Asymptotic Methods for Differential Equations

Lecture Notes
—
Mastermath Curriculum
The Netherlands
Spring 2013

adrian muntean university of eindhoven [email protected]

antonios zagaris university of twente [email protected]

1
 part i — classic asymptotics
a. zagaris

a brief foreword

This set of notes is by no means meant as a stand-alone text from which one can hope to ‘learn asymptotics’.
It is, instead, a summary (and critical discussion) of the material presented in class. For a proper read, please
consult our excellent coursebook (M. Holmes’s ‘Introduction to Perturbation Methods,’ Springer, 1995). Note,
though, that an integral part of the course to which these notes correspond is gleaning information from the
relevant literature. The transition from coursebook to books and research papers is a subtle (that’s a euphemism
for hard ) yet necessary step in your development as a researcher. I make an effort below to provide an adequate
(yet by no means comprehensive!) list of references, each of which treats appropriately certain parts of the
material presented here. Please do your part and do consult this literature.

My aim in teaching this class is neither to be rigorous nor to present asymptotics in a mathematically consistent
way—books that attempt to do exactly this do exist, and some of them are included in the bibliography. In fact,
there are nearly as many flavors to asymptotics as books on asymptotics; the particular flavor I am interested
in for this class is the intuitive one, and this reflects in the notes below. This is not to say that rigor will be
absent from the class, but, rather, that our point of view will be functional (‘does this work? ’). For this and
other reasons, expect to hear the course’s motto—‘know your problem! ’—more than once.

At any rate, I hope to help you discover the internal logic of (and let you develop more insight in) the subject
at hand. A very good first exercise to that effect is to try and fill in as many gaps in the presentation
below as possible—look out, in particular, for bracketed statements. Not all of these statements are easy to
answer/prove/substantiate, but you will benefit from working on each one of them.

Lecture no.1 February 05, 2013

1.1 A regularly perturbed algebraic equation

[See [Holmes, Sect. 1.5] for a similar example.] We start by considering a problem which is fully tractable
analytically, namely that of finding the solutions to the quadratic equation
f (λ; ε) = λ2 − 2λ + ε = 0. (1.1)
Here, ε is a parameter taking values in an interval [0, ε̄], for some ε̄ > 0. The two solutions of this equation are
exactly given by the formula √
λ± = 1 ± 1 − ε,
see also Fig. 1.
Note, now, the following. If ε̄ < 1, then we may Taylor-expand the square root in the formula above to
obtain the convergent series
∞ Qn−1
!
2 − 12 ε − 81 ε2 − . . .
  
k=1 (2k − 1) n 1 1 2
X
±
λ =1± 1− ε = 1 ± 1 − ε − ε − . . . = 1 1 2 (1.2)
2ε + 8ε + . . . .
n! 2n 2 8
n=1

[Why ε̄ < 1? What goes wrong for ε̄ ≥ 1?] Our fundamental observation is that we can obtain the coefficients
of the series (1.2) by

2
(a) substituting the series
∞
X
λ= εn λn = λ0 + ελ1 + ε2 λ0 + ... (1.3)
n=0

directly into (1.1);

(b) collecting powers of ε in the resulting equation;
(c) setting the coefficient of each such power to zero.
Indeed, here is how the three-point program above can be carried out.

(a) First, substitution of (1.3) into (1.1) yields

2
λ0 + ελ1 + ε2 λ0 + ... − 2 λ0 + ελ1 + ε2 λ0 + ... + ε
  
0 =
= λ20 + 2ελ0 λ1 + ε2 λ21 + 2λ0 λ2 + ... − 2 λ0 + ελ1 + ε2 λ0 + ... + ε.

  

(b) Rearranging terms, we can rewrite this equation as

 2
λ0 − 2λ0 + ε [2λ0 λ1 − 2λ1 + 1] + ε2 λ21 + 2λ0 λ2 − λ2 + ... = 0.
  
(1.4)

(c) Setting the coefficient of each power of ε to zero, one obtains a countably infinite set of algebraic equations;
the first three are

−
• λ20 − 2λ0 = 0, whence λ+
0 = 2 or λ0 = 0;
−
• 2λ0 λ1 − 2λ1 + 1 = 0, whence λ+
1 = −1/2 or λ1 = 1/2, respectively;
−
• λ21 + 2λ0 λ2 − λ2 = 0 whence λ+
2 = −1/8 or λ2 = 1/8, respectively.

Discussion. Plainly, this outcome matches (1.2) up to and including ε2 −terms. The reason underpinning
this agreement is not hard to fathom, once one knows that (1.3) is absolutely convergent. Indeed, squaring a
series and subsequently rearranging its terms (as we did) are both valid under this condition; the same goes
for summing up series. Similarly, (1.4) is true if and only if the coefficient of each power of ε is identically zero
(essentially by the uniqueness property of Taylor expansions). Thus, our work above is rigorously justified,
provided that we know that (1.3) is an absolutely convergent series. To prove absolute convergence, knowledge
of (1.1) is unnecessary (although it does not hurt at all, either...): the same information can be deduced by the
series’ coefficients, assuming that an appropriate formula for the nth (general ) coefficient has been successfully
determined. [How can one prove absolute convergence of a series of which one has the general term?] We will
return to this fine point later on, when we discuss asymptotic expansions.
There are certain fine points related to our discussion above. First, one cannot but observe that, for ε > 1,
(1.1) yields two complex solutions with non-trivial imaginary parts. This seems incompatible with the fact that
(1.3) is a real series, until one notices that this series becomes divergent for such values of ε. That the series is
divergent, though, can only be deduced if one can produce a formula for the general coefficient of the series.
In this particular example, this is possible; this is not always the case, though. Even worse, in cases where it is
possible, one might P find that the series diverges for all nonzero values of ε—a classic example of such a series is
the Stieltjes series n≥0 (−1)n n! xn (see [Bender & Orszag, Sect. 3.8, Example 3]). We will see further down
what such divergent series have to offer.

1.2 A singularly perturbed algebraic equation

[See [Holmes, Sect. 1.5] for a similar example.] Consider, now, the problem

f˜(λ; ε) = −ελ4 + λ2 − 2λ + ε = 0, (1.5)

with ε as before: a “small” parameter. Although this equation can also be solved explicitly, we shall not bother
with the analytic formula. We will, instead, determine series expansions for the roots to this quartic.

3
 9

5 f(λ;ε)
4

f(λ;0)
2

1 λ−(ε) λ+(ε)

−1
ε
−2 −1 0 1 2 3
λ 4

Figure 1: Graphs of the functions f (·; 0) (in black) and f (·; ε) (in red). Note that the latter is a mere vertical
shift by ε of the former.

Equation (1.5) seems to be a perturbation of (1.1), since the additional term, ελ4 , is multiplied by the small
parameter ε. Hence, a (hasty...) first guess would be that the solutions of (1.5) are close to the solutions (1.2)
of (1.1). (The more careful reader, on the other hand, will not fail to notice that (1.5) must have not two but
four roots.) Substituting (1.3) into (1.5), rearranging, and collecting terms, we find
 2
λ0 − 2λ0 + ε −λ40 + 2λ0 λ1 − 2λ1 + 1 + . . . = 0,
  

whence we obtain
15 1
λ+
0 =2+ ε + . . . and λ− 0 = ε + .... (1.6)
2 2
Indeed, these two roots seem to be ε−close to those given in (1.2)—no wonder, as both (1.1) and (1.5) read
λ20 − 2λ0 = 0, for ε = 0. In fact, it can be easily proven that there are two roots of (1.5)—say, λ+ (ε) and
λ− (ε)—which satisfy
lim λ± (ε) − λ±
 
0 = 0.

ε↓0

Even more strongly, it can be shown that there exist ε0 > 0 and constants C ± > 0 such that
λ (ε) − λ±  < C ± ε, for all 0 ≤ ε ≤ ε0 .
 ± 
0

In other words, the distance between λ± (ε) and λ± 0 scales at least as fast as ε. (In fact, it scales exactly as fast
as ε—that is, C ± cannot be made arbitrarily small by lowering the value of ε0 , and hence λ± converge to their
unperturbed values linearly, as ε ↓ 0.) [Try to prove all of this.]
What about the remaining two roots, which neither were captured by the expansion (1.3) nor are present
for ε = 0? To understand our problem at an intuitive level, we return to the idea that f˜(·; ε) is a perturbation
of f (·; ε). This is certainly true for every fixed value of λ, in the sense that

lim f˜(λ; ε) = λ2 − 2λ = lim f (λ; ε), for every λ.

ε↓0 ε↓0

Employing the definition of the limit, we obtain that, for every c > 0, there exists ε0 > 0 such that
 
˜
f (λ; ε) − f (λ; ε) < c, for all 0 ≤ ε ≤ ε0 .


(Using logical quantifiers, we may rewrite this in the form

h i
∀c>0 ∀λ∈R ∃ε̄>0 ∀0≤ε≤ε0 f˜(λ; ε) − f (λ; ε) < c; (1.7)
 

4
 f(λ;ε)

f(λ;ε)

Figure 2: Graphs of the functions f (·; ε) (in black) and f˜(·; ε) (in red).

we will favor the use of quantifiers in the next section, too, so it is suggested you familiarize yourselves with
them.) As is always the case when a statement involves the universal (∀) and the existential (∃) quantifiers,
one has to be careful with interchanging them; in particular, ∃∀ ⇒ ∀∃ but not vice versa. Here, in particular,
it is not true that h i
∀c>0 ∃ε0 >0 ∀λ∈R ∀0≤ε≤ε0 f˜(λ; ε) − f (λ; ε) < c,
 

which would imply that f˜(·; ε) and f (·; ε) are uniformly close over R!1 And indeed, the two functions do not
remain uniformly close over the entire real line, as the calculation

lim f˜(λ; ε) = −∞ =
6 ∞ = lim f (λ; ε), for all ε > 0, (1.8)
|λ|→∞ |λ|→∞

plainly shows. In other words, no matter how small (but positive) ε is, there is always a neighborhood of
infinity where the functions f˜(·; ε) and f (·; ε) diverge: the former approaches −∞ (due to the term −ελ4 ,
which eventually takes over), while the latter approaches ∞ (due to the term λ2 ). Under these circumstances,
there can be no discussion of closeness, see also Fig. 2.
This last remark also offers us the way out of the conundrum. For any fixed value of λ (and for sufficiently
small ε), f˜(λ; ε) and f (λ; ε) are close to each other; for all λ 6∈ [0, 3] and all ε ∈ [0, ε̄] (for some ε̄ > 0), then,
f˜(λ; ε) > 0. Since f˜(·; ε) tends to −∞ in a continuous manner as |λ| increases, it needs to cross zero on its
way there; this yields a root smaller than λ = 0 and a root larger than λ = 3. As argued above, the uniform
closeness of f˜(λ; ε) and f (λ; ε) over any bounded set I implies that these roots cannot be contained in any such
set as ε ↓ 0: hence, they must become arbitrarily large (i.e., unbounded) as ε ↓ 0. To capture this effect, we
rescale λ via λ = ε−a Λ, for some yet undetermined a > 0; the exponent a will be determined by demanding
that the values of Λ corresponding to these unbounded roots remain bounded away from both infinity and zero
as ε ↓ 0. In other words, we assume that these unbounded roots tend to infinity algebraically (as ε−a , for some
positive power a). Under this rescaling, (1.5) becomes

−ε1−4a Λ4 + ε−2a Λ2 − 2ε−a Λ + ε = 0. (1.9)

Now, the last term (ε) limits to zero as ε ↓ 0, whereas the next-to-last term (2ε−a Λ) blows up (recall that
Λ was assumed bounded away from zero). Hence, the former is a mere perturbation to the latter. Similarly,
the second term (ε−2a Λ2 ) blows up much faster than 2ε−a Λ, as ε−2a tends to infinity much faster than ε−a as
1 The reader should try to show that, nevertheless, f˜(·; ε) and f (·; ε) are uniformly close over any bounded subset I of R—that

is, replacing R by I in (1.7) allows us to exchange the two quantifiers. Note that ε0 has to be taken smaller as the diameter of I
grows, with ε0 limiting to zero as the diameter approaches infinity. This last property often acts as the ‘signature’ of a problem
where uniformity cannot be established.

5
ε ↓ 0; hence, the next-to-last term is also perturbative to the second term. To obtain bounded, non-zero values
of Λ, then, one must balance the first and second terms. Since these two terms must balance each other for all
ε ∈ [0, ε̄], it follows that 1 − 4a = −2a or, equivalently, that a = 1/2. Under this condition, we may recast (1.9)
in the form √
−Λ4 + Λ2 − 2 εΛ + ε2 = 0, (1.10)
which is a regularly perturbed problem like the one treated in the previous section. Postulating for Λ a power
series of the form √
Λ = Λ0 + εΛ1 + εΛ2 + . . .
√
(note that the “natural” small parameter in (1.10) is ε and not ε) and working as in that section, we can
produce the expansion √
Λ = ±1 − ε + . . . .
Passing back to the original variable λ, then, we find
1
λ = ±√ − 1 + . . . .
ε
For completeness, note that (1.10) also yields two roots which are zero to leading order; the reader can show
that these correspond to the two bounded roots reported in (1.6).

Remark. The leading order result λ = ±ε−1/2 − 1 + . . ., which suggests the proper rescaling for λ, can be also
obtained by the following considerations. First, there exist no extra roots which remain bounded as ε ↓ 0 for the
reasons outlined above. Second, for “large” values of λ, the dominant term among the second-through-fourth
ones in (1.5) is the second (quadratic) one. For a root to exist, λ needs to have such a value that this term
(which is large and positive) is balanced by the first (quartic) term: for smaller values of lambda, the quadratic
term is dominant and f˜(λ; ε) > 0; for larger values of lambda, the dominant term is the quartic and f˜(λ; ε) < 0.
Thus, −ελ4 + λ2 ≈ 0, whence λ ≈ ±ε−1/2 .

1.3 Order symbols

In this and the next section, we will (briefly) formalize our discussion of approximate solutions and asymptotic
expansions. The reader is encouraged to consult any book in the bibliography for additional information on
the subject. In what follows, D ⊂ RM (for some M ∈ N), I = (0, ε̄) (for some ε̄ > 0), u : D × I → R, and
v : D × I → R. First, we introduce the order symbol o(·) (called ‘little Oh’).

Definition. We write
h i
u(·; ε) = o(v(·; ε)) (ε ↓ 0) over D if ∀x∈D ∀k>0 ∃ε0 >0 ∀0<ε<ε0 |u(x; ε)| < k |v(x; ε)| .  (1.11)

In words, u = o(v) (‘(ε ↓ 0)’ is often omitted) if, for every x ∈ D, u(x; ·) can be bounded by any—arbitrarily
small—multiple k of v(x; ·) over some (k−dependent, naturally) neighborhood of zero. The relation u = o(v)
is often also written u  v, for historical reasons. If the set {ε : v(x; ε) 6= 0} has, for each x, zero as a limit
point, then (1.11) is simply identical to
 
u(x; ε)
∀x∈D lim =0 . (1.12)
ε↓0 v(x; ε)

(In fact, some authors use this relation to define o(·).) Next, we introduce the order symbol O(·) (called ‘big
Oh’).

Definition. We write
h i
u(·; ε) = O(v(·; ε)) (ε ↓ 0) over D if ∀x∈D ∃k>0 ∃ε0 >0 ∀0<ε<ε0 |u(x; ε)| < k |v(x; ε)| .  (1.13)

In words, u = O(v) if, for every x ∈ D, u(x; ·) can be bounded by a—large enough—multiple k of v(x; ·)
over some neighborhood of zero. Note that (1.13) does not imply that u and v exhibit the same asymptotic

6
behavior as ε ↓ 0—for example, u(x; ε) could be approaching zero (e.g., u(x; ε) = ε) while v(x; ε) could be
growing unboundedly (e.g., v(x; ε) = ε−1 ) as ε ↓ 0. Even if both are approaching zero or growing unboundedly,
they could do so at different asymptotic rates—compare u(x; ε) = ε2 and v(x; ε) = ε, in the first case, and
u(x; ε) = ε−1 and v(x; ε) = ε−2 in the second one. To alleviate this, we introduce the stronger symbol Os via
h   i
u(·; ε) = Os (v(·; ε)) (ε ↓ 0) over D if ∀x∈D u(x; ε) = O(v(x; ε)) (ε ↓ 0) ∧ u(x; ε) 6= o(v(x; ε)) (ε ↓ 0) .
(1.14)
(The reader should nevertheless note that, unfortunately, O is often used in the literature where Os should
have been used...)

Example. The following relations can be easily shown to be true:

• For all a < b, εb = o(εa ) (ε ↓ 0).

• For all a ∈ R, (log ε)−1 = o(εa ) (ε ↓ 0); additionally, log ε = o(εa ) (ε → ∞) (the logarithm tends to
infinity slower than any algebraic function).
• For all a ∈ R and c > 0, e−c/ε = o(εa ) (ε ↓ 0) (exponentially small terms are asymptotically smaller than
any algebraic term).
• ε sin ε−1 = Os (ε) (ε ↓ 0), while the converse is not true.

