Qj ∂Pj ∂ 3 Rj
+ = ε2 , j = 1, 2, 3. (34)
∂t ∂x ∂x3
Here
Q1 = u, P1 = u4 , Q2 = u2 , P2 = 85 u5 − 3(εux )2 , (35)
Q3 = (εux ) −
2 2 5
5u , P3 = 16u (εux ) − u − 3(ε uxx ) ,
3 2 8 2 2
(36)
and R1 = R2 = 0, R3 = −Q3 .
In contrast to Definition 1 we define now the asymptotics in the following
manner:
Definition 3. Let u = u(t, x, ε) be a sequence that belongs to the functional space
indicated in Definition 1. Then u is called a 3-phase weak asymptotic mod OD (ε2 )
162 G. Omel’yanov
solution of the gKdV-4 equation if the relations (34) hold uniformly in t with the
accuracy OD (ε2 ) in the sense of Definition 2.
To construct the asymptotics we present the ansatz in the form similar to
(16), that is
3

u= Gi ω βi (x − ϕi )/ε , (37)
i=1
where we use the notation (16). However, the “fast time” τ characterizes now the
distance between the first and third trajectories,

τ = β1 ϕ30 (t) − ϕ10 (t) /ε, and ψ̇0 = β1 (V3 − V1 ). (38)
Next we calculate weak expansions for all terms from Definition 3 and pass to the
following system of model equations:
(0) (0)
Vi aQj ,i − aPj ,i = 0, i = 1, 2, 3, j = 1, 2, 3, (39)
(0)
RQj = 0, j = 1, 2, 3, (40)
d 0 1 (1) 1
3 (0)
aQ ,i 1 (0)
ψ̇0 ϕi1 j + 2 RQj − RPj = 0, j = 1, 2, 3. (41)
dτ i=1 βi β3 β3

∞
(n)
aF,i = η n F Ai ω(η), . . . , Ai βik ω (k) (η) dη, n = 0, 1, (42)
−∞
∞ #
3 3
(n)
RF = η n
F Gi ω(ηi3 ), . . . , Gi βik ω (k) (ηi3 )
−∞ i=1 i=1
8

3
− k (k)
F Ai ω(ηi3 ), . . . , Ai βi ω (ηi3 ) dη, (43)
i=1
where F = F (u, εux, . . . , εk ux ) is a smooth function,

(k)
ηij = θij η − σij , θij = βi /βj , σij = βi (ϕi − ϕj )/ε. (44)

Let us note at the beginning that the system (39) of three equations contains for
each i only two free parameters Ai = Ai (βi ), Vi = Vi (βi ). However, since Ai (βi )
and Vi (βi ) are associated with the exact solution of (3), we obtain
Lemma 1. Let ω(η), Ai = A(βi ), and Vi = V (βi ) be of the form (4). Then the
equalities (39) are satisfied for each βi > 0.
Let us simplify the equations (41). We take into account the equalities (40)
and the following consequence of the definitions (24), (44) of ϕi and σi3 :
σi3 τ
ϕi1 = ϕ31 + − (Vi − V3 ), i = 1, 2. (45)
βi ψ̇0
Then the system (41) can be transformed to the form
d 0 aQj ,i 1 (1) 1
2 (0)
dϕ31
ψ̇0 rj + ψ̇0 σi3 + R = fj , j = 1, 2, 3, (46)
dτ dτ i=1 βi2 β32 Qj
where

3
aQj ,i 1 (0) aQj ,i
2
rj = , fj = RPj + (Vi − V3 ). (47)
i=1
βi β3 i=1
βi
Now it is clear that (46) contains three unknown functions, ϕ31 and σi3 . Moreover,
this system can be transformed easily into a 2 × 2 autonomous system for σi3 ,
i = 1, 2.
Taking into account our hypothesis (19) we supplement (46) by the scattering-
type condition
ϕ31 → 0, σi3 /τ → ξi3 as τ → −∞, i = 1, 2, (48)
where
ξi3 = βi (Vi − V3 )/ψ̇0 . (49)
Let us summarize now the previous constructions as the following formal result:
Theorem 3. Let us suppose that the system (40) has a solution which satisfies the
assumption (18). Assume also that the solution to problem (46), (48) is such that
|ϕ31 | ≤ const and σi3 /τ → ξi3 as τ → ∞. Then the ansatz (37) describes the
elastic scenario of the three wave collision for the gKdV-4 equation.
Again, an analysis of the equalities (39)–(41) allows us to treat the weak
asymptotics (37) in the classical sense
Theorem 4. Let the assumptions of Theorem 3 be satisfied. Then the function u of
the form (37) is a weak asymptotic mod OD (ε2 ) solution of the gKdV-4 equation
(3) if and only if u satisfies the following conservation and balance laws:
∞
d ∞ d ∞
Qj dx = 0, x Qj dx − Pj dx = 0, j = 1, 2, 3. (50)
dt −∞ dt −∞ −∞
3.2. Analysis of the model equations

The next step of the construction is the analysis of (40). This system consists of
algebraic equations for Si of the first, second, and fifth order; and the solution
should be a smooth real function which satisfies the assumption (18). Existence
of such solution for general θij remains unproved, however, in some cases it is
possible to find sufficient conditions of solvability. In particular, let
θ23 = μ3 , θ13 = μ3(3+α)/2 , where α ∈ [0, 1) and μ < 1. (51)
Lemma 2. Let σi3 → ∞ as τ → ∞, i = 1, 2. Then for sufficiently small μ the
system (40) has a unique solution which satisfies the assumption (18).
164 G. Omel’yanov
Proof. We look for the asymptotic solution of the system (40) in the form:
1 cβ1 α 1 cβ2 2/3
S1 = 1/3
μ (y − μ2−α x), S2 = − 1/3 μα (y + μ2−α x), S3 = cβ3 μ2 x, (52)
2β 2β
3 3
where x and y are free functions. Then the equation (40) for j = 1 is satisfied
automatically. Furthermore, the hypothesis (51) allows us to pass to the asymptotic
representation of the equations (40) for j = 2, 3:
(0) (0)
a2 x − 12 μα y = −λ1,0,1 (σ23 ) − μα λ1,0,1 (σ12 ) + OS (μ1+α ), (53)
(1) (0) (0) (0)
2a2 − 5a5 x = 5λ4,0,1 (σ23 ) + 5μ1+α λ4,0,1 (σ13 ) − 2 yλ4,0,1 (σ23 ) + OS (μ2 ).
1

∞
p
η n ω k (η) ω (η) ω l (ηij )dη,
(n)
λk,p,l (σij ) = n = 0, 1, (54)
−∞
and the following definition of the smallness:
Definition 4. A function v(τ, μ) is said to be of the value OS (μκ ) if there exists a
function f (τ ) ≥ 0 from the Schwartz space S such that the inequality
|v(τ, μ)| ≤ μκ f (τ )
holds uniformly in τ ∈ R.
Obviously, the matrix in the left-hand side of (53) is degenerate mod O(μα ).
However, the right-hand side has the same rank.
Lemma 3. For the function ω of the form (4) the relation
(1) (0) (0)
5a5 − 2a2 λ1,0,1 (σln ) = 5a2 λ4,0,1 (σln ) + OS (θln ) (55)
holds for all indices l, n.
Now we set
(0)
x = −λ1,0,1 (σ23 ) + μα x1 /a2 (56)
and transform (53) to the final form:
a2 (0)
x1 − y = −λ1,0,1 (σ12 ) + OS (μ), (57)
2
5 (0) (0)
x1 = μ 2λ4,0,1 (σ13 ) − yλ4,0,1 (σ23 ) + OS (μ2−α ).
4
Since the matrix in the left-hand side of the system (57) is nonsingular, we obtain
the desired assertion of Lemma 3.
Next analyzing the problem (46), (48) and returning to ϕi1 we obtain:
Lemma 4. Let the assumption (51) be verified. Then the equations (41) have a
solution which satisfies the assumptions (19).
Therefore, we can formulate the main result of this section (see [16, 17] for
detail):
Figure 2. Evolution of the soliton triplet with β1 = 0.7, β2 = 1.25,

β3 = 1.62
Theorem 5. Under the assumption (51) the three-phase asymptotic solution (37)
exists and describes mod OD (ε2 ) the elastic scenario of the solitary waves inter-
action.
Figure 2 demonstrates the result of a numerical simulation of problems (3)
and (33) for the gKdV-4 equation [8].
4. Non-uniqueness phenomenon
Let us consider now the next questions: what to do when the quantity of interacting
waves is greater than the number of existing conservation laws; and conversely, if
the quantity of interacting waves is less than the number of conservation laws, how
to select conservation laws between all the existing ones?
In fact, the answer is very simple: instead of deficient conservation laws it is
possible to use balance laws associated with reasonable a priori estimates, whereas
the choice of specific conservation or balance laws is not important. Let us note
that the situation with non-uniqueness here is inverse to the case of shock waves
for hyperbolic equations: for shocks there are many divergent forms for the original
classical equation but we should fix only one of them; conversely, for solitons we
should use both all divergent forms and balance laws.
To discuss shortly the indicated problems let us note that the gKdV-4 equa-
tion has three conservation laws only, namely, the conservation laws (34). So in
order to consider N -soliton asymptotics with N ≥ 4, (34) should be supplemented
by balance laws. Let us choose the simplest one:
∂Q4 ∂P4
+ + ε−1 K4 = OD (ε2 ), (58)
∂t ∂x
166 G. Omel’yanov
where
Q4 = 12 (ε2 uxx )2 + 5 8
21 u −
10 3
3 u (εux ) ,
2
K4 = −(εux )5 , (59)
P4 = 12u (ε uxx ) −
3 2 2
19u(εux)4 − 32 (ε3 uxxx )2 + 160 11
231 u − 100 6 2
3 u (εux ) .
We will write (34) in the form (58) setting Kj = 0 for j = 1, 2, 3. Let us note that,
in contrast to Qj and Pj , the non-divergent term ε−1 K4 (“production”) does not
belong to the so-called “regularly degenerating” functions, so that its value is
O(ε−1 ) in the C-sense, that is the same as the value of the derivatives of Qj and
Pj . At the sametime, the production ε−1 K4 , calculated for single solitary wave
Aω β(x − ϕ0 )/ε , is an odd function and it disappears after the integration.
Let us consider again a two-phase asymptotic solution for the gKdV-4 equa-
tion. In contrast to Definition 1 we define it in the following manner:
Definition 5. Let 1 ≤ k0 < k1 ≤ 4 and let a sequence uk0 ,k1 = uk0 ,k1 (t, x, ε) belong
to the same functional space as u(t, x, ε) in Definition 1. Then uk0 ,k1 is called a
weak asymptotic mod OD (ε2 ) solution of (3) if the relations
∂Qj ∂Pj
+ + ε−1 Kj = OD (ε2 ), j = k0 , k1 , (60)
∂t ∂x
hold uniformly in t.
To construct the asymptotics we present the ansatz uk0 ,k1 again in the form
(16). Calculating weak expansions for all terms and substituting them into (60)
we obtain the following analog of (30)–(32):
(0) (0) (1)
Vi aQj ,i − aPj ,i + aKj ,i /βi = 0, i = 1, 2, j = k0 , k1 , (61)
d (0) (0)
ψ̇0 R + RKj = 0, j = k0 , k1 , (62)
dτ Qj
dϕj1
= Rj (S1 , S2 , σ12 ), j = k0 , k1 . (63)
dτ
(0)
Note that the property aK4 ,i = 0 has been used here essentially.
Lemma 5. The algebraic equations (61) imply again the relations (4) between Ai ,
βi , and Vi .
The system (62) contains two functional equations if k1 < 4 or a functional
equation and an ordinary equation if k1 = 4. Similar to Section 2 we assume:
k0 = 1, θ12 " 1. (64)
Furthermore, let us introduce corrections κi = κi (τ ) of the amplitudes,
−1/3 2/3 2/3
S1 = cβ1 β2 θ12 κ1 , S2 = cβ2γ θ12 κ2 , (65)
where c = (5/2)1/3 . We note that the first assumption (64) and the equation (62)
for j = 1 imply the equality
κ1 + κ2 = 0. (66)
Let us turn to (62) for k1 = 2 or 3. They are algebraic equations of the order 2
and 5, respectively, therefore applying (64) we obtain
2/3
κ1 = Kk1 (σ12 ) + OS (θ12 ), k1 = 2, 3, (67)
(1)
where we use the equality a2 = 3a5 /4, [16], and the notation
(0,1) (0,1)
K2 (σ12 ) = M{20} /a2 , K3 (σ12 ) = 20M{50} /17a5 , (68)
∞
η n ω12 ω m0 −1 (ω )m1 dη.
(n,1)
M{m0 ,m1 } = (69)
−∞
(n,1)
Under the assumption (19) the convolutions M{m0 ,m1 } belong to the Schwartz
space S. Thus κi satisfy the hypothesis (18) in the cases k1 = 2, 3. Moreover, in
both cases κi = κi (σ12 ) are even functions of the same order O(1) in the C-sense.
In the case k1 = 4 we have the differential equation
∞
d0 3 4 γ
c1 κ1 − K4 = ω (η) ω (η12 )dη + OS (θ12 ), (70)
dτ 10 −∞
where
(0,2) 2 3 (2) 2 (0,1) 3 (0,1)
c1 = M{32} − a8 − 3 a2 , K4 = M{80} − M{32} + OS (θ12
γ
), (71)
7 5 7 5
∞
η n ω m0 (ω )m1 dη.
(n,2)
M{m0 ,m1 } = (72)
−∞
The combination of convolutions K4 ∈ S is an even function with respect to σ12 ,
whereas the right-hand side is an odd one. Thus, under the hypothesis
dσ/dτ is an even function (73)
the last term in (70) disappears after the integration over R. Therefore, κ1 turns
out again to be an even function which satisfies the hypothesis (18).
Now let us turn to the equations (63). Analogously to Section 2 we pass from
this system to an autonomous scalar equation
dσ12
Lk0 ,k1 = Fk0 ,k1 , (74)
dτ
where under the assumptions (64)
L1,k1 = 1 + OS (θ12 ), F1,k1 = −1 + OS (θ2nk1 /3 ), (75)
and n2 = 1, n3 = n4 = 1/2. Taking into account our hypothesis (19) we supplement
(74) by the scattering-type condition
σ12 /τ → −1 as τ → −∞. (76)
Obviously, for sufficiently small θ12 the problem (74), (76) is solvable and σ12 /τ →
−1 as τ → ∞.
Finally, to check the property (73) we note that all Qj , Pj , calculated for the
single solitary wave, are even functions, whereas K4 is an odd function. Thus, all
(0) (0) (1)
RQj , RPj and R4 are even functions with respect to σ12 . On the contrary, all
168 G. Omel’yanov
(1) (1) (0)

RQj , RPj and R4 are odd functions. This and representation (75) implies the
fulfilment of (73).
Theorem 6. Under the assumption (64) the asymptotic solution in the sense of
Definition 5 exists for each k1 and describes the elastic collision of the solitary
waves.
Now let us compare the asymptotics which are possible for various choices of
(k ) (k )
k1 . We note at first that the replacement of one ϕi1 1 by another ϕi1 1 implies a
small correction in the sense of Definition 5. Indeed,

(k ) (k )
ω βi (x − ϕi0 − εϕi1 1 )/ε − ω βi (x − ϕi0 − εϕi1 1 )/ε , ψ(x)
ε (k ) (k )

= a1 δ(x − ϕi0 − εϕi1 1 ) − δ(x − ϕi0 − εϕi1 1 ), ψ(x) + O(ε3 ) = O(ε2 ).
βi
This correction is beyond the range of the weak asymptotics accuracy.
Furthermore, according to (62) the shape of the amplitudes Si depends on
the selection of the conservation/balance laws. The difference between them is
O(1) in the C-sense and OD (ε) in the sense of Definition 2. At the same time we
stress that such phenomenon of non-uniqueness is immanently intrinsic to weak
asymptotics. Namely, it is possible to change the shape of the amplitudes adding
to (16) functions which are arbitrarily small in the weak sense (see [3] for detail).
However, this effect is concentrated within O(ε1−ν )-neighborhood of the time-
instant t∗ of the interaction, ν > 0. Thus it is small in the D (Rx,t 2
) sense. It
implies the following
Definition 6. Let functions u1 (x, t, ε) and u2 (x, t, ε) satisfy the problem (3), (6)
in the sense of Definition 5. Then these functions are said to be asymptotically
equivalent if the relation
∞ ∞

ũ1 (x, t, ε) − ũ2 (x, t, ε) ψ(x, t)dx dt = O(ε2 )
−∞ −∞
holds for any test function ψ ∈ D(R2 ). Here ũk are continuations of uk for t < 0
as the sum of solitary waves of the form (4).
Now we can formulate the main result of this section:
Theorem 7. Under the assumptions (64) the weak asymptotic solutions u1,k1 and
u1,k1 of the problem (3), (6) are asymptotically equivalent for all k1 , k1 ∈ {2, 3, 4}.
Acknowledgement
The research was supported by SEP-CONACYT under grant 178690 (Mexico).
References
[1] M. Ablowitz, H. Segur, Solitons and the Inverse Scattering Transform, (SIAM,
Philadelphia, 1981)
[2] V. Danilov, D. Mitrović, Shock wave formation process for a multidimensional scalar
conservation law. Quart. Appl. Math. 69(4), 613–634 (2011)
[3] V. Danilov, G. Omel’yanov, Weak asymptotics method and the interaction of infin-
itely narrow delta-solitons. Nonlinear Analysis: Theory, Methods and Applications.
54, 773–799 (2003)
[4] V. Danilov, G. Omel’yanov, V. Shelkovich, Weak asymptotics method and interac-
tion of nonlinear waves, in: Asymptotic methods for wave and quantum problems,
ed. by M.V. Karasev. AMS Trans., Ser. 2, vol. 208 (AMS, Providence, RI, 2003), pp.
33–164
[5] V. Danilov, V. Shelkovich, Generalized solutions of nonlinear differential equations
and the Maslov algebras of distributions. Integral Transformations and Special Func-
tions. 6, 137–146 (1997)
[6] V. Danilov, V. Shelkovich, Dynamics of propagation and interaction of delta-shock
waves in conservation law systems, Journal of Differential Equations. 211(2), 333–381
(2005)
[7] M. Garcia, G. Omel’yanov, Kink-antikink interaction for semilinear wave equations
with a small parameter. Electron. J. Diff. Eqns. 2009(45), 1–26 (2009)
[8] M. Garcia, G. Omel’yanov, Interaction of solitary waves for the generalized KdV
equation. Communications in Nonlinear Science and Numerical Simulation. 17(8),
3204–3218 (2012)
[9] M. Garcia, G. Omel’yanov, Interaction of solitons and the effect of radiation for
the generalized KdV equation. Communications in Nonlinear Science and Numerical
Simulation. 19(8), 2724–2733 (2014)
[10] K.A. Gorshkov, L.A. Ostrovsky, Interaction of solitons in nonintegrable systems:
direct perturbation method and applications. Physica D. 3, 428–438 (1981)
[11] H. Kalisch, D. Mitrović, Singular solutions of a fully nonlinear 2 × 2 system of
conservation laws. Proceedings of the Edinburgh Mathematical Society II. 55, 711–
729 (2012)
[12] V. Karpman, E. Maslov, Perturbation theory for solitons. JETP 46(2), 281–291
(1977)
[13] D. Kaup, A. Newell, Solitons as particles, oscillators and in slowly changing media:
a singular perturbation theory. Proc. Roy. Soc. London Ser. A 361, 413–446 (1978)
[14] D. Kulagin, G. Omel’yanov, Asymptotics of kink-kink interaction for sine-Gordon
type equations. Math. Notes. 75(4), 563–567 (2004)
[15] V. Maslov, G. Omel’yanov, Asymptotic soliton-form solutions of equations with small
dispersion. Russian Math. Surveys. 36, 73–149 (1981)
[16] G. Omel’yanov, Soliton-type asymptotics for non-integrable equations: a survey.
Mathematical Methods in The Applied Sciences. 38(10), 2062–2071 (2015)
[17] G. Omel’yanov, Interaction of 3 solitons for the GKdV-4 equation. http://
arxiv.org/abs/1504.02167 (2015)
170 G. Omel’yanov
[18] G. Omel’yanov, I. Segundo Caballero, Interaction of solitons for sine-Gordon-type

equations. Journal of Mathematics. 2013, Article ID 845926 (2013)
[19] G. Omel’yanov, M. Valdez-Grijalva, Asymptotics for a C 1 -version of the KdV equa-
tion. Nonlinear Phenomena in Complex Systems. 17(2), 106–115 (2014)
[20] G. Whitham, A general approach to linear and non-linear dispersive waves using a
Lagrangian. J. Fluid Mech. 22, 273–283 (1965)
George Omel’yanov
University of Sonora
Rosales y Encinas s/n
83000, Hermosillo, Sonora, Mexico
Operator Theory:

Microlocal Solvability and Subellipticity

of Several Classes of Pseudodifferential
Operators with Involutive Characteristics
P.R. Popivanov
Abstract. In this paper pseudodifferential operators with involutive charac-

teristics are considered in two different cases: elliptic subprincipal symbol and
subprincipal symbol being a symbol of principal type near some characteristic
point (i.e., vanishing at a part of the characteristic set). We prove (micro)local
non-solvability results as well as subelliptic estimates in the second case when
the loss of regularity is of the following type: 2k+1k+1
= 1 + k+1k
, k ∈ N. For
the operators of subprincipal type interesting results were proved recently by
N. Dencker.
1. Introduction and formulation of the main results

1. We denote by Ψm cl (X) the set of all classical scalar pseudodifferential operators
(Ψdo) of order m on the C ∞ manifold X (see [3, 4, 6]). D (X) stands for the set of
all Schwartz distributions on X, while pom is the positively homogeneous of order m
with respect to ξ principal symbol of the operator P ∈ Ψm cl (X), p ∼ pm +pm−1 +· · · .
0
The wave-front set W F (u) of u ∈ D (X) is defined by W F (u) = {ρ ∈ T ∗ (X)\0 :

a ∈ Ψ0cl (X), a(x, D)u ∈ C ∞ =⇒ a0 (ρ) = 0}, while the s-wave-front set of u is given
by W Fs (u) = {ρ ∈ T ∗ (X)\0 : a ∈ Ψ0cl (X), a(x, D)u ∈ Hloc s
(X) =⇒ a0 (ρ) = 0}.
Certainly, W F (u) and W Fs (u) are closed conical in ξ sets. We denote by
s
Hmcl (ρ) the complement of W Fs (u), i.e., u ∈ Hmcls
(ρ0 ) iff there exists a ∈ Ψ0cl (X),
a(ρ) = 0, near to ρ , such that a(x, D)u ∈ Hloc , H s being the standard Sobolev
0 s
space.
Definition 1 ([4]). The operator P ∈ Ψm cl (X) is locally solvable at a compact set

K ⊂ X if the equation P u = v has a local solution u ∈ D (X) in a neighbour-
hood of K for any v ∈ C ∞ (X) in a set of finite codimension. We can also define
microlocal solvability of P at any compact based cone K ⊂ T ∗ X.
172 P.R. Popivanov
Definition 2 ([4]). If K ⊂ S ∗ X, where S ∗ X is the cosphere bundle over X, S = {ξ :

|ξ| = 1}, is a compact set, then we say that P ∈ Ψm cl (X) is microlocally solvable
at K if there exists an integer N such that for every f ∈ Hloc N
(X) there exists
u ∈ D (X) for which K ∩ W F (P u − f ) = ∅.
The microlocal solvability is invariant under conjugation by elliptic Fourier
integral operators (FIO) and multiplication by elliptic Ψdo.
Let P ∗ be the L2 adjoint operator of P and . k be the L2 Sobolev norm
of order k.
Proposition 1 ([4]). Assume that P ∈ Ψm cl (X) is microlocally solvable at the cone
Γ ⊂ S ∗ Rn . Then for any Y Rn such that Γ ⊂ S ∗ Y there exists an integer ν and
a Ψdo operator A so that W F (A) ∩ Γ = ∅ and
u −N ≤ C( P ∗ u ν + u −N −n + Au 0 ), ∀u ∈ C0∞ (Y ), (1)
where N is given by Definition 2.
Suppose now that V ⊂ T ∗ X\0 is an open conical set in ξ and Γ is a closed
cone contained in T ∗ X\0, Γ ⊂ V .

Proposition 2 ([9]). Suppose that P ∈ Ψm
cl (X) and s < s. Assume that it does not
s
exist a function u ∈ H (X) such that
V ∩ W F (P u) = ∅, V ∩ W Fs (u) = V ∩ W F (u) = Γ. (∗)

Then there exist: a point ρ ∈ Γ, Ψdo ψ, ϕ, ϕ ∈
0
cone supp ϕ ⊂ V \Γ,
Ψ0cl (X),
cone supp ϕ ⊂ V , ψ(ρ) = 1 in a neighbourhood of ρ , C = const > 0, μ ∈ Z+ for
0
which
ψw s ≤ C [ ϕ P w μ + ϕw μ + w s ] , ∀w ∈ C0∞ (X). (2)
Remark 1. Instead of (2) we can write
w s ≤ C [ Pw μ + Aw 0 + w s ] , ∀w ∈ C0∞ (X), (3)
where the full symbol of the ψdo A is identically O near ρ0 . As usual, Char P =
{ρ = (x, ξ) ∈ T ∗ X\0 : p0m (ρ) = 0}. If ρ0 ∈ Γ but ρ0 ∈ Char P then ρ0 ∈ W F (u),
i.e., (*) does not hold.
This way we conclude that the problem of existence of distribution solution
of the equation P u = v with prescribed singularity (∗) is reduced to the violation
of the a priori estime (2) ((3)). The same is true with respect to the (micro)local
nonsolvability of the operator P . Then we must violate the a priori estimate (1).
At the beginning of our paper we shall consider the Ψdo P ∈ Ψ2m cl (X) having
the principal symbol

n−1
n−1
p02m (x, ξ) = εj p2m.j (x, ξ), ε2m.j = 0, 1, ordξ pm.j = m, ε2j > 0 (4)
j=1 j=1
in some conical neighbourhood of the point ρ0 = (x0 , ξ 0 ), ξ 0 = 0.

Microlocal solvability and subellipticity 173
The real-valued functions pm,j (ρ), 1 ≤ j ≤ n − 1, are of principal type, i.e.,

the Hamiltonian vector field

n
∂pm,j (ρ) ∂ ∂pm,j (ρ) ∂
Hpm,j (ρ) = · − ·
∂ξk ∂xk ∂xk ∂ξk
k=1

n
∂
does not have the radial direction ξ · ∂ξ = ξk on the critical manifold
1
∂ξk
H = {ρ : pm,j (ρ) = 0, 1 ≤ j ≤ n − 1} .
Then Hpm,j = 0 and the integral curves of Hpm,j are called bicharacteristics.
Evidently, Char P ⊃ H. If εj = 0 for every 1 ≤ j ≤ n − 1 and they have the
same sign, then Char P = H.
The subprincipal symbol of the operator P ∈ Ψm cl (X) is given by the formula
i ∂ 2 p0m
n
pm−1 = pm−1 + . (5)
2 j=1 ∂xj · ∂ξj
It is well known that (5) is a well-defined function at the critical points of the
principal symbol p0m , i.e., at the points, where p0m (ρ) = ∇x,ξ p0m (ρ) = 0. Moreover,
(5) is invariant under conjugation by elliptic FIO and multiplication by elliptic
Ψdo there. In the case of operator (4) p2m−1 is invariant at H and
∗
p2m−1 = p∗2m−1 .
H H
At first, we shall study (4) under the assumption p2m−1 (ρ0 ) = 0, where
ρ ∈ H, i.e., the subprincipal symbol is elliptic at ρ0 ∈ H.
0
Here are our main assumptions:

(Λ) The Poisson brackets {pm,j , pm,l } ≡ 0 in some conical neighbourhood of ρ0
and ∇ξ pm,j (ρ0 ), 1 ≤ j ≤ n − 1 are linearly independent.
Therefore, H is a smooth manifold near ρ0 , codimH = n − 1.
(Ak0 ) (The sign change condition). Denote by γk0 the zero-bicharacteristic
of pm,k0 ,

εk0 = 1 passing through ρ , i.e., pm,k0
2 0
≡ 0. Then $p2m−1 changes its
γk0 γk0
sign at ρ0 only. Suppose that ρ ∈ H. Then the Hamiltonian vector fields
) pm.j are linearly
H * independent, tangential to H at ρ and each commutator
Hpm,j , Hpm,k , 1 ≤ j, k ≤ n − 1 is a smooth linear combination of the Hamil-
tonian vector fields Hpm.,l , 1 ≤ l ≤ n−1 at H.$The Frobenius theorem asserts
that there exists locally an integral manifold n−1 of those Hamiltonian vec-
tor fields. Therefore, H is foliating in a canonical
$ way and the corresponding
n − 1-dimensional leaves are of the type n−1 . Certainly, H is an involutive
manifold of dimension n + 1.
174 P.R. Popivanov
Geometrically we have
Theorem 1. Suppose that the operator with principal symbol (4) fulfills the con-
ditions (Λ), (Ak0 ), εk0 p2m−1 (ρ0 ) < 0, Hpm,k0 $p2m−1 (ρ0 ) = 0 and there exist n
smooth non-zero real vector fields αj (ρ), 1 ≤ j ≤ n, ρ ∈ γk0 , tangential to H,
transversal to Hpm,k0 (ρ0 ) and linearly independent at ρ0 , such that
−1/2
lim0 αj (ρ), ∇x,ξ p2m−1 (ρ) $p2m−1 (ρ) < ∞. (6)
ρ→ρ ,
ρ∈γk0
Then the operator P (x, D) is microlocally nonsolvable at ρ0 . Moreover, for each

conical neighbourhood W 0 ρ0 and for each pair of real numbers s < s one can
s
find a function u ∈ Hloc (X) having the following properties: W F (P u) ∩ W = ∅,
W F (u) ∩ W = W Fs (u) ∩ W = {tρ0 , t > 0}.
The conditions of Theorem 1 are invariant under conjugation by elliptic
FIO (i.e., under homogeneous symplectic changes of the variables) and by elliptic
Ψdo [2].
(6) is a growth condition guaranteeing the subordination of
1
α, ∇ p2m−1 and α, ∇$p2m−1 to $p2m−1 2
along the zero bicharacteristic γk0 0 ρ0 .

Remark 2. Let $p2m−1 be strictly monotone along γk0 and for each ρ ∈ H such
that $p2m−1 (ρ0 ) = 0
Hpm ,k0 $p2m−1 (ρ0 ) = 0.
Then without loss of generality we can assume that for each α ∈ Tρ0 (H), transver-
sal to Hpm ,k0 ,
1
α, ∇$p2m−1 ≤ C $p2m−1 2
along γk0 .
Example 1. Consider the Ψdo with the symbol

n−1
εj ξj2 + (ψ(xs ) + iϕ(xs )) |ξ|, 1 ≤ s ≤ n − 1,
1
where ψ, ϕ are real-valued smooth functions, εs ψ(0) < 0 and ϕ(xs ) changes its
sign at xs = 0 (say, ϕ(xs ) is strictly monotone, ϕ(0) = 0; functions flat at 0 are
also admissible).
To prove Theorem 1, we violate (1), (3) after using elliptic factorization

of the principal symbol p02m and conjugation by elliptic FIO. Thus, we apply
homogeneous canonical transformation [6] with respect to the operator P . After
the above-mentioned transformations p02m reduces to second order symbol having
the form

n−1
n−1
p02m = εj ξj2 , ε2j = 0, 1, ε2j > 0,
1 1
while H = {ξ1 = · · · = ξn−1 = 0, ξn = 0} is an involutive manifold of dimension

n + 1 in T ∗ Rn and Char p02 ⊃ H. The a priori estimates (1)–(3) are violated by
functions of the type

L

w= ϕj λ−j eiλh , ϕ0 x1 =0 = 1, ϕj x1 =0 = 0 for j ≥ 1.
j=0
The main difficulty is concentrated in solving approximately the following eikonal

equation with initial data:
n−1
2
∂h
εj + p1 (x, ξ 0 ) = O(|t|σ ), t → 0,
j=1 ∂xj (7)

h = i|t| 2
x1 =0
where ε1 = 0, σ ∈ N is arbitrary and the variable x ∈ R1 × Rn−1 is splitted into

two parts: x = (x1 , t), t = (x2 , . . . , xn ). As it concerns the subprincipal symbol,
$p1 (x1 , 0, ξ 0 ) changes its sign only at x1 = 0 (infinite order of vanishing is also
admissible) and condition (6) can be rewritten as:

∇t p1 (x1 , 0, ξ 0 )2 ≤ C|$p1 (x1 , 0, ξ 0 )|.
There are no difficulties in finding the amplitudes ϕj .

The local version of Theorem 1 is proved in [2]. As the proof in the microlocal
case is similar to that in the local one, we omit the details here. The integration
2
of the Cauchy problem (7) can be found in [2]. In the special case p02m = p0m
Theorem 1 holds under the assumptions p2m−1 (ρ0 ) < 0, p0m (ρ0 ) = 0 and the sign
change condition (A1 ) imposed on $p2m−1 . The growth condition (6) is unneces-
sary (see [5]).
176 P.R. Popivanov
2. N. Dencker proposed recently in his preprint [10] an interesting result on the

microlocal nonsolvability of Ψdo P with involative characteristics and subprin-
cipal symbol vanishing at some part of Char P . We shall formulate here a part
of his results as the microlocal form obtained in [10] will be used in our further
investigation – verification of microlocal subelliptic estimates for a subclass of the
Dencker’s class [10]. We give only the microlocal form, avoiding the geometrical
conditions leading to it.
Thus, split the variables into three groups: (t, x, y) ∈ R1 × Rn × Rm and
denote by (τ, ξ, η) the corresponding dual variables. Put Σ2 = {η = 0}. Σ2 is
involutive manifold, of course. Consider the Ψdo operator (Dencker’s form)
P ∗ = F (t, x, y, Dt , Dx , Dy ) (8)
microlocally near the cone Γ = {t, x0 , y0 ; 0, ξ 0 = 0, 0}, t ∈ I = (−ε, ε), where
Ψ2cl 0 F ∼ F20 + F1 + · · · , ord(τ,ξ,η) Fj = j,

F20 = Bjk (t, x, y; τ, ξ, η)ηj ηk ,
j,k
Bjk is positively homogeneous of 0 order with respect to (τ, ξ, η),

F20 Σ2 = ∇F20 Σ2 = 0
and the matrix (Bjk )j,k is smooth.
It is also assumed that
F1 (t, x, y; τ, ξ, η) = τ + if (t, x, ξ) + A(t, x, y; τ, ξ, η), η , ordF1 = 1,
the function f ∈ C ∞ is real-valued, ordξ f = 1 and the mapping t → f (t, x0 , ξ0 )
changes sign from “+” to “−” as t increases on I ⊂ R. If f (t, x0 , ξ0 ) = 0 on
a subinterval I ⊂ I, |I | > 0, then it is supposed that ∂tk ∂xα ∂ξβ f (t, x0 , ξ0 ) = 0,
∀k, α, β and for t ∈ I . Let the sign change of f be of infinite order on Γ. In this
case it is required additionally that
{Bjk }jk + |A| + |df | ≤ C|f |ε1 near Γ on Σ2
for some ε1 > 0.
Theorem 2 (N. Dencker). Consider the operator P ∗ defined by (8). Under the
above-formulated conditions its L2 adjoint operator P is not microlocal solvable
at Γ.
Remark 3. Let t → f (t, x0 , ξ0 ) have a finite-order sign change from “−” to “+”
at t = 0. Then P is (micro)locally nonsolvable at (0, x0 , y0 ; 0, ξ 0 , 0).
Example 2.
(a) P = D1 D2 + if (t, y, Dy ), t = (x1 , x2 ), y ∈ Rn , where ordη f = 1, f (t, y, η) is
real-valued and x1 or x2 → f (t, y, η) changes its sign of finite order from ”−“
to ”+“. Then P is (micro)locally nonsolvable.

