Anda di halaman 1dari 390

Monographs in Mathematics

Vol. 82

Edited by

A. Borel

J. Moser

S.-T. Yau


Boston· Basel· Stuttgart

V. I. Arnold S. M. Gusein-Zade A. N. Varchenko

Singularities of Differentiable Maps

Volume I

The Classification of Critical Points Caustics and Wave Fronts

Under the Editorship ofY. I. Arnold Translated by Ian Porteous Based on a Previous Translation by Mark Reynolds



Boston' Basel' Stuttgart

Originally published as Osobennosty differentsiiruemykh otobrazhenii by Science, Moscow, 1982.

Library of Congress Cataloging in Publication Data

Arnol'd, V. I. (Vladimir Igorevich), 1937- Singularities of differentiable maps.

(Monographs in mathematics ; 82- ) Translation of: Osobennosti differentsiruemykh otobrazheniL Bibliography: p. Includes index. 1. Differentiable mappings. 2. Singularities (Mathematics) I. Varchenko, A. N. (Aleksandr Nikolaevich) II. Guseln-Zade, S. M. (Sabir Medzhidovich)

III. Title. IV. Series: Monographs in mathematics 82, etc.




ISBN -13: 978-1-4612-9589-1


CIP-Kurztitelaufnahme der Deutschen Bibliothek Arnol'd, Vladimir I.:

Singularities of differentiable maps / V. I.



A. N. Varchenko

; S. M. Gusein-Zade.



; Basel

; Stuttgart

: Birkhiiuser

Einheitssacht.: Osobennosti differenciruemych otobrazenij «eng!.»

NE: Varcenko, Aleksandr N.:; Gusejn-Zade, Sabir M.:

Vo!' 1. The classification of critical points, caustics and wave fronts / under the editorship -

of V. I. Arnold.


(Monographs in mathematics ; Vol. 82) ISBN-13:978-1-4612-9589-1 e-ISBN-13:978-1-4612-5154-5 DOl: 10.1007/978-1-4612-5154-5


All rights reserved. No part of this publication may be-reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without prior permission of the copyright owner.

© 1985 Birkhiiuser Boston, Inc. Softcover reprint of the hardcover 1st edition 1985


T he theory of singularities of differentiable maps is a rapidly developing area of contemporary mathematics, being a grandiose generalisation of

the study of functions at maxima and minima and having numerous applications in mathematics, the natural sciences and technology (as in the so-called theory of bifurcations and catastrophes). Chapters of the book are devoted to the theory and stability of smooth maps, critical points of smooth functions, and the singularities of caustics and wave fronts in geometrical optics. The book is the first volume of a large monograph envisaged by the authors. In the second volume will be placed the algebro-topological aspects ofthe theory. The book is intended for mathematicians, from second year students up to research workers, and also for all users of the theory of singularities in mechanics, physics, technology and other sciences.


Introduction to the English Edition

Singularity theory is still in the state of very rapid development and many new results appeared after the Russian edition ofthis book. The reader should consult the two volumes ofthe Arcata singularities conference (Symposia in Pure Math., vol. 40, 1983.); the surveys "Singularities of ray systems", Russ. Math. Surveys (vol. 38, no. 2, 1983); "Singularities in the calculus of variations", Contemporary Problems of Math., (vol. 22, Moscow, 1983); Catastrophe Theory, Springer, 1983; and the Proceedings of the International Congress of Mathematics, Warsaw, 1983. The authors express their gratitude to R. and C. MacPherson and to I. and H. Porteous for their care with the English edition.


there is nothing so enthralling, so grandiose, nothing that stuns or captivates the human soul

... quite so much as a first course in a science. After the first five or six lectures one already holds the

brightest hopes, already sees oneself as a seeker after truth. I too have wholeheartedly pursued science passionately, as one would a beloved woman. I was a slave, and sought no other sun in my life. Day and night I crammed myself, bending my back, ruining myself over my books; I wept when I beheld others exploiting science fot personal gain. But I was not long enthralled. The truth is every science has a beginning, but never an end - they go on for ever like periodic fractions. Zoology, for example, has discovered thirty-five thousand forms of life ...

A. P. Chekhov. "On the road"

In this book a start is made to the "zoology" of the singularities of differentiable maps. This theory is a young branch of analysis which currently occupies a central place in mathematics; it is the crossroads of paths leading from very abstract corners of mathematics (such as algebraic and differential geometry and topology, Lie groups and algebras, complex manifolds, commutative algebra and the like) to the most applied areas (such as differential equations and dynamical systems, optimal control, the theory of bifurcations and catastrophes, short-wave and saddle-point asymptotics and geometrical and wave optics). The main applications of the theory of singularities consist in the listing and detailed examination in each situation of a small collection of the most frequently encountered standard singularities, which are just those which occur for objects in general position: all more complicated singularities decompose into the simplest ones under a small perturbation of the object of study. We give rather complete lists, diagrams and determinators of these simplest singularities for a whole series of objects (functions, maps, varieties, bifurcations, caustics, wave fronts and the like) trying as far as possible to shorten for the reader the path from the beginnings of the theory to its applications. Accordingly our aim is to present the basic ideas, methods and results of the theory of singularities in such a way that the reader may as quickly as possible learn to apply the methods and results of the theory, without being delayed on the more fundamental theological parts. A special effort has been made to ensure that the application of the main ideas and methods is not clogged up with technical details. The most fundamental and the simplest examples are studied in the greatest detail, while the presen- tation of the more specialised and difficult parts of the theory has the character of a review. The reader is supposed to have only a small mathematical knowledge (skill



in differentiating and some linear algebra and geometry)*. The authors have tried to present the material in such a way that the reader can omit parts that he finds difficult without significantly impairing his understanding of what follows. At the present time the theory of singularities is rapidly developing (see, for example, the lists of unsolved problems in [19] and [27]) and we shall not attempt to cover all the many directions of current research (an incomplete bibliography of approximately 500 works is to be found in Poston and Stewart [148] and Brieskorn [41]). The basis of this book is a series of special courses given at the Mechanics- Mathematics Faculty of Moscow State University during the years 1966-1978. In its preparation we have used lecture notes taken by V. A. Vasil'ev, E. E. Landis and A. G. Hovansky; A. G. Hovansky has written Chapter 5. The authors wish to thank these people for their assistance, as well as to thank the participants of a seminar on the theory of singularities, whose assistance they have made great use of, especially A. G. Kushnirenko, E. I. Korkina and V. I. Matov. The complex-analytic and algebro-geometric aspects of singularity theory (monodromy, intersection theory, asymptotics of integrals and mixed Hodge structures) will be discussed in a second volume "Singularities of differentiable maps, Algebro-topological aspects" in course of preparation.

Yasenevo, March 1979

* For

the benefit of the reader we recall the following terminology:

(1) A manifold is an n-dimensional generalisation of a curve or surface, while a map is the analogous generalisation of function. A diffeomorphism is an invertible map, it and its inverse both being differentiable. (2) A transformation of a set is an invertible map of the set to itself. A group of transformations of a set is a set of transformations containing together with each transformation its inverse and together with every two transformations their composite. A group results from axiomatising the properties of a group of transformations. (3) An algebra results from axiomatising the properties of the set of all functions on a set (the elements of an algebra, just like functions, may be added or mUltiplied together or may be multiplied by numbers, and moreover the usual rules of associativity, distributivity and commutativity hold. In an algebra there is a distinguished element 1 with If=f)· (4) A module over an algebra is the result of axiomat ising the properties of the set of all vector fields on a manifold (the elements of a module may be added together and multiplied by elements of the algebra). (5) An ideal in an algebra is a subset of the algebra that is also a module over the algebra. Example: in the algebra of all the functions on a manifold the functions taking the value zero on a given submanifold form an ideaL

Table of Contents



Part I. Basic concepts .


  • 1. The simplest examples


  • 2. The classes ~I


  • 3. quadratic differential of a map .




  • 4. The local algebra of a map and the Weierstrass preparation theorem


  • 5. local


multiplicity of a holomorphic map



  • 6. Stability and infinitesimal stability


  • 7. The proof of the stability theorem



  • 8. Versal deformations



  • 9. The classification of stable germs by genotype.


