Archive for the ‘analysis’ Category

Hilbert’s sixteenth problem

September 14, 2021

 

The International Congress of Mathematicians takes place every four years and is the most important mathematical conference. In 1900 the conference took place in Paris and David Hilbert gave a talk where he presented a list of 23 mathematical problems which he considered to be of special importance. This list has been a programme for the development of mathematics ever since. The individual problems are famous and are known by their numbers in the original list. Here I will write about the 16th problem. In fact the problem comes in two parts and I will say nothing about one part, which belongs to the domain of algebraic geometry. Instead I will concentrate exclusively on the other, which concerns dynamical systems in the plane. Consider two equations \frac{dx}{dt}=P(x,y), \frac{dy}{dt}=Q(x,y), where P and Q are polynomial. Roughly speaking, the problem is concerned with the question of how many periodic solutions this system has. The simple example P(x,y)=y, Q(x,y)=-x shows that there can be infinitely many periodic solutions and that the precise question has to be a little different. A periodic solution is called a limit cycle if there is another solution which converges to the image of the first as t\to\infty. The real issue is how many limit cycles the system can have. The first question is whether for a given system the number N of limit cycles is always finite. A second is whether an inequality of the form N\le H(n) holds, where H(n) depends only on the degree of the polynomials. H(n) is called the Hilbert number. Linear systems have no limit cycles so that H(1)=0. Until recently it was not known whether H(2) was finite. A third question is to obtain an explicit bound for H(n). The first question was answered positively by Écalle (1992) and Ilyashenko (1991), independently.

The subject has a long and troubled history. Already before 1900 Poincaré was interested in this problem and gave a partial solution. In 1923 Dulac claimed to have proved that the answer to the first question was yes. In a paper in 1955 Petrovskii and Landis claimed to have proved that H(2)=3 and that H(n)\le P_3(n) for a particular cubic polynomial P_3. Both claims were false. As shown by Ilyashenko in 1982 there was a gap in Dulac’s proof. After 60 years there was almost no progress on this problem. Écalle worked on it intensively for a long time. For this reason he produced few publications and almost lost his job. At this point I pause to give a personal story. Some years ago I was invited to give a talk in Bayrischzell in the context of the elite programme in mathematical physics of the LMU in Munich. The programme of such an event includes two invited talks, one from outside mathematical physics (which in this case was mine) and one in the area of mathematical physics (which in this case was on a topic from string theory). In the second talk the concept of resurgence, which was invented by Écalle in his work on Hilbert’s sixteenth problem, played a central role. I see this as a further proof of the universality of mathematics.

The basic idea of the argument of Dulac was as follows. If there are infinitely many limit cycles then we can expect that they will accumulate somewhere. A point where they accumulate will lie on a periodic solution, a homoclinic orbit or a heteroclinic cycle. Starting at a nearby point and following the solution until it is near to the original point leads to a Poincaré mapping of a transversal. In the case of a periodic limiting solution this mapping is analytic. If there are infinitely many limit cycles the fixed points of the Poincaré mapping accumulate. It follows that this mapping is equal to the identity, contradicting the limit cycle nature of the solutions concerned. Dulac wanted to use a similar argument in the other two cases. Unfortunately in that case the Poincaré mapping is not analytic. What we need is a class of functions which on the one hand is general enough to include the Poincaré mappings in this situation and on the on the other hand cannot have an accumulating set of fixed points without being equal to the identity. This is very difficult and is where Dulac made his mistake.

What about the second question? It has been known since 1980 that P(2)\ge 4. There is an example with three limit cycles which contain a common steady state and another which does not. It is problematic to find (or to portray) these with a computer since three are very small and one very large. More about the history can be found in an excellent article of Iyashenko (Centennial history of Hilbert’s 16th problem. Bull. Amer. Math. Soc. 39, 301) which was the main source for what I have written. In March 2021 a preprint by Pablo Pedregal appeared (http://arxiv.org/pdf/2103.07193.pdf) where he claimed to have answered the second problem. I feel that some caution is necessary in accepting this claim. A first reason is the illustrious history of mistakes in this field. A second is that Pedregal himself produced a related preprint with Llibre in 2014 which seems to have contained a mistake. The new preprint uses techniques which are far away from those usually applied to dynamical systems. On the one hand this gives some plausibility that it might contain a really new idea. On the other hand it makes it relatively difficult for most people coming from dynamical systems (including myself) to check the arguments. Can anyone out there tell me more?