ξ = (ξ1 , ξ2 , η),
(b) P = D1 D2 + Dt + if (t, x, Dx ), where y ∈ Rn , ∂xj f =
x = (x1 , x2 , y),
O(|f |), j = 1, 2 and t → f (t, x, ξ) changes its sign of finite order from “−” to
“+”. Evidently, Σ2 = {ξ1 = ξ2 = 0}. The operator P is locally nonsolvable.
3. Our next step is to obtain in some special cases of the microlocal form (8),
including f ≥ 0 (f ≤ 0) everywhere, subelliptic estimates for the operators under
consideration. So let the Ψdo
Q = F (t, x, y, Dt , Dx , Dy ) (9)
microlocally near the cone Γ = {(t, x0 , y0 ; 0, ξ 0 = 0, η 0 = 0), t ∈ I}, F20 (η) =
m

Bjl ηj ηl , Σ2 = {η = 0} is involutive submanifold, F20 Σ2 = 0, Bjl are real
j,l=1
constants, ordη F1 = 1. We shall deal with two subcases, namely:
Q = F20 + F1 , F1 = τ + iP (t, ξ, η)b(ξ, η) ≡ τ + if (9a)

k−1
where P = tk + aj (ξ, η)tj ≤ 0 (≥ 0) everywhere and b, aj ∈ C ∞ are real-valued
j=0
functions near Γ, positively homogeneous of order 1, respectively of order 0, with
respect to (ξ, η) and b > 0 near Γ, i.e., b(ξ, η) ∼ c0 (|ξ| + |η|) there; c0 > 0.

k
Q = F20 + F1 , F1 = τ + iP (t)b(ξ, η) ≡ τ + if, P (t) = a j tj , (9b)
j=0
ak = 1, aj are real constants and the polynomial P (t) changes sign from “−” to
“+” for increasing t ∈ I. (Suppose that t ∈ [−ε, ε]. Then P (t, ξ, η) ≥ 0 (≤ 0)
everywhere for (9a), while P (t0 ) = 0 for some t0 ∈ (−ε, ε) ⇒ P (t) < 0 for t < t0
and P (t) > 0 for t > t0 in case (9b).) Evidently, Q2 = F1 = τ + if . Having in
mind that aj (ξ, η) are bounded, we conclude that the coefficients of P (t, ξ, η) are
bounded near Γ and consequently the corresponding roots of the polynomial in t
P (t, ξ, η) = 0, i.e., t1 (ξ, η), . . . , tk (ξ, η) are bounded too. Our first result concerns
the operator with symbol (9a), i.e.,
Q(t, Dt , Dx , Dy )u

m
k
= Bjl Dyj Dyl u + Dt u + i tj aj (Dx , Dy ) b(Dx , Dy ) (10)
j,l=1 j=0
= g(t, x, y), t ∈ I

and u ∈ C0∞ R1t × Rnx × Rm y , supp u ⊂ {|t| < ε}.
Proposition 3. Consider the ψdo with symbols (9a), (9b) near the point ρ0 =
(t0 , x0 , y0 ; 0, ξ 0 = 0, 0), where P (t0 , ξ 0 , 0) = 0, respectively P (t0 ) = 0, |t0 | < ε.
1
s+
Then Qu ∈ Hmcl
s
(ρ0 ) ⇒ u ∈ Hmclk+1 (ρ0 ) for each s.
178 P.R. Popivanov
In the case (9a) k is even, while in the case (9b) k is odd. The rational num-
ber k+11
is called gain of regularity of the operator Q and 2 − k+1 1
= 2k+1
k+1 < 2 is
called loss of regularity of Q. Evidently, k ∞ ⇒ k+1 → 0 and 2 − k=1
1 1
→ 2.
The coefficients of Q are polynomials in t. The case when the coefficients of Q are
flat functions in t is rather complicated. We can expect the appearance of subel-
liptic estimates with loss of regularity equal to 2. In order to obtain more precise
regularity results it is useful to introduce Sobolev spaces with logarithmic weight
logα (2 + |ξ|2 ). The corresponding definitions and results with logarithmic gain of
regularity for operators with Gevrey coefficients will be given in Proposition 5,
immediately after the proof of Proposition 3. It concerns the following class of
operators:
m
R= bjl Dyj Dyl + Dt − iγ(t)b(Dx , Dy ), ord(ξ,η) b = 1, (11)
j,l=1
b ∼ c0 (|ξ| + |η|), γ ∈ C ∞ [−1, 1] and the real-valued function γ is even, γ > 0 for
1
t > 0, γ(0) = 0 and 0 γ(t)dt = 1.
At the end of this paper we will discuss the problem of local nonsolvability
of the operators (9b), more precisely we shall deal with
Qu = F2 (Dy ) + Dt u + if (t, Dx , Dy )u = g(t, x, y), (12)
where f (t, ξ, η) = P (t)b(ξ, η), b(ξ, η) ∼ c0 (|ξ| + |η|), c0 > 0 and the real-valued
function P ∈ C ∞ is such that P (0) = 0, P (t) > 0 for t > 0, P (t) > 0 for t > 0.
t
Put B(t) = 0 P (s)ds. Thus, P (t) can vanish of infinite order at t = 0, changing
sign from “−” to “+” for t – increasing.
Consider the cut-off real function ψ ∈ C0∞ Rm+n ξ,η , 0 ≤ ψ ≤ 1, ψ = 1
in a neighbourhood of the ray (λξ 0 , η 0 = 0), λ ≥ 1, ψ = 0 outside a larger
neighbourhood of the same ray. Define now the integral operator
(E ∗ v)(x, y)
∞
(13)
= ψ(ξ, η)ei(x,ξ)+i(y,η) e−B(t)b(ξ,η)+iF2 (η)t F2 (v)(t, ξ, η)dtdξdη,
−∞ Rm+n
(ξ,η)
v being compactly supported function near (0, x0 , y0 ) (say, v(t, x, y) ∈ C0∞ , supp v
is located near to (0, x0 , y0 )). We denote by v̂ = F (v) the usual Fourier trans-
formation, and by F2 (v) the partial Fourier transformation with respect to the
second variables (x, y) → (ξ, η).
Proposition 4. Assume that Qu = v for the compactly supported distribution u ∈
E (⇒ v ∈ E ) near (0, x0 , y0 ), v being sufficiently smooth. Then (E ∗ v)(x, y) = 0
for each (x, y) near (x0 , y0 ).
Thus, the partial Fourier transformation F2 (v) of the right-hand side v(t, x, y)
should satisfy infinitely many (continuum many) compatibility conditions. This
fact explains the nonsolvability of Qu = v for each v ∈ C0∞ (Ω), Ω being arbitrary
small neighbourhood of (0, x0 , y0 ).
2. Proof of Propositions 3–5

û(τ, ξ, η) stands for the full Fourier transformation of the tempered distribution u:
t → τ , x → ξ, y → η. Put z = (x, y) and ζ = (ξ, η) and let P (s, ζ) ≤ 0.
After a partial Fourier transformation in (10) we get:
(iF2 (η) + ∂t − f (t, ξ, η)) F2 (u)(t, ξ, η) = iF2 (g)(t, ξ, η). (14)
We multiply (14) by F2 (u), integrate with respect to t ∈ (−ε, ε) and take the real
part of both sides. Thus,
t
(∂t F2 (u)(s, ζ) − f (s, ζ)F2 (u)) F2 (u)(s, ζ)ds
−ε
t (15)
= −$ F2 (g)(s, ζ)F2 (u)(s, ζ)ds,
−ε
as F2 (u)(−ε, ζ) = 0. Consequently,
t
1
|F2 (u)(t, ζ)|2 + b(ζ) (−P (s, ζ)) |F2 (u)|2 ds
2 −ε
t ε (16)
≤ |F2 (g)| |F2 (u)| ds ≤ |F2 (g)| |F2 (u)| ds.
−ε −ε
1
Put M = {|t| ≤ ε : |t − tj (ζ)| ≥ |ζ|− k+1 , j = 1, 2, . . . , k}. The parame-
ters (ξ, η) are fixed.Otherwise, M = M (ζ). Then the measure of the complement
1
[−ε, ε]\M : meas ([−ε, ε]\M ) ≤ meas (R1 \M ) ≤ 2k|ζ|− k+1 .
Evidently,
⎧
ε ⎨ |F2 (u)|
2
(t, ζ), ∀t ∈ (−ε, ε)
t
2 |F2 (g)| |F2 (u)| ds ≥
−ε ⎩ b(ζ) (−P ) |F2 (u)|2 ds, ∀t ∈ (−ε, ε),
−ε
i.e.,
ε ε
2 2
4 |F2 (g)| |F2 (u)| ds ≥ max |F2 (u)| (t, ζ) + b(ζ) (−P ) |F2 (u)| ds. (17)
−ε t∈[−ε,ε] −ε
On the other hand,

ε
2 2 2
|F2 (u)| (s, ζ)ds = |F2 (u)| (s, ζ) + |F2 (u)| (s, ζ)ds
−ε M [−ε,ε]\M
and

2 1 1 2
max |F2 (u)| (t, ζ) ≥ |ζ| k+1 |F2 (u)| (s, ζ)ds. (18)
t∈[−ε,ε] 2k [−ε,ε]\M
180 P.R. Popivanov

k
(t − tj (ζ)) ⇒ |P (t, ζ)| ≥ |ζ|− k+1 . Therefore,
k
If t ∈ M ⇒ P (t, ζ) =
1
ε
b(ζ) (−P ) |F2 (u)|2 ds ≥ b(ζ) (−P ) |F2 (u)|2 ds
−ε M
(19)
− k+1
k 2
≥ b(ζ)|ζ| |F2 (u)| ds.
M
Combining (17), (18), (19) and having in mind that b(ζ) ∼ c0 |ζ| we obtain:
ε ε
1 2
|F2 (g)| |F2 (u)| ds ≥ C1 |ζ| k+1 |F2 (u)| ds. (20)
−ε −ε
Applying the Cauchy–Schwarz inequality to the left-hand side of (20) we

conclude that:
ε ε
2
2 −2 − k+1 2
|F2 (u)| ds ≤ C1 |ζ| |F2 (g)| ds. (21)
−ε −ε
Certainly, near ρ0 we have that |τ | ≤ δ|ξ|, |η| ≤ δ|ξ|, 0 < δ " 1. Let Γ
be a conical neighborhood of ρ0 , Γ = V(t0 ,z0 ) × Γ(τ,ζ) , i.e., Γ(τ,ζ) = {(τ, ζ) = 0 :
|τ | ≤ δ|ξ|, |η| ≤ δ|ξ|}. Here |ξ| = |ξ 0 | = 1, ξ is near to ξ 0 . Denote by W a conical
neighbourhood of (x0 , y0 ; ξ 0 = 0, η 0 = 0) : W = Vz0 ×Wζ0 , W = {ζ = 0, |η| ≤ δ|ξ|},
i.e., W is a cone in dual space Rm+nζ with vertex at 0, the axes (ξ 0 , 0) and opening
δ > 0. Integrating (21) in W we have:
ε ε
2 2 2
|F2 (g)| (t, ζ)dtdζ ≥ C12 (1 + |ζ|) k+1 |F2 (u)| (t, ζ)dtdζ, (22)
−ε
W −ε
W
as 1 + |ξ| + |η| ∼ (1 + |ξ|) in W .

Conclusion: Let g = Qu ∈ L2 (I, L2 ). Then
mcl,W
1
u ∈ L2 I, H k+1 , I = (−ε, ε).
mcl,W
t
The case P ≥ 0 is studied in a similar way but then we integrate ε , t ∈ (−ε, ε),
etc. As it concerns (9b) we consider two cases separately: a) P (t) ≤ 0 for t ≤ t0
t t
integrating then −ε , and b) P (t) ≥ 0 for t ≥ t0 . Then we integrate ε , t ∈ (−ε, ε).
Combining the estimates obtained in a) and b) we come to the same subelliptic
estimate (22) for k – being odd.
Remark 4. As the above given proof imitates the proof of Theorem 4.3, Chapter
VIII from [6] and Lemma 3.1 from [1], we do not give details. Applying Parseval’s
equality to (22) we obtain

∞
2
|ĝ(τ, ξ, η)| dτ dξdη
−∞ W
∞
2 2
≥ C12 (1 + |ξ| + |η|) k+1 |û| (τ, ξ, η)dτ dξdη ≥ (23)
−∞ W
∞
2 2
≥ C12
(1 + |ξ|) k+1 |û| dτ dξdη.
−∞
W
1
Thus, g ∈ Hmcl 0
(ρ0 ) ⇒ u ∈ Hmcl
k+1
(ρ0 ) as in (23) we are integrating in a conical
neighbourhood of ρ : {|τ | ≤ δ|ξ|, |η| ≤ δ|ξ|} . By using the Ψdo with the symbol
0
(1 + |ξ|2 + |η|2 )s we deduce from (23) that for each real s, g ∈ Hmcl s
(ρ0 ) ⇒ u ∈
1
s+
Hmclk+1 (ρ0 ).
Remark 5. The subprincipal symbol of the operators with symbols (9a), (9b) is
equal to τ + if (t, ζ) and is microlocally subelliptic and hypoelliptic with sharp loss
k
of regularity , k ∈ N (see [4], [6]). The addition of the second-order operator
k+1
with real constant coefficients and involutive characteristics F2 (Dy ) conserves the
2k + 1
microlocal subellipticity but the loss of regularity becomes 2 > > 1, k ∈ N.
k+1
We point out that the operator (9b) is locally and microlocally nonsolvable at
(t0 , x0 , y0 ), respectively at (t0 , x0 , y0 ; τ 0 = 0, ξ 0 = 0, η 0 = 0) according to Theo-
rem 2. Similar considerations for the Ψdo FΓ (Dy ) + Dt + if (t, Dx , Dy ) = N with
FΓ ≡ 0, ordη FΓ = Γ ≥ 2, FΓ having real constant coefficients, enable us to con-
clude that if P satisfies the conditions (9a), (9b), then N is microsubelliptic with
1
loss of regularity Γ − . Moreover, under the conditions (9b) we can expect
k+1
(microlocal) nonsolvability of N .
We propose below some preparatory definitions and facts on the generalized
Sobolev spaces [8] in order to apply them to the Ψdo of the type (11) having
flat coefficients (mainly coefficients belonging to some Gevrey classes Gs (see [7]).
t
With the notations from (11) we put Γ(t) = 0 γ(s)ds, 0 ≤ t ≤ 1. Then Γ(t) has
the following properties (see [8]):
(i) Γ is strictly monotonically increasing, Γ(0) = 0, Γ(1) = 1, 0 ≤ Γ(t) ≤
tγ(t) ≤ γ(1)t for t ∈ [0, 1], i.e., Γ is a homeomorphism from [0, 1] → [0, 1], a
Γ(t)
diffeomorphism from (0, 1) → (0, 1) and γ(t) ≥ . Therefore, there exists Γ−1 :
t
[0, 1] → [0, 1] and Γ−1 (0) = 0, Γ−1 (1) = 1. As examples we take: (a) γ = |t|s , s > 0,
1 1
(b) γ(t) = 2 e1− |t| , |t| ≤ 1. Certainly, in case (b) γ is flat at t = 0, γ ∈ G2 (R1 ),
t 1
i.e., to a larger class than G2 ([−1, 1]). Let Γ(t) = e1− t = x for 0 ≤ t ≤ 1. Then
1
t = Γ−1 (x) = for 0 ≤ x ≤ 1.
log xe
182 P.R. Popivanov
Remark 6. Let Γ(t) = e1−(1/|t| ) , α > 0. Then Γ ∈ G1+ α1 . Put γ(t) = Γ (t). One
α
1
can easily see that t = Γ−1 (x) = for 0 ≤ x ≤ 1.
e 1/α
ln x
Define now the positive continuous function νγ on [0, ∞):
⎧
⎨ 1
1, z ≥ 1
νγ (z) = Γ −1
⎩ z
1, 0 ≤ z ≤ 1.
We list below several properties of νγ :
νγ (z) 1
(ii) νγ (z) +∞, → 0. Suppose now that ε ≥ |t| ≥ → 0 . Then
z∞ z z→0 νγ (λ) λ→∞
1 1 1
λγ(t) ≥ λγ ≥ ν(λ). In fact, ν(λ) = −1 1 for λ ! 1, i.e., λΓ =
ν(λ) Γ λ
ν(λ)
1
1. Having in mind that γ is even, we get: |t| ≥ (νγ (λ))−1 ⇒ λγ(t) ≥ λγ ≥
ν(λ)
1
λν(λ)Γ (λ) = ν(λ).
ν
Repeating the proof of Proposition 3, case (9a), and defining

1
M = t : |t| ≤ ε, |t| ≥ , λ = |ζ|, λ → ∞
νγ (λ)
(i.e., 1/(νγ (λ)) → 0) we obtain:
ε ε
2
|F2 (g)| |F2 (u)| ds ≥ C1 ν(λ) |F2 (u)| ds
−ε −ε
ε ε (24)
⇒ |F2 (g)|2 ds ≥ C12 νγ2 (|ζ|) |F2 (u)|2 ds,
−ε −ε
as b0 (ξ, η) ∼ c0 λ.
We shall introduce now a generalization of the classical Sobolev spaces having
μ(ξ)
the weight function 0 < μ ∈ C 0 (Rn ) [8]. By definition ≤ C(1+|ξ−η|)l for each
μ(η)
s
ξ, η ∈ Rn and C > 0, l > 0 are some constants. The function μs (ξ) = (1 + |ξ|2 ) 2
is the weight function to the Sobolev space H s (Rn ).
Definition 3. The symbol H μ (Rn ) stands for the set of all tempered distributions
u ∈ S (Rn ) such that û(ξ) is a measurable function and the norm

u μ = |μû|2 dξ < ∞.
2
The spaces H μ can be microlocalized in a standard way. In our case μ = νγ .

Integrating (24) in W we get:
ε ε
2
|ĝ(τ, ξ, η)| dτ dξdη ≥ C1
2
νγ2 (|ζ|)|û(τ, ξ, η)|2 dτ dξdη (25)
−ε
W −ε
W
and |ζ|2 = |ξ|2 + |η|2 . As we know, W = {|η| ≤ δ|ξ|} and |τ | ≤ δ|ξ|. The mono-
tonicity of νγ (λ) w. r. to λ ≥ 1 implies that νγ (|ζ|) ∼ νγ (|ξ|) in W . In our model
1
example case (b) νγ is logarithmic weight of the type [log(e|ξ|)] α , |ξ| ≥ 1, i.e., we
have logarithmic gain of regularity only. This way we come to:
ν
Proposition 5. Consider the Ψdo (11). Then Ru ∈ Hmcl 0
(ρ0 ) ⇒ u ∈ Hmcl
γ
(ρ0 ).
In all cases the full symbol of the operators under consideration vanish at the
characteristic point ρ0 .
The proof of Proposition 4 is simple. In fact, Q∗ u = F2 (Dy )u + Dt u −
if (t, Dx , Dy )u = 0. Therefore, F2 (u)(t, ξ, η) = ψ(ξ, η)ŵ(ξ, η)e−B(t)b(ξ,η)−itF2 (η) ,
ŵ(ξ, η) being arbitrary tempered distribution in Rm+n (ξ,η) and we define

−1
u(t, x, y) = Ew = Fξ→x,η→y ŵ(ξ, η)ψe−B(t)b−itF2 (η) , (26)
F −1 being the inverse Fourier transformation with respect to (ξ, η). Evidently,
Q∗ Ew = 0 (at least for w ∈ C0∞ ).
Assume that
(Ew, g)L2 (Rm+n+1 ) = (w, E ∗ g)L2 (Rm+n
x,y )
t,x,y
∞
m+n+1
for g ∈ C0 Rt,x,y . Easy computations show that
(E ∗ g)(x, y)
∞
= ψ(ξ, η)ei(x,ξ)+i(y,η) e−B(t)b(ξ,η)+iF2 (η)t F2 (g)(t, ξ, η)dtdξdη.
−∞ Rm+n
(ξ,η)
E ∗ is the L2 adjoint operator to E. Suppose now that u ∈ E , Qu = v (i.e., v ∈ E ).

Then E ∗ Qu = E ∗ v ⇒ (Q∗ E)∗ u = E ∗ v ⇒ E ∗ v = 0.
From Theorem 2 it is known that the operator (12) is (micro)locally nonsolv-
able. Proposition 4 explains the nonsolvability of Q by the overdetermination of the
problem (v must satisfy continuum many compatibility-orthogonality conditions).
References
[1] G. Eskin. Elliptic pseudodifferential operators having first order degeneration along
the space variables, Trudy Mosk. Math. Ob-va, vol. 25, 83–118, (1971).
[2] P. Popivanov. Local solvability of pseudodifferential operators having double char-
acteristics, Math. Sbornik, vol. 100, 217–241 (1976).
[3] F. Treves. Introduction to pseudodifferential and Fourier integral operators, volumes
1–2, Plenum Press, NY and London (1982).
[4] L. Hörmander. The analysis of linear partial differential operators, vol. I–IV,
Springer-Verlag, Berlin (1983–1985).
[5] P. Popivanov, Ch. Georgiev. A necessary condition for the local solvability of some
operators with double characteristics, Annuaire de l’Université de Sofia “St. Kl.
Ohridski” – Math., vol. 75, 57–71 (1984).
184 P.R. Popivanov
[6] Yu.V. Egorov. Linear differential equations of principal type, Consultants Bureau,
NY (1986).
[7] L. Rodino. Linear partial differential equations in Gevrey spaces, World Sci., Singa-
pore (1993).
[8] B. Paneah. The oblique derivative problem. The Poincaré problem, Wiley-VCH,
Berlin – NY – Singapore (2000).
[9] P. Popivanov. Hypoellipticity, solvability and construction of solutions with pre-
scribed singularities for several classes of PDE having symplectic characteristics,
Rend. Sem. Univ. Pol. Torino, vol. 66:4, 321–337 (2008).
[10] N. Dencker. Operators of subprincipal type, arXiv: 1507.05594V2 [math. AP], 3 Oct.
2015, 1–26.
P.R. Popivanov
Institute of Mathematics and Informatics
Bulgarian Academy of Sciences
Acad. G. Bonchev str., bl 8
Sofia 1113, Bulgaria
Operator Theory:

An Observation of the Subspaces of S

Yoshihiro Sawano
Abstract. The spaces S /P equipped with the quotient topology and S∞
equipped with the weak-∗ topology are known to be homeomorphic, where

P denotes the set of all polynomials. The proof is a combination of the fact
in the textbook by Treves and the well-known bipolar theorem. In this paper
by extending slightly the idea employed in [5], we give an alternative proof
of this fact and then we extend this proposition so that we can include some
related function spaces.
1. Introduction
It is useful to consider the quotient spaces of S or D when we consider the
homogeneous function spaces. Usually such a quotient space can be identified with
some dual spaces as the following theorem shows:
Theorem 1.1. Let X be a locally convex (Hausdorff) space whose topology is given
by a family of semi-norms {pλ }λ∈Λ . Equip X ∗ with the weak-∗ topology. Let V be
a closed subspace of X ∗ . Define the orthogonal space XV to V by:

XV ≡ ker(x∗ )
x∗ ∈V
and equip XV with the topology induced by X. Then the topological dual XV∗ is
isomorphic to X ∗ /V equiped with the quotient topology.
The proof of Theorem 1.1 is a combination of [9, Propositions 35.5 and 35.6]
and the bipolar theorem.
Theorem 1.2 ([9, Propositions 35.5 and 35.6]). Let X be a locally convex Hausdorff
space, and let N be a closed linear subspace of X. Then the kernel of the restriction
X to N is
N◦ = {x∗ ∈ X : < x , n >= 0}.
n∈N
Furthermore, its quotient mapping is a homeomorphism from X /N ◦ to N .
186 Y. Sawano
Theorem 1.3 (Bipolar theorem, [7, p. 126, Theorem]). Let X be a Hausdorff topo-
logical vector space. Let V be a closed subspace of X ∗ equipped with the weak-∗
topology. Define

◦
V ≡ {x ∈ X : v ∗ , x = 0} = ker(v ∗ ).
v ∗ ∈V v ∗ ∈V
Then

V = {v ∗ ∈ X ∗ : v ∗ , x = 0}(≡ (◦ V )◦ ).
x∈◦ V
If we let V = P(⊂ S ), the linear space of all polynomials, then we can show
that V is a closed subspace of S . One of the ways to check this is to use the
Fourier transform. In fact, f ∈ S belongs to P if and only if the Fourier transform
is supported in {0}.
The isomorphism X ∗ /V → XV∗ is given as follows: Let R be the natural
restriction mapping R : X ∗ ∈ f → f |XV ∈ XV∗ . Denote by ι : XV → X the
natural inclusion. Then ι and R are dual to each other and R is clearly continuous.
The aim of this paper is to give out an alternative proof of Theorem 1.1.
We organize this short note as follows: We prove Theorem 1.1 in Section 2. After
collecting some preliminary facts in Section 2.1 we plan to prove Theorem 1.1. We
shall show that ker(R) = V in Section 2.2 which is essentially the bipolar theorem,
that R is surjective in Section 2.3 and that R is an open mapping in Section 2.4.
We compare Theorem 1.1 with the existing results in Section 3. We devote

Sections 3.1, 3.2, 3.3 and 3.4 to the application of Theorem 1.1 to the spaces S∞ ,

Sm , Se and D respectively. The definition of Sm and Se will be given in Sections
3.2 and 3.3, respectively. For a topological space Y and its dual Y ∗ , we write
y ∗ , y ≡ y ∗ (y) for the coupling of y ∈ Y and y ∗ ∈ Y ∗ .
2. Proof of Theorem 1.1

2.1. A reduction and preliminaries
Let X be a locally convex space whose topology is given by a family of semi-norms
{pλ }λ∈Λ . Let us set
# 8

P≡ aλ pλ : Λ0 is a finite subset of Λ and {aλ }λ∈Λ0 ⊂ N .
λ∈Λ0
∗
Let O ⊂ X be an open set. Then there exists q ∈ P such that
{x ∈ X : q(x) < 1} ⊂ O.
Therefore by replacing {pλ }λ∈Λ with P, we can assume that for any open set O
there exists λ(O) ∈ Λ such that
{x ∈ X : pλ(O) (x) < 1} ⊂ O. (2.1)
We invoke the propositions concerning the Hahn–Banach extension. First, we
recall the following fact:
An Observation of the Subspaces of S 187
Proposition 2.1 (Geometric form, [7, p. 46]). Let M be a linear subspace in a

topological vector space L and let A be a non-empty convex, open subset of L, not
intersecting M . Then there exists a closed hyperplane in L, containing M and not
intersecting A.
Next, we recall the Mazur theorem.
Proposition 2.2 (Analytic form, [14, p. 108, Theorem 3]). Let X be a locally convex
linear topological space, and M be a closed convex subset of X such that a · m ∈ M
whenever |a| ≤ 1 and m ∈ M . Then for any x0 ∈ X \ M there exists a continuous
linear functional f on X such that f (x0 ) > 1 ≥ |f (x)| for all x0 ∈ M .
When M is a linear subspace in the above, f (x) = 0 for all x ∈ M . Thus, we

can deduce the following well-known version:
Proposition 2.3 (Analytic form). Let X be a topological vector space and let Y
be a closed linear space. Then for any continuous linear functional Y and x ∈
X \ Y there exists a continuous linear functional X such that X |Y = Y and that
X (x) = 0.
2.2. The kernel of R

We now specify ker(R). It is easy to see that V ⊂ ker(R) and that V and ker(R) are
weak-∗ closed. Assume that V and ker(R) are different. Then by Proposition 2.1,
we can find a continuous linear functional Φ∗∗ : X ∗ → C such that V ⊂ ker(Φ∗∗ )
and
ker(R) ∩ ker(Φ∗∗ )c = ∅. (2.2)
Since Φ∗∗ : X ∗ → C is continuous, we have
{x∗ ∈ X ∗ : | x∗ , x1 | < 1, | x∗ , x2 | < 1, . . . , | x∗ , xk | < 1}
⊂ {x∗ ∈ X ∗ : | Φ∗∗ , x∗ | < 1}
for some x1 , x2 , . . . , xk ∈ X. This means that

k
ker(Qxj ) ⊂ ker(Φ∗∗ ),
j=1
where Q : X 0 x → Qx ∈ X ∗∗ is a natural inclusion. By the Helly theorem, we

see that
k
Φ∗∗ = aj Qxj ∈ X ∗∗ .
j=1
Let x∗ ∈ V be arbitrary. Then we have

9 :

k
x∗ , aj xj = Φ∗∗ , x∗ = 0,
j=1
188 Y. Sawano
since x∗ ∈ ker(Φ∗∗ ). This means that

k
aj xj ∈ XV .
j=1
Let x∗ ∈ ker(R). Then

9 :
; <
k
∗∗ ∗ ∗ ∗
Φ , x = Qk aj x j , x = x , aj xj = 0,
j=1
j=1
$k
since x∗ |XV = 0 and j=1 aj xj ∈ XV . Thus, ker(R) ⊂ ker(Φ∗∗ ). This contradicts
(2.2).
2.3. The surjectivity of R : X → XV

Let z ∗ ∈ XV∗ . Then | z ∗ , z | ≤ pλ (z) for some λ ∈ Λ by (2.1). Use the Hahn–
Banach theorem of analytic form to have a continuous linear functional x∗ on X
which extends z ∗ . Then R(x∗ ) = z ∗ .
2.4. The openness of R

We need the following lemma:
Lemma 2.4. Let z1 , z2 , . . . , zk ∈ XV and x1 , x2 , . . . , xl ∈ X. Assume that the

system {[x1 ], [x2 ], . . . , [xl ]} is linearly independent in X/XV . Then for all z ∗ ∈ XV
satisfying | z ∗ , zj | < 1 for all j = 1, 2, . . . , k, we can find x∗ ∈ XV so that z ∗ is
a restriction of x∗ and that x∗ , xj = 0 for all j = 1, 2, . . . , l.
Proof. We know that any linear space X has a norm · X although it is not
necessarily compatible with its original topological structure of X. For example,
choose a Hamel basis {xθ }θ∈Θ and define
6 6
6 6
6 6
6 aθ xθ 6 ≡ |aθ |
6 6
θ∈Θ0 X θ∈Θ0
for any finite set Θ0 .

$l
Observe j=1 aj xj ∈ / XV for any (a1 , a2 , . . . , al ) = (0, 0, . . . , 0), which yields
x∗(a1 ,a2 ,...,al ) ∈ X ∗ such that
9 :

l
∗
x(a1 ,a2 ,...,al ) , aj xj = 1. (2.3)
j=1
Since x∗(a1 ,a2 ,...,al ) ∈ X ∗ is a continuous linear functional, we can find an index
λ(a1 , a2 , . . . , al ) ∈ Λ such that
; <
{x ∈ X : pλ(a1 ,a2 ,...,al ) (x) < 1} ⊂ {x ∈ X : | x∗(a1 ,a2 ,...,al ) , x | < 1}. (2.4)
Write
# 8

l
U(a1 ,a2 ,...,al ) ≡ (b1 , b2 , . . . , bl ) ∈ C : pλ(a1 ,a2 ,...,al )
n
(aj − bj )xj <1 .
j=1
(2.5)
Since S 2l−1 ≡ {(b1 , b2 , . . . , bl ) ∈ Cl : |b1 |2 + |b2 |2 + · · · + |bl |2 = 1} is a compact
set, we can find a finite covering {U(a1 ,a2 ,...,al ) }(a1 ,a2 ,...,al )∈A of S 2l−1 , where A is
a finite subset of S 2l−1 .
Let us denote by XA and XA,V the completion of X and XV with respect to
the norm
· A≡ · X+ pλ(a1 ,a2 ,...,al ) (·),
(a1 ,a2 ,...,al )∈A
respectively. Then we can extend x∗(a1 ,a2 ,...,al ) to
a continuous linear functional
X∗(a1 ,a2 ,...,al ) to XA,V .
Note that for any (b1 , b2 , . . . , bl ) ∈ Cn \ (0, 0, . . . , 0), there exists
(a1 , a2 , . . . , al ) ∈ A
such that
1
. (b1 , b2 , . . . , bl ) ∈ U (a1 , a2 , . . . , al ).
|b1 | + |b2 |2 + · · · + |bl |2
2
For such an element (a1 , a2 , . . . , al ) ∈ A, we deduce

k

∗
xλ(a1 ,a2 ,...,al ) aj xj = 0
j=1
from (2.3)–(2.5). Since X∗λ(a1 ,a2 ,...,al ) (x)

= 0 for any x ∈ XV,A , we see that
{[x1 ], [x2 ], . . . , [xl ]} is a linearly independent system in XA /XA,V . Then use the
Hahn–Banach theorem of geometric form (see Proposition 2.3). Then we obtain a
continuous functional X∗ : XA → C so that X∗ (xj ) = 0 for j = 1, 2, . . . , l. If we
set X∗ |X = x∗ , then we have the desired result.
The following proposition shows that R is an open mapping:
Proposition 2.5. Let z1 , z2 , . . . , zk ∈ XV and x1 , x2 , . . . , xL ∈ X. Then there exists
r > 1 depending only on z1 , z2 , . . . , zk ∈ XV and x1 , x2 , . . . , xL ∈ X such that for
all z ∗ ∈ XV satisfying | z ∗ , zj | < 1 for all j = 1, 2, . . . , k, we can find x∗ ∈ XV
so that z ∗ is a restriction of x∗ and that | x∗ , xj | < r for all j = 1, 2, . . . , L.
Namely, the range

k
l
{x∗ ∈ X ∗ : | x∗ , zj | < 1} ∩ {x∗ ∈ X ∗ : | x∗ , xj | < r}
j=1 j=1
by R contains

k
{x∗ ∈ X ∗ : | x∗ , zj | < 1}.
j=1
190 Y. Sawano
In particular, the range

k
l
{x∗ ∈ X ∗ : | x∗ , zj | < 1} ∩ {x∗ ∈ X ∗ : | x∗ , xj | < 1}
j=1 j=1
by R contains

k
{x∗ ∈ X ∗ : | x∗ , zj | < r−1 }.
j=1
Proof. Let us assume that {[x1 ], [x2 ], . . . , [xl ]} is a maximal linearly independent
family in X/XV , where l ≤ L. Then for any j ∈ (l, L] ∩ N and k ∈ [1, l] ∩ N, we
can find μjk ∈ C and z̃j ∈ XV such that

l
xj = z̃j + μjm xm .
m=1
Let
r ≡1+ max | z ∗ , z̃j |
j=1,2,...,l
Note that
| z ∗ , zj |, | z ∗ , z̃m | < r
for all j = 1, 2, . . . , k and m = l + 1, l + 2, . . . , L. According to Lemma 2.4, we can
find x∗ ∈ X ∗ so that x∗ is an extension of z ∗ and that
z ∗ , xj = 0, j ∈ (l, L].
Thus, we obtain the desired result.
3. Applications
3.1. Schwartz space
The Schwartz space S is defined to be the set of all Φ ∈ C ∞ for which the semi-
norm pN (Φ) is finite for all N ∈ N0 ≡ {0, 1, . . .}, where

pN (Φ) ≡ sup (1 + |x|)N |∂ α Φ(x)|.
x∈Rn
|α|≤N
The space S∞ is the set of all Φ ∈ S for which

xα Φ(x) dx = 0
Rn
for all α ∈ N0n .
The topological dual of S and S∞ are denoted by S and S∞
, respectively. The

elements in S and S∞ are called Schwarz distributions and Lizorkin distributions,
respectively. Equip S and S∞
with the weak-∗ topology.
Since S∞ is continuously embedded into S, the dual operator R, called the
restriction, is continuous from S to S∞

. We can generalize the following fact and
refine the proof:
Theorem 3.1 ([5, Theorem 6.18]). The restriction mapping R : F ∈ S → F |S∞

∈
S is open, namely the image R(U ) is open in S∞

for any open set U in S .
The statement can be found in [8], where Triebel applied this theorem to the
definition of homogeneous function spaces. Note also that Holschneider considered
Theorem 3.1 in the context of wavelet analysis in [3, Theorem 24.0.4], where he
applied a general result [9, Propositions 35.5 and 35.6] to this special setting. We
can find the proof of Theorem 3.1 in [15, Proposition 8.1]. But there is a gap
in Step 4, where the openness of R is proved using the closed graph theorem. It
seems that the closed graph theorem is not applicable to the space S . Our proof
reinforces Step 4 in the proof of [15, Proposition 8.1].
According to the proof of Theorem 1.1, there is no need to use the Fourier
transform.