  • 10. Review offurther results



Part II. Critical points of smooth functions




  • 11. A start to the classification of critical points


  • 12. Quasihomogeneous and semiquasihomogeneous singularities


  • 13. The classification of quasihomogeneous functions.


  • 14. Spectral sequences for the reduction to normal forms.


  • 15. Lists of singularities .


  • 16. The determinator of singularities



  • 17. Real, symmetric and boundary singularities



Part III. Singularities of caustics and wave fronts


  • 18. Lagrangian singularities


  • 19. Generating families


  • 20. Legendrian singularities


  • 21. The classification of Lagrangian and Legendrian singularities



Table of contents



Further references


Subject Index



Part I

Basic Concepts

The theory of singularities of smooth maps is a wide-ranging generalisation of the theory of the maxima and minima of functions of one variable. Therefore the singularities that we shall be discussing are not connected with discontinuities and poles, but with the vanishing of certain derivatives and Jacobians. In this Part the basic concepts in the theory of the singularities of differentiable maps will be introduced: singular points, their local algebras and other invariants; concepts concerned with stability will be defined and a start made to the classification of singularities.


The simplest examples

Here we describe the classification due to H. Whitney of the singularities of smooth maps of spaces of small dimensions.

  • 1.1 Critical points of functions

A point







derivative off is zero.

critical point of a function f,

if at


point the

Example: Let f: IR -+ IR be the function, given by the formula y = x 2 • Then the point 0 is a critical point of the function. The critical points of functions are divided into generic or nondegenerate critical points and degenerate critical points.

Definition: A critical point of a smooth function is said to be nondegenerate if the second differential of the function at that point is a nondegenerate quadratic form.

Example: The critical point 0 of the function y = x 2 is nondegenerate, while the critical point 0 of the function y = x 3 is degenerate (Fig. 1).






Consider an arbitrary smooth function, close (along with its derivatives) to

the function y =

x 2 • It is clear that near zero this function will have a critical


Basic concepts

point, similar to the critical point of y = X2. The critical point of y = X2 is stable in the sense that under small perturbations of the function it does not vanish, but simply shifts slightly. The degenerate critical point of the function y = X3 behaves completely dif- ferently under small perturbations.

Example: Consider the family of functions of one variable y = X3 + eX. For small e the functions of this family can be considered as small perturbations of y = X3. We see that under this perturbation the degenerate critical point either

vanishes (for e > 0) or decomposes

into two non degenerate critical points at a

distance of order Jr;T from it (for e < 0). Thus the critical point of y = X2 is stable, while that of y = X3

is unstable.

It is

not difficult to

analyse all possible situations for

a function

of one

variable. We consider the space of all functions of interest to us* We distinguish in this space the set of functions having degenerate critical points or having coincident values at different critical points (Fig. 2.). [In the case where the domain of definition is a line segment we consider an end-point to be criticai, counting it as nondegenerate if the derivative there is nonzero.] It is not difficult to see that such "degenerate" functions form a thin set, to be precise a hypersurface or surface of codimension one "given by one equation" in our function space. We shall later give a precise meaning to this phrase and prove a corresponding theorem in a more general situation. This hypersurface divides our function space into parts, in each of which all the functions are "constructed in the same way": in each part the order relations between the values at successive critical points are the same for each function in the region.



* This could be the space of all infinitely differentiable or sufficiently differentiable functions, or the space of analytic functions, or even the space of polynomials; it is convenient to suppose that the domain of definition of functions is compact, considering functions on a circle or line segment.

The simplest examples


Functions having neither degenerate critical points nor multiple critical values are said to be Morse functions. Under a small perturbation a Morse function "preserves its form" and can be transformed to the original function by smooth changes of the independent and dependent variables x and y. In this sense a Morse function is stable. Thus in the case where the source and target spaces of the maps are one-dimensional the stable maps form an open everywhere dense set in the space of all maps. Moreoever, the stable maps admit a sufficiently clear description and classification, while the unstable maps, though constructed possibly in a much more complicated manner (the set of critical points can be an arbitrary closed set), are transformed into stable ones by a small perturbation:

each complicated singularity breaks up into several nondegenerate stable singularities. The study of maps to the line (Morse theory) provides a special case for which the goals of singularity theory can be attained. The results of Morse theory of interest to us can be stated as follows.

Theorem: (1) The stable maps f: M m ~ [R1 of a closed* manifold M m to the line form an everywhere dense set in the space of smooth maps.

(2) For a map f to be stable two conditions be satisfied:

it is necessary and sufficient that the following



The map f

is stable at each point (in other words all the critical points of

the function are nondegenerate).

M 2. All the critical values of the function are distinct.


A map f: M m ~ [R1 is stable at a point Xo if and only if there are neighbour-

hoodsofxo E M m and Yo = f(xo) E [Rlin which coordinates Xl,"" X m ; ycan be chosen so that the map can be written in one of the following m + 2 forms:

MI. y = Xl



y = xi


+ x~ -

x~+ 1 -

.•• - x~ (k = 0, 1,

, m).

For a proof see, for example, [131]. The question arises as to whether such a result holds true in more dimensions, that is for maps f: M m ~ N n of manifolds of arbitrary dimensions m and


  • 1.2 Critical points and critical values of smooth maps

Consider a

differentiable map f: M m ~ N n.

First of all we must extend to this

case the concept of critical point. The derivative of a map f at


point X



* Here and in the sequel a closed manifold will mean a compact manifold without boundary.


Basic concepts

linear map of the tangent space of the source manifold at the point x to the tangent space of the target manifold at the point f(x):

Example: Let M 2 be the surface of a sphere in three-dimensional space, N 2 the plane and f the projection of the sphere vertically on to the horizontal plane (Fig. 3).




The linear map /.X of the plane TxM, tangent to the sphere at x, to the plane TJ(x)N, tangent to the horizontal plane, is a nondegenerate linear map if the point x does not belong to the horizontal equator of the sphere. If, however, x lies on the equator, then the tangent plane to the sphere at x contains a vertical line. In this case the projection operator has a nontrivial kernel (the subspace mapped to zero). The kernel of /.X at points of the equator is one- dimensional. The rank of/.x at these points is equal to 1. We make now the more general

Definition: A point x of the manifold M is said to be a critical point ofthe smooth map f: M -+ N if the rank of the derivative

at that point is less than the maximum possible value, that is less than the smaller of the dimensions of M and N:

rank/. x < min (dim

M, dim N).

The simplest examples


in M and let Yl,


, Yn be local coordinates in a neighbourhood of the pointf(x)

in N. In terms of these coordinates the map f is given by n smooth functions in m variables:

The matrix (oJi/OXj) is called the Jacobian matrix of the map. In these terms

one can say that the point x is critical if the rank of the Jacobian matrix there is not maximal.

Example: For the projection of the sphere to the horizontal plane the critical points are the points of the horizontal equator. Off the equator the rank of the derivative is equal to 2, while at points x of the equator the rank of the operator /.X falls to The image of a critical point is called a critical value.


Example: The critical values of the projection of the sphere to the plane form the circle of the apparent contour of the sphere.

  • 1.3 Differentiable equivalence

There are several different ways of classifying smooth maps. Clearly the crudest classification is the topological one: two maps are said to be topologically equivalent if there are homeomorphisms (one-to-one continuous maps with continuous inverses) of the source and target manifolds, transforming the one map into the other. The functions Y = x 2 and y = X4 are topologically equivalent. If fp: M p --+ N p, p = 1, 2 are two given maps then to say that they are topologically equivalent means that there exist homeomorphisms h: M 1 --+ M 2 and k:N 1 --+N 2 such thatfz = k/Ih- 1 • In other words a topological equivalence is a commutative diagram

in which the vertical arrows are homeomorphisms.


Basic concepts

For the purposes of analysis topological equivalence, as a rule, is too crude a concept. For example, the function y = X4 with a degenerate unstable singularity is topologically equivalent to a stable one. Therefore in the theory of singularities a second concept is basic, the concept of differentiable equivalence.