 

A model for the Calvin cycle with diffusion

June 29, 2021

In the past I have spent a lot of time thinking about mathematical models for the Calvin cycle of photosynthesis. These were almost exclusively systems of ordinary differential equations. One thing which was in the back of my mind was a system involving diffusion written down in a paper by Grimbs et al. and which I will describe in more detail in what follows. Recently Burcu Gürbüz and I have been looking at the dynamics of solutions of this system in detail and we have just produced a preprint on this. The starting point is a system of five ODE with mass action kinetics written down by Grimbs et al. and which I call the MA system (for mass action). That system has only one positive steady state and on the basis of experimental data the authors of that paper were expecting more than one. To get around this discrepancy they suggested a modified system where ATP is introduced as an additional variable and the diffusion of ATP is included in the model. I call this the MAd system (for mass action with diffusion). Diffusion of the other five chemical species is not included in the model. The resulting system can be thought of as a degenerate system of six reaction reaction-diffusion equations or as a system where five ODE are coupled to one (non-degenerate) diffusion equation. The idea was that the MAd system might have more than one steady state, with an inhomogenous steady state in addition to the known homogeneous one. Experiments which measure only spatial averages would not detect the inhomogeneity. To be relevant for experiments the steady states should be stable.

For simplicity we consider the model in one space dimension. The spatial domain is then a finite closed interval and Neumann boundary conditions are imposed for the concentration of ATP. We prove that for suitable values of the parameters in the model there exist infinitely many smooth inhomogeneous steady states. It turns out that all of these are (nonlinearly) unstable. This is not a special feature of this system but in fact, as pointed out in a paper of Marciniak-Czochra et al. (J. Math. Biol. 74, 583), it is a frequent feature of systems where ODE are coupled to a diffusion equation. This can be proved using a method discussed in a previous post which allows nonlinear instability to be concluded from spectral instability. We prove the spectral instability in our example. There may also exist non-smooth inhomogeneous steady states but we did not enter into that theme in our paper. If stable inhomogeneous steady states cannot be used to explain the experimental observations what alternatives are available which still keep the same model? If experiments only measure time averages then an alternative would be limit sets other than steady states. In this context it would be interesting to know whether the system has spatially homogeneous solutions which are periodic or converge to a strange attractor. A preliminary investigation of this question in the paper did not yield definitive results. With the help of computer calculations we were able to show that it can happen that the linearization of the (spatially homogeneous) system about a steady state has non-real eigenvalues, suggesting the presence of at least damped oscillations. We proved global existence for the full system but we were not able to show whether general solutions are bounded in time, not even in L^1. It can be concluded that there are still many issues to be investigated concerning the long-time behaviour of solutions of this system.

From spectral to nonlinear instability, part 2

May 21, 2021

Here I want to extend the discussion of nonlinear instability in the previous post to the infinite-dimensional case. To formulate the problem fix a Banach space X. If the starting point is a system of PDE then finding the right X might be a non-trivial task and the best choice might depend on which application we have in mind. We consider an abstract equation of the form \frac{du}{dt}=Au+F(u). Here t\mapsto u(t) is a mapping from an interval to X. A is a linear operator which may be unbounded. In fact the case I am most interested in here is that where A is the generator of a strongly continuous semigroup of linear operators on X. Hence A may not be globally defined on X. It will be assumed that F is globally defined on X. This is not a restriction in the applications I have in mind although it might be in other interesting cases. In this setting it is necessary to think about how a solution should be defined. In fact we will define a solution of the above equation with initial value v to be a continuous solution of the integral equation u(t)=e^{tA}v+\int_0^t e^{(t-\tau)A}F(u(\tau))d\tau. Here e^{tA} is not to be thought of literally as an exponential but as an element of the semigroup generated by A. When A is bounded then e^{tA} is really an exponential and the solution concept used here is equivalent to usual concept of solution for an ODE in a Banach space. Nonlinear stability can be defined just as in the finite-dimensional case, using the norm topology. In general the fact that a solution remains in a fixed ball as long as it exists may not ensure global existence, in contrast to the case of finite dimension. We are attempting to prove instability, i.e. to prove that solutions leave a certain ball. Hence it is convenient to introduce the convention that a solution which ceases to exist after finite time is deemed to have left the ball. In other words, when proving nonlinear instability we may assume that the solution u being considered exists globally in the future. What we want to show is that there is an \epsilon>0 such that for any \delta>0 there are solutions which start in the ball of radius \delta about x_0 and leave the ball of radius \epsilon about that point.