As is always the case when dealing with convergence of functions, definitions (1.11) and (1.13) raise the
issue of uniformity. Both of these definitions quantify the notion or relative asymptotic magnitude pointwise:
for every value of x ∈ D, u(x; ε) can be bounded by any (for o(·))/a (for O(·)) multiple of v(x; ε) over an
interval (0, ε0 ). The interval length ε0 in these definitions is x−dependent, ε0 = ε0 (x); in general, no ε0 exists
which applies to all x—equivalently, inf x∈D ε0 (x) = 0. (In definition (1.13), similar concerns apply to the
x−dependent value of k: in general, supx∈D k(x) = ∞.) If such a uniformly valid value of ε0 (and of k, for
O(·)) can be found, then we talk of uniform asymptotic estimates:
h i
u(·; ε) = o(v(·; ε)) (ε ↓ 0) uniformly over D if ∀k>0 ∃ε0 >0 ∀0<ε<ε0 ∀x∈D |u(x; ε)| < k |v(x; ε)| (1.15)

and
h i
u(·; ε) = O(v(·; ε)) (ε ↓ 0) uniformly over D if ∃k>0 ∃ε0 >0 ∀0<ε<ε0 ∀x∈D |u(x; ε)| < k |v(x; ε)| . (1.16)

We will see below that uniformity often characterizes regularly perturbed problems and non-uniformity singu-
larly perturbed ones.

Example. Consider the functions

u1 (x; ε) = εx + e−x/ε , u2 (x; ε) = x + e−x/ε , and v(x; ε) = x, with D = (0, 1) and ε̄ = 1.

• First, u1 (·; ε) = o(v(·; ε)) over D, as

   
|u1 (x; ε)| = εx + e−x/ε ≤ ε + x−1 e−x/ε x = ε + x−1 e−x/ε |v(x; ε)| .

Hence, we can take k(x) = ε + x−1 e−x/ε which, for any x ∈ (0, 1), can be made arbitrarily small
provided that ε > 0 is small enough. More rigorously, given any x ∈ (0, 1) and k > 0, we can use
ε0 = min(k/2, x/ |ln(kx/2)|) to satisfy (1.11). This is a sharp estimate and ε0 ↓ 0 as x ↓ 0 (because
x/ |ln(kx/2)| ↓ 0). It follows that u1 (·; ε) = o(v(·; ε)) non-uniformly over (0, 1). Intuitively, this is to be
expected, as v(x; ε) → 0 and u(x; ε) → 1 for all ε > 0, however small; hence, bounding the latter by the
former uniformly over (0, 1) is not possible.
• Similarly, u2 (x; ε) = O(v(x; ε)) since
   
|u2 (x; ε)| = x + e−x/ε ≤ 1 + x−1 e−x/ε x = 1 + x−1 e−x/ε |v(x; ε)| ,

so that (1.13) is satisfied with k(x) = 1 + x−1 e−x/ε and ε0 = ε̄. This is a sharp estimate and limx↓0 k(x) =
∞, so that the order relation above only holds non-uniformly over D.
• Note, also, that u1 (x; ε) = O(u2 (x; ε)) uniformly over D, since |u1 (x; ε)| = u1 (x; ε) ≤ u2 (x; ε) = |u2 (x; ε)|.

7
Remark. It is worth mentioning here that the pointwise asymptotic relation u(x; ε) = O(v(x; ε)) over a
compact set D does not imply that u(x; ε) = O(v(x; ε)) uniformly over D, as one would expect based on the
argument ‘ε0 (x) and k(x) should depend continuously on x, so the former is bounded below and the latter above
by positive constants which can be used when employing (1.16).’ See [Holmes, Thm. 1.3] for details [and provide
a counterexample].

1.4 Asymptotic expansions

Having defined the order symbols O(·) and o(·), we now turn to formalizing the notion of successive approxi-
mations. We start with making precise what we mean with (asymptotic) approximation. Here, D and I remain
defined as in the previous section, while f : D × I → R and φ : D × I → R are functions.

Definition. We write
h i
f (·; ε) ∼ φ(·; ε) (ε ↓ 0) over D if ∀x∈D f (x; ε) − φ(x; ε) = o(φ(x; ε)) (ε ↓ 0) .  (1.17)

In words, f ∼ φ (‘φ is an asymptotic approximation to f ’) if, for every x ∈ D, the difference f (x; ·) − φ(x; ·)
of the two is asymptotically smaller than φ(x; ·) over some neighborhood of zero. If, in particular, the ratio
(f − φ)/φ is well-defined over D × I 0 , with I 0 ⊂ I a set having zero as its limit point, then we can employ (1.12)
to rewrite (1.17) as
 
f (x; ε) − φ(x; ε)
f (·; ε) ∼ φ(·; ε) (ε ↓ 0) over D if ∀x∈D lim =0 . (1.18)
ε↓0 φ(x; ε)

Note that neither (1.17) nor (1.18) imply that the difference f − φ is o(1) (“small”); they do imply, instead,
that this difference becomes (arbitrarily) small compared to φ (which asymptotically approximates f ) as ε ↓ 0.
[Can you formulate conditions under which ∼ becomes an equivalence relation?]
We now proceed with the notion of an asymptotic sequence.

Definition. We call the sequence of functions {φn : I → R}n≥0 an asymptotic sequence if

h i
∀m≥0 ∀n≥m φn (ε) = o(φm (ε)) (ε ↓ 0) . 

Equivalently, {φn }n≥0 is an asymptotic sequence if

h i
∀m≥0 φm+1 (ε) = o(φm (ε)) (ε ↓ 0) . (1.19)

An interesting fact related to asymptotic sequences is described by a theorem due to du Bois-Reymond [Verhulst,
Sect. 15.1] which states that, for every asymptotic sequence {φn }n≥0 , there exists a function φ : I → R satisfying
∀n≥0 [φ(ε) = o(φn (ε)) (ε ↓ 0)]. Note that a similar result is not true for the real numbers, as there is no positive
number which is smaller than all members of a positive sequence converging to zero; similarly, there is no
positive function which is smaller than all members of a sequence of positive functions converging pointwise to
zero.

Remark. Verhulst [Verhulst, Sect. 2.5] gives another definition of an asymptotic sequence involving a second
sequence, {δn : I → R}n≥0 , of functions which are positive, continuous, monotone and for each of which either
limε↓0 δn (ε) or 1/ limε↓0 δn (ε) exists. In these notes, we will mostly only use asymptotic sequences satisfying
these conditions, so his alternative definition coincides with (1.19).
Next, we define the notion of asymptotic equality between two functions and with respect to a given asymp-
totic sequence.

8
Definition. Let f and φ be as above, {φn }n≥0 be an asymptotic sequence, and N ∈ N∗ = N ∪ {0}. We write

f (·; ε) = φ(·; ε) (ε ↓ 0) ({φn }n≥0 ) (N ) if ∀x∈D [f (x; ε) − φ(x; ε) = o(φN (ε)) (ε ↓ 0)] .  (1.20)
In words, f ang φ are asymptotically equal with respect to {φn }n≥0 up to and including order N if their
difference is asymptotically smaller than φN ([Wong, Sect. I.3]).2
Using this definition, we can now easily introduce the notion of an asymptotic expansion for a given function
f and with respect to a given asymptotic sequence {φP n }n≥0 . In particular, let {an : D → R}n≥0 be a sequence of
functions over D. According to (1.20), then, f and 0≤n≤N an (x)φnP (ε) (for some N ∈ N∗ ) are asymptotically
equal (with respect to {φnP }n≥0 ) up to and including order N if f − 0≤n≤N an (x)φn (ε) = o(φN ). This leads
naturally to the idea that n≥0 an (x)φn (ε) is the asymptotic expansion of f if the two are asymptotically equal
to all orders.

Definition. Let {φn }n≥0 be an asymptotic sequence, f : D × I → R, and {an : D → R}n≥0 . We write
" N
#
X X
f (·; ε) = an (·)φn (ε) (ε ↓ 0) over D if ∀x∈D ∀n∈N∗ f (x; ε) − an (x)φn (ε) = o(φN ) (ε ↓ 0) . 
n≥0 n=0

P (1.21)
That is,
P a φ
n≥0 n n is the asymptotic expansion of f if the difference between f and every truncated sum
a φ
0≤n≤N n n is asymptotically smaller than the asymptotically smallest term in that truncated sum.
Note that we talk here of the asymptotic expansion of f . Indeed, given an asymptotic sequence, there is at
most one asymptotic expansion for f with respect to that sequence. The coefficients {an }n≥0 are calculated
iteratively using the formula
P
f (x; ε) − 0≤n≤N an (x)φn (ε)
aN +1 (x) = lim , for N ≥ 0, (1.22)
ε↓0 φN +1 (ε)

that is,

f (x; ε) f (x; ε) − a0 (x)φ0 (ε) f (x; ε) − (a0 (x)φ0 (ε) + a1 (x)φ1 (ε))
a0 (x) = lim , a1 (x) = lim , a2 (x) = lim ,
ε↓0 φ0 (ε) ε↓0 φ1 (ε) ε↓0 φ2 (ε)

P limits exist and are finite, then f has an asymptotic expansion in terms of {φn }n≥0 ,
and so on. If all of these
and this expansion is n≥0 an φn .

Remark. Some authors (e.g., [Wong]) use, in (1.21), the notation ‘∼’ instead of ‘=’. This use of ∼ will be
avoided in these notesP as, first, it is less common in the literature; and second, it clashes with (1.17). Indeed,
the relation ‘f (·; ε) ∼ n≥0 an (·)φn (ε) (ε ↓ 0) over D’ introduced in (1.21) can be rewritten as

−1
" N
#
X
∀N ∈N∗ f (·; ε) − an (·)φn (ε) ∼ aN (·)φN (ε) (ε ↓ 0) over D (1.23)
n=0

PN
[prove this]. In particular, (1.21) does not imply the relation f (·; ε) ∼ n=0 an (·)φn (ε), except for N = 0.
PN
To see this, note that—as long as a0 6= 0 (which is generically true)—the relation f (·; ε) ∼ n=0 an (·)φn (ε)
translates into !
N
X N
X
f (·; ε) − an (·)φn (ε) = o an (·)φn (ε) = o(φ0 (ε)),
n=0 n=0

where in the last step we used that φn (ε) = o(φ0 (ε)), for all 1 ≤ n ≤ N . This relation matches (1.23) if and
only if N = 0. In that sense, the notion of an asymptotic sequence generalizes the notion of an asymptotic
approximation.
2 In the same section, Wong treats shortly (but sufficiently critically) the theory of generalized asymptotic expansions; the

interested reader is referred there for details.

9
Remark. The reader is strongly encouraged to look up the excellently concise discussion on the differences
between
P convergent and asymptotic sequences in [Bender & Orszag, Sect. 3.8]. In short, a convergent series
n≥0 cnPconverges without reference to a specific function f ; in fact, a function f can be
P defined nby this series,
f (x) = n≥0 cn (x). A characteristic example would be a convergent power series, n≥0 an x ; such series
are often used to express solutions to regular ODEs, implicitly-defined analytic functions, et cetera. To the
contrary, a series cannot be ‘asymptotic’ without reference to such a specific function. Take, for example, a
non-convergent power series n≥0 an xn as above.
P
As Bender and Orszag show, no matter how {an }n≥0 is
chosen, a function f can be constructed with n≥0 an xn as its asymptotic series. Since, then, all power series
P
are ‘asymptotic’ if no reference to a specific function is made, employing the term ‘asymptotic’ without such
a reference is nonsensical. An interesting corollary to this is that asymptotics cannot guarantee the existence
of a solution to a given problem; this usually has to be ascertained in another manner, before its asymptotic
expansion is determined.

1.5 A singularly perturbed transcendental equation

[See [Kevorkian & Cole, Sect. 1.2] for a very similar example.] Consider now the problem of solving the
transcendental equation
ε
g(x, y; ε) = y − x − ε sin(y) + = 0, with x ∈ [0, 2], (1.24)
y−1
in order to express y in terms of x. Note that, compared to the problems we treated in Sect. 1.1–1.2, our problem
contains an additional variable x (acting as a parameter in this problem); further, we have no information on
the number of branches y = y(x) (1.24) admits as solutions. Note, though, that this problem enjoys certain
properties which make it analytically tractable, see the remark below for details.
For ε = 0, the equation becomes y − x = 0 and admits the unique solution (branch) y = y0 (x) = x. The
Implicit Function Theorem guarantees, then [work out the details], that (1.24) has a solution y = y(x) for all
ε in a neighborhood of zero and for any x ∈ (0, 1) ∪ (1, 2). To derive an asymptotic expansion for the solution
y(x) to the perturbed problem, we write

y(x) = y0 (x) + φ1 (ε) y1 (x) + o(φ1 (ε)), (1.25)

where the order function φ1 —the next member of the asymptotic sequence {φn }n≥0 starting with φ0 ≡ 1—is
yet to be determined; at this point, our only requirement is that φ1  φ0 = 1. Substituting (1.25) into (1.24),
we obtain
ε
[x + φ1 (ε) y1 (x) + o(φ1 (ε))] − x − ε sin(x + φ1 (ε) y1 (x) + o(φ1 (ε))) + = 0.
(x − 1) + φ1 (ε) y1 (x) + o(φ1 (ε))

Taylor-expanding the sinusoidal term and the fraction, we find further

 
1
0 = [φ1 (ε) y1 (x) + o(φ1 (ε))] − ε [sin(x) + O(φ1 (ε))] + ε + O(φ1 (ε))
x−1
 
1
= φ1 (ε) y1 (x) + ε − sin(x) + o(φ1 (ε)),
x−1

where we have also absorbed all ε O(φ1 (ε)) = O(ε φ1 (ε)) terms into o(φ1 (ε)). The only possible matching is
φ(ε) = O(ε), and for simplicity we take φ(ε) = ε; then, y1 (x) = sin(x) − 1/(x − 1), and hence
 
1
y(x) = x + ε sin(x) − + o(φ1 (ε)). (1.26)
x−1

This process can be repeated as many times as needed to determine higher order terms in the asymptotic
expansion (and the asymptotic sequence {φn }n≥0 itself).
It should already be apparent that the two-term expansion (1.26) is neither defined over [0, 2]—x∗ = 1
proves problematic—nor has any chance of being uniformly valid in any set with x∗ = 1 as a limit point—e.g.,
in (0, 1). Indeed, the term 1/(x − 1) blows up near x∗ = 1 and the asymptotic expansion stops being well-
ordered, as ε/(x − 1) can easily become comparable to x ≈ 1 or even much larger. In retrospect, the function

10
of two variables g(x, y; ε) can also not be said (for any ε > 0) to be uniformly close to y − x due to this same
term. Recalling Sect. 1.2—where such a non-uniformity gave rise to two additional roots—we decide to zoom in
near the root of all evil (x∗ = 1) and examine the situation closer. To achieve this, we rescale the independent
variable x via
x = x∗ + χ(ε) X = 1 + χ(ε) X, for some o(1) order function χ(ε).
(Under this rescaling, x = 1 is mapped to X = 0, and any O(1) change in X yields O(χ(ε)) = o(1) changes in
x—whence the zooming in.) Since the Implicit Function Theorem fails at (x∗ , y∗ ) = (1, 1), we also rescale y
around y∗ = 1,
y = y∗ + ψ(ε) Y = 1 + ψ(ε) Y, for some o(1) order function ψ(ε).
We emphasize here that we will focus on O(1) values of both X and Y that remain bounded away from zero.3
Plugging into (1.24), we find

ψ(ε) Y − χ(ε) X − ε sin(1 + ψ(ε) Y ) + (ψ(ε))−1 εY −1 = 0

We will first investigate the leading order behavior, Y (x) = Y0 (x) + o(1). To this effect, we may omit the
sinusoidal term as it is asymptotically smaller than the last term by virtue of ε  ε(ψ(ε))−1 (indeed, sin(y) in
the original equation remains bounded near y∗ = 1, while 1/(y − 1) grown unboundedly); the resulting equation
is a quadratic in disguise,
ψ(ε) Y0 − χ(ε) X + ε(ψ(ε))−1 Y0−1 = 0. (1.27)

We consider the following distinct cases:

• χ(ε)  ψ(ε):
√ then, the only possible matching in (1.27) is between the first and third terms, whence
ψ(ε) = O(√ ε). Since both are of the same sign, though, we find that there exist no real solutions. Hence,
when o( ε)−close to x∗ , there are no solution branches at a distance asymptotically larger than χ(ε)
from y∗ .
• χ(ε)  ψ(ε): then, the only possible matching is between the second and third terms, whence χ(ε)ψ(ε) =
O(ε) and Y0 (x) = 1/X—equivalently,
ε
y(x) = 1 + + o(ε). (1.28)
x−1
√
This solution branch appears, then, at a distance asymptotically larger than ε from x∗ —recall that
χ(ε)ψ(ε) = O(ε) and χ(ε)  ψ(ε)—and lies asymptotically closer to y∗ than x does to x∗ . Note that, for
x close to one, it suffers from the same uniformity problems as the branch identified in (1.26).
• χ(ε) = Os (ψ(ε)): in that case, we can let χ(ε) = ψ(ε), so that (1.27) becomes

Y0 − X + ε(ψ(ε))−2 Y0−1 = 0.

There are now three sub-cases to consider:

√
– the case εψ −2 (ε)  1—equivalently, ψ(ε)  ε—is untenable, as the leading order term √ is the third
one which cannot be matched by the remaining two; hence, there are no solutions o( ε)−close to
x∗ (this extends our earlier result);
√
– the case εψ −2 (ε)  1—equivalently, ψ(ε)  ε, on the other hand, yields the solution Y = X. This
corresponds to the branch x = y—recall that x = 1 + χ(ε) X, y = 1 + ψ(ε) Y , and χ(ε) = ψ(ε)—and
√ we establish that this branch extends to a point which√approaches x∗ , as ε ↓ 0, no faster than
so
ε. Additionally, at distances asymptotically larger than ε from x∗ , the only solution branch for
which x − 1 and y − 1 are commensurate is the one we had already identified;
3 The reader can think of this requirement as a way to make the decompositions of x and y more concrete. In particular, each

one of these variables is decomposed into two parts—X and χ for x, Y and ψ for y. Naturally, there is no unique way to do this.
The constraints 0 6= X = O(1) and 0 6= Y = O(1) ensure that χ(ε) and ψ(ε) measure the proximity of x and y to x∗ = 1 and
y∗ = 1, respectively.