Definition: A differentiable equivalence of differentiable maps fl : M 1 ---+ N 1 and fz : M 2 ---+ N 2 is a commutative diagram

whose vertical arrows are diffeomorphisms (differentiable one-to-one maps whose inverses also are differentiable*).

Remark 1: In the language of local coordinates the map y = f(x) is a set of functions, the diffeomorphism h is a change of the independent variables x and the diffeomorphism k is a change of the dependent variables y. From this point of view the question of differentiable equivalence is the question whether one can transform one map into the other by means of smooth changes of the independent and dependent variables.

Remark 2: The commutative diagram depicted above represents the identity

In this formula

h - 1



the right of fl



to the left.

Therefore the

diffeomorphism h -1 of the source space (and the change of the independent

variables x) is said to be a right change. Likewise the diffeomorphism k of the

target space (and the change of the dependent variables y) is said to


be a


* Here and in the sequel unless there is explicit mention to the contrary the word differentiable or smooth means "continuously differentiable the necessary number of times", for example infinitely differentiable.

The simplest examples


Remark 3: Yet another way of expressing the same thing consists in the following. Consider the set O(M, N) of all smooth maps from M to N. Consider the group Diff M of all diffeomorphisms of the source manifold M with itself and the group Diff N of all diffeomorphisms of the target manifold N with itself. The direct product of groups

Diff M x Diff N

consists of all pairs (h, k) of diffeomorphisms of the source space (h: M -+ M)

and the

target space (k: N -+ N ).

The group Diff M x Diff N

acts on the set O(M, N) in the following way: if

fEO(M, N), hEDiffM, kEDiffN, then (h, k)f= kofoh- l


It is not difficult to verify that this is a genuine group action, that is that

This action is said to be left-right (the action of DiffM is said to be right and the action of Diff N to be left). * In these terms we can reformulate the definition of differentiable equivalence

as follows: two maps from M to N are differentiably equivalent if and only if they belong to the same orbit of the left-right action (Fig. 4).


Fig. 4.

Example: The connected components of the set of all Morse functions (defined in section 1.1) are the orbits of the left-right action.

1.4 Stability Consider a smooth map f: M -+ N of a closed manifold M to a manifold N.

* Not to be confused with left action in algebraic terminology.


Basic concepts

Definition: A map fis said to be difJerentiably stable (or more precisely left-right- difJerentiably stable, or briefly simply stable), if every map sufficiently close to itt is differentiably equivalent to it. In other words, f is stable, if its left-right orbit is open.

Example: The projection map of the sphere to the plane is stable. The map of

the circle



mod 2n}


the line

{y}, given

by the formula

y = sin 2x,


Remark: If one replaces differentiable equivalence in the preceding definition by topological then one obtains the definition of topological stability.

Example: The projection map of the sphere to the plane is topologically stable, as is any differentiably stable map. Topologically stable but differentiably unstable maps exist, but it is not easy to give an example (see [128]). The formula y = sin 2x defines a topologically unstable map from the circle to the line.

There exist also local versions of the concepts that we have introduced. For example, the singularity at zero of the function y = x 2 is stable, while the singularity at zero of the function y = x 3 is unstable. To give a formal definition of the stability of a map at a point we employ the following terminology.

Definition: A map-germ M -+ N at a point x of M is an equivalence class of

maps <p: U -+ N (each of which is defined on some neighbourhood U of x


M, not necessarily the same for each); here two maps are regarded as equivalent

if they coincide on some neighbourhood of the point x (this neighbourhood may well be smaller than the intersection of the neighbourhoods on which the two maps are defined). Two maps of the same class are said to have the same

germ at the point x

(Fig. 5).

In other words the germ

of the map f

at x

is what remains of it when "its

domain of definition is made infinitely small".

t Differing from f by a sufficiently small amount, provided that a sufficiently large number of derivatives are taken into account.

The simplest examples







Definition: Two smooth map-germs are said to be (left-right, dif.ferentiably) equivalent if there are germs of diffeomorphisms of the source and target spaces transforming the first germ into the second (if the map-germfl at Xl is equivalent to the map-germf2 at X2 then there exist a diffeomorphism-germ h at Xl sending Xl to X2 and a diffeomorphism-germ k at fl(xd sending /1(xd to f2(X2), such that k(fl(h-l(x))) =h(x) in a sufficiently small neighbourhood of X2). The equivalence class of a germ at a critical point is said to be a singularity.

Definition: A smooth map-germ f: M -t N at a point X of M (Fig. 6) is said to be (left-right, dif.ferentiably) stable if for a sufficiently small neighbourhood V of X there is a neighbourhood E of the map* f in Q(M, N), such that for any mapfin E there is a point x in V such that the germ offat x is equivalent to the germ off at x.








* A neighbourhood of a given map is the set of all maps, differing only slightly from the given one when derivatives up to a fixed order are taken into account. In the local situation of interest to

us at the moment we can suppose that M c

IR m and

N c

IR" are

regions of Euclidean spaces and

that E is

given by the inequalities 11(1-1) Ilk < B, where II g Ilk = sup lo"g/ox"l.

1"1 ';k


Basic concepts

It is not difficult to

verify that stability at a point is a property of the germ

and not of the map: the property is not destroyed by changes in f that leave alone some neighbourhood of x.

The topological equivalence of germs and the topological stability of germs

are defined in similar fashion.

Example: The map-germs y = x 2 and y = X4



point 0

of the



are topologically equivalent. The germ y = x 2 is topologically (and even

differentiably) stable at zero. The germ topologically) unstable at zero.

y =


is differentiably (and even

  • 1.5 The stability of maps from two-dimensional manifolds to two-dimensional manifolds.

We begin with an example that we looked at earlier - the projection map of the sphere to the plane (Fig. 3). The singularities of the projection are the points of the equator of the sphere. It is not difficult to verify that the map-germ at each point of the equator is stable. It is difficult to imagine how other stable singularities of maps of two- dimensional manifolds to two-dimensional manifolds can exist. In fact what we see when we look at the smooth surfaces of three-dimensional bodies are their apparent contours, consisting of the critical values of the projection map of the surface to the retina of the eye. It usually seems to us that these apparent contours consist of smooth curves. However, if we look about us more atten- tatively (for example, at the faces of the people around us) then we may discern alongside singularities like the singularities of the projection of the equator of the sphere singularities of yet another type. These singularities were discovered by H. Whitney, who in 1955 [192] gave a complete description ofthe singularities of a generic map from a two-dimensional manifold to a two-dimensional manifold. This work of Whitney is the foundation of singularity theory whose date of birth can therefore be taken to be 1955.

Whitney established that every smooth map of a compact two-dimensional manifold to a two-dimensional manifold can be approximated arbitrarily closely (for any number of derivatives) by a stable map.

Moreover he studied the structure of stable maps. At each point the germ of such a map is stable. Whitney described all stable map-germs of two-dimensional manifolds (there are three of them, up to differentiable equivalence). Finally he proved that a map of a compact two-dimensional manifold to a two-dimensional manifold is stable if its germ at each point is stable and the critical values are

The simplest examples


disposed "in a general manner" (this being a generalisation of the condition of non-coincidence of critical values at different critical points for Morse functions, see Section 1.1).

Whitney's Theorem: A map of a two-dimensional manifold to a two-dimensional manifold is stable at a point if and only if the map can be described with respect to local coordinates (Xl, X2) in the source and (YI, Y2) in the target in one of the three forms:



= Xl,


= X2 (a regular point);




xi, Y2 =






= xi + XIX2,


= X2



(the point under consideration has the coordinates Xl = X2 = 0).

In other words, every stable map-germ of a two-dimensional manifold to a two-dimensional manifold is differentiably equivalent to one of the three map- germs at zero of the given list. The first of the given germs is a diffeomorphism-germ. Every smooth map of a two-dimensional manifold reduces to this form in the neighbourhood of a

noncritical point. A singularity of a map of the second kind is said to be a fold.

This map of

the plane to the plane can be considered as a family of maps of the line to the line (YI = xi), depending trivially on one parameter (Y2 = X2).

Example: The projection of the sphere to the horizontal plane has a singularity of fold type on the equator, as one may easily verify, by choosing suitable local coordinates (X2 and Y2 as longitude and Xl as latitude, Fig. 7).