3.2. The space Sm
We recall the definition of S /Pm , where Pm denotes the set of all polynomials
of degree less than or equal to m. Following Bourdaud [1], we denote by Sm the

orthogonal space of Pm in S and by Sm its topological dual. See [4, 12, 13, 16] for
applications to homogeneous function spaces defined in [10, 12, 15].
3.3. The Hasumi space Se

Let N ∈ N and α ∈ N0 n . Write temporarily ϕ(N ;α) (x) ≡ eN |x| ∂ α ϕ(x) (x ∈ Rn ) for
ϕ ∈ C ∞ . Define Se as follows:

Se ≡ ϕ ∈ C ∞ : ϕ(N ;α) ∈ L∞ .
N ∈N,α∈N0 n
The topological dual is denoted by Se is called the Hasumi space [2]. An analogy
to the spaces S and Sm

is available. We refer to [6] for function spaces contained

in Se .
3.4. The space D

A similar thing to Sections 3.1–3.3 applies to D . If we define

Dm ≡ ϕ ∈ D : xα ϕ(x) dx = 0 for all α ∈ N0 n with |α| ≤ m ,
Rn
then we have

Dm ∼ D /Pm .
Acknowledgement
The author is thankful to Professor Kunio Yoshino and the anonymous reviewer
for their pointing out the references [3, 9].
192 Y. Sawano
References
[1] G. Bourdaud, Realizations of homogeneous Sobolev spaces, Complex Var. Elliptic
Equ. 56 (2011), no. 10-11, 857–874.
[2] M. Hasumi, Note on the n-dimension tempered ultradistributions, Tôhoku Math. J.
13, (1961), 94–104.
[3] M. Holschneider, Wavelets: an anlysis tool, Oxford Science Publications, 1995.
[4] Y. Liang, Y. Sawano, T. Ullrich, D. Yang and W. Yuan, New characterizations of
Besov–Triebel–Lizorkin–Hausdorff spaces including coorbits and wavelets, J. Fourier
Anal. Appl. 18 (2012), no. 5, 1067–1111.
[5] S. Nakamura, T. Noi and Y. Sawano, Generalized Morrey spaces and trace operators,
Science China Mathematics 59, no. 2, 281–336.
[6] V.S. Rychkov, Littlewood–Paley theory and function spaces with Aloc p weights,
Math. Nachr. 224 (2001), 145–180.
[7] H.H. Schaefer, Topological vector spaces. Third printing corrected. Graduate Texts
in Mathematics, Vol. 3. Springer-Verlag, New York-Berlin, 1971. xi+294 pp.
[8] H. Triebel, Theory of Function Spaces. Birkhäuser, Basel, Boston, 1983.
[9] F. Treves, Topological Vector Spaces, Distributions and Kernels, Academic Press,
1967.
[10] D. Yang and W. Yuan, A new class of function spaces connecting Triebel–Lizorkin
spaces and Q spaces, J. Funct. Anal. 255 (2008), 2760–2809.
[11] D. Yang and W. Yuan, New Besov-type spaces and Triebel–Lizorkin-type spaces
including Q spaces, Math. Z. 265 (2010), 451–480.
[12] D. Yang and W. Yuan, Characterizations of Besov-type and Triebel–Lizorkin-type
spaces via maximal functions and local means, Nonlinear Analysis 73 (2010), 3805–
3820.
[13] D. Yang, W. Yuan and C. Zhuo, Fourier multipliers on Triebel–Lizorkin-type spaces,
J. Funct. Spaces Appl. 2012, Art. ID 431016, 37 pp.
[14] K. Yoshida, Functional Analysis, Sixth edition. Grundlehren der Mathematischen
Wissenschaften [Fundamental Principles of Mathematical Sciences], 123. Springer-
Verlag, Berlin-New York, 1980. xii+501 pp.
[15] W. Yuan, W. Sickel and D. Yang, Morrey and Campanato Meet Besov, Lizorkin and
Triebel, Lecture Notes in Mathematics, 2005, Springer-Verlag, Berlin, 2010, xi+281
pp.
[16] C. Zhuo, D. Yang and W. Yuan, Hausdorff Besov-type and Triebel–Lizorkin-type
spaces and their applications, J. Math. Anal. Appl. 412 (2014), no. 2, 998–1018.
Yoshihiro Sawano
Department of Mathematics and Information Science
Tokyo Metropolitan University
1-1 Minami-Ohsawa
Hachioji, Tokyo 192-0397, Japan
Operator Theory:

Ultradifferentiable Functions of Class Mpτ,σ

and Microlocal Regularity
Nenad Teofanov and Filip Tomić
Dedicated to Professor Pilipović on the occasion of his 65th birthday.
Abstract. We study spaces of ultradifferentiable functions which contain Gev-

rey classes. Although the corresponding defining sequences do not satisfy Ko-
matsu’s condition (M.2) , we prove appropriate continuity properties under
the action of (ultra)differentiable operators. Furthermore, we study conve-
nient localization procedure which leads to the concept of wave-front set with
respect to our regularity conditions. As an application, we identify singular
supports of suitable spaces of ultradifferentiable functions as standard pro-
jections of intersections/unions of wave-front sets.
1. Introduction
Gevrey classes were introduced in order to describe regularity properties of the fun-
damental solution, of the heat operator in [13], and thereafter used in the study
of different aspects of general theory of linear partial differential operators such as
hypoellipticity, local solvability and propagation of singularities. We refer to [3, 26]
for a detailed exposition of Gevrey classes and their applications to the theory of
linear partial differential operators. The intersection (projective limit) of Gevrey
classes is strictly larger than the space of analytic functions while its union (induc-
tive limit) is strictly contained in the class of smooth functions. Therefore, it is of
interest to study the intermediate spaces of smooth functions which are contained
in those gaps by introducing appropriate regularity conditions. On one hand, this
may serve to describe hypoellipticity properties between smooth/analytic hypoel-
lipticity and Gevrey hypoellipticity. On the other hand, such spaces can be used
in the study of corresponding microlocal regularity.
In this paper we continue and complement our research initiated by Professor
Stevan Pilipović and recently published in [22, 23], and show further properties of
classes of ultradifferentiable functions which contain Gevrey classes. Recall, in [22]
194 N. Teofanov and F. Tomić
σ
we introduced sequences Mpτ,σ = pτ p , p ∈ N, τ > 0 and σ > 1, and used them
to define and study test function spaces for Roumieu type ultradistribution. That
approach is further developed in [23] where, together with a more detailed analysis
of ultradifferentiable functions of class Mpτ,σ , we perform microlocal analysis with
respect to the regularity of such classes. More precisely, we proved there that
WF0,∞ (P (D)u) ⊆ WF0,∞ (u) ⊆ WF0,∞ (P (D)u) ∪ Char(P ),
where u is a Schwartz distribution, P (D) is a partial differential operator with

constant coefficients and characteristic set Char(P ), and WF0,∞ denotes the wave-
front set described in terms of new regularity conditions, see Section 4. By Corol-
lary 4.1 and (4.12), WF0,∞ (u) ⊂ WFA (u), where WFA denotes the analytic wave-
front set, see Subsection 1.2. For a recent result on “stronger” singularities and
the extension of WFA we refer to [24]
Different types of wave-front sets are introduced in connection to the equa-
tion under investigation. For example, the Gabor wave-front set from [16] and
[27] is recently successfully applied to different situations including the study of
Schrödinger equations, see [1, 2, 4, 5, 25, 28, 29, 30] and the references therein.
Such wave-front set can be characterized in terms of rapid decay of its Gabor coef-
ficients. That idea is introduced and exploited in [17, 20, 21], and extended in [6, 7]
to more general Banach and Fréchet spaces. The main tool used there are methods
of time-frequency analysis and modulation spaces. We refer to [8, 9, 10, 11, 14]
for details on modulation spaces and their role in time-frequency analysis, and
remark that a version of Gabor wave-front set adapted to regularity proposed in
this paper, will be the subject of our future investigation.
This section ends with fixing the notation and recalling the standard defini-
tion of ultradifferentiable functions and wave-front sets, and the reader familiar
to the subject may proceed to Section 2 which is devoted to the definition and
basic properties of regularity classes Eτ,σ (U ). In particular, we study their embed-
dings with respect to parameters τ > 0 and σ > 1 (Proposition 2.1), and show
the stability under differentiation (Theorem 2.1), although its defining sequence
Mpτ,σ = pτ p does not satisfy (M.2) (cf. Subsection 1.2). Furthermore, we study
σ
the continuity of certain ultradifferentiable operators (Theorem 2.2).

For the purpose of local analysis in Section 3 we introduce particular admis-
sibility condition for sequences of cut-off functions see Definition 3.1, and discuss
regularity of Schwartz distributions in Propositions 3.1 and 3.2.
In Section 4 we first recall the definition of the wave-front set from [23] and
prove Lemma 4.1, an important auxiliary result which is used in the proof of the
pseudolocal property in Subsection 4.1. We conclude the paper by identifying the
standard projections of intersections and unions of wave-front sets with singular
supports of appropriate projective/inductive limits of test function spaces, Theo-
rem 4.2.
We note that Propositions 3.1 and 3.2 and Lemma 4.1 are stated in [23]
without proofs.
Ultradifferentiable Functions of Class Mpτ,σ 195
1.1. Notation
Sets of numbers are denoted in a usual way, e.g., N (resp. Z+ ) denotes the set of
nonnegative ( resp. positive) integers. For x ∈ R+ the floor function is denoted by
1x2 := max{m ∈ N : m ≤ x}. For a multi-index α = (α1 , . . . , αd ) ∈ Nd we write
∂ α = ∂ α1 · · · ∂ αd and |α| = |α1 | + · · · + |αd |. We will often use Stirling’s formula:
√
N ! = N N e−N 2πN eθN /12N ,
for some 0 < θN < 1, N ∈ Z+ . By C ∞ (K) we denote the set of smooth functions
on a compact set K ⊂⊂ U with smooth boundary, where U ⊆ Rd is an open
∞
set, CK are smooth functions supported by K. The closure of U ⊂ Rd is denoted
by U . A conic neighborhood of ξ0 ∈ Rd \ 0 is an open cone Γ ⊂ Rd such that
ξ0 ∈ Γ. The Fourier transform of a locally integrable function f is defined as
f(ξ) = Rd f (x)e−2πixξ dx, ξ ∈ Rd , and the definition extends to distributions by
duality. Open ball of radius r, centered at x0 ∈ Rd is denoted by Br (x0 ).
For locally convex topological spaces X and Y , X → Y means that X is
dense in Y and that the identity mapping from X to Y is continuous, and we use
lim and lim to denote the projective and inductive limit topologies respectively. By
←− −→
X we denote the strong dual of X and by ·, ·X the dual pairing between X and
X . As usual, D (U ) stands for Schwartz distributions, and E (U ) for compactly
supported distributions.
1.2. Ultradifferentiable functions and wave-front sets

For the sake of the clarity of our exposition, in this subsection we recall Komatsu’s
approach to the theory of ultradifferentiable functions, see [18], and the notion of
wave-front set in the context of Gevrey regularity.
By Mp = (Mp )p∈N we denote a sequence of positive numbers such that the
following conditions hold:
(M.0) M0 = 1;
(M.1) Mp2 ≤ Mp−1 Mp+1 , p ∈ Z+ ;
(M.2) (∃C > 0) Mp+q ≤ C p+1
Mp Mq , p, q ∈ N;
∞
Mp−1
(M.3) < ∞.
p=1
Mp
Then Mp also satisfies weaker conditions: (M.1) Mp Mq ≤ Mp+q and (M.2)

(∃C > 0) Mp+q ≤ Cqp+1 Mp , p, q ∈ N.
Let the sequence Mp satisfy the conditions (M.0)–(M.3) and let U ⊆ Rd
be an open set. A function φ ∈ C ∞ (U ) is an ultradifferentiable function of class
(Mp ) (resp. of class {Mp }) if for each compact subset K ⊂⊂ U and each h > 0,
there exists C > 0 (resp. for each compact subset K ⊂⊂ U there exists h > 0 and
C > 0) such that
sup |∂ α φ(x)| ≤ Ch|α| M|α| , α ∈ Nd . (1.1)
x∈K
For a fixed compact set K ⊂ Rd and h > 0, φ ∈ E {Mp },h (K) if φ ∈ C ∞ (K)
and if (1.1) holds for some C > 0. If φ ∈ C ∞ (Rd ) and supp φ ⊂ K, then φ ∈
{M },h
DK p . These spaces are Banach spaces under the norm
|∂ α φ(x)|
φ E {Mp },h (K) = sup .
α∈Nd ,x∈K h|α| M|α|
The spaces of ultradifferentiable functions of class {Mp } and of class (Mp )
are respectively given by

E {Mp } (U ) = lim lim E {Mp },h (K) = E {Mp },h (K),
←− −→
K⊂⊂U h→∞ K⊂⊂U h→∞

{Mp },h
E (Mp )
(U ) = lim lim E (K) = E {Mp },h (K),
←− ←−
K⊂⊂U h→0 K⊂⊂U h→0
and their strong duals are respectively called the space of ultradistributions of
Roumieu type of class Mp and the space of ultradistributions of Beurling type of
class Mp .
The space of ultradifferentiable functions of class {Mp } (resp. of class (Mp ))
with support in K is given by
{M },h
{M },h
D{Mp } (U ) = lim lim DK p = DK p
−→ −→
K⊂⊂U h→∞ K⊂⊂U h→∞

{Mp },h
{M },h
resp. D (Mp )
(U ) = lim lim DK = DK p
−→ ←−
K⊂⊂U h→0 K⊂⊂U h→0
and its strong dual is the space of compactly supported ultradistributions of

Roumieu type of class Mp (resp. of Beurling type of class Mp ).
In what follows, E ∗ (U ) and D∗ (U ) stand for E {Mp } (U ) or E (Mp ) (U ), and for
{Mp }
D (U ) or D(Mp ) (U ), respectively.
t
In particular, if Mp is the Gevrey sequence, Mp = p!t , t > 1, then E {p! } (U )
t
and E (p! ) (U ) are the Gevrey classes of ultradifferentiable functions commonly
denoted by Et (U ). Note that p!t , t > 1, satisfies (M.0)–(M.3) . We refer to [18] for
a detailed study of different classes of ultradifferentiable functions and their duals.
Next we recall the notion of a wave-front set in the context of the Gevrey
regularity.
Let there be given t ≥ 1 and (x0 , ξ0 ) ∈ U × Rd \{0}. Then the Gevrey wave-
front set W Ft (u) of u ∈ D (U ) can be defined as follows: (x0 , ξ0 ) ∈ W Ft (u) if and
only if there exists an open neighborhood Ω of x0 , a conic neighborhood Γ of ξ0
and a bounded sequence uN ∈ E (U ), such that uN = u on Ω and
h N N !t
uN (ξ)| ≤ A
| , N ∈ Z+ , ξ ∈ Γ,
|ξ|N
for some A, h > 0. If t = 1, then the Gevrey wave-front set is sometimes called
the analytic wave-front set and denoted by W FA (u), u ∈ D (U ). The classical
C ∞ wave-front set of u ∈ D (U ) can be also defined through its complement:
(x0 , ξ0 ) ∈ W F (u) if and only if there exists an open neighborhood Ω of x0 , a conic

neighborhood Γ of ξ0 and a smooth compactly supported function φ, equal to 1
on Ω and
CN
|φu(ξ)| ≤ , N ∈ Z+ , ξ ∈ Γ, CN > 0.
(1 + |ξ|)N
We refer to [15, 26, 12] for details.
2. Mpτ,σ sequences and the corresponding regularity classes

σ
In this section we observe the sequence Mpτ,σ = pτ p , p ∈ N, where τ > 0, σ > 1
and study its basic properties. Although Mpτ,σ fails to satisfy the condition (M.2),
the flexibility obtained by introducing the two-parameter dependence enables us
to define ultradifferentiable functions which are less regular than Gevrey functions,
see Subsection 2.2.”
2.1. The defining sequence Mpτ,σ

σ
The following lemma captures the basic properties of the sequence Mpτ,σ = pτ p ,
p ∈ N, τ > 0, σ > 1, M0τ,σ = 1. We refer to [22] for the proof.
σ
Lemma 2.1. Let τ > 0, σ > 1 and Mpτ,σ = pτ p , p ∈ Z+ , M0τ,σ = 1. Then, apart
from (M.1) and (M.3) the sequence Mpτ,σ satisfies the following properties.
M τ,σ ≤ C pσ M τ,σ , for some sequence Cq ≥ 1, p, q ∈ N,

(M.2) p+q q p
M τ,σ ≤ C pσ +qσ M τ 2σ−1 ,σ M τ 2σ−1 ,σ , p, q ∈ N, for some C > 0.

(M.2) p+q p q
Furthermore, there exist A, B, C > 0 such that

σ
Mpτ,σ ≤ AC p 1pσ 2!τ /σ and 1pσ 2!τ /σ ≤ BMpτ,σ . (2.1)
The property (M.3) implies that the corresponding spaces of ultradiffer-
entiable functions (see Subsection 2.2) are non-quasianalytic, i.e., they contain
nontrivial compactly supported smooth functions. Let us now prove (M.3) by
modifying the proof given in [22]. Since the first summand in the series given
in (M.3) is equal to 1, it is enough to observe the summation for p ≥ 2. Since
pσ ≥ (p − 1)σ−1 p = (p − 1)σ + (p − 1)σ−1 , p ∈ Z+ , we have
∞
∞
∞

σ
(p − 1)τ (p−1)
σ
(p − 1)τ (p−1)
σ
(1 − p1 )τ (p−1)
≤ = .
p=2
pτ pσ p=2
pτ ((p−1)σ +(p−1)σ−1 ) p=2
pτ (p−1)σ−1
1 1 p
Since τ pσ ln 1 + = τ pσ−1 ln 1 + ≥ τ pσ−1 ln 2, p ≥ 2, we obtain
p p
σ−1
1 τ p
σ
2τ p ≤ 1+ , p ≥ 2,
p
which gives
∞

σ
∞

(1 − p1 )τ (p−1) 1
τ (p−1)σ−1
≤ < ∞.
p=2
p p=2
(2p)τ (p−1)σ−1
2.2. Regularity classes Eτ,σ

Let τ > 0, σ > 1, h > 0, and K ⊂⊂ U , where U is an open set in Rd . A function
φ ∈ C ∞ (U ) is ultradifferentiable function of class Mpτ,σ if there exists A > 0 such
that
|∂ α φ(x)| ≤ Ah|α| |α|τ |α| , α ∈ Nd , x ∈ K.
σ σ
The space of ultradifferentiable functions of class Mpτ,σ denoted by Eτ,σ,h (K) is a

Banach space with the norm given by
|∂ α φ(x)|
φ Eτ,σ,h (K) = sup sup |α|σ τ |α|σ , (2.2)
α∈Nd x∈K h |α|
and Eτ1 ,σ1 ,h1 (K) → Eτ2 ,σ2 ,h2 (K), 0 < h1 ≤ h2 , 0 < τ1 ≤ τ2 , 1 < σ1 ≤ σ2 . By (2.1)
the norm in (2.2) is equivalent to
∼ |∂ α φ(x)|
φ Eτ,σ,h (K) = sup sup < ∞, h > 0.
α∈Nd x∈K h|α|σ 1|α|σ 2!τ /σ
Let Dτ,σ,h
K
be the set of functions in Eτ,σ,h (K) with support contained in K.
Then, in the topological sense, we set
E{τ,σ} (U ) = lim lim Eτ,σ,h (K),
←− −→
K⊂⊂U h→∞
E(τ,σ) (U ) = lim lim Eτ,σ,h (K),
←− ←−
K⊂⊂U h→0
D{τ,σ} (U ) = lim D{τ,σ}

K
= lim lim Dτ,σ,h
K
,
−→ −→ −→
K⊂⊂U K⊂⊂U h→∞
D(τ,σ) (U ) = lim D(τ,σ)

K
= lim lim Dτ,σ,h
K
.
−→ −→ ←−
K⊂⊂U K⊂⊂U h→0
We will use abbreviated notation τ, σ for {τ, σ} or (τ, σ).

Remark 2.1. If τ > 1 and σ = 1, then Eτ,1 (U ) = Eτ (U ) are the Gevrey classes
and Dτ,1 (U ) = Dτ (U ) are the corresponding subspaces of compactly supported
functions in Eτ (U ). When 0 < τ ≤ 1 and σ = 1 such spaces are contained in the
corresponding spaces of quasianalytic functions. In particular, Dτ (U ) = {0} when
0 < τ ≤ 1.
Clearly, compactly supported Gevrey functions belong to D{τ,σ} (U ). How-
ever, one can find a compactly supported function in D{τ,σ} (U ) which is not in
Dτ (U ), for any τ > 1. We refer to [22] for the proofs.
It is known that the spaces E {Mp } (U ) are nuclear if the defining sequence Mp
satisfies (M.2) , cf. [18, Theorem 2.6 ]. Although Mpτ,σ = pτ p , τ > 0, σ > 1, does
σ
not satisfy (M.2) , it can be proved that the spaces Eτ,σ (U ), Dτ,σ
K
and Dτ,σ (U ) are
nuclear as well. Again, we refer to [22] for the proof.
The basic embeddings between the introduced spaces with respect to σ and
τ are given in the following proposition.
Proposition 2.1. Let σ1 ≥ 1. Then for every σ2 > σ1 and τ > 0

lim Eτ,σ1 (U ) → lim Eτ,σ2 (U ). (2.3)
−→
τ →∞
←−+
τ →0
Moreover, if 0 < τ1 < τ2 , then for every σ ≥ 1 it holds

E{τ1 ,σ} (U ) → E(τ2 ,σ) (U ) → E{τ2 ,σ} (U ), (2.4)
and
lim E{τ,σ} (U ) = lim E(τ,σ) (U ), lim E{τ,σ} (U ) = lim E(τ,σ) (U ).
−→ −→ ←− ←−+
τ →∞ τ →∞ τ →0+ τ →0
For the proof of (2.3) we refer to [22] and the complete proof of Proposition
2.1 can be found in [23].
We are also interested in projective (when τ → 0+ or when σ → 1+ ) and
inductive (when τ → ∞ or when σ → ∞) limits which we denote as follows:
E0,σ (U ) := lim Eτ,σ (U ), E∞,σ (U ) := lim Eτ,σ (U ),
←−+ −→
τ →0 τ →∞
Eτ,1 (U ) := lim Eτ,σ (U ), Eτ,∞ (U ) := lim Eτ,σ (U ),
←−+ −→
σ→∞
σ→1
E0,1 (U ) := lim E0,σ (U ), E0,∞ (U ) := lim E0,σ (U ), (2.5)
←−+ −→
σ→1 σ→∞
E∞,1 (U ) := lim E∞,σ (U ), E∞,∞ (U ) := lim E∞,σ (U ). (2.6)
←−+ −→
σ→∞
σ→1
The proof of the following corollary can be found in [23].
Corollary 2.1. With the notation from (2.5) and (2.6) the following strict embed-
dings hold true:
lim Et (U ) → E0,1 (U ) → E∞,1 (U ) → E0,∞ (U ) → E∞,∞ (U ).
t→∞
−→
Recall that the Komatsu’s condition (M.2) , also known as “stability under
differential operators”, is sufficient to ensure that the spaces E ∗ (U ) are closed
under the differentiation, cf. [18, Theorem 2.10]. In the next proposition we show
that Eτ,σ (U ) is closed under the finite-order differentiation, although the condition
(M.2) is violated.
Theorem 2.1. Let U be open in Rd , and let τ > 0 and σ > 1. Then the space
Eτ,σ (U ) is closed under pointwise multiplications and finite-order differentiation.
Proof. We leave to the reader to prove that the spaces are closed under translations
and dilations and show the algebra property first.
σ−1
Let K ⊂⊂ Rd and for h > 0 set ch = min{h, h2 }. Then for φ, ψ ∈
Eτ,σ,ch (K), by the Leibnitz formula we obtain
α ch
|α−β|σ |β|σ
ch |α − β|τ |α−β| |β|τ |β|
σ σ
||φψ||Eτ,σ,2h (K) ≤ sup

α∈Nd β≤α β (2h)|α|σ |α|τ |α|σ
· ||φ||Eτ,σ,ch (K) ||ψ||Eτ,σ,ch (K) .
If h ≥ 1, then we put ch = h. Note that |α − β|σ + |β|σ ≤ |α|σ when β ≤ α.
By (M.1) property of Mpτ,σ we then have
|α−β|σ |β|σ
α ch ch |α − β|τ |α−β| |β|τ |β|
σ σ
2|α| h|α|
σ
≤ ≤ 1, α ∈ Nd .
β (2h)|α|σ |α|τ |α|σ (2h)|α|σ
β≤α
σ−1
If 0 < h < 1, then ch = h2 , and
(1/h)|α| ≤ (1/h)2 |α−β|σ |β|σ
σ σ−1 σ−1
(1/h)2 , β ≤ α.
which gives
α ch
|α−β|σ |β|σ
ch |α − β|τ |α−β| |β|τ |β|
σ σ
2|α|
≤ ≤ 1, α ∈ Nd ,
β≤α
β (2h)|α|σ |α|τ |α|σ 2|α|σ
that is
φψ Eτ,σ,2h (K) ≤ φ Eτ,σ,ch (K) ψ Eτ,σ,ch (K) ,
wherefrom the algebra property holds.

To prove that Eτ,σ (U ) is closed under differentiation we fix β ∈ Nd , and set
σ−1
c = max{h, h2 }, h > 0. Then, for every x ∈ K, from (M.2) it follows that
h
|(∂ α+β φ(x))| ≤ ||φ||Eτ,σ,h (K) h|α+β| |α + β|τ |α+β|

σ σ
|β|σ
(C|β| ch )|α| |α|τ |α| ,
σ σ
≤ ||φ||Eτ,σ,h (K) Ch
where Ch = max{1, h2 } and C|β| is the constant from (M.2)

σ−1
(see Lemma 2.1
for q = |β|). This implies that for every h > 0 there exists Ch,β > 0 such that
||∂ β φ||Eτ,σ,C (K) ≤ Ch,β ||φ||Eτ,σ,h (K) ,
|β| ch
and the statement follows.
We conclude this section with the continuity properties of certain ultradif-

ferentiable operators P (x, ∂) acting on Eτ,σ (U ). Note that Komatsu’s condition
(M.2) provides the stability of E ∗ (U ) under the action of ultradifferentiable opera-

tors, cf. [18, Theorem 2.12]. The following theorem shows that the condition (M.2)
provides only the continuity of certain ultradifferentiable operators from Eτ,σ (U )
into Eτ 2σ−1 ,σ (U ). We refer to [23, Theorem 2.1] for a more general result which
involves ultradifferentiable operators with non constant coefficients.
$
∞
Theorem 2.2. Let U be open in Rd , τ > 0 and σ > 1. If P (∂) = aα ∂ α is a
|α|=0
constant coefficient differential operator of infinite order such that for some L > 0
and A > 0 (resp. every L > 0 there exists A > 0) such that
L|α|
σ
|aα | ≤ A , (2.7)
|α|τ 2σ−1 |α|
σ
then E∞,σ (U ) is closed under action of P (∂). In particular,
P (∂) : Eτ,σ (U ) −→ Eτ 2σ−1 ,σ (U ) ,

is continuous linear mapping, where Eτ,σ (U ) denotes E(τ,σ) (U ) or E{τ,σ} (U ).

Proof. Let φ ∈ Eτ,σ,h (K), for some h > 0. Then, for x ∈ K, using (2.7) and (M.2)
property of Mpτ,σ we obtain
L|α|
σ
h|α+β| (|α + β|)τ |α+β|

σ σ
|∂ β (aα ∂ α φ(x))| ≤ A||φ||Eτ,σ,h (K)
|α|τ 2σ−1 |α|
σ
L|α|
σ
h|α+β| C |α| C |β| |α|τ 2 |α|σ |β|σ

σ σ σ σ−1 σ−1
≤ A||φ||Eτ,σ,h (K) |β|τ 2
|α|τ 2σ−1 |α|
σ
≤ A||φ||Eτ,σ,h (K) (LCch )|α| (Cch )|β| |β|τ 2 |β|σ

σ σ σ−1
, (2.8)
where for the last inequality we have used that for σ > 1,
|α|σ + |β|σ ≤ |α + β|σ ≤ 2σ−1 (|α|σ + |β|σ ),
σ−1
Note that ch = h when
ch = max{h, h2 } and C > 1 is the constant from (M.2).
2σ−1
0 < h ≤ 1 and ch = h when h > 1.
Now, we may choose either h > 0 or L > 0 such that LCch < 1/2 holds true.
∞

(1/2)|α| < ∞, by taking the sum with respect to α and the supremum
σ
Since
|α|=0
with respect to β and x ∈ K, from (2.8) it follows that
||P (∂)φ||Eτ 2σ−1 ,σ,Cc (K) ≤ C ||φ||Eτ,σ,h (K) ,
h
for some C > 0 and the theorem is proved, since the result for E∞,σ (U ) follows
immediatelly.
3. Local analysis of distributions with respect to Eτ,σ

In this section we study local behavior of distributions through appropriate local-
ized versions of their Fourier transforms. The localization is defined by the means
of τ, σ-admissible sequences of smooth functions, see Definition 3.1. We first de-
scribe the process of enumeration.
Let τ > 0, σ > 1, Ω ⊆ K ⊂⊂ U ⊆ Rd , where Ω and U are open in Rd , and

the closure of Ω is contained in K. Let u ∈ D (U ). We observe the nature of its
regularity with respect to the condition
hN N !τ /σ
|
uN (ξ)| ≤ A , N ∈ N, ξ ∈ Rd \{0}, (3.1)
|ξ|N 1/σ
where {uN }N ∈N is bounded sequence in E (U ) such that uN = u in Ω and A, h
are some positive constants.
One of the main ingredients of the following proofs is the procedure which
we call enumeration and which consists of a change of variables in indices which
“speeds up” or “slows down” the decay estimates of single members of the cor-
responding sequences, while preserving their asymptotic behavior when N → ∞.
In other words, although estimates for terms of a sequence before and after enu-
meration are different, the asymptotic behavior of the whole sequence remains
unchanged. Therefore, the condition (3.1) is equivalent to another condition ob-
tained after replacing N with positive, increasing sequence aN such that aN → ∞,
N → ∞. We then write N → aN and uN instead of uaN .
For example, applying Stirling’s formula to (3.1) we obtain
τ
hN1 N
σN
|
uN (ξ)| ≤ A1 , N ∈ N, ξ ∈ Rd \{0}, (3.2)
|ξ|N 1/σ
for some positive constants A1 , h1 . After enumeration N → N/τ and writing uN
instead of the uN/τ , (3.2) becomes
N/τ τ
h1 (N/τ ) σ (N/τ ) hN2 N!
1/σ
|
uN (ξ)| ≤ A1 1/σ
≤ A2 , N ∈ N, ξ ∈ Rd \{0}
|ξ|(N/τ ) |ξ|(N/τ )1/σ
for some A2 , h2 > 0. Moreover, if {uN }N ∈N is bounded sequence in E (U ), then
{uN/τ }N ∈N is also bounded in E (U ) (with respect to the strong topology).
In [15, Proposition 8.4.2] Hörmander used a sequence of carefully chosen
cutoff functions {χN }N ∈N to define the analytic wave-front set W FA . We modify
that construction to define and analyze a new type of wave-front sets in D (U )
related to (3.1) or (3.2).
Definition 3.1. Let τ > 0, σ > 1, and Ω ⊆ K ⊂⊂ U , such that Ω is strictly
∞
contained in K. A sequence {χN }N ∈N of functions in CK is said to be τ, σ-
admissible with respect to K if
a) χN = 1 in a neighborhood of Ω, for every N ∈ N,
b) there exists a positive sequence Cβ such that
|α|+1
sup |Dα+β χN (x)| ≤ Cβ 1N 1/σ 2|α| , |α| ≤ 1(N/τ )1/σ 2, (3.3)
x∈K
for every N ∈ N and β ∈ Nd .

Remark 3.1. When τ = σ = 1 we recover the sequence {χN }N ∈N used by
Hörmander to analyze the analytic behavior of distributions. Moreover, note that
for σ > 1 and 0 < τ ≤ 1, {χτ N σ }N ∈N gives another sequence with the same
asymptotic properties as {χN }N ∈N . This implies that, for σ > 1 and 0 < τ ≤ 1,
the enumeration N → τ N σ in (3.3) may be used to define the analytic wave-front
sets.
Remark 3.2. From (3.3) it follows that

|α|+1
|
χN (ξ)| ≤ Aβ 1N 1/σ 2|α| ξ−|α|−|β| , |α| ≤ 1(N/τ )1/σ 2, (3.4)
for every N ∈ N, ξ ∈ Rd , where ξ = (1+|ξ|2 )1/2 . If α = 0 in (3.3), then {χN }N ∈N

is bounded sequence in C ∞ (U ) and from (3.4), it follows that {χN }N ∈N is bounded
in the Schwartz space S(Rd ). From the boundedness of {χN }N ∈N in C ∞ (U ), it
follows that {χN u}N ∈N is bounded in E (U ) for every u ∈ D (U ).
Recall that if {uN }N ∈N is a bounded sequence in E (U ) then Paley–Wiener
type theorems and e−ix·ξ ∈ C ∞ (Rdx ), for every ξ ∈ Rd , imply
uN (ξ)| = |uN , e−i·ξ | ≤ CξM ,
| (3.5)
for some C, M > 0 independent of N .
The existence of the τ, σ-admissible sequences is given by the following lemma.

We refer to [23] for the proof, see also [15, Theorems 1.3.5 and 1.4.2].
Lemma 3.1. Let there be given r > 0, τ > 0, σ > 1 and x0 ∈ Rd . There exists
τ, σ-admissible sequence {χN }N ∈N with respect to B2r (x0 ) such that χN = 1 on
Br (x0 ), for every N ∈ N.
Next we show that (3.1) implies local regularity related to E{τ,σ} (U ). For the
opposite direction, if u ∈ E{τ,σ} (Ω) we need to observe τ̃ , σ-admissible sequences,
where τ̃ = τ σ/(σ−1) . The precise statements are the following.
Proposition 3.1. Let u ∈ D (U ), τ > 0, σ > 1, Ω ⊆ U with the closure contained

in U and let {uN }N ∈N be a bounded sequence in E (U ), uN = u on Ω and such
that (3.1) holds. Then u ∈ E{τ,σ} (Ω).
Proof. After the enumeration N → N σ and by Lemma 2.1, condition (3.1) is

equivalent to
σ σ
kN N τ N
uN (ξ)| ≤ A
| , N ∈ N, ξ ∈ Rd \{0}. (3.6)
|ξ|N
for some A, k > 0.
By the Fourier inversion formula and the fact that uN = u in Ω we obtain

(h|α| |α|τ |α| )−1 |Dα u(x)| = (h|α| |α|τ |α| )−1 N (ξ)e2πixξ dξ
σ σ σ σ
+ ξα u
|ξ|≤1 |ξ|>1
≤ I1 + I2 , N ∈ N, α ∈ N , x ∈ Ω,
d
where h > 0 will be chosen later on. Using (3.5) we estimate I1 by

I1 = (h|α| |α|τ |α| )−1
σ σ
N (ξ)e2πixξ dξ
ξα u
|ξ|≤1

σ σ
≤ C(h|α| |α|τ |α| )−1 ξM dξ.
|ξ|≤1
If h ≥ 1 we conclude that I1 ≤ C1 where C1 does not depend on α. To estimate

I2 , note that by (3.6) we have

I2 = (h|α| |α|τ |α| )−1
σ σ
N (ξ)e2πixξ dξ
ξαu
|ξ|>1

σ σ σ σ σ−1 σ
≤ A(h|α| |α|τ |α| )−1 k N N τ N |ξ||α|−N dξ ≤ C(k 2 /h)|α| ,
|ξ|>1
property
where for the last inequality we chose N = |α| + d + 1, and use (M.2)
2σ−1
of Mp , p ∈ N. Now, for h > k
τ,σ
we conclude that I2 ≤ C2 , and C2 does not
σ−1
depend on α. Hence, if we take h > max{1, k 2 }, we conclude that u ∈ E{τ,σ} (Ω),
and the statement is proved.
Proposition 3.2. Let Ω ⊆ K ⊂⊂ U , Ω ⊂ K, u ∈ D (U ), and let {χN }N ∈N be the

τ̃ , σ-admissible sequence with respect to K, where τ̃ = τ σ/(σ−1) , τ > 0, σ > 1. If
u ∈ E{τ,σ} (Ω), then {χN u}N ∈N is bounded in E (U ), χN u = u on Ω, and
−1/σ
hN N !τ̃ /σ
N u(ξ)| ≤ A
|χ= 1/σ
, N ∈ N, ξ ∈ Rd \{0}. (3.7)
|ξ|(N/τ̃ )
That is, after enumeration N → τ̃ N , {χN u}N ∈N satisfies (3.1). for some A, h > 0.
Proof. Put uN = χN u, N ∈ N. By the Remark 3.2, {uN }N ∈N is bounded in
E (U ). Note also that uN = u on Ω and supp uN ⊆ K.
Since u ∈ E{τ,σ} (Ω), from (3.3) for |α| ≤ 1(N/τ̃ )1/σ 2, x ∈ Ω, and for some
k > 1 we obtain
α
|Dα uN (x)| ≤ |Dα−β χN (x)||Dβ u(x)|
β
β≤α
α |α−β| |β|σ
A|α−β|+1 1N 1/σ 2 |β|τ |β|
σ
≤ ||u||Eτ,σ,k (Ω) k
β
β≤α
1/σ 1/σ
≤ A||u||Eτ,σ,k (Ω) (2A)(N/τ̃ )
1N 1/σ 2(N/τ̃ ) N/τ̃ τN
k 1N 1/σ 2 τ̃
1 1/(σ−1)
1
N 1/σ 1 1 1/(σ−1)
≤ A||u||Eτ,σ,k (Ω) B N N σ ( τ ) N σ(τ ) N
(3.8)
for some B > 0, where for the last inequality we have used that τ̃ = τ σ/(σ−1) .
Next we note that there exists c > 0, such that
N 1/σ ln N ≤ cN 1/σ N 1−1/σ = cN,
1 1 1/(σ−1) 1/σ
wherefrom N σ ( τ ) N
≤ C N for some C > 1 (which depends on τ and σ).
Hence (3.8) can be estimated by
1 1 1/(σ−1)
|Dα uN (x)| ≤ A||u||Eτ,σ,k (Ω) hN N σ ( τ ) N
, (3.9)
for some h > 0. Applying the Fourier transform to (3.9) for |α| = 1(N/τ̃ )1/σ 2 we
obtain
1 1 1/(σ−1)
hN N σ ( τ ) N
|
uN (ξ)| ≤ A||u||Eτ,σ,k (Ω) , N ∈ N, ξ ∈ Rd \{0}. (3.10)
|ξ|(N/τ̃ )1/σ
Finally, after the enumeration N → τ̃ N , we note that (3.10) and Stirling’s formula
imply (3.7), and the proposition is proved.
Remark 3.3. Sequences of functions {ϕN }N ∈N “analytic up to the order N ” are
used to extended results from [15] to Gevrey type ultradistributions, cf. [26, Propo-
sition 1.4.10, Corollary 1.4.11]. When τ > 0, σ > 1 and β = 0 in (3.3) we obtain
1N 1/σ 2|α| 1 N r/σ 1
sup |∂ α χN (x)| ≤ A|α|+1 1 |α| σ |α| ≤ A|α|+1 sup |α| σ |α|
x∈K |α| σ |α| r>0 rr/σ
1 1
|α|+1 σ |α|
=A e eσ N |α| , |α| ≤ 1(N/τ )1/σ 2,
so χN might be called “quasi-analytic up to the order 1(N/τ )1/σ 2”. When σ → 1+
the order of quasi-analyticity of χN tends to infinity (for fixed N ∈ N) for 0 < τ <
1, while for τ > 1 it tends to zero. Therefore the study of the “critical” behavior
when σ → 1+ is possible only if τ depends on σ.
In particular, when σ = 1 and τ = 1, the proof of Proposition 3.2 fails, while
for τ = σ = 1, Proposition 3.2 coincides with necessity part of [15, Proposition
8.4.2.].
4. Wave-front sets with respect to Eτ,σ

In this section we define wave-front sets which detect singularities that are
“stronger” then classical C ∞ singularities and “weaker” then Gevrey type sin-
gularities, which is done within the framework of the regularity classes Eτ,σ (U ).
Definition 4.1. Let τ > 0 and σ > 1, u ∈ D (U ), and (x0 , ξ0 ) ∈ U × Rd \{0}. Then
(x0 , ξ0 ) ∈ WF{τ,σ} (u) (resp. WF(τ,σ) (u)) if there exists open neighborhood Ω ⊂ U
of x0 , a conic neighborhood Γ of ξ0 , and a bounded sequence {uN }N ∈N in E (U )
such that uN = u on Ω and (3.1) holds for some constants A, h > 0 (resp. for every
h > 0 there exists A > 0).
Note that WF{τ,σ} (u), u ∈ D (U ), is a closed subset of U × Rd \{0} and for
τ > 0 and σ > 1 we have
WF{τ,σ} (u) ⊆ WFσ (u) ⊆ WF{1,1} (u) = WFA (u),
where WFσ (u) is the Gevrey wave-front set. If 0 < τ < 1 and σ = 1 then
WFA (u) ⊆ WF{τ,1} (u).
Next we prove an important fact on microlocal regularity.