11
 √
– finally, the case εψ −2 (ε) = O(1)—equivalently, ψ(ε) = O( ε)—yields the quadratic
s 
2
X X
Y02 − X Y0 + 1 = 0, with solutions Y0± = ± − 1.
2 2
√
Note that the regime X  1 corresponds to the regime ψ(ε)X  ε, and the solutions above
−1
become (upon Taylor-expanding them—recall that X = o(1) in this regime)
 s  2 
± X X  2 X + o(X −1 ),
Y0 = ±  1− = (1.29)
X + o(X −3 ).
−1
 
2 2 X

The first among these corresponds to the branch reported in (1.26), whereas the second of these
corresponds to the branch reported in (1.28).

Discussion. To summarize the results of our investigation near (x∗ , y∗ ), we√note the following (see also Fig. 3):
first, the solution branch with asymptotic expansion as in (1.26) extends ε−close to x∗ . In fact, √ it extends
√
all the way to two points x− < x∗ (from the left) and x+ > x∗ (from the right), where x± = 1 ± 2 ε + o( ε).
(These points correspond to the X−values for which the root in (1.29) disappears. The higher order terms in
their expansions derive√from the higher order sinusoidal term which we omitted in writing (1.27).) The branch
remains uniformly O( ε)−close over [0, x− ] ∪ [x+ , 2] to the leading order approximation y = x, albeit only in a
C 0 −fashion. [Why does C 1 −proximity fail? How can one deduce this from Fig. 3?] Additionally, we identified
a second solution branch with asymptotic expansion
√ reported in (1.28) and extending up to the same points
x± . This second branch remains uniformly O( ε)−close (also in a C 0 −fashion) over [0, x− ] ∪ [x+ , 2] to the
leading order approximation y = 1. At x± , these two solution branches meet and annihilate each other.
Note, further, that all cases we considered
√ above turned out to be subcases of the final one. Indeed, both
branches we identified for x − x∗  ε—namely Y (x) ∼ X and Y (x) ∼ X −1 —are limiting cases of the
two-branched solution reported in (1.29). Kevorkian and Cole [Kevorkian & Cole] call this the principle of
least degeneracy, since this final case is the only one in which all three terms play a role (i.e., are of the
same asymptotic magnitude). Additionally, they work out the matching of each branch for their example:
that is, they show that the formulas for Y (X)—which are valid Os (ε)−close to x∗ , in our case—match their
counterparts y(x) at an Os (1)−distance from x∗ over an intermediate lengthscale. We will discuss the role of
matching in the context of singularly perturbed ODEs at a later section; until then, we remark that expressing
the solutions identified in (1.29) in terms of x and y leads to (1.26) and (1.28), up to and including Os (ε)−terms.

Remark. Note that (1.24) is explicitly solvable for x,

ε
x(y) = y − ε sin(y) + , (1.30)
y−1

which facilitates the analysis of (1.24). Further, one can shed more light into the singularly perturbed nature
of (1.24) by rewriting it as

g̃(x, y; ε) = (y − 1)(y − x) + ε [1 − (y − 1) sin(y)] = 0, with x ∈ [0, 2].

In this form, it is directly apparent that the solution y = y(x) has two branches. Note that ∂y g̃(1, 1; 0) = 0,
and hence
√ the Implicit Function Theorem remains inapplicable at x∗ —as it should, since for every ε > 0 there
is a O( ε) neighborhood of x∗ where the equation above has no solutions (recall our work above).

12
 first second
branch branch

x(y)

x(y)

second first
branch branch

0
0 x 1 x+
−

Figure 3: The exact solution to (1.24) as given in (1.30) (in black), together with the two-term asymptotic
expansions of the lower (in red) and upper (in blue) branches, see (1.26) and (1.28), respectively.

Homework set no.1 due February 19, 2013

01. Consider the quadratic equation

(1 − ε)x2 − 2x + 1 = 0.
This is a regularly perturbed problem, and thus we anticipate all roots to remain bounded as ε ↓ 0. Set, then,
X
x= cn εn ,
n≥0

substitute into the equation, and work out the equations for c0 , c1 , and c2 . What goes wrong? Explain the
reason this approach fails and devise a way to rectify it.

02. Consider the transcendental equation

x
− ex−1/ε − ε = 0, where x ∈ R.
x−1
(a) Derive an asymptotic expansion for any bounded roots x(ε) this equation might have. Find the coefficient
of the nth (‘general ’) term and show that your asymptotic expansion converges for all ε in a neighborhood of
zero. Does this asymptotic expansion converge to x(ε) for ε 6= 0?
(b∗ ) Does the equation above also have unbounded roots? If it does, can you derive asymptotic expansions for
them?

03. [Holmes, Sect. 1.5, Ex. 9] Let A, D be real, n × n matrices.

(a∗ ) Suppose A is symmetric and has n distinct eigenvalues. Find a two-term asymptotic expansion of the
eigenvalues of the perturbed matrix A + εD, where D is positive definite. What you are finding is known as a
Rayleigh–Schrödinger series for the eigenvalues.
(b) Suppose A is the identity and D is symmetric. Find a two-term asymptotic expansion of the eigenvalues of
the matrix A + εD.
(c) By considering    
0 1 0 0
A= and D = ,
0 0 1 0

13
show that a O(ε) perturbation of a matrix need not result in a O(ε) perturbation of the eigenvalues. This
example also demonstrates that a smooth (smooth here means differentiable) perturbation of a matrix need
not result in a smooth peturbation of the eigenvalues.

04. Let D ⊂ RM (for some M ∈ N) and I = (0, ε0 ) (for some ε0 > 0), and consider functions f , g, φ, and γ
all of which are defined on D × I and real-valued. Assume that f (x, ε) = O(φ(x, ε)) and g(x, ε) = O(γ(x, ε)),
both as ε ↓ 0.
(a) Show that f (x, ε)g(x, ε) = O(φ(x, ε)γ(x, ε)), as ε ↓ 0.
(b) Show that f (x, ε) + g(x, ε) = O(|φ(x, ε)| + |γ(x, ε)|) but f (x, ε) + g(x, ε) 6= O(φ(x, ε) + γ(x, ε)), in general
(both as ε ↓ 0).
(c∗ ) Recall our definition of the symbol Os . Verhulst [Verhulst, Sect. 2.1] also introduces the symbol ≈ with
the meaning

f ≈ g (ε ↓ 0) if and only if [f (ε) = O(g(ε)) (ε ↓ 0)] ∧ [g(ε) = O(f (ε)) (ε ↓ 0)].

Are the two relations equivalent? If not, does one of the relations f = Os (g) and f ≈ g (both as ε ↓ 0) imply
the other? If yes, prove it; if not, provide a counterexample.

14
Lecture no.2 February 12, 2013

2.1 A regularly perturbed ODE

We now turn to perturbed ODEs and consider a simple boundary-value problem involving a linear ODE,

y 00 − εy 0 − y = 0, with y(0) = −1, y(1) = 1. (2.1)

The general solution to the linear ODE is

y(x; ε) = c+ eλ+ x + c− eλ− x , where c+ and c− are arbitrary constants.

p
Here, λ± = ε/2± 1 + ε2 /4 are the roots of the characteristic polynomial. Employing the boundary conditions,
we find
1  λ−

λ+ x λ+

λ− x 
y(x; ε) = λ+ λ
1 + e e − 1 + e e . (2.2)
e −e −
We remark for later use that, Taylor-expanding λ± , eλ± , and eλ± x and retaining all terms up to and including
Os (ε), we find
1 n h i h i
2

o
y(x; ε) = 2 sinh(x) + sinh(x − 1) + ε (x − 1) sinh(x) + x sinh(x − 1) + O ε . (2.3)
e − e−1

The same asymptotic result can be obtained by postulating the asymptotic expansion
X
y(x) = φn (ε) yn (x) (2.4)
n≥0

for the solution, substituting into the ODE (2.1), matching terms to derive ODEs for each one of the components
y0 , y1 , . . ., and solving these ODEs with the help of the boundary conditions to find these components.

• Substitution yields
X X X
0 = φn yn00 − ε φn yn0 − φn yn
n≥0 n≥0 n≥0

= [φ0 y000 + φ1 y100 + o(φ1 )] − ε [φ0 y00 + φ1 y10 + o(φ1 )] − [φ0 y0 + φ1 y1 + o(φ1 )] ,

where we have assumed that the asymptotic expansion (2.4) can be differentiated term-by-term (recall
that this is not possible, in general).
• The boundary conditions yield1

X X y0 (0) = −1 and y0 (1) = 1,
φn (ε) yn (0) = −1 and φn (ε) yn (1) = 1, whence φ0 (ε) = 1 and
yn (0) = yn (1) = 0, n ≥ 1.
n≥0 n≥0

• The only leading order terms that can be matched in the ODE above are those in the first and third
terms, since εφ0 (ε) = o(φ0 (ε)). Hence,

y000 − y0 = 0, whence y0 (x) = c+,0 ex + c−,0 e−x . (2.5)

• Employing the boundary conditions y0 (0) = −1 and y0 (1) = 1, we find

2 h i
y0 (x) = sinh(x) + sinh(x − 1) . (2.6)
e − e−1
This concludes the determination of φ0 and y0 , and hence also of the leading order behavior of y.
1 Note that, strictly speaking, the boundary conditions only yield that lim
ε↓0 φ0 (ε) = 1—equivalently, that φ0 = O(1)—and
not that φ0 (ε) = 1. Since each member of an asymptotic sequence is characterized solely by its limit behavior for ε ↓ 0, we may
choose φ0 (ε) = 1 without loss of generality. Note that this choice simplifies considerably the boundary conditions for the elements
of {yn }n≥0 . [Work out the boundary conditions for y0 and y1 for the choices φ0 (ε) = 1/(1 + ε) and φ1 (ε) = ε/(1 + ε).]

15
 • We now match the next-order terms. If φ1 (ε)  ε, then the only possible matching is between φ1 y100 and
φ1 y1 . The outcome is an ODE for y1 identical to that for y0 (see (2.5)) and equipped with homogeneous
boundary conditions; it follows that y1 ≡ 0, and hence the term φ1 y1 is, in fact, absent from the asymptotic
expansion. Therefore, the next-order correction must be at most O(ε). If φ(ε)  ε, now, then the
next-order correction is −εφ0 (ε)y00 = −εy00 . This cannot be matched by any other term, as these are
asymptotically smaller, and thus it must be identically zero. We thus reach a contradiction, as y00 6= 0 by
(2.6), and hence φ(ε) 6 ε either. We thus confidently choose φ(ε) = Os (ε) (note that this scaling satisfies
the aforementioned principle of least degeneracy, as all three terms—φ1 y100 , εφ0 y00 , and φ1 y1 —now play a
role). For simplicity, we let φ(ε) = ε, whence

y100 − y1 = y00 , with y1 (0) = y1 (1) = 0. (2.7)

This problem is explicitly solvable using variation of parameters,

ˆ x 0 ˆ x
y0 (s) e−s y00 (s) es
y1 (x) = c+,1 ex + c−,1 e−x − ex s −s
ds + e −x
s −s )
ds
0 W (e , e ) 0 W (e , e
1 + e−1 x 1 x
   
1 + e x 1 −x
= c+,1 + + e + c −,1 − − e ,
e − e−1 2 4 e − e−1 2 4

where
e−s 
 

 es
W es , e−s =  s 0 = −2
(e ) (e−s )0 
is the Wronskian of the fundamental solutions es and e−s .
• Employing the homogeneous boundary conditions, we obtain
1 h i
y1 (x) = (x − 1) sinh(x) + x sinh(x − 1) . (2.8)
e − e−1

This process can be repeated as many times as needed to obtain higher order terms in the asymptotic
expansion of the solution y(·; ε).

Discussion. Note that (2.6) and (2.8) match the two-term Taylor expansion (2.3); in fact, it can be shown
that the asymptotic expansion for y matches the Taylor expansion of the exact solution (2.2) to all orders in ε.
A corollary of this observation is that (2.4) can, indeed, be differentiated term-by-term, since it is convergent
for t ∈ R̄+ and ε in a neighborhood of zero. Additionally, the Os (ε) component of our two-term expansion
(2.3) remains Os (ε) uniformly over any bounded subset of R—a property which was advertised in Sect. 1.5 as
characteristic of regularly perturbed problems.
A plot of the solution together with the associated phase plane analysis is given in Fig. 4. The effect of the
perturbative term εẏ on any bounded subset of the phase plane—such as the one depicted in that figure—is
limited to a (uniformly) Os (ε)−perturbation of the orbits and of their parameterizations by time.

2.2 A singularly perturbed ODE

[See [Kevorkian & Cole, Sect. 2.1] for another discussion of this example.] Consider, now, the damped linear
oscillator with small mass,

d2 Y dY dY (0)
m +b + kY = 0, with Y (0) = 0 and = v0 . (2.9)
dT 2 dT dT
Here, Y is the displacement of the oscillator from the equilibrium state, m its (“small”) mass, v0 its initial
velocity, b > 0 the damping coefficient, and k > 0 the spring constant, see also Fig. 5.

16
 y y

0 0

−1

−2

−1 0 1 y 0 1 t

Figure 4: Left panel : the exact, perturbed (ε > 0) solution to (2.1) as given in (2.2) (solid red) and plotted
in the (y, ẏ)−plane (phase plane) together with its extension outside the unit time interval (dashed red). Also
shown: the exact, unperturbed (ε = 0) solution (solid black) together with its extension (dashed black); the
perturbed (solid blue) and unperturbed (solid black) eigenspaces; several other orbits corresponding to the
same ODE (dashed blue) but failing to satisfy the boundary conditions. The solid blue part of those orbits
residing in the top quarter of the phase plane corresponds to the unit time interval: note that these orbits fail
either because they are too close to the saddle point and thus too slow or because they are too far removed
from that point and thus too fast. Right panel : the perturbed—y(·; ε), in red—and unperturbed—y(·; 0), in
black—solutions plotted as functions of time.

displacement
Y

velocity
restoring force dY/dT damping
−kY −b dY/dT

Figure 5: Schematic depiction of the mass–damper–spring system described by (2.9).

2.2.1 Scaling the model

Our first task is to non-dimensionalize (scale) the independent (T ) and dependent (Y ) variables. To quote
Segel and Slemrod [Segel & Slemrod], some of the considerations entering scaling are the following: the scale
of a dependent variable should provide an estimate of the order of magnitude of that variable; the scale of
an independent variable should provide an estimate of the range of that variable over which the dependent
variables change significantly. A crucial third consideration is that ‘[. . .] scaling requires prior knowledge of
the solution. This can be provided by experiments, “physical intuition,” and/or numerical analyses of special
cases.’ 2
2 We leave out, for the time being, an equally crucial fourth consideration—namely, ‘It may be necessary to choose different
scales in different domains of the independent variable. If so, the appropriate approximation methods are of singular perturbation
type.’— It will be inadvertently brought to the spotlight in one of the later sections.