Basic concepts

  • 1.6 The Whitney pleat

The Whitney pleat, frequently called the

Whitney cusp, is


third stable

singularity of the above list, that is the singularity of the map

at zero.




singularity clearly we realise it as

a singularity of the

vertical projection of a smooth surface in three-dimensional space to the

horizontal plane.





With this in view, consider the graph of the function Yl = xi + X1X2 in three dimensional space with coordinates (Xl, X2, yt} (Fig. 8). This graph is diffeomorphic to the plane (every graph of a smooth map is diffeomorphic to its domain): as coordinates on the graph we may take Xl and X2. Consider the intersection of the graph with the vertical plane X2 = const. For a fixed value of X2 the equation Yl = xi + X1X2 determines a cubic curve, lying in the vertical plane. If X2 > 0 then along the corresponding cubic curve Yl increases monotonically with Xl. If however X2 < 0 then Yl has two critical points - a local maximum and a local minimum. Consider now the projection of our graph to the horizontal plane, (Xl, X2, yd~(X2' Yl). The resulting smooth map of the surface to the plane has a singularity of pleat type at the origin. In fact consider the following systems of coordinates: (Xl, X2) on the graph and (Yl, Y2 = X2) on the horizontal

The simplest examples


plane. With respect to these coordinates the projection is described exactly by the formulas of the Whitney pleat. Whitney's theorem asserts that the singularity is stable. In particular under a small perturbation of our surface in three- dimensional space a surface is formed whose projection to the horizontal plane has a singularity of the same type at some point near to the origin.

Problem: Find the critical points of the Whitney map


= xf + XIX2, Yz = Xz·

Solution: The Jacobian matrix has the form

At critical points the rank of this matrix is less than two, that is its determinant

is equal



3xi + Xz = O.

Consequently the set

of critical




smooth curve. In the (Xl, xz) plane the equation 3xi + Xz = 0 determines a parabola (Fig. 9).


X 2





On our graph in three-dimensional space the critical points of the projection are those points where the tangent plane to the graph contains a vertical line. It is clear that these are just the critical points of the functions YI on the cubic curves described above. We conclude that all these points form a smooth curve on the graph (this is not obvious without calculation).


Basic concepts

Problem: Find the critical values of the Whitney map.

Solution: At critical points X2 = - 3xi. Substituting for X2 its expression in terms of Xl, we obtain the parametric equations of the set of critical values:



xi + X1X2


-2xi, Y2 = X2 = -3xI.

Thus the set of critical values is a semicubical parabola on the (Yl, Y2) plane. This curve has a singular point (a cusp, also called a point of regression) at the origin. It divides the plane into two parts. Under the Whitney map every point in the smaller part has three preimages while a point in the larger has only one.

Remark: If one looks down on the surface of the graph then one sees only the upper fold, ending at the pleat point, and it has the form of one half of a semi cubical parabola. Most of the surfaces that we see are opaque surfaces. Therefore we usually see that the fold terminates at a pleat point, but we do not notice the cusp.

Example: Consider the surface of a torus in three-dimensional space as depicted in Fig. 10. Two pleat points are clearly seen. If the torus was transparent then we would see the picture drawn in Fig. 11, with four pleats. We seldom come across transparent tori. More frequently we come across the smooth surface of





The simplest examples




a glass bottle. If one looks at the neck it is easy to make out two pleats (Fig. 12). By moving the bottle one can convince oneself of their stability.

1.7 Catastrophes

Whitney's theorem asserts that folds and pleats are not destroyed by small perturbations, while all more complicated singularities break up under a small perturbation into folds and pleats and therefore do not occur for generic smooth maps of two-dimensional surfaces. One comes across smooth maps everywhere. By Whitney's theorem we must therefore everywhere come across fold contours and cusps of semicubical parabolas on them. This remarkable theorem has given birth to much speculation, associated principally with the names of R. Thorn and C. Zeeman. They gave the name of catastrophe theory to the whole subject of applications of Whitney's theorem. Hundreds of papers have been published and are being published on catastrophe theory, being mainly concerned with applications. The usual form of argument in catastrophe theory is as follows. One considers a smooth surface in three-dimensional space about which usually almost nothing is known, together with a projection map to the plane. Since nothing is known about the surface it is supposed to be a surface in general position. In that case the singularities of the projection are folds and pleats. Already this alone leads to some information about the jumps or "catastrophes" which the objects of study can undergo. We give below an example of such an "application". It is more or less borrowed from the work of Zeeman. Let us characterise the creative personality (say of a scholar) by three parameters, which we call "technique", "enthusiasm" and "achievement". Clearly these parameters must be interdependent. In fact a surface is obtained in three- dimensional space with coordinates (T, E, A). We project this surface to the (T, E)


Basic concepts



plane. For a surface in general position the singularities of the projection are folds and pleats. It is asserted that a pleat, placed as depicted in Fig. 13, satis- factorily describes the phenomena under investigation. Consider, in fact, how under these hypotheses the achievement of a scholar varies with his technique and enthusiasm. If his enthusiasm is weak then his achievement will increase monotonically and rather slowly with his technique. If, however, his enthusiasm is sufficiently great then qualitatively new phenomena appear. In this case his achievement may increase with a jump as his technique increases. The area of high achievement which we reach in this way is indicated on our diagram by the word "geniuses". On the other hand, an increase in enthusiasm not reinforced by a correspond- ing increase in technique leads to an area indicated on our diagram by the word "maniacs". It is noteworthy that the "catastrophes", the jumps from the state of "genius" to the state of "maniac" and back again, take place on different lines, so that with a sufficiently high degree of enthusiasm geniuses and maniacs may have the same amount of technique but differing amounts of achievement. The defects in the model we have described are too obvious for it to be necessary for us to speak of them in more detail (see, for example, [148]). There are serious applications of the theory of singularities, for example in the theory of elasticity, in optics (the singularities of caustics and wave fronts), in the theory of oscillating integrals (the method of stationary phase), and so on. We return to these applications after we have developed the appropriate techniques.

The simplest examples


  • 1.8 The kernel field of the derivative for folds and pleats. The disintegration of more complicated singularities

We return to the Whitney pleat (Fig. 9) f: 1R2 -+ 1R2:

Problem: To find the kernel of the derivative /.X of the map f at the point x.

Solution: The Jacobian matrix

has rank 2 at all points apart from the critical ones. The critical points form the parabola 3xI + X2 = O. At those the rank of the Jacobian is equal to 1. Consequently the kernel of the derivative has dimension 1 at all these points (not only at the fold points, but also at the pleat point). Therefore, there is a field of lines on the parabola of critical points, the kernel field of the derivative. As regards the Jacobian matrix we remark that the kernel of the derivative is at each point parallel to the Xl axis (Fig. 14).


X 2



Remark: From the solution of the problem it is clear that the kernel of the derivative touches the curve of singular points only at the pleat point. This remark frequently enables us quickly to find the pleat points of a generic map of two-dimensional manifolds. According to Whitney's theory for such maps:

(1) the set of critical points is a smooth curve, (2) the kernel of the derivative at each point of the curve is one-dimensional, (3) at general points of the curve


Basic concepts

of singular points the kernel of the derivative is transversal to the curve, but at particular points of the curve the kernel touches it. These latter points are the pleat points.

Example: Consider the map of the plane of the complex variable z = Xl + iX2 to the plane of the complex variable w = Yl + iY2, given by the formula w = Z2, as a smooth map from the two-dimensional real plane to the two-dimensional real plane:

The rank of the derivative is equal to two everywhere except at the point z = 0, where the derivative is equal to zero and where, consequently, the dimension of the kernel is equal to two. We conclude, that the map under consideration has a singularity at the origin other than a fold or a pleat. Consequently by Whitney's theorem it is unstable and under a small perturbation its singularity at zero must break up into folds and pleats.

In the class of germs of holomorphic maps C 1


C 1 the map w = Z2 is stable.

Therefore to put the map in general position in the class of real maps it is absolutely necessary to make an addition that breaks the holomorphicity. The simplest perturbation of all is to take the map

w = Z2 + ez.