The discussion which follows is based on a paper called ‘Spectral Condition for Instability’ by Jalal Shatah and Walter Strauss. At first sight their proof looks very different from the proof I presented in the last post and here I want to compare them, with particular attention to what happens in the finite-dimensional case. We want to show that the origin is nonlinearly unstable under the assumption that the spectrum of A intersects the half of the complex plane where the real part is positive. In the finite-dimensional case this means that A has an eigenvalue with positive real part. The spectral mapping theorem relates this to the situation where the spectrum of e^A intersects the exterior of the unit disk. In the finite-dimensional case it means that there is an eigenvalue with modulus \mu greater than one. We now consider the hypotheses of the Theorem of Shatah and Strauss. The first is that the linear operator A occurring in the equation generates a strongly continuous semigroup. The second is that the spectrum of e^A meets the exterior of the unit disk. The third is that in a neighbourhood of the origin the nonlinear term can be estimated by a power of the norm greater than one. The proof of nonlinear instability is based on three lemmas. We take e^\lambda to be a point of the spectrum of e^A whose modulus is equal to the spectral radius. In the finite-dimensional case this would be an eigenvalue and we would consider the corresponding eigenvector v. In general we need a suitable approximate analogue of this. Lemma 1 provides for this by showing that e^\lambda belongs to the approximate point spectrum of e^A. Lemma 2 then shows that there is a v whose norm grows at a rate no larger than e^{{\rm Re}\lambda t} and is such that the norm of the difference between e^Av and e^{{\rm Re}\lambda t}v can be controlled. Lemma 3 shows that the growth rate of the norm of e^A is close to e^{{\rm Re}\lambda t}. In the finite-dimensional case the proofs of Lemma 1 and Lemma 2 are trivial. In Lemma 3 the lower bound is trivial in the finite-dimensional case. To get the upper bound in that case a change of basis can be used to make the off-diagonal entries in the Jordan normal form as small as desired. This argument is similar to that used to treat A_c in the previous post. In the theorem we choose a ball B such that instability corresponds to leaving it. As long as a solution remains in that ball the nonlinearity is under good control. The idea is to show that as long as the norm of the initial condition is sufficiently small the contribution of the nonlinear term on the right hand side of the integral equation will remain small compared to that coming from the linearized equation, which is growing at a known exponential rate. The details are complicated and are of course the essence of the proof. I will not try to explain them here.

From spectral to nonlinear instability

April 28, 2021

An important property of a steady state of a dynamical system is its stability. Let x(t) be the state of the system at time t and let x_0 be a steady state. For a system of ODE these are points in Euclidean space while for more general dynamical systems they are functions which can be thought of as points in suitably chosen function spaces. In general it may be useful to have more than one function space in mind when considering a given dynamical system. First I will concentrate on the ODE case. It is possible to linearize the system about x_0 to get a linear system \dot y=Ay. The steady state x_0 is said to be linearly stable when the origin is stable for the linearized system. Since the linear system is simpler than the nonlinear one we would ideally like to be able to use linear stability as a criterion for nonlinear stability. In general the relation between linear and nonlinear stability is subtle even for ODE. We can go a step further by trying to replace linear stability by spectral stability. There are relations between eigenvalues of A with positive real parts and unstable solutions of the linearized system. Again there are subtleties. Nevertheless there are two simple results about the relation between spectral stability and nonlinear stability which can be proved for ODE. The first is that if there is any eigenvalue of A with positive real part then x_0 is nonlinearly unstable. The second is that if all eigenvalues of A have negative real parts then x_0 is nonlinearly stable, in fact asymptotically stable. These two results are far from covering all situations of interest but at least they do define a comfortable region which is often enough. In what follows I will only consider the first of these two results, the one asserting instability.