17
 We begin to glean information on the solution by energy considerations. Initially, the mass is at equilibrium
(Y (0) = 0) and thus the system only has kinetic energy, E(0) = mv02 /2. At the point of maximum displacement,
Y (T∗ ) = YM , velocity—and thus also kinetic energy—is zero and the spring has stored potential energy equal
2
to E(T∗ ) = kYM /2. By conservation of energy,
ˆ YM ˆ T∗  2
1 1 2 dY (T ) dY (T )
mv 2 = kYM + Wd , where Wd = b dY (T ) = b dT > 0 (2.10)
2 0 2 0 dT 0 dT

is the work done by the damping force (i.e., lost to friction). Hence,
r
1 2 1 m
kY < mv02 , or equivalently, YM < v0 .
2 M 2 k

If v0 remains bounded as m ↓ 0, then YM ↓ 0 by the above inequality. In this example, we choose to examine
motion the maximum displacement YM of which remains bounded and nonzero as m ↓ 0; hence, v0 must scale
at least as m−1/2 for YM to be nonzero; consequently, v0 must be unbounded as m ↓ 0.
A first consequence of this result is that the damping force b dY /dT initially dominates over the restoring
force kY : indeed, b dY /dT 6 m−1/2 while kY 6 1, since Y ranges in the bounded interval [0, YM ]. Hence,
one expects that the last term in the left member of (2.9) may be omitted “to leading order” to obtain the
approximate ODE
0
d2 Y dY dY (0)
m 2 +b = 0, with Y (0) = 0 and = v0 . (2.11)
dT dT dT
The solution of this approximate problem is
mv0  
Y (T ) = 1 − e−bT /m , (2.12)
b
a fact which suggests the scalings ψ = Y /[Y ] and τ = T /[T ], with [Y ] = mv0 /b and [T ] = m/b.
Note that, under this scaling, v0 must scale like m−1  m−1/2 in order to obtain an Os (1) maximum
displacement YM = mv0 /b! This apparent discrepancy is due to the fact that, for v0  1, essentially all kinetic
energy is lost to friction and only an o(E(0)) amount is left to be stored in the spring. Indeed, the amount of
energy lost to friction can be determined by (2.10), if T∗ is known. Note that (2.12) predicts an exponential
approach to Y = YM , i.e., that the mass will never reach its maximum displacement but, rather, tend to it
asymptotically. The reason for this discrepancy between our intuition and our analytic formula is our omission
in (2.11) of the restoring force kY , which becomes commensurate to damping as the velocity decreases and
the displacement approaches its maximum. Hence, we will estimate T∗ by demanding that the damping force
becomes Os (1), as the restoring force is near YM . Using (2.12), we calculate dY (T )/dT = v0 e−bT /m , and hence
the velocity becomes Os (1) for e−bT∗ /m = Os (v0−1 ) = Os (m) = o(1). (We remark, for later use, that this
entails the asymptotic relation T∗ = Os (m ln m), and hence T∗ ↓ 0 as m ↓ 0.) Using this estimate and (2.10),
we calculate ˆ YM  2 ˆ T∗
dY (T )  
Wd = b dT = bv02 e−2bT /m dT = E(0) 1 − e−2bT∗ /m ,
0 dT 0

with E(0) = mv02 /2 as before. The amount of energy left to be stored in the spring equals E(0)e−2bT∗ /m which
is, indeed, asymptotically smaller than E(0) (by our estimate e−bT∗ /m = o(1) above).
Using the scalings ψ = bY /(mv0 ) and τ = bT /m introduced above, we rewrite (2.9) in the form

mk
ψ 00 (τ ; ε) + ψ 0 (τ ; ε) + εψ(τ ; ε) = 0, with ψ(0; ε) = 0 and ψ 0 (0; ε) = 1, and where ε = >0 (2.13)
b2
is a parameter taking values in a neighborhood of zero, on account of m being allowed to become arbitrarily
small. Additionally, (·)0 = d · /dτ . Note, also, that our scaling has brought down the number of parameters
present in the model from four (m, b, k, and v0 ) to one (ε); this feat alone is justification enough for the
effort we put into scaling the model, as it reveals that systems satisfying (2.9) but equipped with different sets
of parameters may exhibit similar behavior. In particular, the four-dimensional parameter space is foliated :
each foil is defined by the relation ε = mk/b2 = const., and parameter values belonging to the same foil yield
systems with identical behavior up to rescaling time and/or space.

18
2.2.2 The initial transient

The problem presented in (2.13) is regularly perturbed, as setting ε to zero does not cause a reduction in the
order of the ODE. Following our work in Sect. 2.1, we postulate the asymptotic expansion
X
ψ(τ ; ε) = φn (ε)ψn (τ ), (2.14)
n≥0

substitute into (2.13) to find

X X X X X
φn ψn00 + φn ψn0 + ε φn ψn = 0, with φn (ε)ψn (0) = 0 and φn (ε)ψn0 (0) = 1,
n≥0 n≥0 n≥0 n≥0 n≥0

and group terms of equal asymptotic order to find

ψ000 + ψ00 = 0, with general solution ψ0 (τ ) = C+,0 + C−,0 e−τ . (2.15)

Now, the initial condition pertaining to the velocity yields φ0 (ε) = 1—in fact, we could have guessed this
readily from our scaling analysis above, as the approximate solution (2.12) and our scaling for Y suggest that
ψ lies in the unit interval, in the timescale under examination. Hence, we may select φ0 ≡ 1 and thus recast
the boundary conditions as

∀n≥0 [ψn (0) = 0], ψ00 (0) = 1, and ∀n≥1 [ψn0 (0) = 0].

Combining the boundary condition for ψ0 with (2.15), we then find3

ψ0 (τ ) = 1 − e−τ . (2.16)

Higher order terms may be obtained in a manner similar to that we employed in Sect. 2.1. For example,
employing the principle of least degeneracy to set φ1 (ε) = ε (the reader can exclude the cases φ1  ε and
φ1  ε as in the aforementioned section), we find the linear, inhomogeneous problem

ψ100 + ψ10 = −ψ0 , with ψ1 (0) = 0 and ψ10 (0) = 0.

The solution to this problem is ψ1 (τ ) = 2 − τ + (2 + τ )e−τ , so that the two-term expansion for ψ is

ψ(τ ; ε) = 1 − e−τ + ε 2 − τ − (2 + τ )e−τ + Os (ε2 ).

 
(2.17)

As we discussed earlier, the leading order term in the truncated asymptotic expansion (2.17) predicts exponential
decay towards ψ = 1. This picture is significantly altered by the next-order term in that expansion, which grows
unboundedly with τ due to the linear term in it. The expansion only remains well-ordered for τ = o(1/ε); for
larger values of τ , the second term in the asymptotic expansion becomes commensurate with the first; in fact,
this is true of all subsequent terms, as (it is easy to show that) the fastest-growing term in ψn (τ ) grows as τ n .
Recalling the definitions of ε and τ , we find that the relation τ = o(1/ε) becomes kT /b = o(1) (m ↓ 0)—that
is, one can expect that the asymptotic expansion only provides a legitimate solution to the problem over a
time interval [0, T̄ ], with T̄ ↓ 0 as m ↓ 0. (The meaning of the non-dimensional time kT /b appearing above
will be clarified in the next section.) Past that short time interval, our assumption that the restoring force is
higher-order becomes invalid and we have to use a different scaling. This fact lends this initial behavior of the
system its name—transient—as it soon offers its place to phenomena of an entirely different qualitative nature.

2.2.3 Behavior past the initial transient

To describe the behavior of the system past the initial transient analyzed above, we zoom out and consider a
longer timescale. Since the asymptotic expansion derived for (2.13) breaks down for τ = o(1/ε), we will focus
precisely on Os (1/ε) − values of τ . To that effect, we rescale τ and define the new time variable t = ετ = kT /b,
writing also ψ(τ (t); ε) = ψ(t/ε; ε) = y(t; ε) for clarity. Under this rescaling, (2.13) becomes
1
εÿ(t; ε) + ẏ(t; ε) + y(t; ε) = 0, with y(0; ε) = 0 and ẏ(0; ε) = , (2.18)
ε
3 The attentive reader will realize that this is merely the rescaled form of (2.12), just as the problem (2.15) it satisfies is merely

the rescaled version of (2.11).

19
and where we have written (·) ˙ = d · /dt. (Note that (2.18) is of singular perturbation type, since setting ε = 0
in it reduces the order of the ODE.) Here also, we postulate an asymptotic expansion
X
y(t; ε) = fn (ε)yn (t), (2.19)
n≥0

substitute into (2.13) to find X X X

ε fn ÿn + fn ẏn + fn yn = 0,
n≥0 n≥0 n≥0

and group terms of like asymptotic order to find

ẏ0 + y0 = 0, with general solution y0 (t) = C0 e−t . (2.20)

Here, C0 is an arbitrary constant. Higher order terms may be obtained in a similar manner. For example, at
next order, we set f1 (ε) = εf0 (ε) to find the linear, inhomogeneous problem

ẏ1 + y1 = −ÿ0 , with general solution y1 (t) = (C1 − C0 t) e−t .

Here, C1 is another arbitrary constant. Hence, the two-term expansion for y is

y(t; ε) = C0 e−t + ε (C1 − C0 t) e−t + Os (ε2 ), (2.21)

with the constants C0 and C1 remaining undetermined.

At this point, it is not clear how these constants must be selected, as only one initial condition can be
accommodated. The initial condition y(0; ε) = 0 yields y0 (t) ≡ 0, which implies that the displacement y in
this timescale is o(1)—this is in disagreement with our intuition on the behavior of the system. The initial
condition ẏ(0; ε) = 1/ε, on the other hand, yields f0 (ε)y0 (t) = −ε−1 e−t ; this is equally unrealistic, as it implies
that y  1 and, also, that y is negative (!) although y(0; ε) = 0 and ẏ(0; ε) > 0. In truth, we should not
use any of the prescribed initial conditions, as we do not expect (2.19) to describe the behavior of the system
as t ↓ 0. To see this, recall that the behavior of the system close to zero is described, to leading order, by
ψ(τ ; ε) ∼ 1 − e−τ . Passing to t = τ /ε, we rewrite this as y(t; ε) ∼ 1 − e−t/ε ; note that the exponential term in
this relation is exponentially small for each t > 0 but not uniformly over any time interval containing zero. In
fact, this relation suggests that
1 −t/ε 1 −t/ε
ẏ(t; ε) ∼ e and ÿ(t; ε) ∼ − e ,
ε ε2
and hence εÿ in (2.18) is not higher-order but, instead, of the same order as ẏ. (This agrees with our analysis
in the last section, naturally.) An asymptotic expansion such as the one in (2.19), on the other hand, could
never capture this behavior, as the second-order term in (2.18) ends up being considered higher-order. We will
see how to determine the arbitrary constants in (2.21) through matching in the next section.

2.2.4 Matching

The idea behind matching is that the solutions to (2.13) and (2.18) represent the same, unique solution in
different timescales. Since ψ is valid in a neighborhood of zero which vanishes as ε ↓ 0 and y is valid everywhere
except for a neighborhood of zero, it seems natural to try to identify a region where both are equally valid.
Should we manage to locate such a region, the solutions should match (i.e., be identical) in it. In this case, we
will be able to use our explicit knowledge of (2.14) to determine the unknown constants in (2.19).
First, we deal with the leading order results (2.16) and (2.20) for the inner solution ψ and the outer solution
y, respectively. (The terms ‘inner’ and ‘outer’ reflect the presence of a boundary layer for our problem: the
‘inner solution’ is the solution inside that boundary layer, while the ‘outer solution’ is the solution outside it.)
As discussed earlier, limτ →∞ ψ0 (τ ) = 1; additionally, limt↓0 y0 (t) = C0 . Continuity of the solution implies,
then, that C0 = 1. In fact, we can add more content to this fast calculation by showing that there is an entire
range of intermediate timescales—all of them slower than the fast one and faster than the slow one—in which
the two solutions ψ0 and y0 match exactly, for this specific choice of C0 . To locate them, let δ : I → R be a
function such that ε  δ  1, and introduce the variable s = δτ = δt/ε. By virtue of ε  δ  1, Os (1)−values

20
of s correspond to asymptotically large values of τ and asymptotically small values of t. Recasting (2.16) and
(2.20) in terms of this new variable, we find, for s in any fixed interval [s, s̄] (with s > 0),
 
ψ0 (τ (s)) = 1 − e−s/δ = 1 + O e−s/δ and y0 (t(s)) = C0 e−εs/δ = C0 + O (ε/δ) .

Indeed, then, the two formulas match each other exactly, as long as C0 = 1. This match is extended to all
intermediate timescales, i.e., for all ε  δ  1.
This result can be extended to higher orders. Indeed, recasting the two-term expansion (2.17) for ψ in terms
of s, we obtain h s  s  −s/δ i h si
1 − e−s/δ + ε 2 − − 2 + e =1+ε 2− + Os (ε2 /δ 2 ),
δ δ δ
since the exponential term is asymptotically smaller than all algebraic terms. Note that the Os (ε2 /δ 2 )−remainder
derives from ε2 ψ2 —the Os (ε2 )−term in the asymptotic expansion of ψ. (Here, we have used our earlier remark
on the fastest-growing term in ψn .) The two-term expansion (2.21) for y, in turn, becomes
h εs i −εs/δ h si
C0 e−εs/δ + ε C1 − C0 e = C0 + ε C1 − C0 + Os (ε2 /δ 2 ),
δ δ
where we have Taylor-expanded the exponential terms, as their arguments are asymptotically small. The two
formulas match as long as C1 = 2, which fixes the value for the second arbitrary constant, too.
We finally note that, following matching, the two (truncated) asymptotic expansions above can be combined
to derive a uniformly valid (truncated) asymptotic expansion. Indeed, adding them up and subtracting their
common part at each order, as this was identified above, we find

e−t + ε (2 − t) e−t − e−t/ε + (2ε + t)e−t/ε + Os (ε2 ). (2.22)

2.2.5 Discussion

Our results in Sect. 2.2.2–2.2.3 exemplify Segel and Slemrod’s admonition on the necessity of choosing different
timescales to describe different phases of the system evolution. Indeed, to elucidate the behavior of the system
during the short initial transient, we have worked with the fast time τ = bT /m. In that phase, friction
dominates and causes a radical decrease in momentum; hence, the behavior of the system is described, to
leading order, by (2.11)—or, in its rescaled form, (2.15). To analyze the behavior in the much slower timescale,
we have recast the system in the slow time t = bT /m. In this slow phase, friction and the restoring force are
both moderate and approximately counteract each other, so that the momentum barely changes. To leading
order, then, the system evolution is described by
dY
b + kY = 0,
dT
which is the dimensional form of (2.20). Note that this ODE is simply obtained by omitting the inertial term
mY 00 in (2.9)—a “natural” choice, as m is assumed “small”, albeit one that only becomes effective once the
fast initial transient plays itself out.
Another focal point of our work in this section was non-dimesionalizing the original system (2.9). It is worth
remarking here that, since the ODE satisfied by Y is linear, it remains invariant under all scalings y = Y /[Y ]
(the prefactors of [Y ] ‘cancel out’). This is hardly surprising, of course, as the solutions to linear equations form
a vector space and thus there can be no “natural” scaling associated with equations of that type. It is the initial
or boundary conditions that select a unique solution, and hence it is these conditions that determine a natural
scaling. In our work here, the scaling [Y ] was suggested by the condition we imposed that the amplitude of the
motion should remain bounded—both above and below—as the mass m of the body limits down to zero; the
actual scaling was then derived using the initial conditions and physical intuition on the relative magnitudes
of various terms in the ODE.
Regarding matching, we note that the two-term uniform asymptotic expansion in (2.22) is, in this particular
example, the (multiscale) Taylor expansion of the exact solution,
√
eλ+ t − eλ− t −1 ± 1 − 4ε
y(t; ε) = , where λ± = (2.23)
ε (λ+ − λ− ) 2ε

21
 y
y
1/ε

t=t*

t=t
**

y t* t** t

Figure 6: Left panel : the exact, perturbed (ε > 0) solution to (2.9) as given in (2.23) (solid red) and plotted
in the phase plane. Note the points corresponding to the time instants t∗ and t∗∗ , at which ẏ(t∗ ) = 0 (vertical
vector field) and ÿ(t∗∗ ) = 0 (horizontal vector field)—cf. problem 2(b–c) in Homework #02. Also shown:
several other orbits corresponding to the same ODE (dashed blue) but failing to satisfy the initial conditions;
and the perturbed (solid blue) and unperturbed (solid black) eigenspaces. Right panel : the perturbed solution
y(·; ε) plotted as functions of time, together with the points corresponding to the time instants t∗ (the tangent
is horizontal) and t∗∗ (the tangent osculates the graph of the solution) at the edge of the boundary layer.

are the corresponding eigenvalues. (See also Fig. 6, where the solution is plotted both in the corresponding
phase space and as a time trajectory.) To see this, note that
1
λ+ = −1 − ε + O(ε2 ) and λ− = − + 1 + ε + O(ε2 ).
ε
Substituting, now, into (2.23), Taylor-expanding the first exponential, and omitting the second exponential, we
arrive at (2.21) with C0 = 1 and C1 = 2—recall that matching led to the same values for these constants. We
remark that the omission of eλ− t is valid as long as t  ε, as this term is, then, asymptotically smaller than all
algebraic terms—recall that this condition was also employed in matching, while discussing the properties of
the intermediate timescale. Switching, on the other hand, to the fast variable τ = t/ε, we obtain the solution
ψ(τ ; ε) = y(ετ ; ε). In this case, the first exponential can be Taylor-expanded around zero as long as τ  1/ε
(equivalently, t  1), since then λ+ τ  1. The exponent of the second exponential, on the other hand, is O(1)
for values of τ covering an O(1) range, and thus must be retained as is. The result is, here again, (2.17).

22
Homework set no.2 due February 26, 2013

01. Consider the regularly perturbed problem

ÿ(t) + (1 + εf (t)) y(t) = 0, with y(0) = 0 and ẏ(0) = 1.

Here, 0 < ε  1 as usual.

(a) Derive a two-term asymptotic expansion φ0 (ε)y0 (t) + φ1 (ε)y1 (t) for the solution y(t; ε) to this problem in
the case f (t) = e−t . Is this expansion uniformly valid in R̄+ ? Could you have foreseen this from the nature of
the perturbation?
(b) Consider, now, the same problem with f (t) = 1. Derive, here also, a two-term asymptotic expansion for the
solution to this problem, and show that φ1 y1 is not uniformly asymptotically smaller than the leading order
term φ0 y0 over R̄+ . Explain where this non-uniformity comes from.
(c∗ ) Formulate a condition for f which is sufficient for φ1 y1  φ0 y0 to hold uniformly over R̄+ . Explain the
meaning of that condition in physics terms.

02. Consider the rescaled problem (2.13)

ψ 00 (τ ) + ψ 0 (τ ) + εψ(τ ) = 0, with ψ(0) = 0 and ψ 0 (0) = 1,

in which mass, damping coefficient, and initial velocity have been rescaled to one and the spring constant has
been rescaled to ε  1.
(a) Use the asymptotic expansion (2.14), with as many terms as needed, to derive a leading order formula for
the time instant τ∗ at which ψ 0 (τ∗ ) = 0.
(b) Using this leading order formula for τ∗ , calculate the work done by friction over the time interval [0, τ∗ ] and
up to and including terms of Os (ε). Additionally, estimate the energy stored in the spring to leading order.
Do the two results imply conservation of energy up to and including terms of Os (ε)? Explain.
(c∗ ) Repeat the same procedure as in (a) to derive a leading order formula for the time instant τ∗∗ at
which ψ 00 (τ∗∗ ) = 0. Show that, while the point (ψ(τ∗ ), ψ 0 (τ∗ )) in the phaseplane is an Os (1)−distance from
the eigenspace corresponding to the Os (ε) eigenvalue (slow eigenspace), the point (ψ(τ∗∗ ), ψ 0 (τ∗∗ )) is only
Os (ε)−away from it. [Note that this generalizes to n−th order derivatives, thus yielding a method to approxi-
mate points on that slow eigenspace.]