Lemma 4.1. Let τ > 0, σ > 1, u ∈ D (U ), K ⊂⊂ U , and let {χN }N ∈N be a
τ̃ , σ-admissible sequence with respect to K with τ̃ = τ σ/(σ−1) . Then {χN u}N ∈N
is a bounded sequence in E (U ), and if WF{τ,σ} (u) ∩ (K × F ) = ∅, where F is a
closed cone, then there exist A, h > 0 such that
−1/σ
hN N !τ̃ /σ
|χ=
N u(ξ)| ≤ A , N ∈ N,ξ ∈ F . (4.1)
|ξ|(N/τ̃ )1/σ
Proof. Let (x0 , ξ0 ) ∈ K × F , and set r0 := rx0 ,ξ0 > 0. Furthermore, let {χN }N ∈N
be the τ̃ , σ-admissible sequence with respect to Br0 (x0 ), Br0 (x0 ) ⊆ Ω ⊆ K. Bound-
edness of {χN u}N ∈N follows by Remark 3.2.
Since (x0 , ξ0 ) ∈ WF{τ,σ} (u) we choose uN , Ω and Γ as in Definition 4.1 so
that
hN N !τ /σ
|
uN (ξ)| ≤ A N 1/σ , N ∈ N, ξ ∈ Γ, (4.2)
|ξ|
for some A, h > 0. Recall, the condition (4.2) is equivalent to
−1/σ
hN N !τ̃ /σ
uN (ξ)| ≤ A
| 1/σ
, N ∈ N, ξ ∈ Γ, (4.3)
|ξ|(N/τ̃ )
after applying Stirling’s formula and enumeration N → N/τ̃ .
Let Γ0 be an open conical neighborhood of ξ0 with the closure contained in
Γ and choose ε > 0 such that ξ − η ∈ Γ when ξ ∈ Γ0 and |η| < ε|ξ|. Then, since
χN u = χN uN , we write

χ=N u(ξ) = + χ uN (ξ − η) dη = I1 + I2 , ξ ∈ Γ0 , N ∈ N .
N (η)
|η|<ε|ξ| |η|≥ε|ξ|
To estimate I1 , note that for |η| < ε|ξ| we have

|ξ − η| ≥ |ξ| − |η| > (1 − ε)|ξ|.
Thus, by using (3.4) for α = 0 and |β| = d + 1 and (4.3), we obtain
−1/σ
hN N !τ̃ /σ

|I1 | = N (η)
χ uN (ξ − η) dη ≤ |
χN (η)|A dη
|η|<ε|ξ| |η|<ε|ξ| |ξ − η|(N/τ̃ )1/σ
−1/σ τ̃ −1/σ /σ
hN N !τ̃ /σ
−d−1 hN
1 N!
≤A η dη ≤ A 1 , ξ ∈ Γ0 , N ∈ N.
((1 − ε)|ξ|)(N/τ̃ )1/σ Rd |ξ|(N/τ̃ )1/σ
To estimate I2 , note that for |η| ≥ ε|ξ| we have
|ξ − η| ≤ |ξ| + |η| ≤ (1 + 1/ε)|η|,
and thus, using (3.4) for |α| = 1(N/τ̃ )1/σ 2, together with (3.5) and
(N/τ )1/σ 1/σ −1/σ
1N 1/σ 2 ≤ N 1/σ(1/τ ) N
≤ C N N !τ /σ
,
for every β ∈ Nd and some M > 0 we obtain

|I2 | = uN (ξ − η) dη
N (η)
χ
|η|≥ε|ξ|
(N/τ̃ )1/σ +1 (N/τ̃ )1/σ
Aβ 1N 1/σ 2
≤ η−|β| Cξ − ηM dη
(ε|ξ|)(N/τ̃ )1/σ |η|≥ε|ξ|
(N/τ̃ )1/σ
AN +1
1N 1/σ 2
η−|β| (1 + 1/ε)ηM , dη
β
≤
(ε|ξ|)(N/τ̃ )1/σ Rd
−1/σ
AN +1 N !τ̃ /σ
≤ 1/σ
, ξ ∈ Γ0 ,
|ξ|(N/τ̃ )
for some A > 0, where we have chosen |β| = M + d + 1.
Thus, the statement follows for (x, ξ) ∈ Br0 (x0 ) × Γ0 .
In order to extend the result to K × F we use the same idea as in the proof
of [15, Lemma 8.4.4]. Since the intersection of F with the unit sphere is a compact
set, there exists a finite number of balls Brx0 ,ξj (x0 ), and cones Γj that covers F ,
j ≤ n, n ∈ Z+ , and note that (4.1) remains valid if {χN }N ∈N is chosen so that
n
supp χN ⊆ Brx0 := Brx0 ,ξj (x0 ).
j=1
Moreover, since K is compact set, it is covered by a finite number of balls
Brxk , k ≤ n, n ∈ Z+ . By [18, Lemma 5.1.] there exist non-negative functions
$
n
χk ∈ C0∞ (Brxk /2 ), k ≤ n, such that χk = 1 on a neighborhood of K. Next,
k=1
for every N ∈ N we choose a non-negative function φN ∈ C0∞ (Brxk /2 ) such that

φN (x) = 1 and
sup |Dα φN (x)| ≤ C |α| 1N 1/σ 2|α| ,
x∈K
for |α| ≤ 1(N/τ̃ )1/σ 2, where the constant C > 0 depends on τ and σ, cf. [15,
$
n
Theorem 1.4.2.]. Now, for χN,k = φN ∗ χk , we have χN,k = 1 in a neighborhood
k=1
of K, and each χN,k , 1 ≤ k ≤ n, satisfies (3.3).
To conclude the proof we note that if {χN }N ∈N is a τ̃ , σ-admissible sequence
with respect to K, then χN χN,k also satisfies estimate of type (3.3), for 1 ≤ k ≤ n.
This follows by simple application of Leibniz rule. Thus, (4.1) holds if we replace
$n
χN by χN χN,k . Since χN χN,k = χN , the result follows.
k=1
Next we give a short comment on WF(τ,σ) (u), u ∈ D (U ). From our analysis

it follows that the regularity related to the complement of WF{τ,σ} is described
by the (microlocal) regularity of E{τ,σ} . Therefore the following corollary follows
from the embeddings given by (2.4).
Corollary 4.1. Let u ∈ D (U ), t > 1. Then for 0 < τ < ρ and σ > 1 it holds

WF(u) ⊆ WF{ρ,σ} (u) ⊆ WF(ρ,σ) (u) ⊆ WF{τ,σ} (u) ⊆ WFt (u) ⊆ WFA (u) ,
t>1
∞
where WF(u), WFt and WFA are C , Gevrey and analytic wave-front sets, re-
spectively.
4.1. Pseudolocal property of WFτ,σ
We refer to [23] for a more general result, and prove here only the pseudolocal
property of the wave-front set WF{τ,σ} (u), u ∈ D (U ).
Theorem 4.1. Let
P (x, D) = aα (x)Dα
|α|≤m
be a differential operator of order m on U with aα ∈ E{τ,σ} (U ), |α| ≤ m, and let
u ∈ D (U ), τ > 0, σ > 1. Then
WF{τ,σ} (P (x, D)u) ⊆ WF{τ,σ} (u),
The statement directly follows from the next lemma.
Lemma 4.2. Let u ∈ D (U ), τ > 0, σ > 1. Then
WF{τ,σ} (∂j u) ⊆ WF{τ,σ} (u), 1 ≤ j ≤ d.
If, in addition φ ∈ E{τ,σ} (U ), then
WF{τ,σ} (φu) ⊆ WF{τ,σ} (u). (4.4)
Proof. We refer to [23, Lemma 4.1] for the first part and prove here only (4.4).
σ
Set τ̃ = τ σ−1 and fix (x0 , ξ0 ) ∈ W F{τ,σ} (u). Then by the definition, there
exists open conic neighborhood Ω×Γ of (x0 , ξ0 ) and a bounded sequence {uN }N ∈N
in E (U ) such that uN = u on Ω and
hN N !τ /σ
|
uN (ξ)| ≤ A , N ∈ N, ξ ∈ Γ. (4.5)
|ξ|N 1/σ
Choose a compact neighborhood Kx0 ⊂⊂ Ω of x0 , and let {χN }N ∈N be τ̃ , σ-
admissible sequence with respect Kx0 . Set χN = φχN , N ∈ N, and note that
σ
(see Lemma 2.1) for some
χN u = χN uN . Since Mpτ,σ = pτ p satisfies (M.2)
positive increasing sequence Cq , q ∈ N, and h > 1 we obtain
α β
|Dα+β χN (x)| ≤ |Dα−δ+β−γ χN (x)||Dγ+δ φ(x)| (4.6)
δ γ
δ≤α γ≤β
α β |α−δ|+1
1N 1/σ 2|α−δ| h|γ+δ| +1 |γ + δ|τ |γ+δ|
σ σ
≤ Aβ
δ γ
δ≤α γ≤β
α |α−δ|+1
≤ (2h )|β| 1N 1/σ 2|α−δ| (Cβ h )|δ| |δ|τ |δ| ,
σ σ σ
+1
Aβ
δ
δ≤α
for x ∈ Kx0 , |α| ≤ 1(N/τ̃ )1/σ 2, β ∈ Nd , where h = h2

σ−1
. Now it is clear that by
putting |α| = 0 in (4.6) we obtain,
|Dβ χN (x)| ≤ Cβ , x ∈ Kx0 ,
and hence by applying Fourier transform it follows
(ξ)| ≤ C ξ−|β| ,
|χ β ∈ Nd , ξ ∈ Γ,
N β
for suitable Cβ > 0. In particular, since E{τ,σ} (U ) → C ∞ (U ) it follows that

χN = φχN , N ∈ N, is bounded in C ∞ (U ) and hence χN u, N ∈ N, is bounded in
E (U ).
Moreover, note that by the same type of estimates as in (3.8) for |α| =
1(N/τ̃ )1/σ 2 and β ∈ Nd , by (4.6) we obtain that
1 1 1/σ
|Dα+β χN (x)| ≤ Cβ N +1 N σ ( τ̃ ) N
, β ∈ Nd , x ∈ K x 0
and hence after applying the Fourier transform it follows
(ξ)| ≤ C N +1 N σ ( τ̃ )
|χ
1 1 1/σ
N
ξ−|α|−|β| , β ∈ Nd , ξ ∈ Γ, (4.7)
N β
for some constants Cβ > 0.

Now using (4.5) and (4.7) and arguing in the same way as in the proof of
Lemma 4.1, one can find an open cone Γ0 ⊆ Γ such that
τ̃ −1/σ
N N
h N
|χ=
σ
N u(ξ)| ≤ A , N ∈ N, ξ ∈ Γ0 ,
|ξ|(N/τ̃ )1/σ
for suitable A, h > 0. After enumeration N → τ̃ N the statement follows.
4.2. Intersections and unions of WFτ,σ and the corresponding singular supports
It turns out that the regularity related to the complement of the unions and
intersections of wave-front sets WFτ,σ , τ > 0, σ > 1, coincides with the regularity
given by (2.5)–(2.6).
In particular, for u ∈ D (U ), we consider

WF0,1 (u) = WFτ,σ (u), (4.8)
σ>1 τ >0

WF∞,1 (u) = WFτ,σ (u), (4.9)
σ>1 τ >0

WF0,∞ (u) = WFτ,σ (u), (4.10)
σ>1 τ >0

WF∞,∞ (u) = WFτ,σ (u), (4.11)
σ>1 τ >0
where WFτ,σ (u) denotes either WF{τ,σ} (u) or WF(τ,σ) (u).

From Corollary 4.1 we have

WF{τ,σ} (u) = WF(τ,σ) (u) and WF{τ,σ} (u) = WF(τ,σ) (u),
τ >0 τ >0 τ >0 τ >0
so it is sufficient to observe WF{τ,σ} (u) in (4.8)–(4.11).

By [23, Lemma 3.4], i.e.,

WFτ,σ2 (u) ⊆ WFτ,σ1 (u), u ∈ D (U ), σ2 > σ1 ≥ 1,
τ >0 τ >0
we have the following:
WF(u) ⊆ WF0,1 (u) ⊆ WF∞,1 (u)

⊆ WF0,∞ (u) ⊆ WF∞,∞ (u) ⊆ WFτ (u) , (4.12)
τ >1
where WF(u) and WFτ (u) are the classical and the Gevrey wave-front sets, re-
spectively, see also [23, Corollary 3.1].
Next we define singular support of distributions with respect to classes
E{τ,σ} ,τ > 0 and σ > 1, and the corresponding borderline cases τ ∈ {0, ∞} and
σ ∈ {1, ∞} defined by (2.5)–(2.6).
Definition 4.2. Let τ ∈ [0, ∞] and σ ∈ [1, ∞], u ∈ D (U ) and x0 ∈ U . Then

x0 ∈ singsupp{τ,σ} (u) if and only if there exists a neighborhood Ω of x0 such that
u ∈ E{τ,σ} (Ω).
Let π1 : U ×Rd\{0} → U denotes the standard projection given by π1 (x, ξ) =

x. From Propositions 3.1, 3.2, and Lemma 4.1 it follows that for a given u ∈ D (U ),
τ > 0 and σ > 1, we have singsupp{τ,σ} (u) = π1 (WF{τ,σ} (u)).
When τ ∈ {0, ∞} and σ ∈ {1, ∞} the following is true.
Theorem 4.2. Let there be given u ∈ D (U ) and let π1 : U × Rd \{0} → U be the

standard projection. Then
π1 (WF∞,∞ (u)) = singsupp0,1 (u) ,
π1 (WF0,1 (u)) = singsupp∞,∞ (u) ,
π1 (WF∞,1 (u)) = singsupp0,∞ (u) ,
π1 (WF0,∞ (u)) = singsupp∞,1 (u).
Proof. We prove here only π1 (WF0,∞ (u)) = singsupp∞,1 (u) and leave the other
equalities to the reader.
Assume that x0 ∈ π1 (WF0+ ,∞ (u)), so that there is a compact neighborhood
K ⊂⊂ U of x0 such that

K × Rd \{0} ⊆ (WF0,∞ (u))c = (WF{τ,σ} (u))c , (4.13)
σ>1 τ >0
where (WF{τ,σ} (u))c denotes the complement of the set WF{τ,σ} (u) in U ×Rd \{0}.
Therefore, if (x, ξ) ∈ K × Rd \{0} then for every σ > 1 there exist τ0 > 0 such
that (x, ξ) ∈ WF{τ0 ,σ} (u).
σ/(σ−1)
Let σ > 1 be arbitrary but fixed, and set τ̃0 = τ0 . From Lemma 4.1, it
follows that there is a τ̃0 , σ-admissible sequence {χN }N ∈N such that uN = χN u,
N ∈ N is a bounded sequence in E (U ), uN = u on some Ω ⊆ K, and
−1/σ
hN N !τ̃0 /σ
N u(ξ)| ≤ A
|χ= , N ∈ N , ξ ∈ Rd \{0} ,
|ξ|(N/τ̃0 )1/σ
which after enumeration N → τ̃0 N becomes
hN N !τ0 /σ
N u(ξ)| ≤ A
|χ= , N ∈ N , ξ ∈ Rd \{0} . (4.14)
|ξ|N 1/σ
By Proposition 3.1 it follows that u ∈ E{τ0 ,σ} (U ), and since σ can be chosen
arbitrary, we conclude that u ∈ E∞,1 (U ) (see Proposition 2.1). Therefore
singsupp∞,1 (u) ⊂ π1 (WF0,∞ (u)).
For the opposite inclusion, assume that x0 ∈ singsupp∞,1 (u). Then u ∈
E∞,1 (Ω), for some Ω which is a neighborhood of x0 . In particular, for every σ > 1
σ/(σ−1)
there exists τ0 > 0 such that u ∈ Eτ0 ,σ (Ω). Fix σ > 1 and put τ̃ = τ0 . Now
we use a τ̃0 , σ-admissible sequence {χN }N ∈N and Proposition 3.2 implies (4.14). It
follows that (x0 , ξ) ∈ (WF{τ0 ,σ} (u))c for every σ > 1 and for some τ0 > 0. Hence,
by the equality in (4.13) it follows that (x0 , ξ) ∈ WF0,∞ (u) for every ξ ∈ Rd \{0}
and therefore x0 ∈ π1 (WF0,∞ (u)), wherefrom π1 (WF0,∞ (u)) ⊂ singsupp∞,1 (u),
which finishes the proof.
Acknowledgment
This research is supported by Ministry of Education, Science and Technological
Development of Serbia through the Project no. 174024.
References
[1] M. Cappiello, R. Schulz, Microlocal analysis of quasianalytic Gelfand–Shilov type
ultradistributions, Complex Variables and Elliptic Equations 61 (4) (2016), 538–561.
[2] E. Carypis, P. Wahlberg, Propagation of exponential phase space singularities for
Schrödinger equations with quadratic Hamiltonians, Journal of Fourier Analysis and
Applications, Online First, 2016. DOI: 10.1007/s00041-016-9478-6
[3] H. Chen, L. Rodino,General theory of PDE and Gevrey classes in General theory of
partial differential equations and microlocal analysis. Pitman Res. Notes Math. Ser.,
Longman, Harlow, 349 (1996), 6–81.
[4] E. Cordero, F. Nicola, L. Rodino, Schrödinger equations with rough Hamiltonians,
Discrete Contin. Dyn. Syst. 35 (10) (2015), 4805–4821.
[5] E. Cordero, F. Nicola, L. Rodino, Propagation of the Gabor wave front set for
Schrodinger equations with non-smooth potentials, Rev. Math. Phys. 27 (1) (2015),
33 pages
[6] S. Coriasco, K. Johansson, J. Toft, Local Wave-front Sets of Banach and Fréchet
types, and Pseudo-differential Operators, Monatsh. Math. 169 (3-4) (2013), 285–316.
[7] S. Coriasco, K. Johansson, J. Toft, Global Wave-front Sets of Banach, Fréchet and
Modulation Space Types, and Pseudo-differential Operators, J. Differential Equations
254 (8) (2013), 3228–3258.
[8] H.G. Feichtinger, Modulation spaces on locally compact abelian groups, Technical
Report, University Vienna, 1983, and also in Wavelets and Their Applications, 99–
140, Allied Publishers, 2003.
[9] H.G. Feichtinger, K. Gröchenig, Gabor frames and time-frequency analysis of distri-
butions, J. Funct. Anal. 146 (1997), 464–495.
[10] H.G. Feichtinger, T. Strohmer, editors, Gabor Analysis and Algorithms: Theory and
Applications, Birkhäuser, 1998.
[11] H.G. Feichtinger, T. Strohmer, editors, Advances in Gabor Analysis, Birkhäuser,
2003.
[12] G.B. Foland, Harmonic analysis in phase space. Princeton Univ. Press, 1989.
[13] M. Gevrey, Sur la nature analitique des solutions des équations aux dérivées partielle.
Ann. Ec. Norm. Sup. Paris, 35 (1918), 129 –190.
[14] K. Gröchenig, Foundations of Time-frequency analysis, Birkhäuser, Boston, 2001.
[15] L. Hörmander, The Analysis of Linear Partial Differential Operators. Vol. I: Distri-
bution Theory and Fourier Analysis, Springer-Verlag, 1983.
[16] L. Hörmander, Quadratic hyperbolic operators, 118–160, in Microlocal analysis and
applications (Montecatini Terme, 1989), Lecture Notes in Math. 1495, Springer,
1991.
[17] K. Johansson, S. Pilipović, N. Teofanov, J. Toft Discrete Wave-front sets of Fourier
Lebesgue and modulation space types, Monatshefte für Mathematik, 166 (2) (2012),
181–199.
[18] H. Komatsu, Ultradistributions, I: Structure theorems and a characterization. J. Fac.
Sci. Univ. Tokyo, Sect. IA Math., 20 (1) (1973), 25–105.
[19] H. Komatsu, An introduction to the theory of generalized functions. Lecture notes,
Department of Mathematics Science University of Tokyo 1999.
[20] S. Pilipovic, N. Teofanov, J. Toft, Micro-local analysis in Fourier Lebesgue and mod-
ulation spaces. Part I, Journal of Fourier Analysis and Applications, 17 (3) (2011),
374–407.
[21] S. Pilipovic, N. Teofanov, J. Toft, Micro-local analysis in Fourier Lebesgue and mod-
ulation spaces. Part II, Journal of Pseudo-Differential Operators and Applications,
1 (3) (2010), 341–376.
[22] S. Pilipović, N. Teofanov, and F. Tomić, On a class of ultradifferentiable functions.
Novi Sad Journal of Mathematics, 45 (1) (2015), 125–142.
[23] S. Pilipović, N. Teofanov, and F. Tomić, Beyond Gevrey regularity. Journal of
Pseudo-Differential Operators and Applications 7 (1) (2016), 113–140.
[24] S. Pilipović, J. Toft, Wave-front sets related to quasi-analytic Gevrey sequences

(2015). arXiv:1210.7741v3 [math.AP]
[25] K. Pravda-Starov, L. Rodino, P. Wahlberg, Propagation of Gabor singularities
for Schrödinger equations with quadratic Hamiltonians (2016). arXiv:1411.0251v6
[math.AP]
[26] L. Rodino, Linear Partial Differential Operators in Gevrey Spaces. World Scientific,
1993.
[27] L. Rodino, P. Wahlberg, The Gabor wave front set, Monatsh. Math. 173 (4) (2014),
625–655.
[28] R. Schulz, P. Wahlberg, The equality of the homogeneous and the Gabor wave front
set, Comm. PDE (accepted for publication), arXiv:1304.7608v2[math.AP] 2016.
[29] R. Schulz, P. Wahlberg, Microlocal properties of Shubin pseudodifferential and local-
ization operators, J. Pseudo-Differ. Oper. Appl. 7 (1) (2016), 91–111.
[30] P. Wahlberg, Propagation of polynomial phase space singularities for Schrödinger
equations with quadratic Hamiltonians, Mathematica Scandinavica (accepted for
publication), arXiv:1411.6518 [math.AP], 2016.
Nenad Teofanov
Department of Mathematics and Informatics
Faculty of Sciences
University of Novi Sad
Novi Sad, Serbia
Filip Tomić
Faculty of Technical Sciences
University of Novi Sad
Novi Sad, Serbia
Operator Theory:

Matrix Parameterized Pseudo-differential

Calculi on Modulation Spaces
Joachim Toft
Abstract. We consider a broad matrix parameterized family of pseudo-dif-

ferential calculi, containing the usual Shubin’s family of pseudo-differential
calculi, parameterized by real numbers. We show that continuity properties
in the framework of modulation space theory, valid for the Shubin’s family
extend to the broader matrix parameterized family of pseudo-differential cal-
culi.
0. Introduction
A pseudo-differential calculus on Rd is a rule which takes any appropriate function
or distribution, defined on the phase space T ∗ Rd R2d into a set of linear opera-
tors acting on suitable functions or distributions defined on Rd . There are several
other situations with similar approaches. For example, a main issue in quantum
mechanics concerns quantization, where observables in classical mechanics (which
are functions or distributions on the phase space) carry over to corresponding ob-
servables in quantum mechanics (which usually are linear operators on subspaces
of L2 (Rd )). A somewhat similar situations can be found in time-frequency anal-
ysis. Here the phase space corresponds to the time-frequency shift space, and the
filter parameters for (non-stationary filters) are suitable functions or distributions
on the time-frequency shift space, while the corresponding filters are linear op-
erators acting on signals (which are functions or distributions, depending on the
time).
A common family of pseudo-differential calculi concerns a → Opt (a), param-
eterized by t ∈ R. If a ∈ S (R2d ), then the pseudo-differential operator Opt (a) is
defined by

Opt (a)f (x) = (2π)−d a(x − t(x − y), ξ)f (y)eix−y,ξ dydξ,
R2d
when f ∈ S (R ) (cf., e.g., [28]).

d
216 J. Toft
In the paper we consider as in [3] a slightly larger family of pseudo-differential

calculi, compared to the situations above, which are parameterized by matrices
instead of the real number t. More precisely, if a ∈ S (R2d ) and A is a real d × d
matrix, then the pseudo-differential operator OpA (a) is defined by

OpA (a)f (x) = (2π)−d a(x − A(x − y), ξ)f (y)eix−y,ξ dydξ,
R2d
when f ∈ S (R ) (cf., e.g., [28]). We note that OpA (a) = Opt (a) when A = t · I,
d
where I is the d × d identity matrix. On the other hand, in [1], D. Bayer considered
a more general situation, where each pseudo-differential calculus is parameterized
by four matrices instead of one.
The definition of OpA (a) extends in several directions. In Section 2 we dis-
cuss such extensions within the theory of modulation spaces. That is, we deduce
continuity for such operators between different modulation spaces, when a belongs
to (other) modulation spaces. Similar analysis and results can be found in, e.g.,
[17, 19, 20, 30, 32, 35, 38] in the more restricted case A = t · I, and we emphasize
that all results are obtained by using the framework of these earlier contributions.
Furthermore, some results here are in some cases contained in certain results in
Chapters 1 and 2 in [1].
In Section 3 we also give examples on how these operators might be used
in quantization, by taking the average of OpA (a) with A = 12 · I + r · U , over all
r ∈ [0, 1] and unitary matrices U with real entries.
1. Preliminaries
In this section we introduce some notations and discuss basic results. We start by
recalling some facts concerning Gelfand–Shilov spaces. Thereafter we recall some
properties about pseudo-differential operators. Especially we discuss the Weyl
product and twisted convolution. Finally we recall some facts about modulation
spaces. The proofs are in general omitted, since the results can be found in the
literature.
We start by considering Gelfand–Shilov spaces. Let 0 < h, s ∈ R be fixed.
Then Ss,h (Rd ) consists of all f ∈ C ∞ (Rd ) such that
|xβ ∂ α f (x)|
f Ss,h ≡ sup
h|α|+|β| α!s β!s
is finite. Here the supremum should be taken over all α, β ∈ Nd and x ∈ Rd .
Obviously Ss,h → S is a Banach space which increases with h and s. Here
and in what follows we use the notation A → B when the topological spaces
A and B satisfy A ⊆ B with continuous embeddings. Furthermore, if s > 1/2
an s0 = 1/2 , then Ss,h and ∪h>0 Ss0 ,h contain all finite linear combinations of
Hermite functions. Since such linear combinations are dense in S , it follows that
the dual (Ss,h ) (Rd ) of Ss,h (Rd ) is a Banach space which contains S (Rd ).
Matrix Parameterized Pseudo-differential Calculi 217
The Gelfand–Shilov spaces Ss (Rd ) and Σs (Rd ) are the inductive and projec-
tive limits respectively of Ss,h (Rd ). This implies that

Ss (Rd ) = Ss,h (Rd ) and Σs (Rd ) = Ss,h (Rd ), (1.1)
h>0 h>0
and that the topology for Ss (R ) is the strongest possible one such that the in-
d
clusion map from Ss,h (Rd ) to Ss (Rd ) is continuous, for every choice of h > 0.
The space Σs (Rd ) is a Fréchet space with semi norms · Ss,h , h > 0. Moreover,
Σs (Rd ) = {0}, if and only if s > 1/2, and Ss (Rd ) = {0} if and only if s ≥ 1/2 (cf.
[16, 24, 26]).
The Gelfand–Shilov distribution spaces Ss (Rd ) and Σs (Rd ) are the projective
and inductive limit respectively of Ss (Rd ). This means that

Ss (Rd ) =
Ss,h (Rd ) and Σs (Rd ) =
Ss,h (Rd ). (1.1)
h>0 h>0
We remark that in [24, 26] it is proved that Ss (Rd )

is the dual of Ss (Rd ), and
d d
Σs (R ) is the dual of Σs (R ) (also in topological sense).
For each ε > 0 and s > 1/2 we have
S1/2 (Rd ) →Σs (Rd ) → Ss (Rd ) → Σs+ε (Rd )
(1.2)
and Σs+ε (Rd ) → Ss (Rd ) →Σs (Rd ) → S1/2

(Rd ).
The Gelfand–Shilov spaces are invariant under several basic transformations.
For example they are invariant under translations, dilations and under (partial)
Fourier transformations. We also note that the map (f1 , f2 ) → f1 ⊗f2 is continuous
from Ss (Rd1 ) × Ss (Rd2 ) to Ss (Rd1 +d2 ), and similarly when each Ss are replaced
by Σs , Ss or by Σs .
We let F be the Fourier transform which takes the form

(F f )(ξ) = f (ξ) ≡ (2π) −d/2
f (x)e−ix,ξ dx
Rd
when f ∈ L1 (Rd ). Here · , · denotes the usual scalar product on Rd . The map
F extends uniquely to homeomorphisms on S (Rd ), Ss (Rd ) and Σs (Rd ), and
restricts to homeomorphisms on S (Rd ), Ss (Rd ) and Σs (Rd ), and to a unitary
operator on L2 (Rd ).
1.1. An extended family of pseudo-differential calculi

Next we discuss some issues in pseudo-differential calculus. Let M(d, Ω) be the set
of all d × d-matrices with entries in the set Ω, and let s ≥ 1/2, a ∈ Ss (R2d ) and
A ∈ M(d, R) be fixed. Then the pseudo-differential operator OpA (a) is the linear
and continuous operator on Ss (Rd ), given by

(OpA (a)f )(x) = (2π)−d a(x − A(x − y), ξ)f (y)eix−y,ξ dydξ, (1.3)
218 J. Toft
when f ∈ Ss (Rd ). For general a ∈ Ss (R2d ), the pseudo-differential operator

OpA (a) is defined as the linear and continuous operator from Ss (Rd ) to Ss (Rd )
with distribution kernel given by
Ka,A (x, y) = (2π)−d/2 (F2−1 a)(x − A(x − y), x − y). (1.4)
Here F2 F is the partial Fourier transform of F (x, y) ∈ Ss (R2d ) with respect to
the y variable. This definition makes sense, since the mappings
F2 and F (x, y) → F (x − A(x − y), x − y) (1.5)
are homeomorphisms on Ss (R2d ).
In particular, the map a → Ka,A is a homeo-
morphism on Ss (R2d ).
An important special case appears when A = t · I, with t ∈ R. Here and in
what follows, I ∈ M(d, R) denotes the d × d identity matrix. In this case we set
Opt (a) = Opt·I (a).
The normal or Kohn–Nirenberg representation, a(x, D), is obtained when t = 0,
and the Weyl quantization, Opw (a), is obtained when t = 12 . That is,
a(x, D) = Op0 (a) and Opw (a) = Op1/2 (a).
For any K ∈ Ss (Rd1 +d2 ), we let TK be the linear and continuous mapping
from Ss (Rd1 ) to Ss (Rd2 ), defined by the formula
(TK f, g)L2 (Rd2 ) = (K, g ⊗ f )L2 (Rd1 +d2 ) . (1.6)
It is well known that if A ∈ M(d, R), then it follows from the Schwartz kernel
theorem that K → TK and a → OpA (a) are bijective mappings from S (R2d ) to
the set of linear and continuous mappings from S (Rd ) to S (Rd ) (cf., e.g., [23]).
Furthermore, by, e.g., [25, Theorem 2.2] it follows that the same holds true
if each S and S are replaced by Ss and Ss , respectively, or by Σs and Σs ,
respectively.
In particular, for every a1 ∈ Ss (R2d ) and A1 , A2 ∈ M(d, R), there is a unique
a2 ∈ Ss (R2d ) such that OpA1 (a1 ) = OpA2 (a2 ). The following result explains the
relations between a1 and a2 .