This map is near to

the original one within an arbitrary fixed

circle if Ie I is

sufficiently small. Multiplication of z and w by a real number allows us to vary e, and so it is sufficient to take any number we please for e in our map - it does not have to

be small. We take e =

2 and study the singularities of the real plane z to the plane w,

given by the formula

w = Z2 + 2z.

Problem: To find the set of critical points of this map.

Solution: The derivative of our map takes the


dw(~) = 2z~ + 2;; on


The simplest examples


vector~. The derivative is degenerate if the equation 2z~ + 2'; = 0 has a nonzero

solution. We get z =

-,;/~, from which it follows that Izl = 1. Thus the set of

critical points is the circle of radius 1.

Problem: To find the kernel field of the derivative on the circle of critical points.

Solution: From the formula z =


~g obtained

above it

is clear that

when z

goes round the circle once in the positive direction the kernel completes a turn through the angle n in the negative direction. Moreover the kernel touches the circle at the point z = 1. Therefore the kernel field takes the form depicted in Fig. 15. The kernel field touches the circle of critical points at three points.



Problem: To find the set of critical values of the map under consideration.




Solution: Let the critical point be z = ei'P. The corresponding critical value is w = 2e- i 'P + e 2i 'P. This is the parametric equation of a hypocycloid with three


Basic concepts

cusps (Fig. 16). In fact consider a large stationary circle of radius 3, inside which a small circle of radius 1 is rolling, touching it. Then the centre of the small circle describes a circle of radius 2 with angular velocity half that of the speed of rotation of the small circle, and in the opposite sense. A point of the small circle describes the hypocycloid. This is expressed by the given formula.

Remark: It can be proved that the map under consideration has only stable singularities. Then from the results of the preceding problems it follows, by Whitney's theorem, that our map has three pleats, joined together by the fold circle. Thus the complicated singular point of the map w = Z2 has under a small deformation (w = Z2 + eZ) broken up in such a way that three pleats have been formed.

Notice that for small lei the radius of the circle of critical points is so small

(it is equal to

Ie 1/2) that the three pleat points

are all close to one another.

Problem: describe the images of circles Iz I = r of various radii under the map

w = Z2 + 2z.

Solution: Since 0 maps to 0 the image of a circle of small radius differs slightly from a small circle with centre at O. As r increase from 0 to 1 we obtain a family of smooth curves, terminating with the hypocycloid with three cusps (Fig. 17, (a)). The image of a circle of slightly greater radius in the neighbourhood of the images of the pleat points may be drawn from our knowledge of the pleat model (Fig. 8). In the neighbourhood of the image of a pleat point this image of the circle is on the same side of the hypocycloid as the image of the circle whose radius is slightly less than one. We obtain a curve with' three loops, close to the hypocycloid (Fig. 17, (b)).




A ,



























1 '"











The simplest examples


With further increase of the radius of the preimage circle the loops grow in size and the curve moves away from the hypocycloid. For a particular value

of r there is a

triple point (Fig.

17, (c)), and then a curve passing through the

images of the three pleats (Fig. 17 , (d)); finally we get a curve not intersecting the set of critical values, with three points of self-intersection (Fig. 17, (e)).

Remark: The last curve is close to a circle and goes round w = 0 twice as z goes round zero once. In fact the map w = Z2 + 82 in any bounded domain tends to

the function w = Z2 as 8 --+ 0; therefore "near infinity" behave "almost like" w = Z2.

the map w = Z2 + 22 must

Problem: How many preimages does the map w = Z2 + 22 have at points of the regions inside and outside the hypocycloid of critical values?

Answer: Inside 4, outside 2.

It can be proved

that the singularity of the map




at zero breaks up

into three pleats not only for the perturbation w = Z2 + 82, but also for any sufficiently small generic perturbation.

Problem: Consider the map of the plane to the plane which is the direct product of two folds:

Yl = xi, Y2 = x~.

Study the singularities of its small perturbation

Answer: See Fig. 18.

  • 1.9 Singularities of maps of two-dimensional manifolds to three-dimensional manifolds

Whitney also described the singularities of generic maps of two-dimensional


Basic concepts



manifolds to three-dimensional manifolds. It turns out that there is also a finite number of these. The image of such a map forms a surface in three-dimensional space. It goes without saying that this surface may have lines where two sheets of the surface intersect each other at a nonzero angle and particular points where three sheets intersect. It turns out that apart from these obvious singular- ities generic maps may only have one other type of singularity. This singularity is stable. The image of the corresponding map is a remarkable surface in three- dimensional space. This surface is called the Whitney (or Cayley) umbrella. This umbrella is depicted in Fig. 19.






The illustrated surface intersects the



= const


pairs of lines

yi = ay~




Y2 = const




= byI.



equation has the form YI = Y3Y~.

Remark: This last equation is satisfied in 1R3 not only by points of the depicted surface, which has a line of self-intersection along the positive Y3 axis, but also by all the remaining points of the Y3 axis. Therefore the set given by this equation

The simplest examples


indeed does have the form of an eccentric umbrella, whose handle is the negative Y3 axis.

Problem: Find a smooth map 1R2 -+ 1R3 whose image is the Whitney umbrella (without its handle).

Definition: The Whitney singularity of a map 1R2 -+ 1R3 is the germ at zero of the

map of the plane to space, given by the above formula.

Whitney proved that this singularity is stable, that every map of a compact two-dimensional manifold to a three-dimensional manifold can be approximated by stable maps and that stable maps have no other singularities (see [195], [196]).

  • 1.10 Other dimensions

A generic map of the circle to three-dimensional space does not have any singularities: by a small shift one can get rid of them. Whitney proved (as long ago as the 30's) that a generic smooth map f: M m -+ N" has no singularities (that is, is an embedding) if the dimension of the target is sufficiently great, to be precise if n > 2m. * For n = 2m the singularities can also easily be enumerated (only transversal selfintersections of the image occur). See [193]-[196]. The opposite case where the target is of small dimension is the case of a single function, n = 1. In this case also all the singularities of a generic map are stable, all the stable maps being given locally by a finite list (y = x or

y = ± xi ±


± x;').

* The dimension of the space of chords of a manifold M m embedded in IRk is equal to 2m, the space of tangent lines has dimension 2m - 1 while the space ofline directions in IRk has dimension k - 1. Therefore for almost all directions of projection the projection of M m to the complementary

(k -

l)-dimensionalspace is an embedding, if k > 2m + 1. For sufficiently large k a smooth map of M m

to IRk is easily approximable by an embedding. By the series of projection IRk -+ IRk - 1 -+ .•• -+ 1R 2m + 1


Basic concepts






Fig. 20.



In Fig. 20 the values of the dimensions of the source and target spaces that we have considered up to now are indicated on the (m, n) plane. For all these dimensions generic maps are stable. It turns out that such a good situation does not hold for all (m, n). For example for m = n = 9 the stable maps are not dense in the space of all maps and there is not a finite list of singularities of generic maps. The classes of differentiable equivalence of singularities of generic maps at a point do not in this case form a discrete set but a continuous set (there are continuous invariants of singularities, the so-called moduli). Moreover for large m and n the number of moduli becomes infinite and the singularity types begin to depend on arbitrary functions, the number of whose arguments increases with the growth of dimensions. We return to this question in Chapter 3 (Section 3.7).


The classes LI

Here singularities are classified according to the rank of the first differential of the map and the ranks of its restrictions to submanifolds of singularities.

  • 2.1 The classification according to the degeneracy of the first differential

Let f: M m --+ N n be a smooth map, /.x: TxMm -+ TJ(x)N n its derivative at x.

Definition: The point x


said to

be a point of class :E i for J, if the dimension

of the kernel of /.X is

equal to


All the points of class :E i for f

form a subset

of M, called the set :E i for f and denoted by :EV).

Example: For the map of Whitney pleat (Fig. 21)

Fig. 21.

all the critical points are of class :El, while the noncritical points are of class :Eo.

Remark: In particular, the fold points and the pleat points are of the same class :E 1.