03. [Kevorkian & Cole, Sect. 2.2, Ex. 2] Consider the boundary-value problem

1 d·
εy 00 + √ y 0 − y = 0, with y(0) = 0 and y(1) = e2/3 , and where (·)0 = .
x dx

(a) Calculate a leading order (as ε ↓ 0) approximation to the solution of this problem which is uniformly valid
over the interval [0, 1].
(b∗ ) Can you interpret your results geometrically using a plot of the dynamics in the (x, y, y 0 )−space (extended
phase space)?

23
Lecture no.3 February 19, 2013

The topic of this lecture is (Introductory) Fenichel Theory, also known as Geometric Singular Perturbation
Theory. An excellent introduction to the subject is given in [Kaper] (available here: http://wwwhome.math.
utwente.nl/~antoniosa/teach/mm13/kaper_ams99.pdf). These notes are partially new material and par-
tially a reworking of an older, handwritten set of notes (see http://wwwhome.math.utwente.nl/~antoniosa/
teach/mm13/Fenichel_Notes.pdf) which, in turn, were based on [Jones] (available here: http://wwwhome.
math.utwente.nl/~antoniosa/teach/mm13/jones_montecatiniterme94.pdf). A brief presentation of the
theory alongside a well-researched collection of applications in the Life Sciences can be found in the more
recent [Hek].

3.1 The linear problem

Seeing as our work in the preceding two lectures has been confined to linear models, we briefly present an
overview of singularly perturbed such models to introduce notation and concepts. Let us consider, then, the
linear, constant-coefficient problem in RN

w0 = Aε w, where Aε is an ε−dependent, constant N × N matrix and w = (w1 , . . . , wN )T . (3.1)

Here, (·)0 = d · /dτ denotes differentiation with respect to time, measured by the independent variable τ . The
superscript ‘ε’ in Aε denotes the dependence of that matrix on the small parameter; it is not a power. The
same is true of all superscripts below, with the exception of (·)−1 which denotes the inverse of the parenthesized
matrix.

3.1.1 Basic assumptions and definitions

Assume that we know the following for N × N matrix A0 = Aε=0 :

• It possesses a zero eigenvalue of algebraic and geometric multiplicity equal to n < N .

• All other eigenvalues of A0 may be complex but are bounded away from the imaginary axis iR; that is,
they all have nonzero real parts.

Alongside these assumptions, let us define certain quantities of interest:

• We denote by U 0 the eigenspace of A0 associated with the zero eigenvalue—i.e., U 0 = Ker(A0 )—and let
{u01 , . . . , u0n } be any basis of vectors in RN for that space. That is, U 0 = span{u01 , . . . , u0n } and each basis
vector is mapped to zero by the matrix, A0 u01 = . . . = A0 u0n = 0.
• We write λ01 , . . . , λ0m (with m ≤ ` := N − n) for the remaining eigenvalues. Without loss of generality, we
arrange these in non-descending order of their real parts:

Re(λ01 ) ≤ . . . ≤ Re(λ0m− ) < 0 < Re(λ0m− +1 ) ≤ . . . ≤ Re(λ0m ), for some 0 ≤ m− ≤ m.

The case m− = 0 is taken to mean that all eigenvalues have positive real parts (are unstable); the
case m− = m corresponds to all eigenvalues having negative real parts (being stable); and any two
eigenvalues are allowed to have the same real part only if they are complex conjugates (that is, we do
not list eigenvalues according to their multiplicities). We also introduce the generalized eigenvectors
v10 , . . . , vn0 1 for λ01 , vn0 1 +1 , . . . , vn0 2 for λ02 , and so on all the way through vn0 m−1 +1 , . . . , vn0 m for λ0m ; here,
n1 < . . . < nm = `.1
• We further denote by V 0 = span{v10 , . . . , v`0 } the `−dimensional, invariant under the action of A0 subspace
of RN associated with these eigenvalues. We similarly write V− 0
for the invariant eigenspace associated
with the stable eigenvalues, which is of (some) dimension `− = nm− ≥ m− and spanned by {v10 , . . . , v`0− }.
0
Its unstable counterpart is denoted by V+ , of dimension `+ = ` − `− ≥ m − m− =: m+ and spanned by
{v`0− +1 , . . . , v`0 }.
1 Clearly, the numbers n1 , n2 − n1 , . . . , nm − nm−1 are the algebraic multiplicities of the eigenvalues.

24
 Given these definitions, we can write RN = U 0 ⊕ V− 0 0
⊕ V+ . That is, every vector w ∈ RN has a unique
0 0 0
decomposition of the form w = u + v− + v+ , with u ∈ U , v− ∈ V− and v+ ∈ V+ . Using the bases for these
three spaces, we can write
X X X
u= xi u0i , v− = yi vi0 , and v+ = 0
yi+`− vi+`−
.
1≤i≤n 1≤i≤`− 1≤i≤`+

This decomposition amounts to a change of coordinates w 7→ z, written more compactly in the form
 
x1
 .. 

 . 

 xn 
   
w1 
x
  y
 1


0  ..  .
 and P 0 = u01 , . . . , u0n , v10 , . . . , v`0 = U 0 , V−0 , V+0 .
   
w = P z, where w =  .  , z =  y−  = 
 .. 
y+
 
wN  y`
 −


 y` +1 
 − 
 .. 
 . 
y`
(Note that P 0 here is a N × N matrix with its columns listed explicitly; the expressions for the N × n, N × `−
and N × `+ matrices U 0 , V−0 and V+0 are hopefully clear.) The inverse change of coordinates z 7→ w reads
 ⊥,0    V ⊥,0 w 
V 
x
 ⊥,0   ⊥,0 
z = (P 0 )−1 w, where (P 0 )−1 =  U+  ; hence, z =  y−  =  U+ w  .
U ⊥,0 y+ U ⊥,0 w
− −

The notation here is suggestive (if somewhat cluttered): the columns of the N × n block U 0 span the space U 0
and those of the N × `± blocks V±0 span V± 0
. The rows, on the other hand, of the n × N block V ⊥,0 span the
0 0 ⊥ ⊥,0 0 ⊥
orthogonal complement of V —written (V ) —and those of the `∓ ×N blocks U± span the spaces (U ⊕V± ) .
−1
This becomes evident by writing out blockwise the identity P P = I,
 ⊥,0 0
V ⊥,0 V−0 V ⊥,0 V+0
  
V U I 0 0
 ⊥,0 0 ⊥,0 0 ⊥,0 0 
 U+ U U+ V− U+ V+  =  0 I 0  , (3.2)
 
⊥,0 0 ⊥,0 0 ⊥,0 0 0 0 I
U− U U− V− U− V+
and interpreting the equations involving zero blocks as orthogonality relations.

3.1.2 The ε = 0 dynamics

Using this information, we can understand the dynamic behavior of our ε = 0 problem. A standard calculation
yields, for ε = 0,
 ⊥,0 0 0
V ⊥,0 A0 V−0 V ⊥,0 A0 V+0

V A U
 ⊥,0 0 0 ⊥,0 0 0 ⊥,0 0 0 
z 0 = (P 0 )−1 A0 P 0 z, where (P 0 )−1 A0 P 0 =  U+ A U U+ A V− U+ A V+  .
⊥,0 0 0 ⊥,0 0 0 ⊥,0 0 0
U− A U U− A V− U− A V+
Let us now derive concrete expressions for the blocks in this matrix.

• First, the leftmost (block-)column of the matrix is identically zero by virtue of the identity A0 U 0 = 0.
Indeed, recall that the columns of U 0 are associated with the zero eigenvalue.
• Second, note for immediate use below that A0 V±0 = V±0 Λ0± , for some `± × `± matrices Λ0± , because
V±0 is invariant under the action of A0 and associated with the unstable/stable eigenvalues.2 Since we
selected the columns of V±0 to be generalized eigenvectors, Λ0± will be in Jordan canonical form with the
unstable/stable eigenvalues along its diagonal arranged according to their algebraic multiplicity.
2 The reader should convince themselves that A0 V 0 = V 0 Λ0 is indeed an expression of invariance of the spaces spanned by the
± ± ±
columns of V±0 .

25
 • It now follows that the top (block-)row of the matrix is also identically zero. Indeed, we calculate
V ⊥,0 A0 V±0 = V ⊥,0 V±0 Λ0± = 0, where the last step follows from (3.2).
⊥,0 0 0 ⊥,0 0
• By a similar token, the two remaining off-diagonal blocks are also zero, U± A V± = U± V± Λ± = 0.
⊥,0 0 0 ⊥,0 0 0
• Finally, the bottom diagonal blocks become U± A V∓ = U± V∓ Λ∓ = Λ0∓ .

It follows from the above that the ε = 0 ODE system for the z−components reads

x0 = 0,
0
y− = Λ0− y− , (3.3)
0
y+ = Λ0+ y+ .

The dynamic behavior of (3.3) is easy to describe in its entirety. First, the fixed points of this system are
precisely the points on the x−coordinate space (y = 0), since the matrices Λ0± are invertible;3 we will be
writing M0 = {(x, 0) | x ∈ Rn } for this subspace of fixed points. Next, for every fixed x ∈ Rn , the affine linear
space F 0 (x) = {(x, y) | y ∈ R` } is invariant under the flow generated by (3.3). This follows trivially by x0 = 0,
and the space is called an ε = 0 fast fiber. Finally, the subspaces M0 and F 0 (x) intersect transversally (in
fact, orthogonally) at the point (x, 0)—called the base point of that fiber—which is the unique fixed point on
the invariant set F 0 (x). In particular, the phase space Rn+` is foliated by these fast fibers,
[
Rn+` = F 0 (x).
(x,0)∈M0

0
For each x, the dynamic behavior on F 0 (x) is dictated by the ODE system y± = Λ0± y± , which makes no
0
reference to x. Hence, all fibers have identical dynamics. The evolution on a fiber F (x) is explicitly determined
by integration of the ODEs for y. Writing φ0τ (x, y) for the flow on that plane carrying the initial condition
(x, y) to the corresponding solution at time τ , we find

x
 
τ Λ0−
φ0τ (x, y) =  e y−  . (3.4)
τ Λ0+
e y+

It follows that
 0   0 
y−  < C− e−τ σ− ||y− ||
 τ Λ−   τ Λ+ 
e (for τ > 0) and e y+  < C+ eτ σ+ ||y+ || (for τ < 0),
   
for some constants C± ≥ 1, 0 < σ− < Re(λ0m− ) and 0 < σ+ < Re(λ0m− +1 ).4 In other words, the y− −
   
and y+ −components of the solution contract exponentially in forward and backward time, respectively. In
particular, initial conditions on an affine copy of the y− −coordinate plane (or y+ −coordinate plane) based at
the fixed point (x, 0) ∈ F 0 (x) limit down to that fixed point in forward (or backward) time. These are the only
subsets with those properties: any initial condition lying outside these subspaces yields an unbounded solution
both in forward and backward time, so that we can write
0
 
W− (x, 0) = (x, y)  limτ →∞ φ0τ (x, y) = x = {x} × R`− × {0}`+ ⊂ F 0 (x),
0 (3.5)
W+ (x, 0) = (x, y)  limτ →−∞ φ0τ (x, y) = x = {x} × {0}`− × R`+ ⊂ F 0 (x).

These subspaces are termed the stable (-) and unstable (+) subspaces of the fixed point (x, 0), and they organize
the dynamics on the fast fiber F 0 (x). This is abundantly evident in (3.4), which describes the dynamics of an
arbitrary initial condition and only involves, one, the projection of that initial condition on those subspaces;
and two, the generators of the dynamics, Λ0± , on them.
Note, finally, that, for any fixed x, the `−dimensional fiber F 0 (x) is foliated by the `+ −dimensional family
0
{z +W− (x, 0)}z∈W+0 (x,0) of `− −dimensional manifolds. In plain words, this family consists of copies of the stable
manifold (i.e., of the y− −coordinate plane), cf. (3.5). The family has been parameterized by the basepoints z
3 Recall that σ(Λ0 ) ∪ σ(Λ0 ) = σ(Λ0 ) and that 0 6∈ σ(Λ0 ), since σ(Λ0 ) ∩ iR = ∅ by assumption. Hence, Λ0 have no zero
+ − ±
eigenvalues and can thus be inverted.
4 Recall the form assumed by the exponential of a Jordan block (without forgetting to account for secular terms).

26
of these copies located on the unstable manifold. None of these copies is independently invariant, except for
the one going through the fixed point at the origin. The family is, nevertheless, invariant as a whole in the
sense that each of its members is mapped to another member by the flow,

φ0τ z + W−0
(x, 0) = φ0τ (z) + W−
0


(x, 0), for all τ .

In that sense, the evolution of each such copy (fiber ) is dictated by the evolution of its basepoint on the unstable
0
manifold. Identical, mutatis mutandis, results hold for the `− −dimensional family {z + W+ (x, 0)}z∈W−0 (x,0) of
`+ −dimensional manifolds. The invariance condition for these two families, together with the observation that
a member of each family goes through any given point in z ∈ F 0 (x), makes precise the statement that the
dynamics are organized by the stable/unstable manifolds. Indeed, to flow z = (x, y− , y+ ) forward, one only
0 0
has to obtain its basepoints z− = (z, y− , 0) ∈ W− (x, 0) and z+ = (z, 0, y+ ) ∈ W+ (x, 0) by projecting z on each
manifold along the other (here: along the coordinate planes). Then,
   
φ0τ (z) = φ0τ (z− ) + W+ 0
(x, 0) ∩ φ0τ (z+ ) + W− 0
(x, 0) . (3.6)

Interpreting these conclusions in terms of the original w−coordinates is a straightforward exercise. The
subspace of fixed points, M0 , corresponds to U 0 . Associated with any point p ∈ U 0 is a fast fiber F 0 (p), which
is an affine copy of V 0 ; namely, F 0 (p) = p + V 0 := {p + v | v ∈ V 0 }. The dynamics on F 0 (p) are organized
around the stable and unstable (affine) subspaces of p; these are defined as the ω− and α−limit sets of that
0 0 0
point and are affine copies of V± ; in particular, W± (p) = p + V± . The phase space is foliated by the family of
n+`
fast fibers based on the subspace of fixed points, R = ∪p∈U 0 F 0 (p), so that any initial condition evolves on a
0
specific fiber according to the hyperbolic dynamics associated with (V+ 0
, Λ0+ ) and (V− , Λ0− ). Additionally, each
0
fiber F (p) is itself foliated by affine copies of the stable and unstable manifolds,
[   [  
F 0 (p) = q + V−0
= q + V+ 0
, (3.7)
0
q∈V+ 0
q∈V−

with each such collection forming an invariant family:

φ0τ q + V−0
= φ0τ (q) + V−
0
and φ0τ q + V+
0
= φ0τ (q) + V+
0



, for all τ . (3.8)

The flow in RN is obtained by that on the stable and unstable eigenspaces through the counterpart of (3.6),
   
φ0τ (q) = φ0τ (q− ) + V+
0
∩ φ0τ (q+ ) + V−
0
.

3.1.3 The ε > 0 dynamics

We now investigate how much of the structure presented above for the ε = 0 dynamics carries over to the far
more interesting ε > 0 scenario.
We begin by briefly characterizing σ(Aε ).5 The eigenvalues of A0 perturb as eigenvalues of Aε , under
the assumptions for A0 laid out in the previous section and as long as Aε − A0 =: εRε with Rε a bounded
perturbation. In concrete terms,

σ(Aε ) = {εµεi }1≤i≤k ∪ {λεj0 }1≤j 0 ≤m0 ,

where k ≤ n and m0 ≤ ` are numbers,6 the coefficients µε1 , . . . , µεk are O(1) (i.e., bounded) as ε ↓ 0 and, for
every j 0 < m0 , there exists j < m such that λεj0 → λ0j as ε ↓ 0. For all small enough values of ε, these two sets of
eigenvalues can be kept disjoint. At the same time, the ε = 0 eigenspaces also perturb to U ε , V− ε
and V+ε
with
dimensions equal to their ε = 0 counterparts. In particular, although the number of distinct eigenvalues with
positive (or negative) real parts may change due to branchings, the numbers n and `± are ε−independent, for
5 The interested reader can find much more information in [Stewart & Sun, Ch. IV].
6 Ingeneral, m0 ≥ m. This is so because an eigenvalue with algebraic multiplicity greater than one may branch off into distinct
eigenvalues, if its geometric multiplicity is strictly smaller than its algebraic one. Distinct eigenvalues, on the other hand, will
remain distinct for small enough ε.

27
ε within a neighborhood of zero. The reason for that is, essentially, that dimension is a discrete entity whereas
these subspaces change continuously—i.e., limit to their ε = 0 counterparts as ε ↓ 0.7
Working as in the previous section, we can introduce a matrix P ε = [U ε , V−ε , V+ε ], whose columns span
the three invariant eigenspaces introduced above, and use its inverse to define new coordinates z = (P ε )−1 w.
Choosing the columns of U ε to be generalized eigenvectors of Aε , we arrive at the system

x0 = εM ε x,
0
y− = Λε− y− , (3.9)
0
y+ = Λε+ y+ .