Proposition 1.1. Let a1 , a2 ∈ S1/2 (R2d ) and A1 , A2 ∈ M(d, R). Then
OpA1 (a1 ) = OpA2 (a2 ) ⇔ eiA2 Dξ ,Dx a2 (x, ξ) = eiA1 Dξ ,Dx a1 (x, ξ). (1.7)
Note here that the latter equality in (1.7) makes sense since it is equivalent to
eiA2 x,ξ
a2 (ξ, x) = eiA1 x,ξ
a1 (ξ, x),
and that the map a → eiAx,ξ a is continuous on Ss (cf., e.g., [3, 40]).
Passages between different kinds of pseudo-differential calculi have been con-
sidered before (cf., e.g., [23, 40].) On the other hand, except for [1, 3], it seems
that the representation a → OpA (a) for general matrix A ∈ M(d, R), has not been
considered in the literature before.
A proof of Proposition 1.1 for matrices of the form A = t · I, t ∈ R, can

be found in, e.g., [28], and the proof of the result for general A follows by similar
arguments. In order to be self-contained we here present the arguments. Here and
in what follows, A∗ denotes the transpose of the matrix A.
Proof of Proposition 1.1. It is no restriction to assume that A2 = 0. Let a = a1 ,

b = a2 and A = A1 . We also prove the result only in the case a, b ∈ S (R2d ). The
general case follows by similar arguments and is left for the reader.
The equality OpA (a) = Op(b) is the same as
F (b(x, · ))(y − x) = F (a(x − A(x − y), · ))(y − x) ⇔

b(x, η)eiy,η dη = a(x − Ay, η)eiy,η dη ⇔

−d
b(x, ξ) = (2π) a(x − Ay, η)eiy,η−ξ dydη
By Fourier’s inversion formula we get

−d
b(x, ξ) = (2π) a(x − Ay, η)eiy,η−ξ dydη

= (2π)−2d a(η1 , y1 )ei(x−Ay,η1 +y1 ,η+y,η−ξ) dy1 dη1 dydη

∗
= (2π)−d a(η1 , y1 )ei(x,η1 +y1 ,ξ+A
η1 )
dy1 dη1
= eiADξ ,Dx a(x, ξ),

which gives the result.
Let a ∈ Ss (R2d ) be fixed. Then a is called a rank-one element with respect
to A ∈ M(d, R), if OpA (a) is an operator of rank-one, i.e.,
OpA (a)f = (f, f2 )f1 , f ∈ Ss (Rd ), (1.8)
for some f1 , f2 ∈ Ss (Rd ). By straightforward computations it follows that (1.8) is
d
fulfilled if and only if a = (2π) 2 WfA1 ,f2 , where WfA1 ,f2 is the A-Wigner distribution,
defined by the formula

WfA1 ,f2 (x, ξ) ≡ F f1 (x + A · )f2 (x + (A − I) · ) (ξ), (1.9)
which takes the form

WfA1 ,f2 (x, ξ) = (2π)− 2 f1 (x + Ay)f2 (x + (A − I)y)e−iy,ξ dy,
d
when f1 , f2 ∈ Ss (Rd ). By combining these facts with (1.7), it follows that

eiA2 Dξ ,Dx WfA12,f2 = eiA1 Dξ ,Dx WfA11,f2 , (1.10)
220 J. Toft
for every f1 , f2 ∈ Ss (Rd ) and A1 , A2 ∈ M(d, R). Since the Weyl case is particularly
important, we set WfA1 ,f2 = Wf1 ,f2 when A = 12 I, i.e., Wf1 ,f2 is the usual (cross-)
Wigner distribution of f1 and f2 .
For future references we note the link
(OpA (a)f, g)L2 (Rd ) = (2π)−d/2 (a, Wg,f

A
)L2 (R2d ) ,
a ∈ Ss (R2d ) and f, g ∈ Ss (Rd ) (1.11)

between pseudo-differential operators and Wigner distributions, which follows by
straightforward computations (see also, e.g., [8, 9]).
For any A ∈ M(d, R), the A-product, a#A b between a ∈ Ss (R2d ) and b ∈

Ss (R2d ) is defined by the formula
OpA (a#A b) = OpA (a) ◦ OpA (b), (1.12)
provided the right-hand side makes sense as a continuous operator from Ss (Rd )
to Ss (Rd ).
1.2. Modulation spaces

Next we discuss basic properties for modulation spaces, and start by recalling the
conditions for the involved weight functions. A function ω on Rd is called a weight
(on Rd ), if ω > 0 and ω, ω −1 ∈ L∞
loc (R ). Let ω and v be weights on R . Then ω
d d
is called moderate or v-moderate if

ω(x + y) ω(x)v(y), x, y ∈ Rd . (1.13)
Here and in what follows we write A B when A, B ≥ 0 and A ≤ cB for a
suitable constant c > 0. We also let A B when A B and B A. The weight v
is called submultiplicative, if v is even and (1.13) holds when ω = v. We note that
if (1.13) holds, then
v(−x)−1 ω(x) v(x).
Furthermore, for such ω it follows that (1.13) is true when
v(x) = Cec|x| ,
for some positive constants c and C (cf., e.g., [18]). In particular, if ω is moderate
on Rd , then
e−c|x| ω(x) ec|x|,
for some constant c > 0.
The set of all moderate functions on Rd is denoted by PE (Rd ). Furthermore,
if v in (1.13) can be chosen as a polynomial, then ω is called a weight of polynomial
type, or polynomially moderated. We let P(Rd ) be the set of all polynomially
moderated weights on Rd . If ω(x, ξ) ∈ PE (R2d ) is constant with respect to the
x-variable (ξ-variable), then we sometimes write ω(ξ) (ω(x)) instead of ω(x, ξ). In
this case we consider ω as an element in PE (R2d ) or in PE (Rd ) depending on
the situation.
Let φ ∈ Ss (Rd ) be fixed. Then the short-time Fourier transform Vφ f of f ∈

Ss (Rd )
with respect to the window function φ is the Gelfand–Shilov distribution
on R , defined by
2d
Vφ f (x, ξ) ≡ (F2 (U (f ⊗ φ)))(x, ξ) = F (f φ( · − x))(ξ),

where (U F )(x, y) = F (y, y − x). If f, φ ∈ Ss (Rd ), then it follows that

Vφ f (x, ξ) = (2π)−d/2 f (y)φ(y − x)e−iy,ξ dy.
We recall that the short-time Fourier transform is closely related to the Wigner
distribution, because
Wf,φ (x, ξ) = 2d e2ix,ξ Vφ̌ f (2x, 2ξ), (1.14)
which follows by elementary manipulations. Here φ̌(x) = φ(−x).
Let ω ∈ PE (R2d ), p, q ∈ (0, ∞] and φ ∈ S1/2 (Rd ) \ 0 be fixed. Then the
mixed Lebesgue space Lp,q
(ω) (R ) consists of all measurable functions F on R
2d 2d
such that F Lp,q

(ω)
< ∞. Here
F Lp,q
(ω)
≡ Fp,ω Lq , where Fp,ω (ξ) ≡ F ( · , ξ)ω( · , ξ) Lp . (1.15)
We note that these quasi-norms might attain +∞.
p,q
The modulation space M(ω) (Rd ) is the quasi-Banach space which consist of

all f ∈ S1/2 (Rd ) such that f M(ω)p,q < ∞, where
f p,q
M(ω) ≡ Vφ f Lp,q
(ω)
. (1.16)
p,q
We remark that the definition of M(ω) (Rd ) is independent of the choice of φ ∈
S1/2 (Rd ) \ 0 and different φ gives rise to equivalent quasi-norms. (See Proposition
1.2 below.)
p p,p p,q
For convenience we set M(ω) = M(ω) . Furthermore we set M p,q = M(ω) when
ω ≡ 1.
The proof of the following proposition is omitted, since the results can be
found in [4, 10–15, 17, 32–35, 37, 39]. Here we recall that p, p ∈ [1, ∞] satisfy
1 1
p + p = 1.
Proposition 1.2. Let p, q, pj , qj ∈ (0, ∞] for j = 1, 2, r ≤ min(p, q, 1), and ω, ω1 , ω2 ,

v ∈ PE (R2d ) be such that v is submultiplicative, ω is v-moderate and ω2 ω1 .
Then the following is true:
p,q
(1) f ∈ M(ω) (Rd ) if and only if Vφ f Lp,q(ω)
< ∞ holds for any φ ∈ M(v) r
(Rd ) \ 0.
p,q
Moreover, M(ω) is a quasi-Banach space under the quasi-norm in (1.16) and
different choices of φ give rise to equivalent quasi-norms. Furthermore, if
p,q
p, q ≥ 1, then M(ω) is a Banach space;
(2) if p1 ≤ p2 and q1 ≤ q2 then
p1 ,q1
Σ1 (Rd ) → M(ω 1)
p2 ,q2
(Rd ) → M(ω 2)
(Rd ) → Σ1 (Rd );
222 J. Toft
(3) if in addition p, q ≥ 1, then the L2 product ( · , · )L2 on S1/2 (Rd ) extends

p,q p ,q
uniquely to a continuous map from M(ω) (Rn ) × M(1/ω) (Rd ) to C. On the
other hand, if a = sup |(a, b)|, where the supremum is taken over all b ∈
S1/2 (Rd ) such that b M p ,q ≤ 1, then · and · M(ω) p,q are equivalent norms;
(1/ω)
p,q
(4) if 1 ≤ p, q < ∞, then S1/2 (Rd ) is dense in M(ω) (Rd ) and the dual space of

p,q p ,q
M(ω) (Rd ) can be identified with M(1/ω) (Rd ), through the L2 -form ( · , · )L2 .

p ,q
Moreover, S1/2 (Rd ) is weakly dense in M(ω) (Rd ) with respect to the L2 -
form.
Proposition 1.2 (1) allows us be rather vague concerning the choice of φ ∈

r
M(v) \ 0 in (1.16). For example, if C > 0 is a constant and A is a subset of S1/2 ,
then a M(ω) ≤ C for every a ∈ A , means that the inequality holds for some
p,q
choice of φ ∈ M(v)r
\ 0 and every a ∈ A . Evidently, a similar inequality is true
for any other choice of φ ∈ M(v)r
\ 0, with a suitable constant, larger than C if
necessary.
Remark 1.3. By Theorem 3.9 in [37] and Proposition 1.2 (2) it follows that

p,q
M(ω) (Rd ) = Σ1 (Rd ), p,q
M(ω) (Rd ) = Σ1 (Rd ).
ω∈PE ω∈PE
1/s
+|η|1/s )
More generally, let s ≥ 1, vc (y, η) = ec(|y| , and let P respectively P0 be
the set of all ω ∈ PE (R2d ) such that
ω(x + y, ξ + η) ω(x, ξ)vc (y, η),
for some c > 0 respectively for every c > 0. Then

p,q
M(ω) (Rd ) = Σs (Rd ), p,q
M(1/ω) (Rd ) = Σs (Rd ),
ω∈P ω∈P

p,q
M(ω) (Rd ) = Ss (Rd ), p,q
M(1/ω) (Rd ) = Ss (Rd ),
ω∈P0 ω∈P0
p,q
Σs (Rd ) → M(vc)
(Rd ) → Ss (Rd ) and Ss (Rd ) → M(1/v
p,q
c)
(Rd ) → Σs (Rd ).
(cf. Proposition 4.5 in [6], Proposition 4. in [21], Corollary 5.2 in [27] or Theorem
4.1 in [31]. See also [37, Theorem 3.9] for an extension of these inclusions to broader
classes of Gelfand–Shilov and modulation spaces.)
1.3. Schatten–von Neumann classes
Next we recall some properties on Schatten–von Neumann classes. Let H1 and H2
be Hilbert spaces, and let T be a linear map from H1 to H2 . For every integer
j ≥ 1, the singular number of T of order j is given by
σj (T ) = σj (H1 , H2 , T ) ≡ inf T − T0 H1 →H2 ,
where the infimum is taken over all linear operators T0 from H1 to H2 with rank at
most j − 1. Therefore, σ1 (T ) equals T H1→H2 , and σj (T ) is non-negative which
decreases with j.
For any p ∈ (0, ∞] we set
T Ip = T Ip (H1 ,H2 ) ≡ {σj (H1 , H2 , T )}∞
j=1 lp
(which might attain +∞). The operator T is called a Schatten–von Neumann

operator of order p from H1 to H2 , if T Ip is finite, i.e., {σj (H1 , H2 , T )}∞j=1
should belong to lp . The set of all Schatten–von Neumann operators of order p
from H1 to H2 is denoted by Ip = Ip (H1 , H2 ). We note that I∞ (H1 , H2 ) agrees
with B(H1 , H2 ) (also in norms), the set of linear and bounded operators from H1
to H2 . If p < ∞, then Ip (H1 , H2 ) is contained in K(H1 , H2 ), the set of linear
and compact operators from H1 to H2 . The spaces Ip (H1 , H2 ) for p ∈ (0, ∞]
and K(H1 , H2 ) are quasi-Banach spaces which are Banach spaces when p ≥ 1.
Furthermore, I2 (H1 , H2 ) is a Hilbert space and agrees with the set of Hilbert-
Schmidt operators from H1 to H2 (also in norms). We set Ip (H ) = Ip (H , H ).
The set I1 (H1 , H2 ) is the set of trace-class operators from H1 to H2 , and
· I1 (H1 ,H2 ) coincide with the trace-norm. If in addition H1 = H2 = H , then
the trace
TrH (T ) ≡ (T fα , fα )H
α
is well defined and independent of the orthonormal basis {fα }α in H .
Now let H3 be another Hilbert space and let Tk be a linear and continuous
operator from Hk to Hk+1 , k = 1, 2. Then we recall the Hölder relation
T2 ◦ T1 Ir (H1 ,H3 ) ≤ T1 Ip1 (H1 ,H2 ) T2 Ip2 (H2 ,H3 )
1 1 1 (1.17)
when + =
p1 p2 r
(cf., e.g., [29, 38]).
In particular, the map (T1 , T2 ) → T2∗ ◦ T1 is continuous from Ip (H1 , H2 ) ×
Ip (H1 , H2 ) to I1 (H1 ), giving that
(T1 , T2 )I2 (H1 ,H2 ) ≡ TrH1 (T2∗ ◦ T1 ) (1.18)
is well defined and continuous from Ip (H1 , H2 ) × Ip (H1 , H2 ) to C. If p = 2,
then the product, defined by (1.18) agrees with the scalar product in I2 (H1 , H2 ).
The proof of the following result is omitted, since it can be found in, e.g., [2, 29].
Proposition 1.4. Let p ∈ [1, ∞], H1 and H2 be Hilbert spaces, and let T be a linear
and continuous map from H1 to H2 . Then the following is true:
(1) if q ∈ [1, p ], then
T Ip (H1 ,H2 ) = sup |(T, T0 )I2 (H1 ,H2 ) |,
where the supremum is taken over all T0 ∈ Iq (H1 , H2 ) such that
T0 Ip (H1 ,H2 ) ≤ 1;
224 J. Toft
(2) if in addition p < ∞, then the dual of Ip (H1 , H2 ) can be identified through
the form (1.18).
Later on we are especially interested of finding necessary and sufficient con-
ditions of symbols, in order for the corresponding pseudo-differential operators to
belong to Ip (H1 , H2 ), where H1 and H2 satisfy
Σ1 (Rd ) → H1 , H2 → Σ1 (Rd ).
Therefore, for such Hilbert spaces and p ∈ (0, ∞], let

sA,p (H1 , H2 ) ≡ { a ∈ S1/2 (R2d ) ; OpA (a) ∈ Ip (H1 , H2 ) }
and
a sA,p (H1 ,H2 ) ≡ OpA (a) Ip (H1 ,H2 ) . (1.19)
Since the map a → OpA (a) is bijective from Ss (R2d ) to the set of all linear and
continuous operators from Ss (Rd ) to Ss (Rd ), when s ≥ 12 , it follows from the
definitions that the map a → OpA (a) restricts to a bijective and isometric map
from sA,p (H1 , H2 ) to Ip (H1 , H2 ).
Usually it is assumed that H1 and H2 are tempered in the sense of Definition
3.1 in [38], or more restricted that Hj = M(ω 2
j)
(Rd ), for some ωj ∈ PE (R2d ),
j = 1, 2. For conveniency we therefore set
sA,p (ω1 , ω2 ) ≡ sA,p (M(ω
2
1)
2
, M(ω 2)
).
We remark that the reader who is not interested in the most general setting may
only consider the case when Hj = M(ω 2
j)
(Rd ), with ωj ∈ PE (R2d ). In this case,
the L -dual of Hj is given by M(1/ωj ) (R ).
2 2 d
The latter bijectivity implies that Proposition 1.4 carries over to analogous
properties for sA,p (H1 , H2 ) spaces. In the next section we show that related results
can be proved when the sA,2 product and sp ,A (H1 , H2 ) can be replaced by the L2
product and sp ,A (H1 , H2 ) when Hj are tempered with L2 -duals Hj , j = 1, 2.
2. Algebraic and continuity properties

In this section we deduce basic results for pseudo-differential operators with sym-
bols in modulation spaces, where the corresponding weights belong to PE . The
arguments are in general similar as corresponding approaches in [33, 36].
The continuity results that we are focused on are especially Theorems 2.2
and 2.3. Here Theorem 2.2 deals with pseudo-differential operators with symbols
in modulation spaces, which act on modulation spaces. Theorem 2.3 gives necessary
and sufficient conditions on symbols such that corresponding pseudo-differential
operators are Schatten–von Neumann operators of certain degrees.
In Propositions 2.4 and 2.5 we deduce preparatory results on Wigner distri-
butions and pseudo-differential calculus in the context of modulation space theory.
In the last part we deduce composition properties for pseudo-differential op-

erators with symbols in modulation spaces. Especially we extend certain results
in [5, 22].
Let s ≥ 12 and let K ∈ Ss (Rd2 +d1 ). Then recall that K gives rise to a linear
and continuous operator T = TK from Ss (Rd1 ) to Ss (Rd2 ), defined by the formula
T f (x) = K(x, · ), f , (2.1)
which should be interpreted as (1.6) when f ∈ Ss (R ) and g ∈ Ss (R ).
d1 d2
Before presenting the continuity properties of operators with kernels or sym-

bols in modulation spaces, we present relations between the involved weight func-
tions. The weights ω, ω0 ∈ PE (R4d ) and ω1 , ω2 ∈ PE (R2d ) are in general related
to each others by the formulae
ω2 (x, ξ)
ω(x, y, ξ, −η), x, ξ ∈ Rd2 , y, η ∈ Rd1 (2.2)
ω1 (y, η)
or
ω2 (x, ξ)
ω(x, y, ξ, −η), x, ξ ∈ Rd2 , y, η ∈ Rd1 , (2.2)
ω1 (y, η)
and
ω(x, y, ξ, η) ω0 (x − A(x − y), A∗ ξ − (I − A∗ )η, ξ + η, y − x),
x, y, ξ, η ∈ Rd , (2.3)
or equivalently,
ω0 (x, ξ, η, y) ω(x − Ay, x + (I − A)y, ξ + (I − A∗ )η, −ξ + A∗ η),

x, y, ξ, η ∈ Rd . (2.3)
We note that (2.2) and (2.3) imply
ω2 (x, ξ)
ω0 (x − A(x − y), A∗ ξ + (I − A∗ )η, ξ − η, y − x), (2.4)
ω1 (y, η)
and that (2.2) and (2.3) imply
ω2 (x, ξ)
ω0 (x − A(x − y), A∗ ξ + (I − A∗ )η, ξ − η, y − x), (2.4)
ω1 (y, η)
The Lebesgue exponents of the modulation spaces should satisfy conditions
of the form
1 1 1 1 1 1
− = − = 1 − − , q ≤ p2 , q2 ≤ p, (2.5)
p1 p2 q1 q2 p q
or
p1 ≤ p ≤ p2 , q1 ≤ min(p, p ) and q2 ≥ max(p, p ). (2.6)
The first result is essentially a fundamental kernel theorem for operators in
the framework of modulation space theory, and corresponds to Schwartz kernel
theorem for. The result goes back to [11] in the unweighted case (see also [17]).
226 J. Toft
The general case follows by combining Theorem A.1 in [38] with Proposition 2.5
below. The details are left for the reader.
Theorem 2.1. Let ωj ∈ PE (R2dj ) for j = 1, 2 and ω ∈ PE (R2d2 +2d1 ) be such
that (2.2) holds. Also let T be a linear and continuous map from S1/2 (Rd1 ) to

S1/2 (Rd2 ). Then the following conditions are equivalent:
1 ∞
(1) T extends to a continuous mapping from M(ω 1)
(Rd1 ) to M(ω 2)
(Rd2 );
∞
(2) there is a unique K ∈ M(ω) (Rd2 +d1 ) such that (2.1) holds for every f ∈
S1/2 (R );
d1
(3) if in addition d1 = d2 = d, A ∈ M(d, R) and (2.3) holds, then there is a

∞
unique a ∈ M(ω 0)
(R2d ) such that T f = OpA (a)f when f ∈ S1/2 (Rd ).
Furthermore, if (1)–(2) are fulfilled, then T 1
M(ω )
∞
→M(ω K ∞ ,
M(ω) and if
1 2)
∞
in addition d1 = d2 , then K M(ω) a M(ω∞ .
) 0
The next two results extend Theorems A.2 and A.3 in [38].
Theorem 2.2. Let A ∈ M(d, R) and p, q, pj , qj ∈ [1, ∞] for j = 1, 2, satisfy
(2.5). Also let ω0 ∈ PE (R2d ⊕ R2d ) and ω1 , ω2 ∈ PE (R2d ) satisfy (2.4). If

a ∈ M(ω)
p,q
(R2d ), then OpA (a) from S1/2 (Rd ) to S1/2 (Rd ) extends uniquely to a
p1 ,q1 p2 ,q2
continuous mapping from M(ω1 ) (Rd ) to M(ω2 ) (Rd ), and
OpA (a) M(ω1
p ,q1 p ,q
→M(ω2 )2 a p,q .
M(ω )
(2.7)
1) 2 0
p,q
Moreover, if in addition a belongs to the closure of S1/2 under the M(ω 0)
p1 ,q1 p2 ,q2
norm, then OpA (a) : M(ω1 ) → M(ω2 ) is compact.
Theorem 2.3. Let A ∈ M(d, R) and p, q, pj , qj ∈ [1, ∞] for j = 1, 2, satisfy (2.6).
Also let ω0 ∈ PE (R2d ⊕ R2d ) and ω1 , ω2 ∈ PE (R2d ) satisfy (2.4) . Then
p1 ,q1 p2 ,q2
M(ω 0)
(R2d ) → sA,p (ω1 , ω2 ) → M(ω 0)
(R2d ).
For the proofs we need the following extensions of Propositions 4.1 and 4.8 in [35].
Proposition 2.4. Let A ∈ M(d, R), and let pj , qj , p, q ∈ (0, ∞] be such that p ≤
pj , qj ≤ q, for j = 1, 2, and
1 1 1 1 1 1
+ = + = + . (2.8)
p1 p2 q1 q2 p q
Also let ω1 , ω2 ∈ PE (R2d ) and ω ∈ PE (R2d ⊕ R2d ) be such that
ω0 (x − A(x − y), A∗ ξ + (I − A∗ )η, ξ − η, y − x) ω1 (x, ξ)ω2 (y, η). (2.9)

Then the map (f1 , f2 ) → WfA1 ,f2from S1/2 (Rd ) × S1/2 (Rd ) to S1/2 (R2d ) restricts
p1 ,q1 p2 ,q2 p,q
to a continuous mapping from M(ω 1)
(Rd ) × M(ω 2)
(Rd ) to M(ω 0)
(R2d ), and
WfA1 ,f2 p,q
M(ω )
f1 p ,q1
M(ω1 f2 M(ω2
p ,q2 (2.10)
0 1) 2)

when f1 , f2 ∈ S1/2 (Rd ).
Proposition 2.5. Let p ∈ (0, ∞], ωj ∈ PE (R2dj ), j = 1, 2, ω ∈ PE (R2d2 +2d1 ),

and let T be a linear and continuous operator from S1/2 (Rd1 ) to S1/2 (Rd2 ) with

distribution kernel K ∈ S1/2 (R d2 +d1
). Then the following is true:
(1) if (2.2) holds, then T ∈ I2 (M(ω
2
1)
(Rd1 ), M(ω
2
2)
(Rd2 )), if and only if K ∈
2 d2 +d1
M(ω) (R ), and then
T I2 K 2 ;
M(ω) (2.11)
(2) if d1 = d2 = d, A ∈ M(d, R) and ω0 ∈ PE (R2d ) satisfy (2.3) , a ∈ S1/2

(R2d )
p
and K = Ka,A is given by (1.4), then K ∈ M(ω) (R2d ) if and only if a ∈
p
M(ω 0)
(R2d ), and
K p
M(ω) a p
M(ω .
0)
We need some preparations for the proofs. First we note that (2.9) is the
same as

ω0 (x, ξ, η, y) ω1 (x − Ay, ξ + (I − A∗ )η)ω2 (x + (I − A)y, ξ − A∗ η). (2.9)
Lemma 2.6. Let A ∈ M(d, R), s ≥ 1
2 f, g ∈ Ss (Rd ), φ, ψ ∈ Ss (Rd ), and let
A
Φ = Wφ,ψ . Then
A
(VΦ Wf,g )(x, ξ, η, y)
= e−iy,ξ (Vφ f )(x − Ay, ξ − (A∗ − I)η)(Vψ g)(x − (A − I)y, ξ − A∗ η)
The proof of the preceding lemma follows by similar arguments as for Lemma
14.5.1 in [17]. In order to be self-contained, we here present the arguments.
Proof. We only consider the case when f, g ∈ Ss (Rd ), leaving the modifications of
the general case to the reader.
Let
H(x, y1 , y2 , y3 ) = f (y1 + Ay2 )g(y1 + By2 )φ(y1 − x + Ay3 )ψ(y1 − x + By3 ),
where B = A − I. Then Fourier’s inversion formula gives
(2π)d (VΦ Wf,g
A
)(x, ξ, η, y)

= (2π)−d H(x, y1 , y2 , y3 )e−i(y2 +y,η1 +y3 ,ξ−η1 +y1 ,η) dy1 dy2 dy3 dη1

−iy,ξ
=e H(x, y1 , y2 , y2 + y)e−i(y2 ,ξ+y1 ,η) dy1 dy2

= e−iy,ξ F (y1 + Ay2 , x − Ay)G(y1 + By2 , x − By)e−i(y2 ,ξ+y1 ,η) dy1 dy2 ,
where
F (x, y) = f (x)φ(x − y)
228 J. Toft
and
G(x, y) = g(x)ψ(x − y).
By taking (y1 + Ay2 , y1 + By2 ) as new variables of integration, we obtain
(2π)d (VΦ Wf,g
A
)(x, ξ, η, y)

= e−iy,ξ F (z1 , x − Ay)G(z2 , x − By)e−i(z1 −z2 ,ξ+z1 −A(z1 −z2 ),η) dz1 dz2
= (2π)d e−iy,ξ I(x, ξ, η, y)J(x, ξ, η, y),

where

∗
−d
I(x, ξ, η, y) = (2π) 2 f (z)φ(z − (x − Ay))e−iz,ξ−B η
dz
∗
= Vφ f )(x − Ay, ξ − B η)
and
∗
−d
J(x, ξ, η, y) = (2π) 2 g(z)ψ(z − (x − By))eiz,ξ−A η
dz
= (Vψ g)(x − By, ξ − A∗ η),

and the result follows by combining these equalities.
Proof of Proposition 2.4. We only prove the result when p, q < ∞. The straight-
forward modifications to the cases p = ∞ or q = ∞ are left for the reader. Let
φ1 , φ2 ∈ Σ1 (Rd ) \ 0, and let Φ = WφA1 ,φ2 . Then Fourier’s inversion formula gives
(VΦ (WfA1 ,f2 ))(x, ξ, η, y)

= e−iy,ξ F1 (x − Ay, ξ + (I − A∗ )η)F2 (x + (I − A)y, ξ − A∗ η),
where Fj = Vφj fj , j = 1, 2. By applying the Lp,q
(ω) -norm on the latter equality, and

using (2.9) , it follows from Minkowski’s inequality that

1/p 1/q
p,q ∗ ≤
A
Wf1 ,f2 M(ω )
G1 G2 L r H(η) dη ,
0
where Gj = |Fj ωj |p , r = q/p ≥ 1 and

r 1/r r
H(η) = G1 (y − x, η − ξ)G2 (x, ξ) dx dy dξ .
Now let rj , sj ∈ [1, ∞] for j = 1, 2 be chosen such that

1 1 1 1 1
+ = + = 1+ .
r1 r2 s1 s2 r
Then Young’s inequality gives
r
H(η) ≤ G1 ( · , η − ξ) L r1 G2 ( · , ξ) L r2 dξ
Hence another application of Young’s inequality gives

1/q 1/p
WfA1 ,f2 M(ω
p,q
)
H(η) dη G1 Lr1 ,s1 G2 Lr2 ,s2
0
By letting pj = prj and qj = qsj , the last inequality gives (2.10). The proof is
complete.
Proof of Proposition 2.5. We only prove (2), since (1) is a restatement of Propo-
sition A.5 (2) in [38].
Let Φ, Ψ ∈ S1/2 (R2d ) \ 0 be such that
Φ(x, y) = (F2 Ψ)(x − A(x − y), x − y).
Then it follows by straightforward applications of Fourier’s inversion formula that
|(VΦ Ka,A )(x, y, ξ, η)| |(VΨ a)(x − A(x − y), A∗ ξ + (A∗ − I)η, ξ + η, y − x)|.
The assertion now follows by applying the Lp(ω) quasi-norm on the last equality,
and using the fact that modulation spaces are independent of the choices of window
functions in the definition of the modulation space quasi-norms (cf. Propositions
3.1 and 3.4 in [39]).
Proof of Theorem 2.1. The equivalence between (1) and (2) follows from Theorem
A.1 in [38], and the equivalence between (2) and (3) follows immediately from
Proposition 2.5.
Proof of Theorem 2.2. The conditions on pj and qj implies that
1 1 1 1 1 1
p ≤ p1 , q1 , p2 , q2 ≤ q , + = + = + .
p1 p2 q1 q2 p q
Hence Proposition 2.4 and (2.4) show that
A
Wg,f
p ,q
M(1/ω)
f p ,q1
M(ω1 g p ,q
2 2
1) M(1/ω )
2
p1 ,q1 p2 ,q2

when f ∈ M(ω 1)
(Rd ) and g ∈ M(1/ω 2)
(Rd ).
The continuity is now an immediate consequence of (1.11) and Proposition
1.2 (4), except for the case p = q = ∞, which we need to consider separately.
Therefore assume that p = ∞, and q = 1, and let a ∈ S1/2 (R2d ). Then
p1 = p2 and q1 = q2 , and it follows from Proposition 2.4 and the first part of the
1,∞
proof that Wg,f ∈ M(1/ω 0)
, and that (2.7) holds. In particular,
|(OpA (a)f, g))| f p ,q1
M(ω1 g p ,q
1 1
,
1) M(1/ω ) 2
∞,1
and the result follows when a ∈ S1/2 . The result now follows for general a ∈ M(ω 0)
,
by taking a sequence {aj }j≥1 in S1/2 , which converges narrowly to a. (For narrow
convergence see Theorems 4.15 and 4.19, and Proposition 4.16 in [37].)
p,q
It remains to prove that if a belongs to the closure of S1/2 under M(ω) norm,
p1 ,q1 p2 ,q2
then OpA (a) : M(ω1 ) → M(ω2 ) is compact. As a consequence of Theorem 2.3, it
230 J. Toft
follows that OpA (a0 ) is compact when a0 ∈ S1/2 , since S1/2 → M(ω
1
0)
when ω0 ∈
PE , and that every trace-class operator is compact. The compactness of OpA (a)
now follows by approximating a with elements in S1/2 . The proof is complete.
For the proof of Theorem 2.3 we need the following extension of Theorem
4.12 in [38].
Proposition 2.7. Let A ∈ M(d, R), p ∈ [1, ∞) and that H1 , H2 are tempered
Hilbert spaces on Rd . Then the L2 form on S (R2d ) extends uniquely to a duality
between sA,p (H1 , H2 ) and sA,p (H1 , H2 ), and the dual space for sA,p (H1 , H2 ) can
be identified with sA,p (H1 , H2 ) through this form. Moreover, if ∈ sA,p (H1 , H2 )∗
and a ∈ sA,p (H1 , H2 ) are such that (b) = (a, b)L2 when b ∈ sA,p (H1 , H2 ), then
= a sA,p (H1 ,H2 ) .
Proof. The result follows from Theorem 4.12 in the case A = 0. For general A, the
result now follows from Proposition 1.1 and the fact that eiADξ ,Dx is unitary on
L2 (R2d ).
Proof of Theorem 2.3. The first inclusion in
∞,1 ∞
M(ω 0)
→ sA,∞ (ω1 , ω2 ) → M(ω 0)
follows from Theorem 2.2, and the second one from Proposition 1.2 (2) and The-
orem 2.1.
By Propositions 1.2 (3), 1.4 and 2.7, (1.19), and duality, the latter inclusions
give
1 1,∞
M(ω 0)
→ sA,1 (ω1 , ω2 ) → M(ω 0)
,
and we have proved the result when p = 1 and when p = ∞. Furthermore, by
2
Proposition 2.5 we have M(ω 0)
= sA,2 (ω1 , ω2 ), and the result also holds in the case
p = 2. The result now follows for general p from these cases and interpolation.
(See, e.g., Proposition 5.8 in [37].) The proof is complete.
The next result shows that the operator eiADξ ,Dx is bijective between suit-
able modulation spaces. (See also [32, 36, 38] for similar results in restricted cases.)
Proposition 2.8. Let s ≥ 1
2, A ∈ M(d, R), p, q ∈ (0, ∞], φ, a ∈ Ss (R2d ) and let
TA = eiADξ ,Dx . Then
(VTA φ (TA a))(x, ξ, η, y) = eiAy,η (Vφ a)(x + Ay, ξ + A∗ η, η, y). (2.12)
Furthermore, if ω ∈ PE (R ) and
4d
ωA (x, ξ, η, y) = ω(x + Ay, ξ + A∗ η, η, y),

then TA from Ss (R2d ) to Ss (R2d ) extends uniquely to a homeomorphism from
p,q p,q
M(ω) (R2d ) to M(ω A)
(R2d ), and
TA a p,q
M(ω a p,q .
M(ω) (2.13)
A)
Proof. The formula (2.12) follows by a straightforward application of Fourier’s

inversion formula (cf. [33, Proposition 1.5] and its proof). The estimate (2.13)
then follows by first choosing φ in Σ1 (R2d ) \ 0 in (2.12), then multiplying this
equation by ωA and thereafter applying the mixed Lp,q quasi-norm. This gives the
result.
2.1. Composition properties
Let s ≥ 12 , A ∈ M(d, R) and a1 , . . . , aN ∈ Ss (R2d ). Then the N -linear product
(a1 , . . . , aN ) → a1 #A · · · #A aN (2.14)
is defined by the formula
OpA (a1 ) ◦ · · · ◦ OpA (aN ) = OpA (a1 #A · · · #A aN ).
The N -linear product (2.14) extends in different ways. Here below we give examples
on extensions in the framework of modulation space theory.
By a straightforward application of Proposition 1.1 we have
−1

a1 #A · · · #A aN = TA−B (TA−B a1 )#B · · · #B (TA−B aN ) , TA ≡ eiADξ ,Dx ,
(2.15)
for A, B ∈ M(d, R) and suitable a1 , . . . , aN .
We have now the following results on compositions. Here it is assumed that
the weight functions should obey

N
1 ω0 (TA (XN , X0 )) ωj (TA (Xj , Xj−1 )), X0 , . . . , XN ∈ R2d , (2.16)
j=1
where
TA (X, Y ) = (y + A(x − y), ξ + A∗ (η − ξ), η − ξ, x − y),
(2.17)
X = (x, ξ) ∈ R2d , Y = (y, η) ∈ R2d .
As in [7] we also let RN (p) with p = (p0 , . . . , pN ) ∈ [1, ∞]N +1 be the Hölder-Young
functional ⎛ ⎞
N
1
RN (p) = (N − 1)−1 ⎝ − 1⎠ ,
p
j=0 j (2.18)
p = (p0 , p1 , . . . , pN ) ∈ [1, ∞]N +1 .
Theorem 2.9. Let s ≥ 1

2, A ∈ M(d, R), pj , qj ∈ [1, ∞], j = 0, 1, . . . , N , and
suppose
1 1
max (RN (q ), 0) ≤ min , , RN (p) . (2.19)
j=0,1,...,N pj qj
Let ωj ∈ PE (R4d ), j = 0, 1, . . . , N , and suppose (2.16) holds. Then the map
(2.14) from Ss (R2d ) × · · · × Ss (R2d ) to Ss (R2d ) extends uniquely to a continuous
p1 ,q1 pN ,qN p0 ,q0
and associative map from M(ω 1)
(R2d ) × · · · × M(ω N)
(R2d ) to M(1/ω 0)
(R2d ).
232 J. Toft
Theorem 2.10. Let s ≥ 1

2, A ∈ M(d, R), pj , qj ∈ [1, ∞], j = 0, 1, . . . , N , and
suppose
1 1 1
RN (p) ≥ 0 and ≤ ≤ . (2.20)
qj pj 2
Let ωj ∈ PE (R4d ), j = 0, 1, . . . , N , and suppose (2.16) holds. Then the map
(2.14) from Ss (R2d ) × · · · × Ss (R2d ) to Ss (R2d ) extends uniquely to a continuous
p1 ,q1 pN ,qN p0 ,q0
and associative map from M(ω 1)
(R2d ) × · · · × M(ω N)
(R2d ) to M(1/ω 0)
(R2d ).
Proof of Theorems 2.9 and 2.10. The result follows immediately from Theorems
0.1 and 2.9 in [7] in the Weyl case, A = 12 I. For general A the result now follows by
from the Weyl case and straightforward applications of (2.15) and Proposition 2.8.