The singularity of the map w = Z2 of the real plane at zero is of class :E 2 By Whitney's theorem generic maps of two-dimensional manifolds do not have


Basic concepts

singularities of class I:Z. The question arises: what is the structure of·the set I,i(f) for a generic map f: M m -> N n? In particular, what is its dimension and when is it nonempty? To give an answer we require the

Definition: Let A: IR m -> IR n be a linear operator of rank r. The differences m - r

and n -

r are called the coranks of A at the source and target respectively.

Remark: The coranks are related to the dimension of the kernel i by the obvious

formulas: m -

r = i, n -

r = n -

m + i.

Theorem: ("the corank product formula"). For a generic* map all the sets I,i(f) are smooth submanifolds of the source space. The codimension of the manifold 'f.i(f) is moreover equal to the product of the coranks:

dimM -

dim'f.V) = (m -

r)(n - r)

(if this number is negative then the set is empty).

To understand where this formula comes from we consider first the cor- responding problem in linear algebra.

  • 2.2 The stratification of the space of linear operators

Consider the set of all linear operators A: IRm -> IRn. This is a linear space of finite dimension mn (if a basis is chosen one can identify the operators with matrices of order m x n). We shall denote this space by L(m, n). The groups of linear changes of coordinates in the source space GL(m) and in the target space GL(n) act on the space of matrices L(m, n) and give rise to a left-right action of the direct product of the two groups. Two matrices lie in the same orbit of this action if they are matrices of one and the same operator for different choices of bases in the source and target spaces. The matrix of any operator A can for a suitable choice of bases be put into the special form:

* The set of maps not satisfying the conclusion of this theorem is at worst a countable union of closed nowhere dense sets in the space of smooth maps; moreover if M is compact then the set of "generic" maps in this context is open and everywhere dense.

The classes :E I


(Ao) = ( ~rl~




where Er is the unit matrix of rank r = rank A. Therefore the set of all matrices of order m x n and of rank r is one orbit of the left-right action of the group

GL(m) x GL(n) on L(m, n).

Lemma: The set of all matrices of rank r in

L(m, n) is a smooth submanifold, whose

codimension is equal to the product of the coranks.

Proof: Since the matrix of any operator of rank r can be put in the form Ao for an appropriate choice of bases it is enough to prove the lemma in a neighbour-

hood of this matrix. We put a matrix close to Ao in the form A = Ao + (ak.I). For small (ak.l) the rank of A is not less than r. It is equal to r if and only if all the minors of order r + 1 bordering Er are equal to zero. We get a system of equations in the elements of the matrix (ak.I), determining a manifold Lr in a neighbourhood of the matrix Ao. The number of these equations is equal to the number of minors of order r + 1 bordering E r , that is, is equal to the product of the coranks (m - r) x (n - r). These equations are independent. Indeed consider the bordering minor, obtained by adding in the row and column which intersect at ak,l. The expansion of this minor as a Taylor series in a begins with ak,1 + O(a 2 ). Therefore the differentials of our (m - r)(n - r) minors are independent at zero. Once the differentials of the minors are independent then, by the implicit function theorem, setting the minors to zero defines a submanifold whose

co dimension is equal to the number of

equations, as was claimed.

The partition of the space of all linear operators (matrices) L(m, n) into the submanifolds L, of operators (matrices) of different ranks is called the natural stratification and the manifolds L, are the strata. We have computed above the codimensions of these strata: they are equal to the products ofthe coranks. For

example, for m = n the codimensions of the strata of corank 1, 2, 3,

respectively to 1, 4, 9, ...


are equal


Basic concepts

Problem: Find the codimension of the set of symmetric matrices of corank k in the space of all symmetric matrices of order n.

Answer: k(k + 1)/2, that is

1, 3, 6,


for k =

1, 2, 3, ...

  • 2.3 The transversality theorem

To deduce the theorem formulated in Section 2.1 from the algebraic lemma of Section 2.2 it is convenient to employ some general concepts and theorems.

Definition: Two linear subspaces of a finite-dimensional linear space are said to be transversal if their sum is the whole space (Fig. 22).

Fig. 22.

Example: In three-dimensional space two one-dimensional subspaces are never transversal, while two non-coincident two-dimensional subspaces are always transversal, a one-dimensional and a two-dimensional being transversal only if the one-dimensional does not lie in the two-dimensional.

Definition: Let f: A --+ B be a smooth map of a manifold A

to a manifold B,

containing a smooth submanifold C. The map f is said to be transversal to C at the point a of A if either f(a) does not belong to C or (Fig. 23) the image of the tangent space to A at a under the derivative f.a is transversal to the tangent

space to C:

The classes 1/






Fig. 23.

The map f is said to be transversal to C if it is transversal to C at every point of A.

Proposition: Iff: A ~ B is transversal to C then f


l( C) is a smooth submanifold in

A, having the same codimension in A


C has

in B.

Example: Let C be a curve in a three-dimensional space B and let A be one- dimensional. Thenf: A ~ B is transversal to C if and only if the image of A does not intersect C.

Remark: The image of A under the mapfmay be a submanifold of B, transversal


C, and

yet the map f

may not be transversal to C.

Example: The map of the line A the submanifold C (b 1 = 0), given

= {a}

to the plane B = Ub 1 , b 2 )},

by the equations b 1 = a 3 , b 2 = O.


Remark: The most important case is the case where B is a linear space and C a subspace of B. Denote by D the quotient space and by p the natural projection p: B ~ D sending C to zero. In these circumstances the map f: A ~ B is transversal

to C if and only if zero is

a non-critical value of the map pof:A ~ D. (A critical

value of a map is its value at a critical point; a critical point is a point where the derivative is not "onto"; for a map f: M m ~ N n with m < n all the points of Mare

critical and the critical values are the points of f(M).)

The weak transversality theorem: In the space of smooth maps of a closed* manifold

* Compact without boundary


Basic concepts

A to a manifold B, containing .a closed submanifold C, the maps. transversal toC form an open everywhere dense set.

Remark: The possibility of putting the map into general position by a small shift and so eliminating non-transversality is intuitively sufficiently obvious (Fig. 24).

Fig. 24.

It is convenient to prove this with the help ofthe following theorem, proved in the setting of algebraic geometry by Bertini and in the case of smooth functions by Sardo

The Bertini-8ard Theorem: The measure of the set of critical values of a sufficiently smooth map is equal to zero.

Remark: The measure of the set of critical points may be positive. Example:


== 0.

Proof of the Bertini-8ard Theorem: We begin with the simplest particular case.

Proposition 1: Let y = f(x) be a smoothfunction on [0, 1]. Then the measure of the set of critical values off is equal to zero.

Proof: Divide [0, 1] into N segments of equal length. Note the segments contain- ing critical points. The measure of the image of one of these segments is bounded

above by C/N 2 ,

where C is independent of N and the segment (the modulus

of the derivative off on the segment being bounded above by C/N). The sum

of the measures

of these segments therefore does not exceed N . (C/N 2). As

The classes L I


Proposition 2: Let f be a smooth map of the m-dimensional cube [0, 1]m to m- dimensional Euclidean space. Then the measure of the set of critical values off is equal to zero.

Proof: Arguing as in the proof of Proposition 1 we cover the set of critical values by no more than (N)m sets, the measure of each of which is no greater than C(l/N)m+ 1, which proves Proposition 2.

Proposition 3: Let f be a smooth map of [0, 1]m to n-dimensional Euclidean space.

Then for sufficiently large k the measure of the image of the set of points where all

the derivatives off of orders 1, (k + l)n > m).


, k are zero is equal to zero (it is sufficient to take

Remark: In particular, for n > m the measure of the whole image of the cube is equal to zero.

Proof: Arguing as in Propositions

1 and


we cover the

set of values

by no

more than (N)m sets of diameter not greater than C 1 (1/N)k+l. This gives an

upper estimate for the measure of the form CN m (l/N)<k+lJn, which proves Proposition 3.

For maps into a space of the same or a greater number of dimensions as the number of dimensions of the manifold under consideration the theorem is indeed proved. In the case of maps to a space of a smaller number of dimensions it is necessary to have some further discussion. We consider the simplest case of a single function.

Proposition 4: Let f be a smooth function of two variables. Then the measure of the set of critical values off is equal to zero.