Up to this point I have avoided giving precise definitions. So what does nonlinear instability of x_0 mean? It means that there is a neighbourhood U of x_0 such that for each neighbourhood W of x_0 there is a solution satisfying x(0)\in W and x(t)\notin U for some t>0. In other words, there are solutions which start arbitrarily close to x_0 and do not stay in U. How can this be proved? One way of doing so is to use a suitable monotone function V defined on a neighbourhood of x_0. This function should be continuously differentiable and satisfy the conditions that V(x_0)=0, V(x)>0 for x\ne x_0 and \dot V>0 for x\ne x_0. Here \dot V is the rate of change of V along the solution. Let \epsilon be sufficiently small so that the closed ball \overline{B_\epsilon (x_0)} is contained in the domain of definition of V. We will take this ball to be the neighbourood U in the definition of instability. Let M be the maximum of V on \overline{B_\epsilon (x_0)}. Thus in order to show that a solution leaves U it is enough to show that V exceeds M. Consider any solution which starts at a point of V other than x_0 for t=0. The set where V(x)<V(x_0) is open and the solution can never enter it for t>0. The intersection of its complement with U is compact. Thus \dot V has a positive minimum there. As long as the solution does not leave U we have \dot V(x(t))\ge m. Hence V(t)\ge V(0)+mt. This implies that if the solution remains in U for all t>0 then V(x(t)) eventually exceeds M, a contradiction. This result can be generalized as follows. Let Z be an open set such that x_0 is contained in its closure. Suppose that we have a function V which vanishes on the part of the boundary of Z intersecting U and for which \dot V>0 on Z except at x_0. Then x_0 is nonlinearly unstable with a proof similar to that just given.

Now it will be shown that if A has an eigenvalue with positive real part a function V with the desired properties exists. We can choose coordinates so that the steady state is at the origin and that the stable, centre and unstable subspaces at the origin are coordinate subspaces. The solution can be written in the form (x,y,z) where these three variables are the projections on the three subspaces. Then A is a direct sum of matrices A_+, A_{\rm c} and A_-, whose eigenvalues have real parts which are positive, zero and negative respectively. It can be arranged by a choice of basis in the centre subspace that the symmetric part of A_c is as small as desired. It can also be shown that because of the eigenvalue properties of A_+ there exists a positive definite matrix B_+ such that A_+^TB_++B_+A_+=I. For the same reason there exists a positive definite matrix B_- such that A_-^TB_-+B_-A_-=-I. Let V=x^TB_+x-y^Ty-z^TB_-z. Then \dot V=x^Tx+z^Tz-y^T(A_c^T+A)y+o(x^Tx+y^Ty+z^Tz). The set U is defined by the condition V>0. There y^Ty\le Cx^Tx for a positive constant C. On this region \dot V\ge\frac12(x^Tx+z^Tz)+o(|x|^2+|z|^2), where we profit from the special basis of the centre subspace mentioned earlier. The quadratic term in y which does not have a good sign has been absorbed in the quadratic term in x which does. This completes the proof of nonlinear instability. As they stand these arguments do not apply to the infinite-dimensional case since compactness has been used freely. A discussion of the infinite-dimensional case will be postponed to a later post.

Strongly continuous semigroups

March 19, 2021

An evolution equation is an equation for a function u(t,x) depending on a time variable t and some other variables x which can often be thought of as spatial variables. There is also the case where there are no variables x, which is the one relevant for ordinary differential equations. We can reinterpret the function u as being something of the form u(t)(x), a function of x which depends on t. I am thinking of the case where u(t,x) takes its values in a Euclidean space. Then u(t) should be thought of as taking values in a function space. Different regularity assumptions on the solutions naturally lead to different choices of function space. Suppose, for instance, I consider the one-dimensional heat equation u_t=u_{xx}. Then I could choose the function space to be C^0, the space of continuous functions, C^2, the space of twice continuously differentiable functions or L^p. For some choices of function spaces we are forced to consider weak solutions. It is tempting to consider the evolution equation as an ODE in a function space. This can have advantages but also disadvantages. It gives us intuition which can suggest ideas but the analogues of statements about ODE often do not hold for more general evolution equations, in particular due to loss of compactness. (In what follows the function space considered will always be a Banach space.) We can formally write the equation of the example as u_t=Au for a linear operator A. If we choose the function space to be X=L^p then this operator cannot be globally defined, since A does not map from X to itself. This leads to the consideration of unbounded operators from X to X. This is a topic which requires care with the use of language and the ideas which we associate to certain formulations. An unbounded operator from X to X is a linear mapping from a linear subspace D(A), the domain of A, to X. Just as there may be more than one space X which is of interest for a given evolution equation there may be more than one choice of domain which is of interest even after the space has been chosen. To take account of this an unbounded operator is often written as a pair (A, D(A)). In the example we could for instance choose D(A) to be the space of C^\infty functions or the Sobolev space W^{2,p}.