This is the ε > 0 analog of (3.3), with εM ε = V ⊥,ε Aε U ε the Jordan block associated with the small, O(ε)
eigenvalues. The flow generated by this linear system is
 ετ M ε 
e − x
ε
φετ (x, y) =  eτ Λ− y−  ; (3.10)
τ Λε+
e y+

this is the ε > 0, direct counterpart of (3.4). These dynamics are also easy to fathom and can be contrasted
directly to those of (3.3); we do this directly in terms of the original w−coordinates. First, the subspace
M0 = U 0 of fixed points perturbs to a nearby subspace Mε = U ε . The dynamics on that subspace are not
trivial (stationary) but, rather, dictated by the x−component of the flow φετ . Next, Mε is equipped with a
foliation of fast fibers, {F ε (p)}p∈Mε , with each fiber being an affine copy of the complementary eigenspace:
F ε (p) = p + V ε . The base point p ∈ Mε of each such fiber evolves in the way outlined above. As a result, the
fiber F ε is, in general, no longer invariant. The flow does, nevertheless, map each fiber to another one through
the rule
φετ (F ε (p)) = F ε (φετ (p)), (3.11)
as the reader can easily (and must) show. In words, the fiber based at p ∈ Mε , at time zero, is mapped after
time τ to the fiber based at φετ (p) ∈ Mε ; the slow dynamics on Mε entirely determines the evolution of the
entire fiber by acting on its basepoint. The fast dynamics transversal to Mε —that is, the precise point on
F ε (φετ (p)) that q ∈ F ε (p) is mapped to—are determined from the evolution of the y−components in (3.10).
Finally, note that each individual fiber F ε (p) is itself foliated by affine copies of the eigenspace V−ε (or V+ε ),
with each copy based on p + V+ε (or on p + V−ε ). These lower-dimensional subfibers are not individual either,
but their union over the entire family {F ε (p)}p∈Mε is: each member is mapped to another member by the
flow, much like our ε = 0 result (3.8).
Perhaps the most notable difference between the ε > 0 and ε = 0 cases concerns the loss of invariance of each
individual fiber. Although no isolated, invariant, ε = 0 fiber can be said to perturb as an independent entity—it
is not clear which member of {F ε (p)}p∈Mε corresponds to a given F 0 (p0 ), since p0 6∈ Mε in general—the family
{F 0 (p)}p∈M0 of fast fibers perturbs collectively as an invariant family. An immediate consequence of this loss
of invariance is that the definitions (3.5) now become inapplicable. In particular, one cannot demand that the
basepoint p of a fiber F ε (p) is the α− or ω−limit set of any point on that fiber, since p itself is not invariant.
ε ε
We can, nevertheless, define the affine subspaces W± (p) = p + V± and observe that they have the following
property: " #
∀q∈W±ε (p) lim ||φετ (q) − φετ (p)|| = 0 . (3.12)
τ →±∞

In other words, all points on the stable/unstable manifold of a point p ∈ Mε approach in forward/backward
time the forward/backward flow of that point. Note that this property does not suffice to recover the relation
W±ε
(p) = p + V±ε
. If, for example, the origin is a globally attracting fixed point for the system w0 = Aε w,
then—irrespectively of p—all points in Rn+` have the forward-contraction property since they limit—together
with p itself—to the origin in forward time.
7 The notion of a subspace limiting to another one requires equipping the set of subspaces with a specific topology. This can

be, for example, the metric topology associated with the Grassmannian metric, but we will refrain from elaborating further.

28
3.1.4 A rescale of time

We conclude our discussion of the linear problem w0 = Aε w by looking at another distinguished limit. We start
by rescaling time via t = ετ , which corresponds to a compression of time by a factor of ε: an O(1) interval of
˙ = d · /dt, we can
time τ corresponds to an O(ε) such interval in terms of t. Using this rescaling and writing (·)
recast (3.9) in the form
ẋ = M ε x,
εẏ− = Λε− y− , (3.13)
εẏ+ = Λε+ y+ ;
this corresponds to the ODE system εẇ = Aε w. We denote the flow generated by this system by ϕεt . Naturally,
for ε > 0, (3.9) and (3.13) produce flows related through ϕεt = φεt/ε . In phase space terms, the two flows
produce identical trajectories: the rescaling of time merely yields a reparameterization of those trajectories.
The situation is markedly different for ε = 0, since then (3.13) reduces to

ẋ = M 0 x,
0 = Λ0− y− , (3.14)
0 = Λ0+ y+ .

Using that the matrices Λ0± are invertible, we can rewrite this system as

ẋ = M 0 x and y = 0.

Clearly, this is a system of differential–alegbraic equations (DAEs) composed of an algebraic constraint (y = 0)

and an ODE (ẋ = M 0 x). The algebraic equation constrains the evolution on the ε = 0 slow subspace M0 ,
while the ODE equips this subspace with nontrivial dynamics. This situation is in start contrast with our
interpretation of the dynamics associated with (3.3), which included nontrivial dynamics outside M0 and
trivial (stationary) dynamics on it. Given the picture of the ε > 0 dynamics which we painted above, this
nontrivial flow on M0 can be geometrically interpreted as the ε ↓ 0 limit of the dynamics on Mε (recall that
M ε → M 0 , as ε ↓ 0). From the reverse point of view, the leading order, ε > 0 dynamics can be seen as the
concatenation of, first, the fast, hyperbolic dynamics along the fast fibers—associated with (3.3)—and, second,
the slow, drifting dynamics on M0 —associated with (3.14).
In either case, the fact that M0 appears fixed in the ε = 0 fast formulation makes sense as long as we
remember that ε measures the timescale disparity between fast (on F ε ) and slow (on Mε ) dynamics. In the
τ −timescale, the dynamics along the fast fibers evolve in an O(1) timescale, whereas those on the slow manifold
in a much longer O(1/ε) timescale. Hence, these slow dynamics appear entirely stationary, when ε is reduced
to zero. Similarly, in the t−timescale, it is the slow dynamics that evolve in an O(1) timescale, with their fast
counterparts evolving on an O(ε) timescale. Hence, these fast dynamics are instantaneously exhausted and the
evolution constrained on M0 already from the start.

3.2 Setting of the nonlinear problem

In this section, we concern ourselves with ODE systems of the form

x0 = εf (x, y; ε),
(3.15)
y0 = g(x, y; ε).

Here, the following quantities are at play:

• The independent variable τ ∈ R+ = [0, ∞), supposedly measuring time.

• The dependent variables x : R+ → Rn and y : R+ → R` , which are multidimensional functions of time
of dimension n and `, respectively. For the purposes of these notes, we will be assuming that the state
vector (x, y) is constrained to lie within an open set U ⊂ Rn+` . Typically, U will be forward invariant,
that is, initial conditions (x, y) ∈ U will generate trajectories {φετ (x, y) | τ ∈ R+ } ⊂ U .
• The parameter ε ∈ I = [0, ε̄], which is a small parameter (ε  1) measuring the timescale disparity
between the dynamics of x and those of y.

29
 • And finally, the functions f : U × I → Rn and g : U × I → R` , which are smooth functions of their
arguments (including in ε). Technically speaking, f, g ∈ C ∞ (U × I).

Note that this explicit fast–slow form is not the most general one can consider, since it assumes that the state
variables are already partitioned in fast (y) and slow (x) classes. This, for example, was not the case for our
linear system w0 = Aε w, where all state variables are generically fast.8 It is clearly the case for the transformed
system (3.13)—for which f and g do not even depend on y and x, respectively—and it would also be the case
had we defined new coordinates z̃ = (P 0 )−1 w (compare to z = (P ε )−1 w).
Here also, we can plainly rescale time by t = ετ to put (3.15) into the form

ẋ = f (x, y; ε),
(3.16)
εẏ = g(x, y; ε),

˙ = d · /dt. Clearly, (3.15) and (3.16) are equivalent as long as ε > 0, in the sense that they produce the
with (·)
same phase portrait. The sole effect of switching from τ to t is a reparameterization of the phase space trajec-
tories, that is, a rescaling of the speed with which these trajectories are traversed. In particular, trajectories
are traversed faster (by a factor of 1/ε) in the formulation (3.16)—the so-called slow formulation—compared
to (3.15)—the so-called fast formulation. We will see imminently that this equivalence between the two phase
portraits breaks down when ε = 0, just like it did for our linear system.

3.3 The ε = 0 dynamics

The fast and slow systems, (3.15) and (3.16) respectively, have two distinct distinguished limits for ε = 0:
 0   
x 0 ẋ = f0 (z), f0 (z) := f (x, y; 0),
z0 = = =: G0 (z) and with (3.17)
y g0 (z) 0 = g0 (z), g0 (z) := g(x, y; 0).

The ε = 0 limit of the fast system (left) exhibits a specific set of dynamical characteristics, which we summarize
below and demonstrate in Fig. [coming up]. The reader should make the correspondence between these and
analogous entities introduced in our linear context of Section 3.1.

• For each fixed value of x ∈ Rn , the `−dimensional, affine linear space F 0 (x) = {(x, y) | y ∈ R` } is
invariant; initial conditions on F 0 (x) remain on F 0 (x) for all times. That is so because x0 = 0, and each
F 0 (x) is called an (unperturbed) fast fiber.
• The fixed points of the system are determined by g0 (x, y) = 0. Since these are ` algebraic equations
in n + ` unknowns, they generically (but not automatically) define an n−dimensional set of solutions
M0 = {(x, y) | g0 (x, y) = 0} called the (unperturbed) slow manifold for our system. Below, we will be
wanting to formulate conditions ensuring that M0 is an n−dimensional manifold embedded in Rn+` .
• For every x ∈ Rn for which g0 (x, y) = 0 has at least one solution y, the point (x, y) ∈ M0 ∩ F 0 (x) is
a fixed point of the flow on F 0 (x). Below, we will be wanting to characterize the transversal dynamics
around that point; in particular, we will be assuming that these dynamics are hyperbolic when constrained
on the the invariant set F 0 (x).

The ε = 0 limit of the slow system (right), on the other hand, is a system of differential–alegbraic equations
(DAEs) exhibiting entirely different characteristics (see, here also, Fig. [coming up]).

• The flow is defined at most for points (x, y) satisfying the algebraic condition g0 (x, y) = 0. In other
words, the flow is certainly not defined outside the manifold M0 .
• As is evident from (3.17), the DAE in question only dictates the evolution of the x−variables. That of
the y−variables is dictated by the demand that M0 be invariant, that is, that the algebraic equation
g0 (x, y) = 0 be satisfied at all times.
8 To see this, work your way from (3.13) backwards to (3.1) by an application of the transform z 7→ w.

30
 The pictures sketched by these two distinguished limits appear complementary at best. The fast system
yields a foliation of the phase space into an `−dimensional family of n−dimensional fast fibers. Each fast fiber
is invariant and contains nonlinear, n−dimensional dynamics. Additionally, each such fiber contains at least
one fixed point—assumed hyperbolic—and the union of these fixed points yields an n−dimensional manifold
M0 .9 Plainly, the dynamics on M0 are trivial: all of its points are fixed under the flow. The slow system, on
the other hand, paints a picture where dynamics—typically nonlinear—only exist on M0 . Naively speaking,
one’s expectation is that the two pictures can be combined in one, as the case was for our linear problem.
Below, we work out the conditions under which this is true and explain in which sense this combination is
valid. Perhaps unsurprisingly, they all turn out to concern the local behavior of the system near M0 .

3.4 A fundamental hypothesis: normal hyperbolicity

To motivate our discussion below, we consider an arbitrary point z = (x, y) ∈ M0 and the linearized dynamics
around it in the ε = 0, fast formulation:
 ˙    
δx 0 0 δx ˙ = DG (z) δz.
= or, compactly, δz 0 (3.18)
δy Dx g0 (z) Dy g0 (z) δy
The eigenvalues of the Jacobian matrix DG0 (z) are rather easy to characterize by means of the following:

Statement #0. The spectrum of the Jacobian is given by

σ (DG0 (z)) = {0} ∪ σ (Dy g0 (z)) .
The zero eigenvalue is of algebraic and geometric multiplicities no smaller than n. 
This statement leads to the following fundamental hypothesis:

Normal Hyperbolicity Assumption. For every z ∈ M0 ,

0 6∈ σ (Dy g0 (z)) or, equivalently, det (Dy g0 (z)) 6= 0.

Under this assumption, we can refine the above statement to read:

Statement #1. The spectrum of the Jacobian is given by

σ (DG0 (z)) = {0} ∪ σ (Dy g0 (z)) .
Both the algebraic and geometric multiplicities of the zero eigenvalue are equal to n. The corresponding
invariant subspace is Tz M0 —the space tangent to M0 at z. The eigenvalues contained in σ (Dy g0 (z)) are all
associated with the the `−dimensional subspace {0} × R` , which DG0 (z) leaves invariant. 
You are asked to prove this statement in Homework #03. For the time being, note that this statement is the
precise analog of the statement, in our linear problem, that Ker(A0 ) is n−dimensional and that the eigenvalues
of the mapping induced by A0 on Rn+` /Ker(A0 ) are bounded away from the imaginary axis. Note, additionally,
that one could have guessed the result. Plainly, the zero eigenvalue is due to all points on M0 being fixed: this
information is picked up by the linearized dynamics, which dictates that perturbations ‘along M0 ’ (i.e., in the
tangent directions) stay put (i.e., do not evolve in time). Similarly, the remaining eigenvalues are associated
with the fast fiber F 0 (z) which remains invariant under the ε = 0 fast flow; indeed, F 0 (z) is but an affine copy
of the invariant subspace {0} × R` based at z. Just as g0 (x, ·) governs the full, nonlinear dynamics on that
fiber, so does Dy g0 (z) govern the linearized dynamics on the same fiber around the fixed point z.
The assumed property det(Dy g0 (z)) 6= 0 is called normal hyperbolicity, as it is equivalent to the demand
that z ∈ M0 is a hyperbolic fixed point for G0 |F 0 (z) —the restriction of G0 on F 0 (z).10 An interesting and
related corollary to that condition is the following:
9 This is purposefully vague. In reality, each fiber might contain more than one fixed points, in which case one will be forced to

choose which points to glue together to obtain a manifold.

10 Recall the definition of hyperbolicity: given a continuously differentiable vector field v : Rk → Rk (for some k) and a point

z ∈ Rk which remains fixed under v (i.e., v(z) = 0), we will be calling z hyperbolic if all eigenvalues of the Jacobian Dv(z) have
nonzero real parts (i.e., σ(Dv(z)) ∩ iR = ∅).

31
Statement #2. Let x ∈ Rn be fixed and assume that z = (x, y) ∈ M0 . Then, the point z is an isolated
solution of g0 (x, y) = 0. Additionally, for small enough ε, there exists a point (x, y(ε)) satisfying g(x, y(ε); ε) = 0
and y(ε) → y as ε ↓ 0. Finally, the ε = 0 slow manifold M0 and the ε = 0 slow fiber F 0 (x) intersect
transversally: Tz M0 ⊕ Tz F 0 (z) = Rn+` . 
The first part of the above statement is a direct consequence of the Implicit Function Theorem, the main
condition for the applicability of which is precisely what we called normal hyperbolicity. The second part is
the precise analog of the statement Rn+` = U 0 ⊕ V 0 for our linear problem; you are called to prove it in
Homework #03.

Connectivity and Compactness Assumptions. Before we proceed, we assume further that M0 is a

connected and compact manifold. Statement #2 dictates, then, that there exists a compact set K ⊂ Rn and a
unique function h0 : K → R` such that

M0 = graph(h0 ) = {(x, h0 (x)) | x ∈ K} .


0
The final corollary of our normal hyperbolicity assumption concerns the evolution, on M and for the ε = 0
slow system, of the y−variables.

Statement #3. The evolution, in the ε = 0 slow formulation, on the manifold M0 is dictated by the ODE
system
ẋ = f0 (x, y),
−1
ẏ = − (Dy g0 (x, y)) Dx g0 (x, y) f0 (x, y).

This can be proven by writing ϕ0t (x, y) 0
for the flow on M , substituting into the ε = 0 DAE (3.17), taking
the total time derivative of (both sides of) the algebraic constraint, and solving the resulting equation for y.
Note that the role of normal hyperboliticy is to ensure that (Dy g0 )−1 exists. An alternative way of describing
evolution on M0 remains, of course, (3.17) with the algebraic constraint solved for y,

ẋ = f0 (x, h0 (x)) and y = h0 (x).

3.5 Straightening it all out

Before elucidating further the ε = 0, fast dynamics in an O(1) neighborhood of M0 , we change coordinates
in (3.17) from (x, y) to a set which is dynamically more meaningful. We undertake this task in the form of a
roadmap to facilitate reading.

Straightening out M0 . Our first step is to coordinatize K × R` via

ẑ = (x̂, ŷ) = (x, y − h0 (x)). (3.19)

The introduction of these new coordinates flattens out the unperturbed slow manifold, in the sense that this
becomes M0 = {(x, 0) | x ∈ K} = K × {0}` ; it lies fully within the x̂−coordinate plane. The unperturbed fast
fibers, on the other hand, remain of the form x̂ = const.
The ODEs dictating the evolution in these new coordinates are obtained by differentiating both sides of
(3.19) and employing the corresponding ODEs (3.17) for the fast system. We find
 0  
x̂ 0
= =: Ĝ0 (ẑ). (3.20)
ŷ g0 (x̂, ŷ + h0 (x̂))

Note that M0 retains its normally hyperbolic character under this change of coordinates, since
   
0 0 0 0
   
σ DĜ0 (x, 0) = σ DG0 (x, h0 (x)) by virtue of DĜ0 = = .11
Dx g0 + Dy g0 Dh0 Dy g0 0 Dy g0
11 Here, we have used that Dx g0 + Dy g0 Dh0 = 0 on M0 , which is obtained by differentiating the identity g0 (x, h0 (x)) = 0.