Proof: The measure of the set of values at points where all three second derivatives are equal to zero is equal to zero by Proposition 3 (for m = 2, n = 1 it is sufficient to take k = 2). Consider the critical points at which one of the second derivatives, say o2f/ox 2 , is different from zero. The equation of/ox = 0 determines a smooth curve (by the implicit function theorem). The critical points


Basic concepts

of interest to us lie on this curve and are critical for the restriction off to this curve. By Proposition 1 the measure of the set of values offat these critical points is equal to zero. For the other second derivatives the argument is the same. Proposition 4 is thus proved.

Proposition 5: Let f be a smooth function of m variables. Then the measure of the set of critical values is equal to zero.

Proof: Arguing as in Proposition 4, we prove Proposition 5 by induction on m. Consider the set of points where all the derivatives up to order q inclusive are equal to zero while one of the derivatives of order q + 1 (say o<p/ox, where <p is a derivative of order q) is different from zero. This set is contained in the set of critical points of the restriction off to the submanifold of dimension m - 1, given by the equation <p = O. In this way Proposition 5 is reduced to Propositions 3 and 1.



case of maps

to space of dimension n> 1 the same induction is

involved, but here, however, there is a new difficulty: at a critical point, generally speaking, the first derivatives are not all equal to zero.

Proposition 6: Let the map f be given by n functions Ji in m variables Xj, and let oft! ox 1 =I 0 at some point. Then in a neighbourhood of this point one can choose coordinates ~ such that the map can be put into the form of a one-parameter family g~, of maps of(m - I)-dimensional space to (n - I)-dimensional space:

Y1 = ~1, y' = gd~')(~' = ~z,


, ~m, y' = Yz,


, Yn).

Proof: We introduce local coordinates ~1 = iI, ~' = x'. By the implicit function theorem this is a coordinate system. Proposition 6 is thus proved.

It is clear that the critical values of the map g obtained in Proposition 6 is

the union of sets S = u( ~ 1, S( ~ 1)), where S( ~ d is the set of critical values of the

map g~,. ThereforemeasS = 0, if (for each

~dmeasS(~d = o (Fubini's Theorem).

Now the Bertini-Sard Theorem is proved by induction on the dimension of the mapped space: in neighbourhoods of those critical points where not all the first derivatives are equal to zero we lower the dimension by Proposition 6 and at critical points where all the first derivatives are zero we lower it by Propositions 5 and 3.

The classes I/


Proof of the weak transversality theorem: We consider the special case where

C is a linear subspace in B. We embed the given map f:A --. B in the family of maps!e, where !e(x) = f(x) - S (SE B). Let p:B --. BIC be the natural projection and let SEB be such that p(s) is a non- critical value of the map pof Such S exists by the Bertini-Sard Theorem applied to pof Then!e is transversal to C. This proves the assertion concerning every- where denseness for the transversality theorem in the special case considered. The general case may be reduced to the special case, something that we do not stop to do here; the openness of the set of maps of a closed manifold, transversal to a closed submanifold, is obvious. The weak transversality theorem cannot be applied directly in the proof of the co rank product theorem (Section 2.1) since the map to which it should be applied is the derivative of the given smooth map M --. N. The weak transversality theorem guarantees the possibility of putting it in general position in the class of all maps, by a small shift, and by no means guarantees that the perturbed map will be any kind of derivative, that is that putting the derivative into general position can be achieved by a perturbation of the original map M --. N. An analogous difficulty arises in many other problems. In turns out, however, that this difficulty is always surmountable: the integrability condition does not prevent the move to general position. To formulate the corresponding theorem, which is called the strong transversality theorem (or simply the transversality theorem) we need some concepts, which we shall constantly be using in the sequel. Letj:Mm--.N n be a smooth map and x a point of M (Fig. 25).




Fig. 25.


Definition: The map Jhas at x contact order k with!, if

where PN and PM are any Riemannian metrics on Nand M respectively.


Basic concepts

Example: Let m = n = 1, that is f and] are functions of one variable. Contact order k = 0 implies the coincidence of the values at the point x: ](x) = f(x). First order contact implies the coincidence (at x) of the values f(x) and the derivatives f'(x); contact of order k the coincidence of the parts of the Taylor series up to the terms of degree k inclusive at x. Note that the concept of contact does not depend on the choice of metrics referred to in the definition. Contact of order k at x is an equivalence relation.

Definition: The equivalence class of smooth maps with respect to the equivalence relation of "contact of order k at x" is called the k-jet at x.

Notation: The k-jet off at x is denoted by j~f

Example: The I-jet of a function of one variable is determined by three numbers


y, p = dy/dx).

If one fixes a system of coordinates then the k-jet may be represented as the Taylor polynomial of degree k.

Definition: The set of all k-jets of smooth maps of a fixed manifold M m



fixed manifold N n (at all possible points of Mm) is called the space of k-jets of

maps from

M m to


n •

Notation: The space of k-jets of maps from M m to N n is denoted by jk(Mm, Nn).

Example: The space of I-jets of maps

of the

line to

the line






three-dimensional space with coordinates (x, y, p).

In the general case the space of k-jets jk(Mm, Nn) is a differentiable manifold.

In fact, let (Xl,


, x m) and (Yl,


, Yn) be local coordinates on M m and N n in

the neighbourhoods of points x andf(x), respectively. The mapfis given locally by the formulas

The classes I/


The k-jet is determined by the following numbers:

These numbers give local coordinates in the space of k-jets. We obtain in this

way the smooth manifold structure of Jk(M m


N n). Locally the manifold

Jk(M m


N n) may be represented as the space of Taylor polynomials of degree k.

The dimensions of the manifolds of k-jets are easily computed:

JD(M, N) = M m x N n , dimJ1(M, N) = m + n + mn, ...

The manifold of k-jets possesses a series of naturally arising additional structures.

The segment of the Taylor series of length k determines the segment of the

Taylor series of length k -

1. Therefore there arises a chain of projections





JD(M, N)+-J 1 (M, N)+- ...




The maps of this chain (except the two on the left) - maps "forgetting the terms of degree k in the Taylor series" - are smooth fibrations. The fibers are diffeo- morphic to linear spaces. In the case of the map J1(M, N)-dD(M, N) the fiber over the point (x, y) is the space oflinear operators Hom(TxM, TyN). This fiber has a natural linear space structure. For the remaining k the fiber of the fibration Jk -+ Jk -1 does not have a natural linear structure ("the theorem of the non- invariance of the higher differentials"). Indeed for fixed systems of coordinates the fiber in question is identifiable with the space of sets of n homogeneous polynomials of degree k in m variables and therefore is diffeomorphic to a linear space. But this diffeomorphism is not canonical (it depends on the choice of the system of coordinates) and therefore there is not a natural linear structure on the fiber. Letf:M m -+ N n be a smooth map.

Definition: The k-jet extension of the map f is the map from M m to the space of k-jets from M m to Nn, associating to each point x of M the k-jet offat that point.


Basic concepts

Notation: The k-jet extension offis denoted by /J:Mm-+Jk(Mm, Nn), xl----+j~f

Note that /fis a smooth section of the natural fibration Jk(M, N) -+ M.

The strong transversality theorem (Thorn): Let M m be a closed manifold and C

a closed submanifold of the space of jets Jk(M, N). Then the set of maps J: M -+ N, whose k-jet extensions are transversal to C, is an open everywhere dense set in the space of all smooth maps from M to N.

This theorem means that by a small shift of a smooth map it is possible to put it into general position not only with respect to an arbitrary submanifold in the target space but also with respect to an arbitrary condition placed on the derivatives up to some finite order.

Remark: The weak transversality theorem is obtained from the one just formu- lated, by taking k = O. The strong one does not follow directly from the weak one, however, for the following reason. One can apply the weak theorem to the map /J:M -+J k and obtain a map close to /f and transversal to C. However this nearby map, generally speaking, will not be the k-jet extension of any smooth map from M to N.

The strong transversality theorem asserts that a transversal deformation may be chosen in a narrower class of deformations: it is sufficient to limit the deformations of k-jet extensions to the space of k-jet extensions and not to the space of all sections M -+ J k • Therefore the theorem implies that the integrability conditions (the satisfying of which distinguishes k-jet extensions of maps from M to N from arbitrary sections M -+ Jk) do not hinder the attainment of transversality.