In the finite-dimensional case we know the solution of the equation u_t=Au with initial datum u_0. It is u(t)=e^{tA}u_0. It is tempting to keep this formula even when A is unbounded, but it must then be supplied with a suitable interpretation. There are general ways of defining nonlinear functions of unbounded linear operators using spectral theory but here I want to pursue another direction, which uses a kind of axiomatic approach to the exponential function S(t)=e^{tA}. It should have the property that S(0)=I and it should satisfy the semigroup property S(s+t)=S(s)S(t) for all non-negative s and t. It remains to require some regularity property. One obvious possibility would be to require that S is a continuous function from [0,\infty) into the space of bounded operators L(X) with the topology defined by the operator norm. Unfortunately this is too much. Let us define an operator Ax=\lim_{t\to 0}\frac{S(t)x-x}{t} whenever this limit exists in X and D(A) to be the linear subspace for which it exists. In this way we get an unbounded operator on a definite domain. The problem with the continuity assumption made above is that it implies that D(A)=X. In other words, if the operator A is genuinely unbounded then this definition cannot apply. In particular it cannot apply to our example. It turns out that the right assumption is that \|S(t)x-x\|\to 0 for t\to 0 and any x\in D(X). This leads to what is called a strongly continuous one-parameter semigroup. A is called the infinitesimal generator of S(t). Its domain is dense and it is a closed operator, which means that its graph is a closed subset (in fact linear subspace) of the product X\times X with the topology defined by the product norm. In a case like the example above the problem with continuity is only at t=0. The solution of the heat equation is continuous in t in any reasonable topology on any reasonable Banach space for t>0 but not for t=0. In fact it is even analytic for t>0, something which is typical for linear parabolic equations.

In this discussion we have said how to start with a semigroup S(t) and get its generator (A,D(A)) but what about the converse? What is a criterion which tells us for a given operator (A,D(A)) that it is the generator of a semigroup? A fundamental result of this type is the Hille-Yosida theorem. I do not want to go into detail about this and related results here. I will just mention that it has to do with spectral theory. It is possible to define the spectrum of an unbounded operator as a generalization of the eigenvalues of a matrix. The complement of the spectrum is called the resolvent set and the resolvent is (A-\lambda I)^{-1}, which is a bounded operator. The hypotheses made on the generator of a semigroup concern the position of the resolvent set in the complex plane and estimates on the norm of the resolvent at infinity. In this context the concept of a sectorial operator arises.

My interest in these topics comes from an interest in systems of reaction-diffusion equations of the form u_t+D\Delta u=f(u). Here u is vector-valued, D is a diagonal matrix with non-negative elements and the Laplace operator is to be applied to each component of u. I have not found it easy to extract the results I would like to have from the literature. Part of the reason for this is that I am interested in examples where not all the diagonal elements of D are positive. That situation might be described as a degenerate system of reaction diffusion equations or as a system of reaction-diffusion equations coupled to a system of ODE. In that case a lot of results are not available ‘off the shelf’. Therefore to obtain an understanding it is necessary to penetrate into the underlying theory. One of the best sources I have found is the book ‘Global Solutions of Reaction-Diffusion Systems’ by Franz Rothe.

Degree theory

January 26, 2020

Degree theory is a part of mathematics which I have had little to do with up to now. The one result related to degree theory which has come up repeatedly in connection with my research interests in the past is the Brouwer fixed point theorem. I had an early contact with it through Hirsch’s book ‘Differential Topology’, which I mentioned in a previous post. One formulation of the theorem is that any continuous mapping from the closed unit ball centred at the origin in n-dimensional Euclidean space to itself has a fixed point. The statement is obviously topologically invariant and so we can replace the closed ball by any topological space homeomorphic to it. Since any closed bounded convex subset of Euclidean space is homeomorphic to a closed ball we get an apparently more general formulation concerning convex sets. For n=1 the theorem is easily proved using the intermediate value theorem. I find that it is already not intuitive for n=2. There are various different proofs.