3. An idea of quantization
In this section we make a suitable average of OpA (a) with respect to the matrix
A to deduce certain types of operators, related to the symbols a, and which might
be of interests in quantizations.
We recall that a quantization is a rule which takes an observable a(x, ξ) in
classical mechanics to the corresponding observable Op(a) in quantum mechanics.
Usually, a is a suitable function or distribution defined on the phase space R2d ,
and Op(a) is an operator which acts on suitable dense subspaces of L2 (Rd ).
A common quantization is the Weyl quantization, a → Opw (a), explained
earlier. Another quantization rule is the Born–Jordan quantization, a → OpBJ (a),
where
1 1/2
OpBJ (a) = Opt (a) dt = Op(t+ 12 )·I (a) dt, (3.1)
0 −1/2
provided the right-hand side makes sense. By straightforward computations it
follows that
OpBJ (a) = Opw (Φ ∗ a), Φ(x, ξ) = sinc(x, ξ/2).
We shall now consider other candidates of quantization, where the average
on the right-hand side in (3.1) over all matrices t · I, 0 ≤ t ≤ 1, is replaced by
averages over all r · U , U ∈ UN(d, R), and r is fixed or taken over certain interval
I ⊆ R+ . Here UN(d, R) = UNd is the set of all d × d orthonormal matrices with
entries in R.
More precisely, for fixed r ≥ 0, we let
−1
Opr,UN (a) ≡ dU OprU+ 12 I (a) dU,
U∈UNd U∈UNd
r
Op0r,UN (a) ≡ r−1 Opt,UN (a) dt,
0
⎧
⎪
⎨(−i) d−2 J(d−2)/2 (iρ)
2 Γ( ) ·
d
, d>1
ψd (ρ) ≡ 2 d−2
ρ 2
⎪
⎩cosh(ρ), d = 1,
and r
ψ0,d (r) ≡ ψd (t) dt
0
where Jν is the Bessel function of order ν ∈ R.
By straightforward computations it follows that
Opr,UN (a) = Opw (a ∗ Ψr ), where Ψr (x, ξ) = ψd (r|x| |ξ|),
and
ψ0,d (r|x| |ξ|)
Op0r,UN (a) = Opw (a ∗ Ψ0r ), where Ψ0r (x, ξ) = .
r|x| |ξ|
Acknowledgement
I am very grateful to Ville Turunen for careful reading and important advices,
leading to several improvements of the original paper.
References
[1] D. Bayer, Bilinear time-frequency distributions and pseudodifferential operators, The-
sis, Universität Wien, Wien, 2010.
[2] Birman, Solomyak, Estimates for the singular numbers of integral operators (Rus-
sian), Uspehi Mat. Nauk. 32, (1977), 17–84.
[3] M. Cappiello, J. Toft, Pseudo-differential operators in a Gelfand–Shilov setting,
Math. Nachr. (to appear).
[4] E. Cordero, Gelfand–Shilov window classes for weighted modulation spaces, Integral
Transforms Spec. Funct. 18 (2007), 829–837.
[5] E. Cordero, J. Toft, P. Wahlberg, Sharp results for the Weyl product on modulation
spaces, J. Funct. Anal. 267 (2014), 3016–3057.
[6] E. Cordero, S. Pilipović, L. Rodino, N. Teofanov, Quasianalytic Gelfand–Shilov
spaces with applications to localization operators, Rocky Mt. J. Math. 40 (2010),
1123–1147.
[7] E. Cordero, J. Toft, P. Wahlberg, Sharp results for the Weyl product on modulation
spaces, J. Funct. Anal. 267 (2014), 3016–3057.
[8] M. de Gosson, Symplectic covariance properties for Shubin and Born–Jordan pseudo-
differential operators, Trans. Amer. Math. Soc. 365 (2012), 3287–3307.
[9] M. de Gosson, F. Luef, Preferred Quantization Rules: Born–Jordan vs. Weyl; Appli-
cations to Phase Space Quantization, J. Pseudo-Differ. Oper. Appl. 2 (2011), 115–
139.
[10] H.G. Feichtinger, Banach convolution algebras of Wiener’s type, in: Proc. Functions,
Series, Operators in Budapest, Colloquia Math. Soc. J. Bolyai, North-Holland Publ.
Co., Amsterdam Oxford NewYork, 1980.
234 J. Toft
[11] H.G. Feichtinger, Modulation spaces on locally compact abelian groups. Technical
report, University of Vienna, Vienna, 1983; also in: M. Krishna, R. Radha, S.
Thangavelu (eds.) Wavelets and their applications, Allied Publishers Private Lim-
ited, New Delhi Mumbai Kolkata Chennai Hagpur Ahmedabad Bangalore Hyderbad
Lucknow, 2003, pp. 99–140.
[12] H.G. Feichtinger and K.H. Gröchenig, Banach spaces related to integrable group rep-
resentations and their atomic decompositions, I, J. Funct. Anal. 86 (1989), 307–340.
[13] H.G. Feichtinger and K.H. Gröchenig, Banach spaces related to integrable group rep-
resentations and their atomic decompositions, II, Monatsh. Math. 108 (1989), 129–
148.
[14] H.G. Feichtinger and K.H. Gröchenig, Gabor frames and time-frequency analysis of
distributions, J. Functional Anal. (2) 146 (1997), 464–495.
[15] Y. Galperin, S. Samarah, Paper on quasi-Banach Gabor frames
[16] I.M. Gelfand, G.E. Shilov, Generalized functions, I–III, Academic Press, NewYork
London, 1968.
[17] K.H. Gröchenig, Foundations of Time-Frequency Analysis, Birkhäuser, Boston, 2001.
[18] K. Gröchenig, Weight functions in time-frequency analysis in: L. Rodino, M.W.
Wong (eds.) Pseudodifferential Operators: Partial Differential Equations and Time-
Frequency Analysis, Fields Institute Comm., 52 2007, pp. 343–366.
[19] K.H. Gröchenig and C. Heil, Modulation spaces and pseudo-differential operators,
Integral Equations Operator Theory (4) 34 (1999), 439–457.
[20] K.H. Gröchenig and C. Heil, Modulation spaces as symbol classes for pseudodiffer-
ential operators in: M. Krishna, R. Radha, S. Thangavelu (eds.) Wavelets and their
applications, Allied Publishers Private Limited, New Delhi Mumbai Kolkata Chennai
Hagpur Ahmedabad Bangalore Hyderbad Lucknow, 2003, pp. 151–170.
[21] K. Gröchenig and G. Zimmermann, Spaces of test functions via the STFT, J. Funct.
Spaces Appl. 2 (2004), 25–53.
[22] A. Holst, J. Toft, P. Wahlberg, Weyl product algebras and modulation spaces, J.
Funct Anal. 251 (2007), 463–491.
[23] L. Hörmander, The Analysis of Linear Partial Differential Operators, vol. I, III,
Springer-Verlag, Berlin, Heidelberg, New York, Tokyo, 1983, 1985.
[24] H. Komatsu, Ultradistributions. I, J. Fac. Sci. Univ. Tokyo Sect. IA Math. 20 (1973),
25–205.
[25] Z. Lozanov-Crvenković, D. Perišić, M. Tasković, Gelfand–Shilov spaces structural
and kernel theorems, (preprint), arXiv:0706.2268v2.
[26] S. Pilipović, Tempered ultradistributions, Boll. Un. Mat. Ital. B (7) 2 (1988), 235–251
[27] S. Pilipović, N. Teofanov, Wilson Bases and Ultramodulation Spaces, Math. Nachr.
242 (2002), 179–196.
[28] M.A. Shubin, Pseudodifferential operators and spectral theory, Springer-Verlag,
Berlin, 1987.
[29] B. Simon, Trace ideals and their applications, I, London Math. Soc. Lecture Note
Series, Cambridge University Press, Cambridge London New York Melbourne, 1979.
[30] J. Sjöstrand, An algebra of pseudodifferential operators, Math. Res. L. 1 (1994),
185–192.
[31] N. Teofanov, Modulation spaces, Gelfand–Shilov spaces and pseudodifferential oper-

ators, Sampl. Theory Signal Image Process, 5 (2006), 225–242.
[32] J. Toft, Continuity properties for modulation spaces with applications to pseudo-
differential calculus, I, J. Funct. Anal. (2), 207 (2004), 399–429.
[33] J. Toft, Continuity properties for modulation spaces with applications to pseudo-
differential calculus, II, Ann. Global Anal. Geom., 26 (2004), 73–106.
[34] J. Toft, Convolution and embeddings for weighted modulation spaces in: P. Boggiatto,
R. Ashino, M.W. Wong (eds.) Advances in Pseudo-Differential Operators, Operator
Theory: Advances and Applications 155, Birkhäuser Verlag, Basel 2004, pp. 165–186.
[35] J. Toft, Continuity and Schatten properties for pseudo-differential operators on mod-
ulation spaces in: J. Toft, M.W. Wong, H. Zhu (eds.) Modern Trends in Pseudo-
Differential Operators, Operator Theory: Advances and Applications, Birkhäuser
Verlag, Basel, 2007, 173–206.
[36] J. Toft, Continuity and Schatten properties for Toeplitz operators on modulation
spaces in: J. Toft, M.W. Wong, H. Zhu (eds.) Modern Trends in Pseudo-Differential
Operators, Operator Theory: Advances and Applications, Birkhäuser Verlag, Basel,
2007, pp. 313–328.
[37] J. Toft, The Bargmann transform on modulation and Gelfand–Shilov spaces, with
applications to Toeplitz and pseudo-differential operators, J. Pseudo-Differ. Oper.
Appl., 3 (2012), 145–227.
[38] J. Toft, Multiplication properties in Gelfand–Shilov pseudo-differential calculus in:
S. Molahajloo, S. Pilipović, J. Toft, M.W. Wong (eds.) Pseudo-Differential Opera-
tors, Generalized Functions and Asymptotics, Operator Theory: Advances and Ap-
plications 231, Birkhäuser, Basel Heidelberg NewYork Dordrecht London, 2013, pp.
117–172.
[39] J. Toft, Gabor analysis for a broad class of quasi-Banach modulation spaces in:
S. Pilipović, J. Toft (eds.), Pseudo-differential operators, generalized functions, Op-
erator Theory: Advances and Applications 245, Birkhäuser, 2015, 249–278.
[40] G. Tranquilli, Global normal forms and global properties in function spaces for second
order Shubin type operators PhD Thesis, 2013.
Joachim Toft
Department of Mathematics and Systems Engineering
Växjö University, Sweden
Operator Theory:

An Application of Internal Objects to Micro-

local Analysis in Generalized Function Algebras
H. Vernaeve
Abstract. We illustrate the use of internal objects in the nonlinear theory

of generalized functions by means of an application to microlocal analysis in
Colombeau algebras.
1. Introduction
As is well known, the Colombeau algebras G(Ω) [1, 2, 9] are differential algebras
of generalized functions containing the space of Schwartz distributions. They have
found diverse applications in the study of partial differential equations [10, 11, 12],
providing a framework in which nonlinear equations and equations with strongly
singular data or coefficients can be solved and in which their regularity can be
analyzed.
The natural extension of microlocal analysis of Schwartz distributions to the
Colombeau generalized function algebras is the so-called G ∞ -microlocal analysis,
which has been developed using the concept of G ∞ -regularity [3, 5, 6, 7, 11].
Recently, we introduced a refinement called G ∞ -microlocal analysis in which one
can consider microlocal regularity at generalized points (x0 , ξ0 ) in the cotangent
bundle of the domain [21]. Also the neighbourhoods which can be considered have
generalized (infinitesimal) radii. The reason to introduce this refinement is the
following: since equations with strongly singular data or coefficients in Colombeau
algebras are modeled by regularization, the corresponding differential operators
themselves become generalized operators. Hence it is to be expected that the most
suitable setting to study the propagation of singularities under such operators is
by means of generalized objects (generalized characteristic varieties, etc.).
The development of G ∞ -microlocal analysis has been obtained using the prin-
ciples introduced in [19, 20] which originate from nonstandard analysis, although
Supported by grant 1.5.138.13N of the Research Foundation Flanders FWO..

238 H. Vernaeve
[21] does not require the knowledge of these principles from the reader. That writ-
ing style comes with two disadvantages: a loss of intuition, and rather technical
proofs. In this paper, we show that these principles are developed well enough
to write the complete proofs of these recent research results, thus revealing the
underlying intuition more clearly. At several places (especially in Theorem 6), we
were able to significantly simplify the proofs in [21]. We hope that this may serve
as an example for researchers in the field, helping them to use the same techniques
also in their own work.
To keep the paper self-contained, we recall the definitions of the internal
objects and the principles that are used.
2. Internal objects
The internal objects that we will consider are closely related to the approaches in
[4] and [16] (see also [17]).
Let ∅ = A ⊆ Rd and I := (0, 1] ⊆ R. We denote
∗
A := AI /∼
where ∼ is the equivalence relation defined by
(xε )ε∈I ∼ (yε )ε∈I ⇐⇒ (∃ε0 )(∀ε ∈ (0, ε0 ])(xε = yε )
which we read as: xε = yε for (sufficiently) small ε. We denote by [xε ] ∈ ∗ A the
equivalence class of (xε )ε ∈ AI . By definition, elements of ∗ A are called internal.
We denote ∗ a := [a] (the equivalence class of the constant family (a)ε∈I ). Since
this defines an injection ∗ : A → ∗ A, we will identify a ∈ A with ∗ a ∈ ∗ A. It is clear
∗
that (Rd ) and (∗ R)d are isomorphic, and we will identify both.

For any map f : A → Rd , there is a canonical extension

∗
f : ∗ A → ∗ Rd : f ([xε ]) := [f (xε )]. (1)
∗
Since it is a canonical extension, it is customary to write f := f .
If ∅ = Aε ⊆ Rd (for each ε ∈ I), we define a set
[Aε ] := {[xε ] : xε ∈ Aε , for small ε} ⊆ ∗ Rd .
By definition, such subsets are called internal. In particular, ∗ A = [A] is internal.
More generally than in equation (1), if ∅ = Aε ⊆ Rd and fε are maps Aε →
d
R (for each ε ∈ I), we define a map

[fε ] : [Aε ] → ∗ Rd : [fε ]([xε ]) := [fε (xε )].
By definition, such maps are called internal.
Any binary relation R on A has an extension on ∗ A (which is also called
internal):
[xε ] ∗ R [yε ] ⇐⇒ xε R yε , for small ε.
Internal Objects and Microlocal Analysis 239
Since it is a canonical extension, it is customary to write R := ∗ R. Caution:

∗
x (¬R)y is not equivalent with ¬(x ∗ R y), hence we will not drop stars in ∗ =, ∗ ≤,
. . . to avoid ambiguities (in contrast, we use x = y for ¬(x = y)).

If Xε are nonempty sets of maps A → Rd (for each ε ∈ I), we define
[Xε ] := {[fε ] : fε ∈ Xε , for small ε}.
By definition, [Xε ] consists of internal maps only. We define again ∗ X := [X].
Then for any f ∈ X, its canonical extension ∗ f belongs to ∗ X. In this paper, we
will mainly consider ∗ X for a function space X ⊆ C ∞ (Rd ).
We denote by P∅ (A) the set of all nonempty subsets of A. If ∅ = B ⊆ P∅ (Rd ),
we define
∗
B := {[Aε ] : Aε ∈ B, for small ε}.
∗
Thus (P∅ (R )) is the set of all internal subsets of ∗ Rd as defined above (notice
d
that ∅ is not internal by definition).

More generally, if (for each ε) fε are maps X1 × · · · × Xm → Y1 × · · · × Yn ,
where each Xi and Yj either is a nonempty subset of Rd or a nonempty set of

maps A → Rd or a nonempty subset of P∅ (B) (where d, d and A, B ⊆ Rd may
depend on i, j), we define a map

[fε ] : ∗ X 1 × · · · × ∗ X m → ∗ Y 1 × · · · × ∗ Y n : [fε ] [x1,ε ], . . . , [xm,ε ]
) *
:= fε (x1,ε , . . . , xm,ε )
where we identify (Y1 × · · · × Yn ) ∼
∗
= ∗ Y 1 × · · · × ∗ Y n , i.e., we identify

[fε (xε )] = [f1,ε (xε )], . . . , [fm,ε (xε )] .
By definition, also such maps are called internal. We define again ∗ f := [f ].
E.g., in this paper, we will use
∗
∗
A ∪ B, ∗
sup A, f, ∗ ∂ α f and ∗ F (f )
A
for internal A, B, α, f , where we consider
∪ :P∅ (Rd ) × P∅ (Rd ) → P∅ (Rd ) ∂ :Nd × C ∞ (Rd ) → C ∞ (Rd )
sup :{A ⊆ R : A = ∅ is bounded} → R F :S (Rd ) → S (Rd )

:{A ⊆ Rd : A = ∅ measurable} × L1 (Rd ) → C.
As differentiation, integration and Fourier transform

are among the most basic op-
∗
erations in analysis, we will write ∂ := ∗ ∂, := and F := ∗ F (this corresponds
to the usual ε-wise definitions in Colombeau theory). As ∗ sup is the supremum
for the partial order on ∗ R, we will write sup := ∗ sup. However, as ∗ ∪ is not the
set-theoretic union, we will not drop stars here.
The above construction can be inductively extended to larger classes of ob-
jects, all of which are called internal [19], but the above definitions suffice for this
paper.
240 H. Vernaeve
3. Principles from nonstandard analysis

Proving properties about internal objects is considerably simplified through the
following principles [19].
Definition 1. An object a is called transferrable if ∗ a is defined.
Transferrable formulas are formal expressions containing symbols called vari-
ables. Particular kinds of variables are relation variables and function variables.
Inductively, terms are defined by the following rules:
1. A variable is a term.
2. If t1 , . . . , tm are terms (m > 1), then also (t1 , . . . , tm ) is a term.
3. If t is a term and f is a function variable, then also f (t) is a term.
Transferrable formulas are defined by the following rules:
F1. (atomic formulas) If t1 , t2 are terms and R is a relation variable, then t1 = t2 ,
t1 ∈ t2 and t1 R t2 are formulas.
F2. If P , Q are formulas, then P & Q is a formula.
F3. If P is a formula, x is a variable free in P and t is a term in which x does
not occur, then (∃x ∈ t)P is a formula.
F4. If P is a formula, x is a variable free in P and t is a term in which x does
not occur, then (∀x ∈ t)P is a formula.
F5. If P , Q are formulas, then [(∃x ∈ t)P ] & [(∀x ∈ t)(P ⇒ Q)] is a formula.
In practice, we will use rule [F5] in formulas by simply writing (∀x ∈ t)(P ⇒
Q), silently checking that the side condition (∃x ∈ t)P is fulfilled.
Notation 1. We write P (x1 , . . . , xm ) for a formula P in which the only occurring
free variables are x1 , . . . , xm . We denote by P (c1 , . . . , cm ) the formula P in which
the variable xj has been substituted by the object cj (for j = 1, . . . , m). In this
case, cj are called the constants occurring in the formula. Relation variables (resp.
function variables) can only be substituted by (binary) relations (resp. functions).
The slight ambiguity that might result from these notations is similar to the
notation for a function f (x) and its value f (c), and is clarified by the context.
An internal formula is a transferrable formula in which all constants are
internal.
Notice that disjunction (∨) and negation (¬) are not allowed in the formation
rules. Also, we only allow bounded quantifiers (i.e., expressions ‘∀x’ and ‘∃x’ have
to be followed by ‘∈ t’). The reasons to consider this particular class of formulas
are:
Theorem 2 (Transfer Principle [19]). Let P (a1 , . . . , am ) be a transferrable formula,
in which the constants aj are transferrable objects. Then
P (a1 , . . . , am ) is true iff P (∗ a1 , . . . , ∗ am ) is true.
Remark 1. The transferrable formulas that we just defined are called h-formulas
in [15] and Palyutin formulas in [14]. Already [15] mentions the transfer principle
for these formulas.
Theorem 3 (Internal Definition Principle [19]). Let A be an internal set. Let P (x)
be an internal formula. If {x ∈ A : P (x)} = ∅, then {x ∈ A : P (x)} is internal.
Definition 2. We call x ∈ ∗ R infinitesimal if |x| ≤ 1/n for each n ∈ N. Notation:
x ≈ 0.
We call x ∈ ∗ R infinitely large if |x| ≥ n for each n ∈ N. Notation: x ∈ ∗ R∞ .
We also write A∞ := A ∩ ∗ R∞ for A ⊆ ∗ R.
Notice that N, R, ∗ R∞ and ≈ are external (i.e., not internal), and therefore
not allowed in internal formulas.
Theorem 4 (Spilling principles [19]). Let A ⊆ ∗ N be internal.
1. (Overspill) If N ⊆ A, then there exists ω ∈ ∗ N∞ such that {n ∈ ∗ N : n ≤
ω} ⊆ A.
2. (Underspill) If ∗ N∞ ⊆ A, then A ∩ N = ∅.
We will also use the following convenient version of the Saturation princi-
ple [19]:
Theorem 5 (Quantifier switching [19]). Let A be an internal set. For each n ∈ N,
let Pn (x), Qn (x) be internal formulas. If Pn gets stronger as n increases (i.e., for
each n ∈ N and x ∈ A, Pn+1 (x) ⇒ Pn (x)) and if
(∀n, m ∈ N)(∃x ∈ A)(Pn (x) & ¬Qm (x)),
then also
(∃x ∈ A)(∀n ∈ N)(Pn (x) & ¬Qn (x)).
4. Internal subsets of ∗ Rd and internal functions on ∗ Rd

If A ⊆ ∗ Rd , we call the exterior of A (cf. also [8])
ext(A) := {x ∈ ∗ Rd : x ∗ = a, ∀a ∈ A} = {x ∈ ∗ Rd : |x − a| > 0, ∀a ∈ A}.
(Recall that x > 0 means x ∗ > 0, i.e., xε > 0 for small ε.)
∗
If A is internal, then ext(A) = {x ∈ ∗ Rd : x ∈ / A}. By transfer on

(∀X ∈ P∅ (Rd ))(X ∈ P∅ (Rd ) \ {Rd } ⇐⇒ (∃y ∈ Rd )(y ∈ / X)
∗
we find that A ∈ (P∅ (Rd ) \ {Rd }) iff A ⊆ ∗ Rd is internal with ext(A) = ∅.
Let co(A) := Rd \ A. Considering co: P∅ (Rd ) \ {Rd } → P∅ (Rd ) \ {Rd }, by
transfer on
(∀X ∈ P∅ (Rd ) \ {Rd })(∀y ∈ Rd )(y ∈ co(X) ⇐⇒ y ∈
/ X)
we find that ext(A) = ∗ co(A) for each internal A ⊆ ∗ Rd with ext(A) = ∅.
If R is a ring and e ∈ R is idempotent (i.e., e2 = e), then we denote its
complement idempotent ec := 1 − e. We denote RId := {e ∈ R : e2 = e}.
242 H. Vernaeve
For any A ⊆ ∗ Rd , we denote its interleaved closure (cf. also [13])

m
m
∗
interl(A) := aj ej : m ∈ N, aj ∈ A, ej ∈ RId , ej = 1 .
j=1 j=1
∗
Again by transfer, an internal set A ⊆ R is closed under interleaving, i.e., A =
d
interl(A). For internal A, B ⊆ ∗ Rd ,

A ∗ ∪ B = interl(A ∪ B) = {xe + yec : x ∈ A, y ∈ B, e ∈ ∗ RId }
is internal. For internal A ⊆ ∗ Rd with ext(A) = ∅, we have that interl(A∪ext(A)) =
A ∗ ∪ ∗ co(A) = ∗ Rd .
Since, by transfer, ∗ co({x ∈ ∗ R : x ≤ 0}) = {x ∈ ∗ R : x > 0}, we have in
particular (although the order on ∗ R is not total)
(∀x ∈ ∗ R)(∃e ∈ ∗ RId )(xe ≤ 0 & xec > 0). (2)

An internal map u: A ⊆ ∗ Rd → ∗ Rd has interleaved values, i.e.,
(∀x, y ∈ A)(∀e ∈ ∗ RId )(u(xe + yec ) = u(x)e + u(y)ec )
or, equivalently,
(∀x, y ∈ A)(∀e ∈ ∗ RId )(xe = ye ⇒ u(x)e = u(y)e).
For A ⊆ E ⊆ Rd , we also denote extE (A) := E ∩ ext(A), the exterior of
A in E.
Lemma 1. Let E ⊆ ∗ Rd be internal. Let A be a countable union of internal subsets
of E and let extE (A) = ∅. Let P (x) be an internal formula. If P (x) holds for each
x ∈ A ∪ extE (A), then P (x) holds for each x ∈ E.
Proof. As {x ∈ E :P (x)} is internal, P (x) also holds for each x ∈ interl(A ∪
extE (A)). Let A = n∈N An with An ⊆ E internal. Let n ∈ N. If x ∈ extE (A),
then x ∗ ∈
/ A1 , . . . , x ∗ ∈
/ An , and thus, by transfer, x ∗ ∈/ A1 ∗ ∪ · · · ∗ ∪ An . Thus

(∃C ∈ ∗ P∅ (Rd )) (∃x ∈ E)(x ∗ ∈ / C) & A1 ∗ ∪ · · · ∗ ∪ An ⊆ C ⊆ E & (∀x ∈ C)P (x) .
By Quantifier switching, we find an internal C ⊆ E with extE (C) = ∅, A ⊆ C and
P (x), ∀x ∈ C. Then also
P (x), ∀x ∈ interl(C ∪ extE (A)) ⊇ interl(C ∪ extE (C)) = E.
Corollary 1. Let E ⊆ ∗ Rd be internal. Let (rn )n∈N be a monotone sequence in ∗ R

and r ∈ ∗ R. Let u be an internal map E → ∗ Rd . Let
A := {x ∈ E : |x| ≤ rn , for some n ∈ N} and B := {x ∈ E : |x| > rn , ∀n ∈ N}
or
A := {x ∈ E : |x| ≥ rn , for some n ∈ N} and B := {x ∈ E : |x| < rn , ∀n ∈ N}.
If A, B = ∅ and |u| ≤ r on A ∪ B, then |u| ≤ r on E.
Proof. Consider the first case (the second case is similar). Let An := {x ∈ ∗ Rd :
|x| ≤ rn } and Bn := {x ∈ ∗ Rd : |x| > rn }. As (rn )n is monotone and A = ∅,
w.l.o.g. An ∩ E = ∅, and thus An ∩ E are internal. In order to apply the previous
lemma with P (x) := (|u(x)| ≤ r), we show that B =extE (A). By transfer,
ext(An ) = ∗ co(An ) = Bn . The result follows, since ext( i Xi ) = i ext(Xi ) =
{x ∈ ∗ Rd : |x − y| > 0, ∀i, ∀x ∈ Xi } for any Xi ⊆ ∗ Rd .
In practice, we will often use the following convenient version of over- and
underspill:
Proposition 1 (Overspill). Let P (m) be an internal formula. Then
P (m) holds for sufficiently small m ∈ ∗ N∞
⇔ P (m) holds for sufficiently large m ∈ N,
i.e., (∃M0 ∈ N∞ )(∀m ∈ ∗ N∞ , m ≤ M0 )P (m) ⇐⇒ (∃m0 ∈ N)(∀m ∈ N, m ≥
∗
m0 )P (m).
Proof. Let A := {m ∈ ∗ N : P (m)}.
⇒: by assumption and eq. (2), ∗ N∞ ⊆ A ∗ ∪ {m ∈ ∗ N : m ≥ M0 }. Then also
∗

N∞ ⊆ B := n ∈ ∗ N : (∀m ∈ ∗ N, m ≥ n)(m ∈ A ∗ ∪ {m ∈ ∗ N : m ≥ M0 }) .
By underspill, B ∩ N = ∅. In particular, each sufficiently large m ∈ N belongs to
N ∩ (A ∗ ∪ {m ∈ ∗ N : m ≥ M0 }) = N ∩ A.
⇐: by assumption, N ⊆ A ∗ ∪ {m ∈ ∗ N : m ≤ m0 }. By overspill, each
sufficiently small m ∈ ∗ N∞ belongs to ∗ N∞ ∩ (A ∗ ∪ {m ∈ ∗ N : m ≤ m0 }) =
∗
N∞ ∩ A.
5. Moderateness and M∞ -regularity

We denote the infinitesimal ρ := [ε] > 0. We call x ∈ ∗ Rd moderate (notation:
x ∈ ∗ RdM ) if |x| ≤ ρ−N for some N ∈ N, and negligible (notation: x 0) if
|x| ≤ ρn for each n ∈ N.
In this paper, Ω ⊆ Rd will denote an open set. We call ∗ Ωc := K⊂⊂Ω ∗ K. We
denote K(Ω) := {K ⊆ Ω : K is compact, K = ∅}.
∗
Let K ∈ (K(Rd )). Let u be a map ∗ Rd → ∗ C. We say that u is supported in
∗
K (cf. also [8]; notation: supp(u) ⊆ K) if u = 0 on ext(K). For u ∈ (C ∞ (Rd )), this
∗
means (by transfer) that supp(u) ⊆ K (i.e., supp(uε ) ⊆ Kε for small ε). We say
∗
that u is compactly supported in A ⊆ ∗ Rd if supp(u) ⊆ K for some K ∈ (K(Rd ))
with K ⊆ A. We define
∗
M(A) := {u ∈ (C ∞ (Rd )) : ∂ α u(x) ∈ ∗ CM , ∀α ∈ Nd , ∀x ∈ A}
∗
M∞ (A) := {u ∈ (C ∞ (Rd )) : (∀x ∈ A)(∃N ∈ N)(∀α ∈ Nd )|∂ α u(x)| ≤ ρ−N }
Mc (A) := {u ∈ M(∗ Rd ) : u is compactly supported in A}
M∞ ∞
c (A) := M (A) ∩ Mc (A).
244 H. Vernaeve

Lemma 2. Let A = n∈N An with An ⊆ ∗ Rd internal. If B ⊆ A is internal, then
B ⊆ An for some n ∈ N.
Proof. Seeking a contradiction, suppose that (∀n ∈ N) (∃x ∈ B) (x ∈/ An ). By
Quantifier switching, there would exist x ∈ B such that x ∈
/ A.

Proposition 2. Let A = n∈N An with An ⊆ ∗ Rd internal. Let u be a map ∗ Rd →
∗
C. Then u is compactly supported in A iff u is compactly supported in An for
some n ∈ N.
∗
Proof. If K ∈ (K(Rd )) and K ⊆ A, then K ⊆ An for some n by the previous
lemma.
Corollary 2.
Mc (∗ Ωc ) = {u ∈ M(∗ Rd ) : (∃K ⊂⊂ Ω)(supp(u) ⊆ ∗ K)}.
∗
Proposition 3. Let A = n∈N An with An ⊆ ∗ Rd internal. Let u ∈ (C ∞ (Rd )).
Then
u ∈ M∞ (B) for some internal B ⊇ A
iff (∃N ∈ N)(∀α ∈ Nd )(∀x ∈ A)|∂ α u(x)| ≤ ρ−N .
Proof. ⇒: by Quantifier switching, as in [20, Prop. 7.6].
⇐: for each n ∈ N, there exists B ∈ ∗ P∅ (Rd ) such that
A1 ∗ ∪ · · · ∗ ∪ An ⊆ B & (∀x ∈ B)(∀α ∈ ∗ Nd , |α| ≤ n)|∂ α u(x)| ≤ ρ−N
since {(α, x) ∈ ∗ Nd × ∗ Rd : |∂ α u(x)| ≤ ρ−N } is closed under interleaving and
{α ∈ ∗ Nd : |α| ≤ n} is, by transfer, the interleaved closure of the finite set
{α ∈ Nd : |α| ≤ n}. The result then follows by quantifier switching.
Corollary 3.
M∞ (∗ Ωc )
∗
= {u ∈ (C ∞ (Rd )) : (∀K ⊂⊂ Ω)(∃N ∈ N)(∀α ∈ Nd )(∀x ∈ ∗ K)|∂ α u(x)| ≤ ρ−N }.
Thus (cf. also [20]) the Colombeau algebras G(Ω), G ∞ (Ω), Gc (Ω) and Gc∞ (Ω)
are quotients of M(∗ Ωc ), M∞ (∗ Ωc ), Mc (∗ Ωc ) and M∞ ∗
c ( Ωc ), respectively, modulo
N (∗ Ωc ) := {u ∈ ∗ (C ∞ (Rd )) : ∂ α u(x) 0, ∀α ∈ Nd , ∀x ∈ ∗ Ωc }.
We call x ∈ ∗ Rd fast scale if x belongs to
∗
Rdf s := {x ∈ ∗ Rd : (∃a ∈ R>0 )(|x| ≥ ρ−a )}
and we call x slow scale if x belongs to
∗
Rdss := {x ∈ ∗ Rd : (∀a ∈ R>0 )(|x| ≤ ρ−a )}.
We call x ∈ ∗ Rd a slow scale infinitesimal (notation: x ≈slow 0) if x ≈ 0 and 1
|x| is
slow scale, i.e., if
ρa ≤ |x| ≤ a, ∀a ∈ R>0
and we call x a fast scale infinitesimal (notation: x ≈fast 0) if

|x| ≤ ρa , for some a ∈ R>0 .
We write x ≈fast y (resp. x ≈slow y) for x − y ≈fast 0 (resp. x − y ≈slow 0).
We call a slow scale neighbourhood of x0 ∈ ∗ Rd any set that contains {x ∈
∗ d
R : |x − x0 | ≤ r} for some r ≈slow 0 (r ∈ ∗ R>0 ). We denote S := {x ∈ Rd : |x| =
1}. A conic slow scale neighbourhood of ξ0 ∈ ∗ S is a cone Γ ⊆ ∗ Rd with vertex 0
that contains a slow scale
ξ neighbourhood
of ξ0 (thus there exists some r ≈slow 0
(r ∈ ∗ R>0 ) such that |ξ| − ξ0 ≤ r ⇒ ξ ∈ Γ).
By Corollary 1, we obtain:
∗
Lemma 3. Let u ∈ (C ∞ (Rd )). If u 0 on ∗ Rdf s ∪ ∗ Rdss , then u 0 on ∗ Rd .
6. M∞ -microlocal regularity
Definition 3.
∗
MS (∗ Rd ) = {u ∈ (S (Rd )) : xα ∂ β u(x) ∈ ∗ CM , ∀x ∈ ∗ Rd , ∀α, β ∈ Nd }.
To keep this paper self-contained, we recast some properties concerning MS
and the Fourier transform in this setting (cf. also [18]):
Lemma 4.
1. Mc (∗ RdM ) ⊆ MS (∗ Rd ).
2. The Fourier transform F is a bijection MS (∗ Rd ) → MS (∗ Rd ).
3. Let u ∈ MS (∗ Rd ). Then for each k ∈ N, there exists m ∈ N s.t.

|u| ≤ ρk .
|x|≥ρ−m

4. Let u ∈ MS ( R ) and u(x) 0 for each x ∈ ∗ RdM . Then |u| 0.
∗ d
5. Let φ ∈ M∞ ∗ d ∗ d
c ( RM ). Then φ(ξ) 0 for all ξ ∈ Rf s .
∗ d ∗ d
∈ M∞ (∗ Rd ).
6. Let u ∈ MS ( R ). If u(x) 0 for all x ∈ Rf s , then u
∗ ∗
Proof. 1. By definition, Mc (∗ RdM ) ⊆ (Cc∞ (Rd )) ⊆ (S (Rd )). Let u ∈ Mc (∗ RdM ).
By Proposition 2, supp(u) ⊆ B(0, ρ−M ), for some M β∈ N. Thus if |x| > ρ−M ,
−M α β
x ∂ u(x) = 0. If |x| ≤ ρ
α β
, x ∂ u(x) ≤ ρ −M|α| ∂ u(x) ∈ RM . The result
∗
follows by Corollary 1.
2. Let u ∈ MS (∗ Rd ). As F : S → S is continuous, there exist C ∈ R and
N ∈ N such that for each α, β ∈ Nd (by transfer)
α β
sup ξ α ∂ β u
(ξ) ≤ C sup x ∂ u(x) ∈ ∗ RM .
ξ∈∗ Rd x∈∗ Rd ,|α |,|β |≤N
Hence u ∈ MS (∗ Rd ). The result follows by Fourier inversion.

3. If m ∈ ∗ N∞ , then

|x|≥ρ−m |u| ≤ C |x|≥ρ−m x
−d−2
dx ≤ Cρm ∗ R x−d−1 dx ≤ ρk (C ∈ ∗ RM ).
246 H. Vernaeve
The result follows by overspill.

4. Let k ∈ N and m ∈ N as in part (3). Then

|u| = |x|≥ρ−m |u| + |x|≤ρ−m |u| ≤ ρk + sup|x|≤ρ−m |u| · |x|≤ρ−m 1 ρk .
−M
5. By Proposition 2, supp(φ) ⊆
B(0, ρ ) for some M ∈ N. Let |ξ| ≥ ρ−1/N
d α
(N ∈ N). Then for each α ∈ N , ξ φ(ξ) ≤ |x|≤ρ−M |∂ φ| ≤ C ∈ ∗ RM (C is
α

independent of α). Then for any m ∈ N, |φ(ξ)| ≤ Cρ−1 |ξ|
−m
≤ Cρm/N −1 .
6. By overspill, there exists k ∈ ∗ N∞ such that |u(x)|
≤ ρk for each x ∈ ∗ Rd
with |x| ≥ ρ−1/k . Let α ∈ Nd . For a suitable m ∈ N, |x|≥ρ−m |xα u(x)| dx ≤ 1 by

part (3). Further, ρ−1/k ≤|x|≤ρ−m |xα u(x)| dx 0 and

|x|≤ρ−1/k
|xα u(x)| dx ≤ ρ|α|/k sup|x|≤ρ−1/k |u(x)| ≤ ρ−1 sup|x|≤ρ−1/k |u(x)| .

Thus for each ξ ∈ ∗ Rd , |∂ α u(ξ)| ≤ |xα u(x)| dx ≤ C ∈ ∗ RM (C is independent
of α).
Definition 4. u ∈ M(∗ Ωc ) is M∞ -microlocally regular at (x0 , ξ0 ) ∈ ∗ Ωc × ∗ S if
there exists v ∈ Mc (∗ Ωc ) such that
ξ
u(x) = v(x), ∀x ≈fast x0 and v(ξ) 0, ∀ξ ∈ ∗ Rdf s with ≈fast ξ0 .
|ξ|
Proposition 4. Let v ∈ Mc (∗ Ωc ), φ ∈ M∞ (∗ Ωc ) and ξ0 ∈ ∗ S. Let v(ξ) 0 for
each ξ ∈ ∗ Rdf s with |ξ|
ξ
≈fast ξ0 . Then also φv(ξ) 0 for each ξ ∈ ∗ Rdf s with
ξ
|ξ| ≈fast ξ0 .
∗ ∗ d
Proof. W.l.o.g., φ ∈ M∞
c ( Ωc ). Fix ξ ∈ Rf s with |ξ| ≈fast ξ0 . Then
ξ

φv(ξ) = φ(ξ − η) v (η) dη.
− η) 0 if ξ − η ∈ ∗ Rd .
By Lemma 4, φ(ξ fs
|ξ−η|
Now let ξ − η ∈ ∗ Rdss . Then also η ∈ ∗ Rdf s . Since |ξ| ≈fast 0,
η η ξ
≈fast ≈fast ≈fast ξ0 .
|η| |ξ| |ξ|
− η)
Hence v(η) 0. By Lemma 3, φ(ξ v (η) 0 for each η ∈ ∗ Rd . As φv ∈ MS ,
also φv
∈ MS . By Lemma 4, φv(ξ) 0.
Corollary 4. Let φ ∈ M∞ (∗ Ωc ). If u ∈ M(∗ Ωc ) is M∞ -microlocally regular at
(x0 , ξ0 ), then also φu is M∞ -microlocally regular at (x0 , ξ0 ).
We use the following notation. We fix φ0 ∈ D(B(0, 1)) with 0 ≤ φ0 ≤ 1 and
with φ0 (x) = 1 for each x ∈ B(0, 1/2). For m ∈ ∗ N and x0 ∈ ∗ Rd , we denote
x − x
0
φm,x0 (x) := φ0 .
ρ1/m
Proposition 5. For u ∈ M(∗ Ωc ) and (x0 , ξ0 ) ∈ ∗ Ωc × ∗ S, the following are equiva-

lent:
1. u is M∞ -microlocally regular at (x0 , ξ0 )
∗
2. there exists φ ∈ M∞
c ( Ωc ) such that
ξ
φ(x) = 1, ∀x ≈fast x0 and φu(ξ) 0, ∀ξ ∈ ∗ Rdf s with ≈fast ξ0 (3)
|ξ|
∗ ∗
3. there exists φ ∈ M∞ c ( Ωc ) and R ∈ Rss such that
#
|∂ α φ(x)| ≤ R ξ
∀x ≈fast x0 , ∀α ∈ Nd and φu(ξ) 0, ∀ξ ∈ ∗ Rdf s , ≈fast ξ0
|φ(x)| ≥ R ,
1 |ξ|

4. there exists m ∈ ∗ N∞ with m sufficiently small, such that φm,x0 ∈ Mc (∗ Ωc )
such that φ ∗ d
m,x0 u(ξ) 0, ∀ξ ∈ R with ξ ≈fast ξ0 .
fs |ξ|
Proof. (1) ⇒ (2): choose v as in the definition of M∞ -microlocal regularity. By

overspill, there exists some m ∈ ∗ N∞ such that u(x) = v(x) for each x ∈ ∗ Rd with
∗
|x − x0 | ≤ ρ1/m . For φ := φm,x0 ∈ M∞
c ( Ωc ), we have φ(x) = 1 for each x ≈fast x0

and φu = φv. By Proposition 4, φu(ξ)
= φv(ξ) 0 for each ξ ∈ ∗ Rdf s , |ξ|
ξ
≈fast ξ0 .
(2) ⇒ (3): trivial.
(3) ⇒ (4): by overspill, there exists m ∈ ∗ N∞ such that |φ(x)| ≥ 1/R and
|∂ α φ(x)| ≤ R for each x ∈ ∗ Rd with |x − x0 | ≤ ρ1/m and for each α ∈ ∗ Nd
∗ ∗
with |α| ≤ m. Then φm,x0 ∈ M∞ c ( Ωc ) (if m ∈ N∞ is sufficiently small) and

φm,x0
∈ M∞ ∗
c ( Ωc ), whence φm,x0 u(ξ) = F
φm,x0 ∗ d
φ φ φu (ξ) 0 for each ξ ∈ Rf s
with ξ
|ξ| ≈fast ξ0 by Proposition 4.
(4) ⇒ (1): let v := φm,x0 u ∈ Mc (∗ Ωc ).
7. Consistency with M∞ -regularity

We now proceed to show that the projection of the wave-front set in the first
coordinate is the singular support (Theorem 7).
Lemma 5. Let u ∈ M(∗ Ωc ) and V ⊆ ∗ Ωc be internal. Then there exists m0 ∈ ∗ N∞
such that φm,x u ∈ Mc (∗ Ωc ) for each x ∈ V and each m ≤ m0 (m ∈ ∗ N).
Proof. By Lemma 2, V ⊆ ∗ K for some compact K ⊂ Ω. Thus we can choose
m0 ∈ ∗ N∞ s.t. ρ1/m0 ≤ 12 d(K, Rd \ Ω) ∈ R>0 .
We first prove the following uniform version of Proposition 5:
Theorem 6. Let V ⊆ ∗ Ωc be internal and let Γ ⊆ ∗ Rd be an internal cone. Let
u ∈ M(∗ Ωc ) be M∞ -microlocally regular at (x0 , ξ0 ), for each x0 ∈ V and each

ξ0 ∈ ∗ S ∩ Γ. Then there exists k ∈ ∗ N∞ such that φk,x u(ξ) 0 for each x ∈ V and
∗ d
ξ ∈ Γ ∩ Rf s .
248 H. Vernaeve
Proof. Let m0 ∈ ∗ N∞ as in the previous lemma.