Proof of the theorem: The essence of the proof lies in a reduction to Sard's lemma similar to that in the weak transversality theorem. The main difference lies in the fact that the transversal deformation is sought not in the class of maps Ie = f - e, but in the wider class of polynomial deformations

Ie = f + elel +


+ eses, where the ei are all possible vector-monomials of degree

not greater than k.

The classes I/


Lemma 1: Consider a smooth map F: A x E -+ B of the direct product of smooth manifolds A and E to a smooth manifold B. We shall consider F as afamily of maps Fe of manifold A to B, depending on the point 8 of the manifold E as on a parameter. Then, if the map F is transversal to the submanifold C of the manifold B, almost every member Fe: A -+ B of the family is transversal to C.

Proof: Consider F -l(C). By the implicit function theorem this is a smooth submanifold of A x E. Consider the projection ofthis submanifold on E along A. By Sard's Lemma almost all the values are non-critical. Let 8 be a non-critical

value. Then Fe:A -+ B is transversal to C (because F is transversal to C, while

A x 8 is

transversal to F - 1( C». The lemma is proved.

Lemma 2: Let f be a smooth map from [Rm to [Rn. We fix systems of coordinates in [Rm and [Rn and consider the smooth map of the direct product of [Rm and [RS to the jet-space ]k([Rm, [Rn) defined by the formula

(x, 8)f-+(j~fe),

wherefe = f + 81e1 +

+ 8s e s (e1,

, e s are all possible products of monomials of

... degree not greater than k in the coordinates of the point x of[Rm with basis vectors


of [Rn).

It is asserted that the map so constructed does not have critical values (and, consequently, is transversal to any submanifold of the space of k-jets).

Proof: The coordinates of the point x in [Rm and the coefficients of the Taylor

jet at that point



degree k

inclusive are the coordinates in the space ]k.

For a suitable choice of the coefficients 81,


, 8 s the vector-polynomial

81e1 +


+ 8se s will have any given set of Taylor coefficients up to degree k

inclusive at any given point x. From this the assertion of the lemma follows at once.




proof of








submanifold in

B = ]k([Rm,


We apply Lemma

1 to

Lemma 2 (with




E = [Rs and

F(x, 8) = j~fe). By Lemma 1 the map Fe = F(·, 8) is transversal to C for almost


Basic concepts

every e. By choosing e sufficiently small we obtain a map f: IRm --. IRS arbitrarily close to f(in any bounded region of IRm) whose k-jet extension is transversal to C. The passage from this local construction to a global one (by replacing IRm, IR n by M, N) does not present difficulties.

Remark 1: If C is not closed then the word "open" in the weak and a fortiori in the strong transversality theorem must be replaced by "a countable intersection

of open sets". Examples: (1) B a torus, Can "irrational line" winding round it and

A a circle; (2) B a plane, A

a circle on it and C one of its tangents (without the

point of contact). In either case the embedding of A in B is transversal to C but

there exist maps non-transversal to C arbitrarily close to this embedding.

Remark 2: If the mapped manifold is not compact then it is convenient to furnish the space of maps with the "fine Whitney topology". In this topology a neigh- bourhood of a map f: A --. B is defined in the following way. We fix an open set G in the jet-space Jk(A, B) for some k. The set of C"'-mapsf: A --. B, whose k-jet at each point belongs to G, is open in the fine topology. Such nonempty open sets form the basic neighbourhoods defining the fine topology in the space of infinitely differentiable maps. Thus the nearness of two maps in the fine topology implies that the maps (with an arbitrary number of their derivatives) approach each other "at infinity" as quickly as we wish; in particular, the graph of a map sufficiently close to f lies in a neighbourhood of the graph of/which thins out "to infinity" as quickly as one wishes. From this it follows that the convergence of sequences in the fine topology implies the complete coincidence outside some compact set ·of all the terms of the sequence from some term onwards. Nevertheless an arbitrary neighbourhood of a given map in the fine topology contains maps which nowhere coincide with the given map. If openness and everywhere dense are taken in the sense of the fine topology then the transversality theorem is true even for noncompact A (for openness C must be closed).

Remark 3: One often has to deal with the case where C is not a smooth manifold, but a manifold with singularities.

The classes 1: 1


Definition: A stratified submanifold of a smooth manifold is the finite union of mutually nonintersecting smooth manifolds (the strata), satisfying the following condition: the closure of each stratum consists of the stratum itself together with a finite union of strata of smaller dimensions.

A map is said to be transversal to a stratified submanifold if it is transversal

to each stratum.

Example: Let C be the union of two planes intersecting along a line in three- dimensional space. Then the stratification is the partition into the line of inter- section and four half-planes. Transversality to C means transversality to each of the planes as well as transversality to the line of intersection. For example a curve that is transversal to the stratified submanifold does not intersect its line of singularities.

The transversality theorems extend in an obvious way to the case of a stratified submanifold C. However in this case the theorem guarantees that transversal maps form not an open everywhere dense set but only an everywhere dense intersection of a countable number of open sets. For the maps transversal to a stratified analytic manifold to form an open everywhere dense set it is sufficient for the stratification to satisfy the following further condition: every embedding, transversal to a stratum of smaller dimension is transversal to all adjacent strata oflarger dimension in some neigh- bourhood of this stratum of smaller dimension*.

Example 1: Let C be a finite union of planes in a linear space, stratified in the natural way (for example, a pair of intersecting planes in 1R 3 ). From transversality to IRk there follows transversality to any 1R' containing it, and therefore our condition is satisfied.

Example 2: Let





x 2



+ Z2


1R 3 , and the stratification its

partition into the point 0 and

the two



is easy


verify that our

condition is satisfied.

* This is also true for stratified manifolds obtained from analytic ones by diffeomorphisms, but is not always true for Coo stratified manifolds (see [168]).


Basic concepts

Example 3: Let C be the Whitney umbrella, given by the equation YI = Y3Y~


1R3 (Fig.


From transversality to the singular line Yl =

Y2 =

0 there does not follow

transversality to the manifold of regular points of the surface close to this line (the plane Y3 = 0 is transversal to the line but is not transversal to the surface).

If the condition (transversality to lower dimensional strata


transversality to higher dimensional ones) is satisfied for the stratification C then transversality to the whole stratification is achieved as follows. 1. Strata of the lowest dimension are smooth and the ordinary theorem is applicable to them. 2. In a neighbourhood of strata of least dimension trans- versality is achievable to all strata. 3. Discard from the containing manifold the closure of a neighbourhood of the strata of least dimension and proceed to consider strata of the next smallest dimension.

Example: Let B be the space of linear operators b: IR m -+ IR n and let C be the set of operators of non-maximal rank. The operators of rank r form a smooth submanifold, whose codimension in B is equal to (m - r)(n - r). The partition into manifolds of operators of different ranks is the stratification on C. A map f: A -+ B is a family of linear operators from IR m to IRn, smoothly depending in parametric fashion on the points of the manifold A. The manifold A is called the base of the family. From the weak transversality theorem there follows at once the

Corollary: In the space of smooth families of matrices of order m x n the families transversal to the stratified manifold C of matrices of non-maximal rank form an everywhere dense set.

In particular, the values of the parameter to which the matrices of rank r correspond form in general (for families from the everywhere dense intersection of open sets in the space of families) a smooth submanifold of codimension

(m -

r)(n -

r) of the base of the family.

Proof of the theorem of Section 2.1: Stratify the manifold of I-jets Jl(M, N)

The classes ~1


according to the rank of the first differential. It is clear that the set of all 1-jets with first differential of rank r form a smooth manifold of co dimension (m - r)(n - r), while its closure consists of the manifolds of jets with differential of rank not greater than r. According to the (strong) transversality theorem one

can by a small shift of f achieve the transversality of the

1-jet extension j 1f to

all the strata of the given stratification, which proves the theorem.

Remark: It is not difficult to verify that the condition "transversality to lower dimensional strata implies transversality to higher dimensional ones" is satisfied for the stratification of the space of matrices by rank. In fact it can be proved that it is possible to decompose any algebraic stratification further in such a way that this