32
Note, also, that the Jacobian is block-diagonal in these new coordinates, as one would expect from knowing
that M0 ⊂ Rn × {0}` and F 0 (x, 0) = (x, 0) + {0}n × R` . Below, we drop the hats in all quantities with a slight
abuse of notation.

Stable and unstable manifolds. Before we proceed further, we define the stable and unstable manifolds
of each point z ∈ M0 and then of the entire slow manifold M0 .
Given z ∈ M0 , the linearized dynamics around it and on F 0 (z) are governed by Dy g0 (z), the spectrum of
which only contains hyperbolic (i.e., non-imaginary) eigenvalues. In general, then, z is a hyperbolic saddle point
under the (nonlinear) dynamics generated by G0 |F 0 (z) , with `− stable and `+ = ` − `− unstable eigendirections.
0
We denote by W∓ (z) ⊂ F 0 (z) the (global) stable and unstable manifolds of z,
n  o   
0
(z) = z0 ∈ F 0 (z)  lim φ0t (z0 ) → z 0
(z) = z0 ∈ F 0 (z)  lim φ0t (z0 ) → z ,

W− and W+

t→∞ t→−∞

where {φ0t }t≥0 denotes the flow generated by the full, nonlinear vector field.
0 0
Note that z ∈ W− (z) ∩ W+ (z), since z is fixed under φ0t . Note, also, that the dimensions `± are necessarily
independent of z ∈ M . Hence, the stable/unstable manifolds are of the same dimension for all z ∈ M0 .
0 12

Using this information, we can define the (n + `± )−dimensional unstable/stable manifolds of M0 ,

[
0
W± (M0 ) = W±0
(z), whence also M0 ⊂ W− 0
(M0 ) ∩ W+ 0
(M0 ).
z

Finally, we glue (tangent) stable and unstable directions to define another pair of manifolds. Specifically, for
each z ∈ M0 , consider the affine subspaces z + Tz W± 0
(z); these are tangent to W±0
(z) at the point z, by the
stable/unstable manifold theorem, and vary smoothly with z. Glueing these subspaces together, we obtain two
other manifolds, [  
0
TW± (M0 ) = z + Tz W± 0
(z) .
z∈M0

At the next stop in our roadmap, we will devise a coordinate change to straighten these manifolds out.

0
Straightening out TW± (M0 ). Let z = (x, 0) ∈ M0 . Although the ε = 0 fast fiber F 0 (z) is itself straight-
13 0
ened out, this is not the case with its submanifolds W± (z) ⊂ F 0 (z). In this step, we change coordinates to
make the tangent spaces to these two manifolds (i.e., the stable/unstable directions) orthogonal to each other
and to M0 ; our work here is heavily indebted to the linear case considered earlier.
Since z is hyperbolic under the dynamics on F 0 (z), we have Tz F 0 (z) = Tz W−
0 0
(z) ⊕ Tz W+ (z). Let, now,
{v10 (z), . . . , v`0 (z)} 0
be a basis for Tz F (z) which respects this partition:
0
Tz W− (z) = span{v10 (z), . . . , v`0− (z)} and 0
Tz W+ (z) = span{v`0− +1 (z), . . . , v`0 (z)}.

Then, every point z ∈ RN can be written as

       
x x 0 x X X
z= = + = + ŷi vi0 (x, 0) + 0
ŷi+`− vi+` (x, 0),
y 0 y 0 −
1≤i≤`− 1≤i≤`+

since  
0
∈ Tz F 0 (z) = {0}n × R` .
y
This decomposition defines a change of coordinates z 7→ ẑ = (x̂, ŷ), with x̂ = x and ŷ = (ŷ− , ŷ+ ) defined
above. Plainly, y = 0 if and only if ŷ = 0, so M0 = {(x̂, 0) | x̂ ∈ K} in these coordinates as well. Similarly,
12 The proof is by contradiction: indeed, `
∓ corresponds to the number of stable or unstable eigenvalues. If these numbers were
to change as z varied, then at least one eigenvalue would have to cross from left to right half-plane or vice versa. Seeing as each
eigenvalue depends continuously on z, as the eigenvalue in question would need to cross through the imaginary axis and as M0 was
assumed compact, we conclude that said eigenvalue would become imaginary at a specific value z ∈ M0 . This is a clear violation
of normal hyperbolicity, whence the contradiction.
13 It is identified with an affine copy of the y−coordinate plane, F 0 (z) = z + {0}n × R` .

33
F 0 (ẑ) = ẑ +{0}n ×R` here as well. Embedded in this fast fiber are, now, the two straightened out submanifolds
0
TW− (M0 ) = K × R`− × {0}`+ and 0
TW+ (M0 ) = K × {0}`− × R`+ (3.21)
coinciding with the ŷ− − and ŷ+ −coordinate planes.
We conclude this step by deriving the governing ODEs in terms of the new coordinate system. Let ej =
(δij )1≤i≤N be the j−th standard basis column vector (1 ≤ j ≤ n) and write
 
x
= P 0 (x, 0) ẑ, with P 0 (x, 0) = e1 , . . . , en , v10 (x, 0), . . . , v`0 (x, 0) = E, V−0 (x, 0), V+0 (x, 0) .
   
z=
y

Then, we calculate

0
0
z 0 = P 0 (x, 0) ẑ 0 + P 0 (x, 0) ẑ = P 0 (x, 0) ẑ 0 , since P 0 (x, 0) = Dx P 0 (x, 0) x0 = 0;

recall (3.20).14 It follows that

−1 0
−1
−1
ẑ 0 = P 0 (x, 0) z = P 0 (x, 0) G0 (z) = P 0 (x, 0) G0 P 0 (x, 0) ẑ =: Ĝ0 (ẑ).

The linearization of this vector field about any point ẑ ∈ M0 may be calculated to be
 
0 0 0
DĜ0 (x, ŷ) =  0 Λ̂0− (x) 0 , (3.22)
0
0 0 Λ̂+ (x)

reflecting the invariance both of M0 and of the coordinate subspaces Rn × R`− × {0}`+ and Rn × {0}`− × R`+
(cf. (3.21)) under the linearized dynamics. Here, the spectrum of the `± × `± blocks Λ̂0± (x) matches the
unstable/stable eigenvalues of Dg0 (x, 0). These blocks can naturally be put in Jordan canonical form by
selecting {v10 (z), . . . , v`0 (z)} to be generalized eigenvectors of Dg0 (x, 0).
Here also, we shall drop hats in what follows.

0
Straightening out W± (M0 ). We are now ready to introduce new coordinates to straighten out the nonlinear
0 0 0
manifolds W± (M ). In the vein of our work for TW± (M0 ) above, we focus on straightening out W± 0
(z) for
0 0 0
any specific z = (x, 0) ∈ M , since W± (M ) are made by glueing such manifolds together.
0
To do this, we first note that the manifolds W± (z) go through z and are tangent to Rn × {0}`− × R`+ and
n `− `+
R × R × {0} , respectively; cf. (3.21). By the stable/unstable manifold theorem, further, these manifolds
are at least C 1 and hence they are graphs over their tangent spaces in a neighborhood of z. It follows that there
exist open sets U± (z) ⊂ R`± containing z and functions η± 0
(·; z) : U± (z) → R`∓ such that
  
0 0 0
W− (z) = graph η− (·; z) = (x, y− , η− (y− ; z)) | y− ∈ U− (z) ,
  
0 0 0
W+ (z) = graph η+ (·; z) = (x, η+ (y+ ; z), y+ ) | y+ ∈ U+ (z) .

0
The functions η± (·; z) satisfy, additionally,
0 0
η± (0; z) = 0 and Dη± (0; z) = 0.
0
The first of these expresses that z ∈ W± (z) and the second the aforementioned tangencies. The change of
coordinates now suggests itself,
0 0
ẑ = (x̂, ŷ− , ŷ+ ) = (x, y− − η+ (y+ ; x, y− , y+ ), y+ − η− (y− ; x, y− , y+ )),

compare to (3.19). Under this—final—change of coordinates, we can straighten out the stable/unstable mani-
folds in a tubular neighborhood of M0 . Indeed, write U± = ∩z∈M0 U± (z) for a neighborhood of z ∈ M0 over
0
which η± (·; z) is defined. Then,

M0 = K × {0}`− × {0}`+ , 0
W− (M0 ) = {0}n × U− × {0}`+ 0
and W+ (M0 ) = {0}n × {0}`− × U+ . (3.23)
14 Intuitively, dP 0 /dt = 0 follows from P 0 being a (matrix) function on M0 which, in turn, is fixed under the flow. Since points

on M0 are fixed, so are functions of these points.

34
Fibration of a neighborhood of M0 . We finally mention, without delving into lengthy details, that there
exists an O(1) neighborhood of M0 which we can equip with a skeleton structure for the ε = 0 dynamics.
This is done the vein of the fibration (3.7) for our linear system. Here, nevertheless, translated copies of
0
the stable/unstable manifolds W± (z) will not suffice to foliate F 0 (z), since they (in general) do not form an
invariant family. Our main tool is, instead, the Hartman–Grobman theorem, which establishes the existence
of a diffeomorphism χz : U ∩ F 0 (z) → V between an `−dimensional neighborhood of z lying on the fiber and
a neighborhood V ⊂ R` of the origin. This diffeomorphism maps the linear flow {δφ0τ }τ ≥0 to its nonlinear
counterpart {φ0τ }τ ≥0 . In particular, evolving a perturbation δz ∈ R` under the linear flow and then mapping
it to F 0 (z) using χz produces the same result as first mapping it to F 0 (z) using χz and then evolving it using
the nonlinear flow on that fiber:

∀δz∈V ∀t≥0 φ0τ (χz (δz)) = χz (δφ0τ (δz)) , that is ∀τ ≥0 φ0τ ◦ χz = χz ◦ δφ0τ .
   

Under this diffeomorphism, the origin of the linear system is mapped to the fixed point z of the nonlinear
0
one, the stable/unstable eigenspaces are mapped to the stable/unstable manifolds W± (z) and the invariant
0
fibrations (3.7) in V are mapped to invariant fibrations in U ∩ F (z). It is this invariant fibration for the
nonlinear system that organizes the dynamics in a neighborhood of M0 .

3.6 A skeleton structure for the ε > 0 dynamics

In this section, we list the persistence results of GSPT for ε in a neighborhood of zero and refer the reader to
the references listed above for more information on the subject.

Persistence of M0 . For ε > 0 in a neighborhood of zero, the manifold M0 perturbs to a manifold Mε of

the same dimension. This perturbed manifold is O(ε)−close to M0 , diffeomorphic to it, locally invariant, and
the graph of a smooth function hε : K → R` .
0
Fenichel theory, in fact, establishes the persistence of the stable/unstable manifolds W± (M0 ); the slow
ε
manifold M is obtained as their intersection.

0
Persistence of W± (M0 ). For ε > 0 in a neighborhood of zero, the manifolds W 0 (M0 ) perturb to manifolds
0 ε
W (M ) of the same dimension. These perturbed manifold are O(ε)−close to their ε = 0 counterparts, diffeo-
ε
morphic to them, locally invariant, and the graphs of smooth functions η± over subsets of the stable/unstable
eigenspaces.
In fact, GSPT also characterizes the flow on these perturbed manifolds by means of establishing expo-
nential contraction estimates (in forward/backward time, respectively, and towards Mε ); see [Jones, pg. 71].
ε
Additionally, it establishes perturbed fibrations of W± (Mε ) by means of n−dimensional families of `± −fibers.

0
Persistence of {W± (z)}z∈M0 . Write z ε = (x, hε (x)) ∈ Mε and z 0 = (x, h0 (x)) ∈ M0 . Then, for ε > 0 in a
0
neighborhood of zero, the fiber W± (z 0 ) perturbs to a fiber W±
ε
(z ε ) of the same dimension. This perturbed fiber
ε
is O(ε)−close to its ε = 0 counterparts and diffeomorphic to it. Each one of the two families {W± (z ε )}zε ∈Mε
is invariant as a family in the sense described earlier.

35
Lecture no.4 February 26, 2013

4.1 Relaxation oscillations in a predator–prey model

As a first application of GSPT, we will consider the appearance of relaxation oscillations in a rather generic
predator–prey model.
We start by formulating the model. Letting x and y denote the populations of predators and prey, respec-
tively, we write
ẋ = F (x, y; ε),
(4.1)
εẏ = G(x, y; ε);
this is the slow formulation of our slow–fast, predator–prey system. A rescale of time via τ = t/ε puts the
system into the fast formulation,
x0 = εF (x, y; ε),
(4.2)
y 0 = G(x, y; ε),
where (·)0 = d · /dτ . The region of interest is the (closed) quadrant R̄+ × R̄+ = {(x, y)|(x ≥ 0) ∧ (y ≥ 0)}, as
negative populations are ecologically meaningless. In fact, we shall remain interested in a compact rectangle
[0, X] × [0, Y ] which, under suitable conditions, will need to be forward invariant: any initial condition (x, y) ∈
[0, X] × [0, Y ] will need to yield a solution {φετ (x, y)}τ ≥0 ⊂ [0, X] × [0, Y ]. Here, ε ∈ [0, ε̄] is a small parameter
(ε  1), dictating that the prey population evolves in a much faster timescale than the predator population.1
Also, F and G are smooth functions, all derivatives of which remain bounded (i.e., are O(1) as ε ↓ 0) over
[0, X] × [0, Y ] × [0, ε̄]. Smoothness and time-invariance constitute our zeroth assumption on the nature of the
model.
Our first assumption is that ‘there is no migration in the ecosystem,’ and hence a zero initial population of
either prey or predator will remain zero at all times. Mathematically, this translates into (forward) invariance
of the x− and y−axes, and in particular into the condition

F (0, y; ε) = 0, for all (y, ε) ∈ [0, Y ] × [0, ε̄], and G(x, 0; ε) = 0, for all (x, ε) ∈ [0, X] × [0, ε̄].

It follows from this assumption that the origin is necessarily a fixed point (also called steady state or equilibrium),
F (0, 0; ε) = G(0, 0; ε) = 0. Below, we use our ecological intuition and basic mathematical arguments to build
up the dynamics of our system step-by-step. Then, following [Hek], we show that systems of this sort can be
expected to exhibit relaxation oscillations (also called slow–fast oscillations).

4.1.1 Detailed modeling

Step 1: directionality of the x−component of the dynamics. First, the populations of predators
should diminish to zero, should prey become extinct (i.e., when y ≡ 0).2 Therefore, F (x, 0; ε) < 0 for all x > 0
(cf. (4.1)). It now follows by continuity that

F (x, y; ε) < 0, for all 0 ≤ y < yε∗ (x); in particular, F (x, yε∗ (x); ε) = 0.

Here, yε∗ : [0, X] → R is some positive—possible infinite—function. We now formulate explicitly our second
assumption: ‘yε∗ (x) = sup{y | F (x, y; ε) < 0} is finite for all (x, ε) ∈ [0, X] × [0, ε̄],’ and hence

F (x, y; ε) < 0, for all 0 ≤ y < yε∗ (x); F (x, yε∗ (x); ε) = 0; and F (x, y; ε) > 0, for all y > yε∗ (x).

This condition plainly formalizes the ecologically sensible condition that predator populations tend to decrease,
for prey populations under a certain threshold, and tend to increase for prey populations over that threshold.3
Note, also, that it appears sensible to incorporate into our second assumption the requirement that yε∗ is non-
decreasing. Indeed, the threshold prey population necessary for predators to thrive should not decrease with
the predator population: the more the predators, the more prey they should require to be sustenance.
1 This happens if the prey adjusts to environmental conditions much faster than the predators, either because it responds much

more quickly to changes in the available resources or because the predators can harvest prey at a very large rate.
2 Built into this argument is the assumption that there are no alternative (called substitutable) food sources for the predators.
3 Note that this behavior is not a necessity, if the response of predators to prey becomes saturated, since then the predator

death rate may be uniformly larger than the reproduction rate, so that yε∗ = ∞. In that case, F < 0 throughout [0, X] × [0, Y ]
and hence all predators are destined to die out. This case results into trivial dynamics, so we do not consider it here.

36
 We further assume that, for small predator populations and in the absence of prey, the decay rate of the
population depends linearly on the population itself: δx ˙ = α δx, for some α < 0 and all small δx. Indeed,
small populations lead to rare, negligible interactions between the individuals, and hence predator decay can
be modeled as a Poisson process; coarse graining leads to the linear ODE above. Since (4.1) and F (0, 0; ε) yield
that α = ∂x F (0, 0; ε), we impose the condition ∂x F (0, 0; ε) < 0.4 It now follows by continuity that, for every
ε ∈ [0, ε̄], there exists a relatively open set Dε ⊂ R2 containing the origin such that
∂x F (x, y; ε) < 0, for (x, y) ∈ Dε . (4.3)
This implies, in particular, that yε∗ (0) > 0. Since yε∗ is non-decreasing, then, yε∗ > 0 uniformly over [0, X].
These considerations fix the directionality of the x−component of the vector field throughout [0, X] × [0, Y ],
as well as its x−nullcline, namely, graph(yε∗ ); see top-left panel of Fig. 7.