(1) For each k ∈ N, x0 ∈ V and ξ0 ∈ ∗ S ∩ Γ, there exists m ∈ ∗ N with
k < m ≤ m0 s.t.

∗ d ξ

|φ u(ξ)| ≤ ρ k
, ∀ξ ∈ R , − ξ
0 ≤ ρ
1/k
, |ξ| ≥ ρ−1/k
m,x0 |ξ|
since every m ∈ ∗ N∞ as in Proposition 5(4) satisfies this condition (m depends on

x0 , ξ0 ).
(2) For each k ∈ N, φk,0 φm,0 = φk,0 , for each m ∈ ∗ N with m > k.
By overspill, (1) and (2) simultaneously hold for some k ∈ ∗ N∞ (k does not
depend on x0 , ξ0 ). Then in particular φ ∗ d
m,x0 u(ξ) 0, for each ξ ∈ Rf s with
∞ ∗
|ξ| ≈fast ξ0 . Since φk,x0 ∈ M ( Ωc ), Proposition 4 shows that also φk,x0 u(ξ) =
ξ

F φk,x0 (φm,x0 u) (ξ) 0 for each ξ ∈ ∗ Rdf s with |ξ|
ξ
≈fast ξ0 . As x0 ∈ V and
∗
ξ0 ∈ S ∩ Γ are arbitrary, the result follows.
Remark 2. The previous proof indicates the need to go beyond the ring R of
generalized Colombeau numbers. Although one can also formulate an overspill
principle in this context [13], one cannot distinguish between ρk (k infinitely large)
and 0 in R.
Theorem 7. Let x0 ∈ ∗ Ωc . For u ∈ M(∗ Ωc ), the following are equivalent:

1. u is M∞ -microlocally regular at (x0 , ξ0 ), for each ξ0 ∈ ∗ S
2. u ∈ M∞ (V ) for some slow scale neighbourhood V of x0 .
Proof. (1) ⇒ (2): by Theorem 6 (with V := {x0 } and Γ := ∗ Rd ), we find k ∈ ∗ N∞
such that φ ∗ d ∗
k,x0 u(ξ) 0 for each ξ ∈ Rf s , and φk,x0 ∈ Mc ( Ωc ). Hence φk,x0 u ∈
∗ ∗ d ∞ ∗ d
Mc ( Ωc ) ⊆ MS ( R ). By Lemma 4, φk,x0 u ∈ M ( R ). As φk,x0 = 1 on a slow
scale neighbourhood V of x0 , also u ∈ M∞ (V ).
∗
(2) ⇒ (1): there exists φ ∈ M∞ c ( Ωc ) with φ = 1 on a slow scale neigh-
∞ ∗
bourhood of x0 and with φu ∈ Mc ( Ωc ) (e.g., φ = φm,x0 for a sufficiently small

m ∈ ∗ N∞ ). By Lemma 4, φu(ξ) 0 for each ξ ∈ ∗ Rdf s .
We can equivalently reformulate the condition in the previous theorem:
Proposition 6. For u ∈ M(∗ Ωc ), the following are equivalent:

1. u ∈ M∞ (V ) for some slow scale neighbourhood V of x0
2. (∃N ∈ N) (∀α ∈ Nd ) (∀x ∈ ∗ Rd , x ≈fast x0 ) (|∂ α u(x)| ≤ ρ−N )
Proof. For n ∈∗ N, let An := B(x0 , ρ1/n ). Then An is internal and {x ∈ ∗ Rd :

x ≈fast x0 } = n∈N An . Each slow scale neighbourhood of x0 and each internal
set containing {x ∈ ∗ Rd : x ≈fast x0 } contains Am for some m ∈ ∗ N∞ by overspill.
Thus the result follows by Proposition 3.
8. Connection with G ∞-microlocal regularity

Definition 5. We denote E(Rd ) := {A ⊆ Rd : A is finite, A = ∅}. Elements of
∗
(E(Rd )) are called hyperfinite subsets of ∗ Rd . Considering the number of elements
as a map #: E(Rd ) → N, we call ∗ #(A) ∈ ∗ N the number of elements of A ∈
∗
(E(Rd )).
Similarly, we can extend other operations
$ to hyperfinite sets, e.g., for A$∈
∗
(E(Rd )) and u: ∗ Rd → ∗ C internal, ∗ x∈A u(x) ∈ ∗ C, where we consider :
E(Rd ) × {u : $ Rd → C} → C. As the sum is one of the most basic operations,
∗$
we will write := . The usual calculation rules hold by transfer. (For A =
[A
$ ε ], A is hyperfinite
$ iff Aε is finite for small ε, and #A = [#Aε ]. For u = [uε ],
x∈A u(x) = [ x∈Aε u ε (x)], ...)
Theorem 8. Let u ∈ M(∗ Ωc ), x0 ∈ Ω and ξ0 ∈ S. Then the following are equiva-
lent:
1. there exists r ∈ R>0 such that u is M∞ -microlocally regular at (x, ξ) for each
x ∈ B∗ Rd (x0 , r) and each ξ ∈ ∗ S with |ξ − ξ0 | ≤ r
∗ ∗
2. there exist φ ∈ M∞ c ( Ωc ), r ∈ R>0 and R ∈ Rss such that |∂ φ(x)| ≤ R
α
and |φ(x)| ≥ 1/R for each α ∈ N and each x ∈ B∗ Rd (x0 , r), and
d

ξ

φu(ξ) 0, ∀ξ ∈ ∗ Rdf s , − ξ0 ≤ r
|ξ|
3. there exist ψ ∈ D(Ω) with ψ(x0 ) = 1 and a conic neighbourhood Γ ⊆ Rd of
ξ0 s.t.

ψu(ξ) 0, ∀ξ ∈ ∗ Γ ∩ ∗ Rdf s .

Proof. (1) ⇒ (2): Let V := B∗ Rd (x0 , r) and Γ := {ξ ∈ ∗ Rd : ξ − |ξ| ξ0 ≤ |ξ| r}.
By Theorem 6, we find k ∈ ∗ N∞ such that φ k,x u(ξ) 0 for each x ∈ V and
ξ ∈ Γ ∩ ∗ Rdf s . For convenience, we use the norm x ∞ := max{|x1 | , . . . , |xd |} on
Rd (and its extension to a map ∗ Rd → ∗ R). Consider a grid
0 r ρ1/k 1
G := x ∈ ∗ Rd : x − x0 ∞ < √ , x ∈ √ ∗ Zd .
d d
Then G is a hyperfinite set with at most (2rρ−1/k + 1)d ≤ ρ−(d+1)/k elements.
$
≤$
Let ψ := y∈G φk,y . Let ξ ∈ Γ ∩ ∗ Rdf s . As G ⊆ V , ψu(ξ)
y∈G |φk,y u(ξ)| ≤

ρn−(d+1)/k for each n ∈ N.Hence ψu(ξ) 0.
∗ d
Now let x ∈ W := x ∈ R : x − x0 ∞ ≤ 2√ r
d
arbitrary. Then there
1/k
exists y0 ∈ G such that x − y0 ∞ ≤ ρ2√d , hence ψ(x) ≥ φk,y0 (x) = 1. On the
other hand, let x ∈ ∗ Ωc . Then there is at most a finite number Cd ∈ N (independent
of x) of elements y ∈ G for which |x − y| ≤ ρ1/k . Hence for each α ∈ Nd ,
√
|∂ α ψ(x)| ≤ Cd sup |∂ α φk,y (x)| ≤ Cd,α ρ−|α|/k ≤ ρ−1/ k
=: R
y∈V
∗
(Cd,α ∈ R). In particular, ψ ∈ M∞
c ( Ωc ).
250 H. Vernaeve
(2) ⇒ (3): as in Proposition 5, we can find ψ ∈ D(B(x0 , r)) with ψ(x0 ) = 1

for which ψφ ∈ M∞ ∗ ∗ ∗ d
c ( Ωc ), and thus ψu(ξ) 0 for each ξ ∈ Γ ∩ Rf s (for some
conic neighbourhood Γ ⊆ Rd of ξ0 ) by Proposition 4.
∗
(3) ⇒ (1): by Proposition 5, as D(Ω) ⊆ M∞ c ( Ωc ).
The following lemma makes the connection with G ∞ - and G ∞ -microlocal
analysis of functions in G(Ω) [21].
For x ∈ ∗ Rd , we denote by x ∈ Rd the equivalence class of x modulo .
Similarly, for u ∈ M(∗ Ωc ), we denote by u ∈ G(Ω) the equivalence class of u
modulo N (∗ Ωc ).
Lemma 6. Let u ∈ M(∗ Ωc ), x0 ∈ ∗ Ωc and ξ0 ∈ ∗ S. Then u is M∞ -microlocally
regular at (x0 , ξ0 ) iff u is G ∞ -microlocally regular at (x0 , ξ0 ).
∗
Proof. u is M∞ -microlocally regular at (x0 , ξ0 ) iff there exists φ ∈ M∞
c ( Ωc )
which satisfies eq. (3). Then clearly also
φ(x) = 1 in C, ∀x ∈ Rd , x ≈fast x0
(4)
and F (φu)(ξ) = 0 in C, ∀ξ ∈ Rdf s , ξ/ |ξ| ≈fast ξ0 ,
i.e., u is G ∞ -microlocally regular at (x0 , ξ0 ).
Conversely, if φ ∈ Gc∞ (Ω) satisfies eq. (4), then we can find a representative
∗ ∗ d
φ ∈ M∞ c ( Ωc ) which satisfies |φ(x)| ≥ 1/R and |∂ φ(x)| ≤ R for each x ∈ R ,
α
x ≈fast x0 and α ∈ Nd , for some R ≈ 1. Further, φu(ξ) 0 for each ξ ∈ ∗ Rdf s ∩ ∗ RdM
∈ MS (Rd ), also φu(ξ)
with |ξ| ≈fast ξ0 . As φu
ξ 0 for each ξ ∈ ext(∗ RdM ). Let

n ∈ N and Bn := {ξ ∈ ∗ Rd : |ξ| ≥ ρ−1/n , |ξ| ξ
− ξ0 ≤ ρ1/n }. By Corollary 1,

φu(ξ)
0 for each ξ ∈ Bn . As n ∈ N is arbitrary, φu(ξ) 0 for each ξ ∈ ∗ Rd with
fs
ξ
|ξ| ≈fast ξ0 . Thus u is M∞ -microlocally regular at (x0 , ξ0 ) by Proposition 5.
Using this lemma, one easily recovers [21, Thms. 4.5 and 5.3] about Colom-
beau generalized functions by factorization modulo negligible elements from The-
orems 7 and 8.
Acknowledgment
We are grateful to P. Giordano for pointing out references to earlier work on
constructive versions of nonstandard principles.
References
[1] J.F. Colombeau, New Generalized Functions and Multiplication of Distributions,
North-Holland, Amsterdam, 1984.
[2] J.F. Colombeau, Elementary Introduction to New Generalized Functions, North-
Holland, Amsterdam, 1985.
[3] N. Djapić, S. Pilipović, D. Scarpalezos, Microlocal analysis of Colombeau’s general-
ized functions – propagation of singularities, J. Anal. Math. (1998) 75: 51–66.
[4] Y.V. Egorov, A contribution to the theory of generalized functions, Russian Math.
Surveys (1990) 45(5): 1–49. Translated from Russian.
[5] C. Garetto, T. Gramchev, M. Oberguggenberger, Pseudodifferential operators with
generalized symbols and regularity theory, Electronic J. Diff. Eqns. (2005) 116, 1–43.
[6] C. Garetto, G. Hörmann, Microlocal analysis of generalized functions: pseudodiffer-
ential techniques and propagation of singularities, Proc. Edinb. Math. Soc. (2005)
48, 603–629.
[7] C. Garetto, M. Oberguggenberger, Generalized Fourier integral operator methods
for hyperbolic equations with singularities, Proc. Edinburgh Math. Soc. (2014) 57,
423–463.
[8] P. Giordano, M. Kunzinger, A convenient notion of compact set for generalized func-
tions, Proc. Edin. Math. Soc., to appear.
[9] M. Grosser, M. Kunzinger, M. Oberguggenberger and R. Steinbauer, Geometric The-
ory of Generalized Functions, Kluwer Academic Publishers, Dordrecht, 2001.
[10] G. Hörmann and M.V. de Hoop, Microlocal analysis and global solutions of some
hyperbolic equations with discontinuous coefficients, Acta Appl. Math. (2001) 67:173–
224.
[11] G. Hörmann, M. Oberguggenberger and S. Pilipović, Microlocal hypoellipticity of
linear partial differential operators with generalized functions as coefficients, Trans.
Amer. Math. Soc. 358 (2006), 3363–3383.
[12] M. Oberguggenberger, Multiplication of Distributions and Applications to Partial
Differential Equations, Pitman Res. Not. Math. 259. Longman Scientific & Technical,
1992.
[13] M. Oberguggenberger, H. Vernaeve, Internal sets and internal functions in Colom-
beau theory, Monatsh. Math. (2008) 341: 649–659.
[14] E. Palmgren, A note on Brouwer’s weak continuity principle and the transfer prin-
ciple in nonstandard analysis, J. Logic Anal. (2012) 4:2, 1–7.
[15] E.A. Palyutin, Categorical Horn classes I, Algebra and Logic (1980) 19, 377–400.
[16] C. Schmieden, D. Laugwitz, Eine Erweiterung der Infinitesimalrechnung, Math.
Zeitschrift 69 (1958) 69: 1–39.
[17] T. Tao, A cheap version of nonstandard analysis,
https://terrytao.wordpress.com/2012/04/02/.
[18] H. Vernaeve, Pointwise characterizations in generalized function algebras, Monatsh.
Math. (2009) 158: 195–213.
[19] H. Vernaeve, Nonstandard principles for generalized functions, J. Math. Anal. Appl.
(2011) 384: 536–548.
[20] H. Vernaeve, Nonstandard principles for generalized functions, Novi Sad Lecture
notes, arxiv:1101.6075.
[21] H. Vernaeve, Microlocal analysis in generalized function algebras based on generalized
points and generalized directions, Monatsh. Math. (2016) 181: 205–215.
H. Vernaeve
Department of Mathematics, Ghent University
Krijgslaan 281, B-9000 Gent, Belgium
Operator Theory:

Rotation Invariant Ultradistributions

- ord̄e Vučković and Jasson Vindas
D
Abstract. We prove that an ultradistribution is rotation invariant if and only

if it coincides with its spherical mean. For it, we study the problem of spherical
representations of ultradistributions on Rn . Our results apply to both the
quasianalytic and the non-quasianalytic case.
Keywords. Rotation invariant, spherical means, ultradistributions, hyperfunc-

tions, spherical representations, spherical harmonics.
1. Introduction
Rotation invariant generalized functions have been studied by several authors,
see, e.g., [3, 18, 19]. The problem of the characterization of rotation invariant
ultradistributions and hyperfunctions was considered by Chung and Na in [3].
They showed there that a non-quasianalytic ultradistribution or a hyperfunction
is rotation invariant if and only if it is equal to its spherical mean. For continuous
functions this result is clear, as a rotation invariant function must be radial and
its spherical mean is given by

1
ϕS (x) = n−1 ϕ(|x|ω)dω.
|S | Sn−1
The approach of Chung and Na to the problem consists in reducing the case
of rotation invariant generalized functions to that of ordinary functions. For ul-
tradistributions, non-quasianalyticity was a crucial assumption for their method
since they regularized by convolving with a net of compactly supported ultrad-
ifferentiable mollifiers. In the hyperfunction case they applied a similar idea but
this time based on Matsuzawa’s heat kernel method.
The aim of this article is to show that the characterization of rotation invari-
ant ultradistributions in terms of their spherical means remains valid for quasian-
alytic ultradistributions. Our approach differs from that of Chung and Na, and we
also recover their results for non-quasianalytic ultradistributions and hyperfunc-
tions.
254 - . Vučković and J. Vindas
D
Our method is based upon the study of spherical representations of ultradis-

tributions, that is, the problem of representing an ultradistribution f on Rn by
an ultradistribution g on R×Sn−1 in such a way that f (x), ϕ(x) = g(r, ω), ϕ(rω).
Spherical representations of distributions were studied by Drozhzhinov and
Zav’yalov in [6]. We shall also exploit results on spherical harmonic expansions
of ultradifferentiable functions and ultradistributions on the unit sphere Sn−1 , re-
cently obtained by us in [20]. We mention that the theory of spherical harmonic
expansions of distributions was developed by Estrada and Kanwal in [7].
The plan of the article is as follows. Section 2 discusses some background
material on spherical harmonics and ultradistributions. Spherical representations
of ultradistributions are studied in Section 3. We show in Section 4 that any ultra-
distribution is rotation invariant if and only if it coincides with its spherical mean.
In the quasianalytic case we go beyond quasianalytic functionals by employing
sheaves of quasianalytic ultradistributions.
2. Preliminaries and auxiliary results

In this section we collect some useful concepts and auxiliary results that will play
a role in our study of rotation invariant ultradistributions and spherical represen-
tations.
2.1. Spherical harmonics

The theory of spherical harmonics is a classical subject in analysis and it is very
well explained in several textbooks (see, e.g., [1, Chap. 5]). A spherical harmonic of
degree j is simply the restriction to the Euclidean unit sphere Sn−1 of a harmonic
homogeneous polynomial of degree j on Rn . Let Hj (Sn−1 ) be the space of spherical
harmonics of degree j. The dimension dj of Hj (Sn−1 ) can be explicitly calculated
[1, Prop. 5.8]; although we will not make use of the explicit value, we need the
growth estimate dj j n−2 . We also point out that each Hj (Sn−1 ) is invariant
under the action of the orthogonal group O(n).
It is well known [1] that
∞
4
L2 (Sn−1 ) = Hj (Sn−1 ),
j=0
where the L -inner product is taken with respect to the surface measure of Sn−1 .
2
dj
Through the rest of the article we fix an orthonormal basis {Yk,j }k=1 of each
Hj (Sn−1 ), consisting of real-valued spherical harmonics. Hence, every function
f ∈ L2 (Sn−1 ) can be expanded as
∞
dj
f (ω) = ck,j Yk,j (ω)
j=0 k=1
with convergence in L2 (Sn−1 ).

Rotation Invariant Ultradistributions 255
2.2. Ultradistributions
We briefly review in the subsection some properties of the spaces of ultradifferen-
tiable functions and ultradistributions [2, 14, 15].
We fix a positive sequence (Mp )p∈N with M0 = 1. We will make use of some
of the following standard conditions on the weight sequence
(M.0) p! ⊂ Mp in the Roumieu case, or p! ≺ Mp in the Beurling case.
(M.1) Mp2 ≤ Mp−1 Mp+1 , p ≥ 1.
(M.2) Mp+1 ≤ AH p Mp , p ∈ N, for some A, H > 0.
(M.2) Mp ≤ AH p min1≤q≤p {Mq Mp−q }, p ∈ N, for some A, H > 0.
∞
Mp−1
(M.3) < ∞.
p=1
Mp
∞
Mp−1
(QA) = ∞.
p=1
Mp
We refer to [14] for a detailed explanation of the meaning of all these con-
ditions. The relations ⊂ and ≺ used in (M.0) are defined as follows. One writes
Np ⊂ Mp (Np ≺ Mp ) if there are C, > 0 (for each there is C = C ) such that
Np ≤ Cp Mp , p ∈ N. If (M.3) holds, we call Mp non-quasianalytic; otherwise it is
said to be quasianalytic. The associated function of the sequence is defined as
tp
M (t) = sup log , t > 0.
p∈N Mp
In the particular case of Gevrey sequences Mp = (p!)s , the associated function is
M (t) t1/s [12].
Let Ω ⊆ Rd be open. The space of all C ∞ -functions on Ω is denoted by E(Ω).
For K Ω (a compact subset with non-empty interior) and h > 0, one writes
E {Mp },h (K) for the space of all ϕ ∈ E(Ω) such that
|ϕ(α) (x)|
ϕ E {Mp },h (K) := sup < ∞,
x∈K h|α| M|α|
α∈Nn
{M },h
and DK p stands for the closed subspace of E {Mp },h (K) consisting of functions
with compact support in K (by the Denjoy–Carleman theorem, its non-triviality
is equivalent to (M.3) ). Set then
E {Mp } (Ω) = lim lim E {Mp },h (K), E (Mp ) (Ω) = lim lim E {Mp },h (K),
←− −→ ←− ←−+
KΩ h→∞ KΩ h→0
and
{Mp },h {Mp },h
D{Mp } (Ω) = lim lim DK D(Mp ) (Ω) = lim lim DK .
−→ −→ −→ ←−+
KΩ h→∞ KΩ h→0
Their duals are the spaces of ultradistributions of Roumieu and Beurling type [14].
In order to treat these spaces simultaneously we write ∗ = {Mp }, (Mp ).
In statements needing a separate treatment we will first state assertions for the
Roumieu case, followed by the Beurling one in parenthesis.
D
In the important case ∗ = {p!}, we write A(Ω) = E {p!} (Ω), the space of real
analytic functions on Ω; its dual A (Ω) is then the space of analytic functionals
on Ω. Note that (M.0) implies that A(Ω) ⊆ E ∗ (Ω), and, if in addition (M.1) and
(M.2) hold, A(Ω) is densely injected into E ∗ (Ω) because the polynomials are dense
in both spaces; in particular, E ∗ (Ω) ⊆ A (Ω) under these assumptions.
If one assumes (M.0) (as we will always do), the pullback of an invertible
analytic change of variables Ω → U becomes a TVS isomorphism between E ∗ (U )
and E ∗ (Ω) [11, Prop. 8.4.1]. Therefore, one can always define the spaces E ∗ (M )
and E ∗ (M ) for σ-locally compact analytic manifolds M via charts if (M.0) holds.
Note that (M.0) is automatically fulfilled if (M.1) and (M.3) hold [14, Lemma 4.1].
2.3. Ultradistributions on Sn−1 and spherical harmonics
The spaces of ultradifferentiable functions and ultradistributions on Sn−1 can be
described in terms of spherical harmonic expansions. A proof of the following theo-
rem will appear in our forthcoming paper [20], which also deals with ultradistribu-
tional boundary values of harmonic functions on the sphere. We point out that the
distribution case goes back to Estrada and Kanwal [7]. See also the forthcoming
article [4] for a treatment of the problem on compact analytic manifolds. We will
apply Theorem 1 in the next subsection to expand ultradifferentiable functions
and ultradistributions on R × Sn−1 in spherical harmonic series.
Theorem 1 ([20]). Suppose that Mp satisfies (M.0), (M.1), and (M.2) .
(i) Let ϕ ∈ L2 (Sn−1 ) have spherical harmonic expansion
∞
dj
ϕ(ω) = ak,j Yk,j (ω). (1)
j=0 k=1
Then, ϕ ∈ E ∗ (Sn−1 ) if and only if the estimate

sup |ak,j |eM( j/h) < ∞ (2)
k,j
holds for some h > 0 (for all h > 0).

(ii) Every ultradistribution f ∈ E ∗ (Sn−1 ) admits a spherical harmonic expansion
∞
dj
f (ω) = ck,j Yk,j (ω), (3)
j=0 k=1
where the coefficients satisfy the estimate

sup |ck,j |e−M( j/h) < ∞ (4)
k,j
for each h > 0 (for some h > 0). Conversely, any series (3) converges in
E ∗ (Sn−1 ) if the coefficients have the stated growth properties.
It is important to point out that Theorem 1 as stated above does not reveal
all topological information encoded by the spherical harmonic coefficients. Denote
{M },h
as Esh p (Sn−1 ) the Banach space of all (necessarily smooth) functions ϕ on Sn−1
having spherical harmonic expansion with coefficients ak,j satisfying (2) for a given
h. One can then show [20]
{Mp },h {Mp },h
E {Mp } (Sn−1 ) = lim Esh (Sn−1 ) and E (Mp ) (Sn−1 ) = lim Esh (Sn−1 )
−→ ←−+
h→∞ h→0
topologically. This for instance yields immediately the nuclearity of E ∗ (Sn−1 ) under
the assumptions of Theorem 1. Observe also that the norm on the Banach space
{M },h
Esh p (Sn−1 ) can be rewritten as

ϕ(ω)Yk,j (ω)dω .
j
M( h )
ϕ E {Mp},h (Sn−1 ) = sup e (5)
sh k,j n−1 S
A similar topological description can be given for the ultradistribution spaces

E ∗ (Sn−1 ) by using the coefficient estimates (4).
2.4. Ultradistributions on R × Sn−1
We also need some properties of the spaces E ∗ (R × Sn ) and E ∗ (R × Sn ). Let us
assume (M.0), (M.1), and (M.2). We have
∗ (Sn−1 ),
E ∗ (R × Sn−1 ) = E ∗ (R, E ∗ (Sn−1 )) = E ∗ (Sn−1 , E ∗ (R)) = E ∗ (R)⊗E
where the tensor product may be equally taken with respect to the π- or -topology
in view of the nuclearity of these spaces. In fact, the first two equalities are com-
pletely trivial, while the third one follows because the linear span of terms of the
form p ⊗ Y , where p is a polynomial on R and Y a spherical harmonic, is dense in
E ∗ (R × Sn−1 ). Moreover, this immediately gives (cf. (4)) that
{Mp },h
E {Mp } (R × Sn−1 ) = lim lim Esh (K × Sn−1 )
←− −→
KR h→∞
and
{Mp },h
E (Mp ) (R × Sn−1 ) = lim lim Esh (K × Sn−1 ),
←− ←−+
KR h→0
{Mp },h
where Esh (K × Sn−1 ) is the space of functions Φ such that
6 6
6 6
Φ E {Mp },h (K×Sn−1 ) = sup eM ( h ) 6 6
j
6 Φ( · , ω)Yk,j (ω)dω 6 {Mp },h < ∞. (6)

sh k,j n−1 S E (K)
These comments yield the following proposition.

Proposition 1. Assume Mp satisfies (M.0), (M.1), and (M.2).
(i) Every Φ ∈ E ∗ (R × Sn−1 ) has convergent expansion
∞
dj
Φ(r, ω) = ak,j (r)Yk,j (ω) in E ∗ (R × Sn−1 ),
j=0 k=1
where ak,j ∈ E ∗ (R) and for each K R

j
sup eM ( h ) ak,j E {Mp },h (K) <∞ (7)
k,j
D
for some h > 0 (for all h > 0). Conversely, any such series converges in the
space E ∗ (R × Sn−1 ) if (7) holds.
(ii) Every ultradistribution g ∈ E ∗ (R × Sn−1 ) has convergent expansion
∞
dj
g(r, ω) = ck,j (r) ⊗ Yk,j (ω) in E ∗ (R × Sn−1 ),
j=0 k=1
∗
where ck,j ∈ E (R) and for any bounded subset B ⊂ E ∗ (R) one has
j
sup e−M ( h ) sup |ck,j , ϕ| < ∞. (8)
k,j ϕ∈B
for each h (for some h). Conversely, any such series converges in the space
E ∗ (R × Sn−1 ) if (8) holds.

Proof. For (i), simply note that ak,j (r) = Sn−1 Φ(r, ω)Yk,j (ω)dω and so (6) is
the same as (7). The convergence of the series expansions of Φ is trivial to check
via the seminorms (6). Part (ii) follows from (i) and the canonical identification
E ∗ (R × Sn−1 ) = E ∗ (Sn−1 , E ∗ (R))(:= Lb (E ∗ (Sn−1 ), E ∗ (R))).
Note that the same proposition holds for D∗ (R × Sn ) if one additionally
assumes (M.3) .
3. Spherical representations of ultradistributions

It is easy to see that any g ∈ E ∗ (R × Sn−1 ) gives rise to an ultradistribution f on
Rn via the formula
f (x), ϕ(x) = g(r, ω), ϕ(rω). (9)
In fact, the assignment g → f is simply the transpose of
ϕ → Φ, Φ(r, ω) := ϕ(rω), (10)
∗ ∗
which is obviously continuous E (R ) → E (R × S
n
).n−1
In this section we study the converse representation problem. That is, the
problem of representing an f ∈ E ∗ (Rn ) as in (9) for some ultradistribution g on
R × Sn−1 . We shall call any such g a spherical representation of f . Naturally, the
same considerations make sense for f ∈ D∗ (Rn ) in the non-quasianalytic case.
In order to fix ideas, let us first discuss the distribution case. The problem of
finding a spherical representation of f ∈ D (Rn ) can be reduced to the determina-
tion of the image of E(Rn ) under the mapping (10). Notice that the range of this
mapping is obviously contained in the subspace of “even” test functions, namely,
Ee (R × Sn−1 ) = {Φ ∈ E(R × Sn−1 ) : Φ(−r, −ω) = Φ(r, ω), ∀(r, ω) ∈ R × Sn−1 }.
In other words, one is interested here in characterizing all those Φ ∈ Ee (R × Sn−1 )
such that
x
ϕ(x) = Φ |x|, (11)
|x|
is a smooth function on Rn . The solution to the latter problem is well known:
Proposition 2 ([6, 9]). Let Φ ∈ Ee (R × Sn−1 ). Then, ϕ given by (11) is an element

of E(Rn ) if and only if Φ has the property that for each m ∈ N
∂ mΦ
(0, ω) is a homogeneous polynomial of degree m. (12)
∂rm
Write
V(R × Sn−1 ) := {Φ ∈ Ee (R × Sn−1 ) : (12) holds for each m ∈ N}.
Hence V(R × Sn−1 ) is precisely the image of E(Rn ) under (10). Since it is ob-
viously a closed subspace of E(R × Sn−1 ), one obtains from the open mapping
theorem that E(Rn ) is isomorphic to V(R × Sn−1 ) via (10). Given f ∈ D (Rn ),
f (x), Φ(|x|, x/|x|) defines a continuous linear functional on D(R × Sn−1 ) ∩ V(R ×
Sn−1 ), and, by applying the Hahn–Banach theorem, one establishes the existence
of a spherical representation g ∈ D (R × Sn−1 ) for f .
We now treat the ultradistribution case. We consider
V ∗ (R × Sn−1 ) := V(R × Sn−1 ) ∩ E ∗ (R × Sn−1 ),
a closed subspace of E ∗ (R × Sn−1 ). It is clear that (10) maps E ∗ (Rn ) continuously
into V ∗ (R × Sn−1 ), but whether this mapping is surjectieve or not is not evident.
The next theorem gives a partial answer to this question, which allows one to
consider spherical representations of ultradistributions. We associate the weight
sequence
.
Np = p!Mp
to Mp . Note that Np ⊂ Mp in the Roumieu case, while Np ≺ Mp in the Beurling
case. The symbol † stands for {Np } if ∗ = {Mp }, while when ∗ = (Mp ) we set
† = (Np ).
Theorem 2. Suppose that Mp satisfies (M.0), (M.1), and (M.2).