A step towards one type of proof is as follows. Let B be the open unit ball in R^n. If the mapping is called f consider for any point x of the closed unit ball \bar B the straight line joining x and f(x). Extend it in the direction beyond x until it meets the boundary \partial B and call the intersection point \phi(x). Then \phi is continuous, it maps \bar B onto \partial B and its restriction to \partial B is the identity. A map of this type is called a retraction. Thus to prove the Brouwer fixed point theorem it is enough the prove the ‘no retraction theorem’, i.e. that there is no retraction of the unit ball onto the unit sphere. My aim here is not to present a proof of the theorem which is as simple is possible but instead to use one proof of it (the one given in Smoller’s book) to try and throw some more light on degree theory.

The degree is an integer d(f,B,y_0) which is associated to a continuous mapping f:\bar B\to R^n and a point y_0\in R^n\setminus f(\partial B). The proof of the no retraction theorem consists of putting together the following three statements. The first is that (i) since y_0 does not belong to the image of \partial B we have d(f,B,y_0)=d(I,B,y_0), where I is the identity mapping. The second is that (ii) d(I,B,y_0)=1. The third is that (iii) the statement (ii) implies that d(f,B,y_0)=1 and y_0\in f(B), a contradiction. It remains to consider how (i)-(iii) are proved. Central properties of the degree is that it varies continuously under certain types of deformations and that it is an integer. These two things together show that it is left unchanged by these deformations. The proof of (i) is as follows. The mappings f and I agree on \partial D and y_0 does not belong to the image of \partial D under either of them. It is possible to join these two mappings by a homotopy given by tf+(1-t)I. The degree at a given point is left unchanged by a homotopy whose image does not meet that point (iv). This property holds in the present case. Thus (i) follows from (iv). In the proof of (iv) in Smoller’s book it is assumed that the restriction of the homotopy to each fixed value of t is C^1. Thus there is a gap in the argument at this point. In the book it is filled by Theorem 12.7 which says that various properties of the degree proved for C^1 functions also hold for continuous functions. The differentiability is used in the proof of (iv) via the fact that the degree can be expressed as the integral of the pull-back of a suitable differential form under the mapping. Property (ii) follows from the fact that at a regular point y_0 the degree is equal to an expression computed from the determinant of the linearization of the mapping at the inverse images of y_0. In the case of the identity mapping every point is regular and the computation is trivial. When y_0\notin f(B) it follows that y_0 is regular and again the computation is trivial and gives the result zero. This completes the proof of (iii).

Sard’s theorem

December 27, 2019

I have recently been reading Smoller’s book ‘Shock Waves and Reaction-Diffusion Equations’ as background for a course on reaction-diffusion equations I am giving. In this context I came across the subject of Sard’s theorem. This is a result which I had more or less forgotten about although I was very familiar with it while writing my PhD thesis more than thirty years ago. I read about it at that time in Hirsch’s book ‘Differential Topology’, which was an important reference for my thesis. Now I had the idea that this might be something which could be useful for my present research, without having an explicit application in mind. It is a technique which has a different flavour from those I usually apply. The theorem concerns a (sufficiently) smooth mapping between n-dimensional manifolds. It is a local result so that it is a enough to concentrate on the case where the domain of the mapping is a suitable subset of Euclidean space and the range is the same space. We define a regular value of f to be a point y such that the derivative of f is invertible at each point x with f(x)=y. A singular value is a point of the range which is not a regular value. The statement of Sard’s theorem is that the set Z of singular values has measure zero. By covering the domain with a countable family of cubes we can reduce the proof to the case of a cube. Next we write the cube as the union of N^n cubes, by dividing each side of the original cube into N equal parts. We need to estimate the contribution to the measure of Z  from each of the small cubes. Suppose that y_0 is a singular value, so that there is a point x_0 where the derivative of f is not invertible with f(x_0)=y_0. Consider now the contribution to the measure of the image from the cube in which x_0 lies. On that cube f can be approximated by its first order Taylor polynomial at x_0. The image is contained in the product of a subset of a hyperplane whose volume is of the order N^{-(n-1)} and an interval whose length is of the order \epsilon N^{-1} for an \epsilon which we can choose as small as desired. Adding over the at most N^n cubes which contribute gives a bound for the measure of the set of singular values of order \epsilon. Since \epsilon was arbitrary this completes the proof. In words we can describe this argument as follows. The volume of the image of a region which intersects the set of singular points under a suitable linear mapping is small compared to the volume of the region itself and the volume of the image under the nonlinear mapping can be bounded by the corresponding quantity for the linear mapping up to an error which is small compared to the volume defined by the linear mapping.