(i) The linear mapping Φ → ϕ, where ϕ is given by (11), maps continuously
V † (R × Sn−1 ) into E ∗ (Rn ).
(ii) Any ultradistribution f ∈ E ∗ (Rn ) admits a spherical representation from

E † (R × Sn−1 ); more precisely, one can always find g ∈ E † (R × Sn−1 ) such
that (9) holds for all ϕ ∈ E † (Rn ).
If Mp additionally satisfies (M.3) , one obviously obtains an analogous version

of Theorem 2 for D∗ (Rn ) and D∗ (Rn ). When ∗ = {p!}, the sequence Np becomes
equivalent to p!. We thus obtain the following corollary for real analytic functions
and analytic functionals.
Corollary 1. The linear mapping (10) is a (topological) isomorphism between the

space the real analytic functions A(Rn ) and V {p!} (R × Sn−1 ). Furthermore, any
analytic functional f ∈ A (Rn ) has a spherical representation g ∈ A (R × Sn−1 ),
so that (9) holds for all ϕ ∈ A(Rn ).
D
The rest of this section is devoted to give a proof of Theorem 2. Note that
(ii) is a consequence of (i) and the Hahn–Banach theorem (arguing as in the
distribution case). In order to show (i) we first need to establish a series of lemmas,
some of them are interesting by themselves.
Lemma 1. The space V ∗ (R × Sn−1 ) consists of all those Φ ∈ E ∗ (R × Sn−1 ) whose
coefficient functions ak,j ∈ E ∗ (R) in the spherical harmonic expansion
∞
dj
Φ(r, ω) = ak,j (r)Yk,j (ω)
j=0 k=1
(m)
satisfy that ak,j (0) = 0 for each m < j, and ak,j is an even function if j is even
and ak,j is an odd function if j is odd.
Proof. Proposition 1 ensures that Φ has the spherical harmonic series expansion.
Since Φ ∈ Ee∗ (R × Sn−1 ) we must necessarily have that ak,j is even when j is even
and ak,j is odd when j is odd. Moreover, the other claim readily follows from the
fact that for each m ∈ N
∞ dj
(m)
ak,j (0)Yk,j (ω)
j=0 k=1
needs to be the restriction to the sphere of a homogeneous polynomial of degree

(m)
m, as for it ak,j (0) needs to be zero if j > m.
The latter suggests to study for each j ultradifferentiable functions having

the same properties as the coefficient functions ak,j from Lemma 1. Define the
closed subspace
Xj∗ = {ϕ ∈ E ∗ (R) : ϕ(m) (0) = 0, ∀m < j}.
Lemma 2. Let j ∈ N and suppose Mp satisfies (M.0), (M.1), and (M.2) . The
mapping
φ(r)
φ → ψ, ψ(r) := j ,
r
is an isomorphism of TVS from Xj∗ onto E ∗ (R). Moreover, giving a compact K ⊂ R
and an arbitrary neighborhood U of K with compact closure, there is a constant ,
only depending on K, U , and Mp (but not on j), such that
ψ E {Mp },h (K) ≤ Ch,U φ E {Mp },h (U ) , ∀φ ∈ Xj∗ . (13)
Proof. The inverse mapping is obviously continuous, so it suffices to prove the last
assertion. In order to treat the non-quasianalytic and quasianalytic cases simulta-
neously via a Paley–Wiener type argument, we use a Hörmander analytic cut-off
sequence [11, 16]. So, find a sequence χp ∈ D(R) such that χp ≡ 1 on K, χp (x) = 0
off U , and
χ(m) L∞ (R) ≤ C(1 p) , m ≤ p.
m
p
By (M.0) and (M.1), we find with the aid of the Leibniz formula a constant 2
such that the Fourier transform of φp = χp φ satisfies
|up φ̂p (u)| ≤ C Mp (2 h)p φ E {Mp },h (U ) , u ∈ R, p ∈ N, (14)

for all φ ∈ E(R) with C = Ch,U . Consider now φ ∈ Xj∗ and the corresponding ψ.
j (j)
Setting ψp = χp ψ, and Fourier transforming r ψp (r) = φp (r), we get ψ̂p (u) =
(i)j φ̂p (u). Thus, using the assumption φ(m) (0) = 0 for m < j, we obtain
u tj−1 t1
ψ̂p (u) = ij ··· φ̂p (t1 )dt1 · · · dtj
−∞ −∞ −∞
∞ ∞ ∞
= (−i)j ··· φ̂p (t1 )dt1 · · · dtj .
u tj−1 t1
Employing this expression for ψ̂p and the fact that ψ = ψp on K, one readily
deduces (13) from (14) after applying the Fourier inversion formula and (M.2) .
Denote as Ee∗ (R) the subspace of even ∗-ultradifferentiable functions.
Lemma 3. Assume Mp satisfies (M.0), (M.1), and (M.2). The linear mapping
.
φ → ψ, ψ(r) = φ( |r|),
maps continuously Ee† (R) into E ∗ (R).
Proof. We only give the proof in the non-quasianalytic case, the quasianalytic
case can be treated analogously by using an analytic cut-off sequence exactly
as in the proof of Lemma 2. Take an arbitrary even function φ ∈ D† (K) with
φ E {√p!Mp },h (K) = 1 and set ψ(r2 ) = φ(r). We have
3
|u2p+1 φ̂(u)| ≤ |K|h2p+1 (2p + 1)!M2p+1 ≤ Ch (h2 )p p!Mp . (15)
√
with Ch = h|K|AH M1 and = (2H)3/2 , because of (M.2). Consider

∞ . ∞
|u ψ̂(u)| = up
p
φ( |r|)e dr = 4 up
iru yφ(y) cos(y u)dy .
2
−∞ 0
Integrating by parts the very last integral, we arrive at

p−1 ∞

|u ψ̂(u)| = 2 u
p
φ (y) sin(y u)dy .
2
0

Note that φ is odd and so φ (0) = 0. Iterating this integration by parts procedure,
we find that

1 ∞ p−1
|u ψ̂(u)| = p−1
p
L (φ )G(y u)dy ≤ |K|21−p Lp−1 (φ ) L∞ (K)
2
(16)
2 0
where G(t) = sin t or G(t) = cos t and the differential operator L is given by
d ϕ(y)
(Lϕ)(y) = .
dy y
D
Note that L and their iterates are well defined for smooth odd functions. Our
problem then reduces to estimate Lp−1 (φ ). Let ηp be the Fourier transform of
Lp−1 (φ ), then
=
|ηp (u)| = |(T p−1 (φ )(u)|,
where # ∞
tκ(t)dt for u > 0
(T κ)(u) = uu
−∞
tκ(t)dt for u < 0 .
The inequality (15) then gives (1 + |u|2 ) ηp L∞ (R) ≤ Ch (h2 )p Mp . Fourier inverse
transforming and using (16), we see that ψ (p) L∞ (R) ≤ Ch (Hh2 )p Mp , which
shows the claimed continuity.
We need one more lemma. We denote as B(0, r) the Euclidean ball with
radius r and center at the origin.
Lemma 4. Given r < 1 there are constants L = Lr and C = Cr such that for any
homogeneous harmonic polynomial Q on Rn one has
∂ αQ L∞ (B(0,r)) ≤ CL|α| α! Q|Sn−1 L2 (Sn−1 ) .
Proof. By a result of Komatsu, one has that there is L, depending only on r, such
that
Δp ϕ L2 (B(0,1))
ϕ E {p!},Lh(B(0,r)) ≤ Ch sup .
p∈N h2p (2p)!
(This actually holds for more general elliptic operators [13].) The estimate then
follows by taking h = 1, ϕ = Q, using that Q is harmonic, and writing out the
integral in polar coordinates.
We are ready to prove Theorem 2:
Proof of Theorem 2. We have already seen that (ii) follows from (i). Let Φ ∈
V † (R × Sn−1 ) and set ϕ as in (11). Since the change of variables (r, ω) → rω is an-
alytic and invertible away from r = 0, it is enough to work with ultradifferentiable
norms in a neighborhood of x = 0. Specifically, we estimate the ultradifferentiable
norms of ϕ on the ball B(0, 1/2). Expand Φ as in Lemma 1 and assume that (cf.
Proposition 1)
j
a √
k,j E { p!Mp },h ≤ e−M ( h ) , ∀j, k.
([−1,1])
Combining Lemma 2 and Lemma 3, we can write

ak,j (r)
= bk,j (r2 ) with bk,j ∈ E ∗ (R)
rj
and
j
bk,j 2
E {Mp },1 h ([−1/2,1/2])
≤ Ch e−M ( h ) , ∀j, k.
where the constant 1 does not depend on h. Therefore,

∞
dj
ϕ(x) = ϕ(rω) = Bk,j (x)Pk,j (x)
j=0 k=1
where Bk,j (x) = bk,j (|x|2 ) and Pk,j is the harmonic polynomial whose restriction
to the unit sphere is Yk,j . Since the mapping x → |x|2 is analytic, the function
Bk,j is ∗-ultradifferentiable and furthermore we can find another constant 2 such
that
j
Bk,j E {Mp },2 h2 (B(0,1/2)) ≤ Ch e−M ( h ) , ∀j, k.
Suppose p! ≤ Ch1 hp1 Mp . By (M.1), Lemma 4, and the Leibniz formula,
∞
j
∂αϕ L∞ (B(0,1/2)) ≤ CCh1 Ch (Lh1 + 2 h2 )|α| M|α| dj e−M ( h )
j=0
which completes the proof of Theorem 2 because log t = o(M (t)) and dj =
O(j n−2 ).
We end this section with two remarks. Remark 2 poses an open question.
Remark 1. The technique from this section leads to a new proof of Proposition 2
as well.
Remark 2. Whether Theorem 2 and Lemma 3 hold true or false with † = ∗ is an
open question. Notice that this holds when ∗ = {p!} (Corollary 1).
4. Rotation invariant ultradistributions

We now turn our attention to the characterization of rotation invariant ultradis-
tributions via spherical means.
We begin with the case of ultradistributions from E ∗ (Rn ). We say that f ∈
∗
E (Rn ) is rotation invariant if f (x) = f (T x) for all T ∈ SO(n), the special
orthogonal group, namely, if for every rotation T and every ϕ ∈ E ∗ (Rn )
f (x), ϕ(x) = f (x), ϕ(T −1 x).
Note that the mapping ϕ → ϕS , where ϕS is its spherical mean, is continuous from
E ∗ (Rn ) into itself. This can easily be viewed from the alternative expression [9]

ϕS (x) = ϕ(T x)dT,
SO(n)
where dT stands for the normalized Haar measure of SO(n). The spherical mean
of f ∈ E ∗ (Rn ) is the ultradistribution fS ∈ E ∗ (Rn ) defined by
fS , ϕ = f, ϕS .
Clearly fS is rotation invariant. All these definitions also apply to f ∈ D∗ (Rn ) if
Mp is non-quasianalytic.
D
Theorem 3. Suppose Mp satisfies (M.0), (M.1), and (M.2) . Then, f ∈ E ∗ (Rn ) is

rotation invariant if and only if f = fS .
Proof. We only need to show that if f is rotation invariant then f = fS . Further-
more, the general case actually follows from that of analytic functionals. In fact,
suppose the theorem is true for ∗ = {p!}. Since A(Rn ) is densely injected into
E ∗ (Rn ), we have that f ∈ E ∗ (Rn ) is rotation invariant if and only if it is rotation
invariant when seen as an analytic functional. Furthermore, taking spherical mean
commutes with the embedding E ∗ (Rn ) → A (Rn ), whence our claim follows.
Suppose that f ∈ A (R) is rotation invariant. Applying Corollary 1 we can
find a spherical representation g ∈ A (R × Sn−1 ) for f . Using Proposition 1 we can
expand g as
∞
dj
g(r, ω) = ck,j (r) ⊗ Yk,j (ω) (17)
j=0 k=1
with convergence in A (R×Sn−1 ) where ck,j are one-dimensional analytic function-
als. Notice that if we also expand the polar coordinate expression of ϕ ∈ E ∗ (Rn ) as
$∞ $dj
ϕ(rω) = j=0 k=1 ak,j (r)Yk,j (ω), we obtain that ϕS (rω) = |Sn−1 |−1/2 a0,0 (r) =

a0,0 (r)Y0,0 (ω). The latter holds because Sn−1 Yk,j (ω)dω = 0 for j ≥ 1, which
follows from the mean value theorem for harmonic functions. Thus, c0,0 ⊗ Y0,0
is a spherical representation for fS . The result would then follow if we show that
c0,0 ⊗Y0,0 is also a spherical representation of f . By Lemmas 1–3 and the expansion
(17), this would certainly be the case if we show that
f (x), |x|2m Q(x) = 0 (18)
for every m ∈ N and every harmonic homogeneous polynomial Q of degree j ≥ 1.
Since every such Q can be written [1, Prop. 5.31] as

Q(x) = Q(ω)Zj (x, ω)dω,
Sn−1
where Zj (x, ω) is the zonal spherical harmonic of degree j, we have that

f (x), |x|2m Q(x) = Q(ω)Pj (ω)dω
Sn−1
with
Pj (ω) := f (x), |x|2m Zj (x, ω), ω ∈ Sn−1 .
So (18) would hold if we show that Pj identically vanishes on Sn−1 if j ≥ 1.
Observe that Pj is a spherical harmonic of degree j ≥ 1. On the other hand,
Zj (T −1 x, ω) = Zj (x, T ω) for every rotation T [1, Prop. 5.27], and using the fact
that f is rotation invariant, we obtain Pj (T ω) = Pj (ω) for all ω ∈ Sn−1 and
T ∈ SO(n). Due to the fact that the group SO(n) acts transitively on Sn−1 , Pj
must be a constant function, and hence a spherical harmonic of degree 0. Since
the spaces of spherical harmonics of different degrees are mutually orthogonal in
L2 (Sn−1 ), one concludes that Pj ≡ 0 if j = 0. This concludes the proof of the
theorem.
In the non-quasianalytic case, we can use Theorem 3 to recover the result [3,
Thm. 4.4] by Chung and Na quoted at the Introduction.
Theorem 4. Suppose Mp satisfies (M.1), (M.2) , and (M.3) . An ultradistribution
f ∈ D∗ (Rn ) is rotation invariant if and only if f = fS .
Proof. Using a partition
$∞ of the unity, we can write any rotation invariant f as a
locally finite sum k=1 fk with each fk ∈ E ∗ (Rn ) being also rotation invariant.
By Theorem 3 we have fk = (fk )S , and, consequently,
∞ ∞

fS = (fk )S = fk = f.
k=1 k=1
We now discuss how one can extend Theorem 3 in the quasianalytic case
(including the hyperfunction case). From now on we assume that Mp satisfies
(M.0), (M.1), (M.2) , and (QA). Our next considerations are in terms of sheaves
of quasianalytic ultradistributions, we briefly discuss their properties following the
approach from [5, 10] (cf. [17] for hyperfunctions).
Let f ∈ E ∗ (Rn ) (referred to as a ∗-quasianalytic functional hereafter). A
compact K ⊆ Rn is called a ∗-carrier of f if f ∈ E ∗ (Ω) for every open neighbor-
hood Ω of K. If f ∈ A (Rn ), it is well known [11, Sect. 9.1] that there is a smallest
compact K ⊆ Rn among all the {p!}-carriers of f , the {p!}-support of f denoted
by suppA f . It was noticed by Hörmander that a similar result basically holds for
quasianalytic functionals [10, Cor. 3.5], that is, for any ∗-quasianalytic functional
there is a smallest ∗-carrier, say suppE ∗ f , and one has suppA f = suppE ∗ f .
Hörmander only treats the Roumieu case in [10], but his proof can be modified to
show the corresponding statement for the Beurling case [5, 8].
Denote as E ∗ [K] the space of ∗-quasianalytic functionals with support in K.
One can show that there is an (up to isomorphism) unique flabby sheaf B∗ whose
space of global sections with support in K is precisely E ∗ [K], for any compact
of Rn . We call B∗ the sheaf of ∗-quasianalytic ultradistributions. When ∗ = {p!},
we simply write B = B∗ , the sheaf of hyperfunctions. Actually, in the Roumieu
case the existence of B∗ can be established exactly as for hyperfunctions with the
aid of Hörmander support theorem by using the Martineau-Schapira method [17].
Details for the Beurling case, which require a subtler treatment, will appear in
the forthcoming paper [5]. Since it is important for us, we mention that on any
bounded open set Ω the sections of B∗ are given by the quotient spaces
B∗ (Ω) = E ∗ [Ω]/E ∗ [∂Ω], (19)
which reduces to the well-known Martineau theorem in the case of hyperfunctions.
Finally, we call the space of global sections B∗ (Rn ) the space of ∗-quasianalytic
ultradistributions on Rn (hyperfunctions if ∗ = {p!}).
The operation of taking spherical mean preserves the space E ∗ [K] if K is a
rotation invariant compact set. Because of (19), we can define the spherical mean
fS ∈ B∗ (Ω) of f ∈ B∗ (Ω) in a canonical manner if Ω is a bounded rotation
invariant open subset of Rn , namely, if f = [g] with g = E ∗ [Ω], we define fS =
D
[gS ]. Using the sheaf property, one extends the definition fS ∈ B∗ (Rn ) for all
f ∈ B∗ (Rn ). We say that f ∈ B∗ (Rn ) is rotation invariant if its restriction to Ω is
rotation invariant for any rotation invariant bounded open set Ω (the latter makes
sense because of (19)). Theorem 3 implies the following generalization:
Theorem 5. Suppose Mp satisfies (M.0), (M.1), (M.2) , and (QA). A quasianalytic
ultradistribution f ∈ B∗ (Rn ) is rotation invariant if and only if f = fS .
We point out that Theorem 5 extends [3, Thm. 5.7], which was obtained for
hyperfunctions.
Acknowledgement
The authors gratefully acknowledge support by Ghent University, through the
BOF-grant 01N01014.
References
[1] Axler, S., Bourdon, P., Ramey, W.: Harmonic function theory. Springer-Verlag, New
York (2001)
[2] Carmichael, R., Kamiński, A., Pilipović, S.: Boundary values and convolution in
ultradistribution spaces. World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ
(2007)
[3] Chung, S.Y., Na, J.Y.: Rotation invariant generalized functions. Integral Transforms
Spec. Funct. 10, 25–40 (2000)
[4] Dasgupta, A., Ruzhansky, M.: Eigenfunction expansions of ultradifferentiable func-
tions and ultradistributions. Trans. Amer. Math. Soc. (2016), doi:10.1090/tran/6765
[5] Debrouwere, A., Vindas, J.: Sheaves of infrahyperfunctions. In preparation
[6] Drozhzhinov, Yu.N., Zav’yalov, B.I.: Asymptotically homogeneous generalized func-
tions and boundary properties of functions holomorphic in tubular cones. Izv. Math.
70, 1117–1164 (2006)
[7] Estrada, R., Kanwal, R.P.: Distributional boundary values of harmonic and analytic
functions. J. Math. Anal. Appl. 89, 262–289 (1982)
[8] Heinrich, T., Meise, R.: A support theorem for quasianalytic functionals. Math.
Nachr. 280, 364–387 (2007)
[9] Helgason, S.: The Radon transform. Birkhäuser Boston, Inc., Boston, MA (1999)
[10] Hörmander, L.: Between distributions and hyperfunctions. Astérisque 131, 89–106
(1985)
[11] Hörmander, L.: The analysis of linear partial differential operators. I. Distribution
theory and Fourier analysis. Springer-Verlag, Berlin (1990)
[12] Gel’fand, I.M., Shilov, G.E.: Generalized functions. Vol. 2. Spaces of fundamental
and generalized functions. Academic press, New York-London (1968)
[13] Komatsu, H.: A characterization of real analytic functions. Proc. Japan Acad. 36,
90–93 (1960)
[14] Komatsu, H.: Ultradistributions I. Structure theorems and a characterization. J. Fac.
Sci. Tokyo Sect. IA Math. 20, 25–105 (1973)
[15] Komatsu, H.: Ultradistributions. II. The kernel theorem and ultradistributions with
support in a submanifold. J. Fac. Sci. Univ. Tokyo Sect. IA Math. 24, 607–628 (1977)
[16] Petzsche, H.-J.: Approximation of ultradifferentiable functions by polynomials and
entire functions. Manuscripta Math. 48, 227–250 (1984)
[17] Schapira, P.: Théorie des hyperfonctions. Springer-Verlag, Berlin-New York (1970)
[18] Tengstrand, A.: Distributions invariant under an orthogonal group of arbitrary sig-
nature. Math. Scand. 8, 201–218 (1960)
[19] Vernaeve, H.: Group invariant Colombeau generalized functions. Monatsh. Math.
153, 165–175 (2008)
- ., Vindas, J.: Ultradistributional boundary values of harmonic functions
[20] Vučković, D
on the sphere. Preprint arxiv:1610.09879 (2016)
- ord̄e Vučković and Jasson Vindas

D
Ghent University
[email protected]
Operator Theory:

Eigenvalue Problems of Toeplitz Operators

in Bargmann–Fock Spaces
Kunio Yoshino
To Stevan Pilipović for his 65th birthday
Abstract. In this paper we will derive a formula for the eigenvalues of Toeplitz
operators with polyradial symbols in Bargmann–Fock spaces. Moreover we
will clarify the relationship between Toeplitz operators in Bargmann–Fock
spaces and Daubechies operators in L2 (Rn ). As application of our results, we
will give a new proof of the formula of the eigenvalues of Daubechies operators
with polyradial symbols.
1. Introduction
We will consider eigenvalue problems of Toeplitz operators in Bargmann–Fock
spaces. Bargmann–Fock spaces were introduced by V. Bargmann in 1961 ([2]).
The elements of a Bargmann–Fock space consist of entire functions with suit-
able estimates. Bargmann–Fock spaces have several applications in mathematical
physics ([1], [2], [6], [9], [14]). The short time (windowed Fourier) transform with
Gaussian window functions is called Gabor transform. The Gabor transform is very
closely related to the FBI transform and the Bargmann transform ([12]). Gabor
transform localized in a phase space is called Daubechies (localization) operator.
In this paper we will show
1. A formula for the eigenvalue of Toeplitz operators in a Bargmann–Fock space.
2. A relationship between Toeplitz operators in Bargmann–Fock spaces and
Daubechies operators in L2 (Rn ).
2. Bargmann transform and Bargmann–Fock space

2.1. Kernel function of Bargmann transform
According to ([2]), we recall the definition of the Bargmann transform and its prop-
erties. First we introduce a kernel function An (z, x) of the Bargmann transform
270 K. Yoshino
as follows:
0 √ 1
An (z, x) = π −n/4 exp − 21 (z 2 + x2 ) + 2z · x , (z ∈ Cn , x ∈ Rn ).
2.2. Bargmann transform B

We define the Bargmann transform B for square integrable functions as

def
B(ψ)(z) = ψ(x)An (z, x)dx, (ψ ∈ L2 (Rn )).
Rn
The inverse Bargmann transform B −1 is given by

1 2
B −1 (g)(x) = g(z)An (z, x)e−|z| dz ∧ dz, (g ∈ BF (Cn )).
def
n
(2πi) Cn
Example.
1. B(An (a, x))(z) = eaz B√
2
2. B(φp,q )(z) = ezw−|w| /2+ipq/2
, w = (p + iq) 2
2
where φp,q (x) = π −1/4 eipx e−(x−q) /2 and n = 1.
zm
3. B(hm )(z) = √ , where hm (x) is Hermite function of degree m and n = 1.
m!
√ −1/2 dm
hm (x) = (−1)m 2m m! π exp(x2 /2) m exp(−x2 ).
dx
Remark 2.1.
C
1. h0 (x) = π −1/4 exp − x2 2 is called coherent state in quantum optics (me-
chanics).
2x2 − 1 C
2. h2 (x) = π −1/4 √ exp − x2 2 is called Mexican hat wavelet in wavelet
2
analysis.
2.3. The Bargmann–Fock space BF (Cn )
Suppose that dm(z) is Lebesgue measure in Cn . We consider the Hilbert space
2
L2 (Cn , dμ) with Gaussian weight dμ(z) = π −n e−|z| dm(z):

L2 (Cn , dμ) = g(z) : |g(z)|2 dμ(z) < ∞ .
Cn

The inner product in L2 (Cn , dμ) is defined by f, g = f (z)g(z)dμ(z).
Cn
The Bargmann–Fock space BF (Cn ) is defined as

BF (Cn ) = g ∈ H(Cn ) : |g(z)|2 dμ(z) < ∞ = H(Cn ) ∩ L2 (Cn , dμ),
Cn
where H(C ) is the space of entire functions on Cn .
n
Since L2 (Cn , dμ) ⊂ L2 loc (Cn ), we can consider the distribution derivatives for
the elements of L2 (Cn , dμ). Then we have BF (Cn ) = L2 (Cn , dμ)∩Ker(∂), ¯ where ∂¯
n 2 n
is the Cauchy–Riemann operator. BF (C ) is a closed subspace of L (C , dμ) ([2]).
Toeplitz Operators in Bargmann–Fock Spaces 271
Example (Elements of a Bargmann–Fock space).

1. Polynomials . . , zn ) belong to BF (Cn ). For example, the polyno-
- of z = (z1 , .2m+1
mial part i<j (zi − zj ) of a Laughlin wave function in quantum Hall
effect is a member of BF (Cn ) ([9]).
2. Entire functions of exponential type belong to BF (Cn ). Since the Fourier–
Laplace transform of Schwartz distributions with compact supports and hy-
perfunctions with compact supports are entire functions of exponential type,
they belong to BF (Cn ).
Similarly, a Fourier–Laplace transform of analytic functionals belong to
BF (Cn ). For instance, the sinc function (sin z)/ z and the prolate spheroidal
d d
function eigenfunction of (τ 2 − z 2 ) − 2z − σ 2 z 2 are entire functions
dz dz
of exponential type. Hence they belong to BF (C).
For the details of prolate spheroidal functions, we refer the reader
to ([15]).
z z z2
3. σ(z) = z 1− exp + 2 is a Weierstrass σ-function
λm,n λm,n 2λ m,n
and λm,n are lattice points in C. (σ(z))/ z is a member of BF (C) under
suitable conditions on lattice points λm,n ([10], [19]).
Proposition 2.2 ([2, 19]).
m ∞
z
1. √ is a complete orthogonal basis in BF (Cn ).
m! m=0
∞
z m wm
2. ezw = √ √ is Bergman kernel function in BF (Cn ), i.e.,
m=0 m! m!

f (z) = ezw f (w)dμ(w), (∀f (z) ∈ BF (Cn )).
Cn
Theorem 2.3 ([2]). The Bargmann transform is a unitary mapping from L2 (Rn )
to the Bargmann–Fock space BF (Cn ).
Remark 2.4.
1. For the relationship between the Bargmann transform and microlocal analy-
sis, we refer the reader to [12].
2. The generalization of the Bargmann transform to generalized functions is
studied in [4] and [13].
3. Orthogonal projection and Toeplitz operators

3.1. Orthogonal projection
Since BF (Cn ) is a closed subspace of L2 (Cn , dμ), we have an orthogonal decom-
⊥
position L2 (Cn , dμ) = BF (Cn ) ⊕ BF (Cn ) .
272 K. Yoshino
Proposition 3.1 ([19]).
1. The orthogonal projection P : L2 (Cn , dμ) −→ BF (Cn ) is given by

(P g)(z) = ezw g(w)dμ(w), (g ∈ L2 (Cn , dμ)).
Cn

2. g(z) = ezw g(w)dμ(w), (g ∈ BF (Cn )).
Cn
3.2. Toeplitz operators

For a bounded function F (z) on Cn , we put MF (f )(z) = F (z)f (z), (f (z) ∈
L2 (Cn , dμ)). The Toeplitz operator TF with the symbol F is a composition of the
multiplication operator MF and the orthogonal projection P , i.e., TF = P ◦ MF .
Proposition 3.2. The Toeplitz operator TF with the symbol F is given as

(TF f )(z) = ezw F (w)f (w)dμ(w), (∀f ∈ L2 (Cn , dμ)).
Cn
3.3. Eigenvalue problems of Toeplitz operators on Bargmann–Fock spaces

Theorem 3.3. Suppose that F (w) is a bounded integrable function and polyradial,
i.e., F (w1 , . . . , wn ) = F (|w1 |2 , . . . , |wn |2 ). Then
1. z m is an eigenfunction of TF .
2. The eigenvalue λm of TF is given by
∞ ∞
n
1
λm = ··· F (s1 , . . . , sn ) e−si si mi dsi ,
m! 0 0 i=1
m = (m1 , . . . , mn ) ∈ Nn .
Proof. For simplicity, we put n = 1.

(TF )(wm )(z) = F (|w|2 )ezw wm dμ(w)
C
1 2
= F (|w|2 )ezw wm e−|w| dm(w)
π C
∞
1 (zw)n 2
= F (|w|2 ) wm e−|w| dm(w)
π C n=0
n!
z 1
∞ n 2
= F (|w|2 )w n wm e−|w| dm(w).
n=0
n! π C
Employing polar coordinates w = reiθ ,

∞
1 z n ∞ 2π 2
= F (r2 )ei(m−n)θ rn rm e−r rdrdθ
π n=0 n! 0 0
∞
1 2
= zm e−r F (r2 )r2m 2rdr
m! 0
∞
m 1
=z e−s sm F (s)ds.
m! 0
Remark 3.4. For the recent development of the theory of Toeplitz operators, we
refer the reader to [3, 5, 11].
4. Gabor transform and resolution of identity

4.1. Gabor transform
2
Put φ(x) = π −n/4 e−x . We define the Gabor transform f (x) ∈ L2 (Rn as
/2

Wφ (f )(p, q) = e−ipx φ(x − q)f (x)dx, (f (x) ∈ L2 (Rn )).
Rn
4.2. Resolution of identity (unity)

We have following inversion formula for the Gabor transform.
Proposition 4.1 ([10]).

n
1
f (x) = φp,q (x)Wφ (f )(p, q)dp dq,
2π R2n
where Wφ (f )(p, q) is the Gabor transform of f (x) and φp,q (x) = eipx φ(x − q).
5. Daubechies localization operator

Definition 5.1 ([7, 8]). Suppose that F (p, q) ∈ L1 (R2n ). For f (x) ∈ L2 (Rn ), we
put

PF (f )(x) = (2π)−n F (p, q)φp,q (x) φp,q , f dpdq.
R2n
PF is called Daubechies operator with symbol F .
Remark 5.2 ([7, 8]).

1. PF is a bounded linear operator from L2 (Rn ) to L2 (Rn ).
2. If the symbol function F is 1, then PF is the identity operator by virtue of
the resolution of identity.
274 K. Yoshino
5.1. Daubechies operator in a Bargmann–Fock space

Theorem 5.3 ([16]).

2
(B ◦ PF ◦ B −1 )(g)(z) = (2πi)−n F (w, w)ezw g(w)e−|w| dw ∧ dw, (∀g ∈ BF ).
Cn
Proof. For simplicity, we put n = 1. Since the Bargmann tranform is unitary, we

have

1
PF (f )(x) = F (p, q)φp,q (x) φp,q , f dpdq
2π

1
= F (p, q)φp,q (x) Bφp,q , Bf dpdq,
2π
By the example in Section 2,

1
B ◦ PF (f )(x) = F (p, q)Bφp,q (z) Bφp,q , Bf dpdq,
2π

1 2
= F (p, q)ezw−1/2|w| +1/2ipq Bφp,q , Bf dpdq,
2π
Hence

−1 1 2
(B ◦ PF ◦ B )(g)(z) = F (p, q)ezw−1/2|w| +1/2ipq
Bφp,q , g dpdq.
2π
By Propositions 3.1 and 3.2,

2 2
Bφp,q , g = et̄w̄−1/2|w| −1/2ipq g(t)dμ(t) = e−1/2|w| −1/2ipq g(w̄).
Therefore we have

−1 1 2
(B ◦ PF ◦ B )(g)(z) = F (w, w)ezw g(w)e−|w| dw ∧ dw
2πi C

1 2
= F (w, w)ezw g(w)e−|w| dw ∧ dw.
2πi C
1
Remark 5.4. If we put w = √ (p + iq), then
2
1 1
p = √ (w + w̄), q = √ (w − w̄).
2 2i
So we can identify F (p, q) with F (w, w̄).
Remark 5.5 ([17, 18]). If the symbol function F is 1, then PF is the identity
operator. This fact corresponds to following reproducing formula in BF (Cn ):

1 2
g(z) = ezw g(w)e−|w| dw ∧ dw, g ∈ BF (Cn )
(2iπ)n Cn
5.2. Eigenvalue problem of the Daubechies operator

Since TF = B ◦ PF ◦ B −1 is a Toeplitz operator, we can apply Theorem 3.3.
We obtain
Theorem 5.6 ([7, 8]). Suppose that F (w) is a bounded integrable function and
polyradial, i.e., F (w1 , . . . , wn ) = F (|w1 |2 , . . . , |wn |2 ). Then
1. PF (hm )(x) = λm hm (x), where hm (x) is a Hermite function with degree m.
∞ ∞ n
1
2. λm = ··· F (s1 , . . . , sn ) e−si si mi dsi , m = (m1 , . . . , mn ) ∈ Nn .
m! 0 0 i=1
zm
Proof. PF (hm )(x) = B −1 ◦ TF ◦ B(hm ) = B −1 ◦ TF √
m!
zm
= B −1 λm √ = λm hm (x)
m!
Example ([16]).
a−1 2
1. F (w) = exp |w| , (0 < a < 1),
a
a−1
F (s) = exp s , λm = am+1 , (m ∈ N).
a
|w|2 s
2. F (w) = , F (s) = s , λm = (m + 1)ζ(m + 2) − (m + 1),
e |w| 2
−1 e −1
ζ(z) is the Riemann zeta function.
References
[1] A. Aftalion, X. Blanc and F. Nier: Lowest Landau level functional and Bargmann
spaces for Bose–Einstein condensates, J. Funct. Anal. Vol. 241, pp. 661–702 (2006).
[2] V. Bargmann: On a Hilbert space of analytic functions and an associated integral
transform, Part I, Comm. Pure Appl. Math., Vol. 14, pp. 187–214 (1961).
[3] W. Bauer, L.A. Coburn and J. Isralowitz: Heat flow, BMO, and the compactness of
Toeplitz operators, J. Funct. Anal. vol. 259, pp. 57–78 (2010).
[4] M. Cappiello, L. Rodino and J. Toft: Radial symmmetric elements and the Bargmann
transform, preprint (2014).
[5] H. Chihara: Bounded Berezin–Toeplitz Operators on the Segal Bargmann space, In-
tegr. Equ. Oper. Theory, Vol. 63, pp. 321–335 (2009).
[6] I. Daubechies and A. Grossmann: Frames in the Bargmann–Fock space of entire
functions, Comm. Pure Appl. Math., Vol. 41, pp. 151–164 (1988).
[7] I. Daubechies: A Time Frequency Localization Operator: A Geometric Phase Space
Approach, IEEE. Trans. Inform. Theory. vol.34, pp. 605–612 (1988).
[8] I. Daubechies: Ten Lectures on Wavelets, SIAM. Philadelphia. Pennsylvania, (1992).
276 K. Yoshino
[9] S.M. Girvin and T. Jach: Formalism of quantum Hall effect, Hilbert space of analytic
functions, Phys. Rev. B, Vol. 29, pp. 5617 (1984).
[10] K. Gröchenig: Foundation of Time Frequency Analysis, Boston, Basel, Birkhäuser,
(2000).
[11] B.C. Hall: Berezin–Toeplitz quantization on Lie groups, J. Funct. Anal. vol. 255, pp.
2488–2506 (2008).
[12] A. Martinez: An Introduction to Semiclassical and Micro local Analysis, New York,
Heidelberg, Springer-Verlag (2001).
[13] J. Toft: Bargmann transform on modulation and Gelfand–Shilov spaces with appli-
cations to Toeplitz and pseudo-differential operators, J. Pseudo-Differ. Oper. Appl.
vol. 3, pp. 145–227 (2012).
[14] A. Voros: Wentzel–Kramers method in the Bargmann representation, Phys. Rev. A,
Vol. 40, pp. 6814 (1989).
[15] Gilbert G. Walter: Wavlet and Other Orthogonal Systems with Applications, CRC
Press, Boca Raton, Florida (1994).
[16] K. Yoshino: Daubechies localization operator in Bargmann–Fock Space, Proceedings
of SAMPTA 2009, Marseille, France, May (2009).
[17] K. Yoshino: Analytic continuation and applications of eigenvalues of Daubechies lo-
calization operator, Cubo A Mathematical Journal, Vol. 12, pp. 203–212 (2010).
[18] K. Yoshino: Spectral analysis of Daubechies localization operator, Operator Theory,
Advances and Applications, Vol. 245, pp. 279–284 (2015).
[19] K. Zhu: Analysis on Fock spaces, Springer-Verlag, New York (2012).
Kunio Yoshino
Tamazutsumi 1-28-1
Setagaya - ku
Tokyo, 158-8557, Japan

Generalized Functions and Fourier Analysis: Michael Oberguggenberger Joachim Toft Jasson Vindas Patrik Wahlberg Editors

Uploaded by

Copyright:

Available Formats

Generalized Functions and Fourier Analysis: Michael Oberguggenberger Joachim Toft Jasson Vindas Patrik Wahlberg Editors

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Generalized Functions and Fourier Analysis: Michael Oberguggenberger Joachim Toft Jasson Vindas Patrik Wahlberg Editors

Uploaded by

Copyright:

Available Formats

Operator Theory

Advances and Applications

Founded in 1979 by Israel Gohberg

Associate Editors: Honorary and Advisory Editorial Board:

More information about this series at http://www.springer.com/series/4850

Jasson Vindas Patrik Wahlberg

ISSN 0255-0156 ISSN 2296-4878 (electronic)

Library of Congress Control Number: 2017939710

© Springer International Publishing AG 2017

Printed on acid-free paper

This book is published under the trade name Birkhäuser, www.birkhauser-science.com

A. Ascanelli, S. Coriasco and A. Süß

F. Baumgartner, M. Oberguggenberger and M. Schwarz

Y. Chen, M. Signahl and J. Toft

G. Debruyne and J. Vindas

P. Giordano and M. Kunzinger

T. Levajković, H. Mena and A. Tuﬀaha

N. Teofanov and F. Tomić

Debrouwere, Teofanov/Tomić, and Vučković/Vindas); Colombeau theory (papers

The editors Michael Oberguggenberger

Abstract. We describe classes of temperate distributions with prescribed de-

2.1. Subspaces of temperate distributions

2.2. SG-symbols and operators

An operator A = Op(a), is called elliptic (or S m,μ -elliptic) if a ∈ S m,μ (R2d )

In 1987, E. Schrohe [12] introduced a class of non-compact manifolds, the

for h ∈ S m+t,μ+τ (R2d ), r ∈ S −∞,−∞ (R2d ).

3. Temperate distributions decaying at inﬁnity and

Deﬁnition 3.1. u ∈ S (Rd ) is a rapidly decreasing distribution if, for all k ∈ Z+ ,

Denoting by F the Fourier transform on S (Rd ), one ﬁnds F OC = OM .

Theorem 3.3. The following three properties are equivalent:

3. for any k ∈ Z+ there exist Mk ∈ N, βjk ∈ Zd+ , fjk ∈ L∞ (Rd ) ∩ C(Rd ),

Since k is arbitrary, this proves the claim.

mapping properties of the SG pseudo-diﬀerential operators on the Sobolev–Kato

By Lemma 3.5, we can then conclude that u ∈ S (Rd )∞ , as claimed.

Proof of Theorem 3.7. By Lemma 3.8,

to zero in E and F , respectively, see, e.g., F. Trèves [14].

(Opϕ (a)u)(ψ) = u(ψ) = (2π)−d u χ g) = (2π)−d (

= (b · Opϕ (1)(u ∗ χ))(ψ),

provided N is chosen large enough, recalling that {λj } ∈ 1 .

Let us introduce, for r ∈ R, the weighted space

In fact, by assumption, · ∂ ψ ∈ L (Rd ) ∩ C ∞ (Rd ), and, by the inductive

In fact, (I) belongs to L1 (Rd ) by hypothesis. Moreover, as above, we have

Proposition 3.17. For any r ∈ R we have

which implies v ∈ (DL1r (Rd )) , as claimed.

Let us examine each summand, by considering functions of type ·s ·[(·−s f )∗

2 In fact, by the deﬁnition of convolution of distributions and test functions,

3 Recall that D σ (K) = {ψ ∈ C σ (Rd ) : supp ψ ⊆ K}.

fj ∈ L∞ (Rd ), j = 1, . . . , M , and ·− 2 −ε ∈ L2 (Rd ) = H 0,0 (Rd ). It also follows,

setting k = s − d2 − ε ≥ 0, κ = − maxj |βj | ≤ 0,

In fact, by the mapping properties of the SG pseudo-diﬀerential operators on

which, by Theorem 3.15, (2), gives the claim.

Alessia Ascanelli and André Süß

Transport in a Stochastic Goupillaud Medium

or, more generally, hyperbolic systems of the form

typically arises in seismology (propagation of acoustic waves) or in material sci-

2. Setting up the Goupillaud medium

2.1. Dyadic deterministic structure

Figure 1. Illustration of τ → γ(τ ; 0, 0) in the Goupillaud medium

for all N, M ∈ N, k ∈ Z. The value of the characteristic curve γ (N ) in the grid

provided N is chosen large enough, recalling that {λj } ∈ 1 .

In fact, by assumption, · ∂ ψ ∈ L (Rd ) ∩ C ∞ (Rd ), and, by the inductive

Let us examine each summand, by considering functions of type ·s ·[(·−s f )∗

fj ∈ L∞ (Rd ), j = 1, . . . , M , and ·− 2 −ε ∈ L2 (Rd ) = H 0,0 (Rd ). It also follows,

when s > 0 and

Moreover, beside the usual ordering in R, the elements σ in R and R are

Deﬁnition 1.3. Let p ∈ (0, ∞], ϑ be a weight on Nd , and let s ∈ R .

H1 → Ht2 (Rd ) and Ht 2 (Rd ) → H2 (2.4)

In order to prove B2 → H[1/ϑ

Proof. By Proposition 2.3 it follows that B2 → H[1/ϑ

which proves that H1 → B1 .

which proves that H1 → B1 .