Step 2: the y−component of the dynamics. We now formulate our third assumption: ‘the life-sustaining
resources available to prey are present but limited ’. The first consequence of this is that G(0, y; ε) > 0, for y
in a right-neighborhood of zero; ecologically speaking, small prey populations should flourish in the absence of
predators and presence of resources. It seems reasonable, further, to incorporate here the condition that the
per capita growth rate of prey in the absence of predators, G(0, y; ε)/y, decreases with y: the larger the prey
population, the fewer the resources available to each individual.5 In fact, we will assume that the derivative of
G(0, y; ε)/y is negative and bounded away from zero; factoring in the positivity of G(0, y; ε) close to zero, one
can show that there exists a unique value ȳε (0) > 0 such that G(0, ȳε (0); ε) = 0. Therefore, (0, ȳε (0)) is an
attracting fixed point for the predator-less model:
G(0, y; ε) > 0, for 0 < y < ȳε (0); G(0, ȳε (0); ε) = 0; and G(0, y; ε) < 0, for y > ȳε (0). (4.4)
(See also top-right panel of Fig. 7.) By assumption, G(0, ·; ε) intersects zero transversally at ȳε (0) and hence
∂y G(0, ȳε (0); ε) < 0. Consequently, prey populations approach the carrying capacity ȳε (0), corresponding to
the prey population that the available, limited resources can sustain in the absence of predators, exponentially
in time. It also now automatically follows that the origin is unstable.
We finally assume that, in the absence of predators, small prey populations follow Malthusian dynamics:
˙ = βδy, for some β > 0 and small δy. Here again, (4.1) and G(0, 0; ε) = 0 yield that β = ε−1 ∂y G(0, 0; ε),
δy
and hence β = ∂y G(0, 0; ε) > 0.6 It now follows by continuity that, for every ε ∈ [0, ε̄] there exists a relatively
open set Eε ⊂ R2 containing the origin such that
∂y G(x, y; ε) > 0, for (x, y) ∈ Eε . (4.5)
A corollary of paramount importance to this statement is that ∂y G(x, 0; ε) > 0, for x in some interval [0, x∗ε );
that is, the (invariant) x−axis is normally repelling in a neighborhood (possibly unbounded) of the origin.
Here, we will assume that x∗ε < X, and in particular that x∗ε is finite.7
These results determine the direction of the y−component of the vector field on the y−axis, as well as in
an O(1) neighborhood of [0, x∗ε ] × {0}. By assuming that ∂y G(x, 0; ε) < 0, for x > x∗ε , we fix the flow in a
neighborhood of the entire interval [0, X] on the x−axis. Note that this last assumption on ∂y G(x, 0; ε) is quite
natural, as is the stronger one that ∂y G(x, 0; ε) is decreasing. Indeed, the larger the predator population small
prey populations are exposed to, the less rapidly they should grow—if at all.

Step 3: upper branch of the slow manifold M0 . We assumed above that ∂y G(0, ȳε (0); ε) < 0, whence
also ∂y G(0, ȳ0 (0); 0) < 0, and remarked that this assumption fits in well with the monotonicity of G(0, y; ε)/y.
It follows by the Implicit Function Theorem that the equation G(x, y; 0) = 0 can be solved to yield y = ȳ0 (x),
for all x ∈ [0, xf ] and some xf > 0 (possibly infinite). Additionally (and by continuity),
∂y G (x, ȳ0 (x); 0) < 0, for all x ∈ [0, xf ), while ∂y G (xf , yf ; 0) = 0 (where we write yf := ȳ0 (xf )).
4 The reader should be able to show that this agrees with—but does not follow from—the condition imposed above: that
F (x, 0; ε) < 0, for x > 0.
5 The minimal such model models G(0, y; ε)/y linearly: G(0, y; ε) = y(c − y), for some constant c > 0—see [Hek, Example 2.1].

This is called the logistic model.

6 Here also, the reader can show that this condition agrees (but cannot be derived from) the condition G(0, y; ε) > 0, for y in a

neighborhood of zero.
7 This amounts to assuming that a predator population kept at a constant, high enough value can lead prey to total extinction.

37
This result implies that the fast formulation (4.2) possesses, for ε = 0, an entire branch M̄0 of normally
attracting fixed points: namely, the graph of the function ȳ0 , see bottom-left panel of Fig. 7. The value ȳ0 (x)
can be thought of as the carrying capacity for a constant predator population x. Since predators harvest on
prey, it is sensible to assume that this carrying capacity decreases with x, i.e., that ȳ0 is a decreasing function.
Our fourth assumption is that 0 < xf < X; that 0 < y0∗ (xf ) < yf ; that G(x, y; 0) = 0 has no solutions
for x > xf ; and that, along every horizontal line y = const. = c ∈ [yf , ȳ0 (0)], the solution x = ȳ0−1 (c) of
G(x, c; 0) = 0 is unique. We include in this assumption a final demand—namely, that ∂x G(xf , yf ; 0) 6= 0, as a
consequence of which the branch of fixed points we identified above must fold at xf , i.e., ȳ00 (xf ) = −∞ [why?].
The first part of our assumption (involving xf ), as well as the second and the last (the branch folds, and does
so before hitting either zero or the x−nullcline) are rather arbitrary and serve the purpose of ‘spicing up’ the
dynamics; hence, they should be interpreted as focusing into a particular parameter regime. In particular, an
intersection of the x−nullcline with this upper branch of M0 would lead to a (globally) stable equilibrium,
which we want to avoid. Further, the assumption that M0 folds will yield the relaxation oscillation we are
after. The third part, on the other hand, is rather sensible; in fact, one expects that G(·, y; 0) is a decreasing
function for every fixed value of y, since the rate at which any given prey population y grows should decrease
with the predator population x to which it is exposed. Should this be true, the third-through-fifth parts of our
assumption follow immediately by virtue of G(x, y; 0) > 0 (for 0 < y < ȳ0 (0)) and as long as ∂x G(·, y; 0) < 0
remains bounded away from zero over [0, X] (and for each y ∈ [0, Y ]).
As a result of the above considerations, we conclude that

G(x, y; 0) > 0 in (x, y)| [yf < y < ȳ0 (0)] ∧ 0 < x < ȳ0−1 (y) ,
  
(4.6)
G(x, y; 0) < 0 in (x, y)| [yf < y < ȳ0 (0)] ∧ ȳ0−1 (y) < x < X .
  
(4.7)

The separatrix between these two regions is the upper branch M̄0 of the slow manifold. This characterization of
the sign of G fixes the directionality of the y−component of the vector field in the rectangle [0, X] × [yf , ȳ0 (0)].
Also, the characterization of the sign of F we offered above fixes the flow (namely, towards increasing x) on
this branch of the slow manifold.

Step 4: lower branch of the slow manifold M0 . Next, we show that there exists a second, lower branch
M0 of M0 which connects (xf , yf ) to (x∗0 , 0). First, (4.4) yields that G(0, ·; 0) > 0 over (0, ȳ0 (0)). At the same
time, (4.7) yields that G(x, ·; 0) < 0 over (yf , ȳ0 (0)) ⊂ (0, ȳ0 (0)), for any xf < x < X. Since, by assumption,
G(x, ·; 0) has no zeros for x > xf , it follows that G(x, ·; 0) > 0 over the entire domain (xf , X) × (0, Y ). By
continuity, along each line y = const. = c ∈ (0, yf ), there exists a value x0 (c) such that G(x0 (c), c; 0) = 0. Each
such point (x0 (c), c) is a point on the lower branch M0 of the slow manifold M0 , see bottom-right panel of
Fig. 7.
As part of our fifth assumption, we impose that the value x0 (c) is unique—recall also our reasoning for the
analogous assumption imposed on the upper branch M̄0 . Further, we assume that x0 is an increasing function
and ∂y G < 0 on it. Under this last assumption, x0 is invertible and y 0 := x−1 0 is an increasing function
defined in a left-neighborhood of xf , the graph of which is M0 . To shed more light into the significance of
this fifth assumption, we remark that G(x, y 0 ; 0) = 0—the vertical component of the vector field vanishes
on M0 —and G changes sign across it (by virtue of ∂y G 6= 0, which means that G crosses zero transversally
across M0 ). It follows that, for any fixed predator population xc smaller than and sufficiently close to xf ,
initial prey populations larger than y 0 (xc ) grow towards the stable steady state ȳ0 (xc ), while initial prey
populations smaller than y 0 (xc ) decay to zero (in particular, M0 is normally repelling). Hence, the value y 0 (xc )
forms a threshold: higher initial populations limit up to the nontrivial steady state; lower initial populations
vanish. The assumption that x0 —and hence also y 0 —is increasing means that this threshold grows with the
predator population—which is ecologically sound, since it should take higher prey populations to overcome
higher predator populations. In particular, our monotonicity assumption excludes a bistable scenario, in which
M0 would fold again and yield another, lower branch of stable steady states. Since, in general, bistable scenarios
cannot be excluded, this choice should again be viewed as focusing into a specific parameter regime.
It remains to determine the left-neighborhood of xf over which y 0 is defined, that is, its left endpoint x0 (0).
As we showed in Step 2, the x−axis is normally repelling in a neighborhood (0, x∗0 ); since G(x, y; 0) < 0 for
0 < y < y 0 (x), the x−axis is normally attracting for all values of x in the domain of definition of y 0 . As a
result, this domain of definition cannot extend below (i.e., to the left of) x∗0 , that is, x0 (0) ≥ x∗0 . Further,
assuming that M0 intersects the x−axis transversally at (x0 (0), 0)—equivalently, that y 00 (x0 (0)) > 0—we find

38
that both ∂x G and ∂y G are zero at that point [why?]. Since x∗0 is the unique zero of ∂y G(x, 0; 0) by assumption,
it follows that y 0 is defined exactly over [x∗0 , xf ]; consequently, y 0 (x∗0 ) = 0 and the lower branch M0 ends up
at the point (x∗0 , 0). Note that, necessarily then, x∗0 < xf .

Step 5: a nontrivial unstable equilibrium. Finally, we remark that the system necessarily possesses a
nontrivial equilibrium (x0 , y0 ) ∈ M0 , as y 0 (x∗0 ) = 0 < y0∗ (x∗0 ) and y 0 (xf ) = yf > y0∗ (x∗0 ). This equilibrium is
also necessarily unstable, since M0 is normally repelling. Additionally, the flow on M0 is towards decreasing
x to the left of the equilibrium and towards increasing x to its right–this fixes the flow on the rest of the slow
manifold.

Remark. Note that (4.3) and (4.5) also imply [why?] that (4.1) can be rewritten as

ẋ = xf (x, y),
(4.8)
εẏ = yg(x, y).

Here, the functions f and g satisfy the same smoothness assumptions as F and G. Additionally,

f (0, y; ε) < 0, for y ∈ [0, yε∗ ), and g(x, 0; ε) > 0, for x ∈ [0, x∗ε ).

By continuity, and for any 0 ≤ ε ≤ ε̄, there exist sets Dε and Eε such that

f (0, y; ε) < 0, for (x, y) ∈ Dε , and g(x, 0; ε) > 0, for (x, y) ∈ Eε .

The signs of f and g can be deduced by our analysis above [do it!].

4.1.2 The relaxation limit cycle

Having constructed our rather generic predator-prey, fast–slow model, we now turn to identifying a limit cycle
in it. This limit cycle is schematically constructed (to leading order) in Fig. 8; as can be seen from that sketch,
it consists of alternating fast and slow components. Fenichel theory guarantees that all three branches of the
slow manifold—M̄0 , M0 , and any compact subset of the positive x−axis perturb for sufficiently small values
of ε > 0, since they are normally hyperbolic. Note that this includes neither the fold point (x̄, ȳ(x̄)) nor the
point of intersection (x∗0 , 0), as M0 loses normal hyperbolicity at both of these points. In particular, one has
to exclude O(1)−neighborhoods around each one of these points to assure persistence of the slow manifold
branches.
The perturbed branches have the same stability properties as their ε = 0 counterparts and the evolution
on them is governed, to leading order, by the dynamics of the ε = 0 slow system. Hence, for any sufficiently
small ε > 0, any fixed initial condition ε−close to M̄0 generates a trajectory which stays ε−close to it (by
invariance and normal attractivity of M̄ε which is ε−close to M̄0 ). Since M̄ε is ε−close to M̄0 , the leading
order component of the trajectory is y0 (·) = ȳ0 (x0 (·)); here, the evolution of x0 —the leading order component
of x—is determined by the system

ẋ0 (t) = F (x0 (t), ȳ0 (x0 (t)); 0) and y0 (t) = ȳ0 (x0 (t)),

up to the point t̄ where x(t̄) = xf . (A leading order formula for tf may be derived from the equation
x0 (t̄) = xf .) Following that time instant, the trajectory is captured by the fast dynamics and drawn—to
leading order vertically—to the x−axis according to the system

x0 (τ ) = xf and y00 (τ ) = G(xf , y0 (τ ); 0), where τ = t/ε.

Following this fast descent to the point (xf , 0) (always to leading order), the trajectory is trapped in an O(ε)
neighborhood of the x−axis—note that this axis remains invariant for ε > 0 and thus represents the lowest
branch of Mε also for ε > 0. The trajectory proceeds towards decreasing values of x following, to leading
order, the dynamics
ẋ0 (t) = F (x0 (t), 0; 0) and y0 (t) = 0. (4.9)
We remark here that the next-order term of the y−component is attracted towards zero according to the
linearized normal dynamics around the x−axis, and thus the trajectory is attracted to that axis up to within

39
an ε−neighborhood of x∗0 and repelled afterwards. At a certain point xl < x∗0 , the trajectory ‘lifts off’ Mε and
follows the fast dynamics,
x0 (τ ) = xl and y00 (τ ) = G(xl , y0 (τ ); 0),
towards the point xl , ȳ0 (xl ). It then follows the slow dynamics towards xf as described in the beginning of this
paragraph.
Naturally, a more careful analysis needs to work locally in neighborhoods around the fold and intersection
points, where Mε is not defined, and establish that the trajectory indeed enters and leaves those neighborhoods.
Additionally, the leading order solutions obtained for the four different parts must be matched properly. Note,
also, that the initial condition for the limit cycle must lie below Mε [why?] and that a proper proof of
its existence can be given with the help of a Poincaré section transversal to Mε . Nevertheless, our—very
tractable—analysis here provides a feeling for the location of (and the dynamics on) this limit cycle, up even
to the location of the lift-off point xl .
To calculate the location of that lift-off point, let us define ‘touch-down’ and ‘lift-off’ as the moments tf
and tl that the trajectory enters and leaves the rectangle [xf , xl ] × [0, yl ], respectively. Here, yl = εYl and Yl is
a fixed value. Now, the first leg of (4.9) implies that x is a monotone function of t (recall that F < 0 on the
x−axis). Hence, it can be used to reparameterize the solution (x(·), y(·)) in the time interval [tf , tl ] according
to the ODE
dy0 (x0 ) y0 (x0 )Gy (x0 , 0; 0)
= ,
dx0 εF (x0 , 0; 0)
where we have Taylor-expanded the right member with respect to y, as y = O(ε) throughout the rectangle
[xf , xl ] × [0, yl ], and recalled that G(x, 0; 0) = 0 for every x ≥ 0. (Note the slight abuse in notation involved in
writing y0 (x0 ) instead of y0 (t(x0 )).) Integrating from x0 = xf to xl , we find then an algebraic equation for xl
with a unique [why?] solution,
h iyl ˆ xl G (x, 0; 0)
y
0 = ε ln |y0 | = dx. (4.10)
yl xf F (x, 0; 0)

Discussion. In the above 2−D example, we used phase plane analysis together with GSPT to (formally)
construct a relaxation limit cycle. A large part of the construction can be made rigorous using persistence
results from GSPT—notably, those for the persistence of a slow manifold, of its stable and unstable manifolds,
and of a fast fibration of each one of these manifolds. The flow around the (fold and intersection) points where
GSPT breaks down, on the other hand, requires further local analysis.8
We finally remark that the relaxation oscillation that we constructed here exhibits catastrophic properties
in that the prey population varies abruptly when the predator population reaches the critical values xf and
xl (to leading order). Naturally, this is far from surprising given the configuration of the phase space, which
exhibits a fold—the hallmark of many catastrophic events. The reader should note, in particular, that the
ε = 0 phase space for the fast system is, in fact, the Cartesian product of a 1−D phase space with R̄+ (the
space in which x takes values). Seen in that light, x acts as a parameter and the (x−dependent) phase line
undergoes a saddle–node (fold) bifurcation at xf , as long as f is quadratic (which it generically is). (This also
casts a different light on the condition ∂x G(xf , yf ; 0) 6= 0, as this becomes one of the two standard conditions
associated with this type of bifurcation.)
As mathematically unsurprising as this behavior is, it has severe ecological significance because it showcases
an all too sudden disappearance of a prey species. Indeed, one would have a hard time predicting this phe-
nomenon using a timeseries {y(tn )}1≤n≤N for the prey population, as the observed behavior is most certainly
not analytic and hence cannot be captured by ordinary data fitting. Additionally, the lift-off at xl exhibits
all the properties pertaining to a delayed bifurcation. Note, in particular, that rebuilding a substantial prey
population requires a predator population much smaller than (the fold point) xf where the catastrophe occurs:
for any value x ∈ (x∗ε , xf ) of the population, the stable prey population corresponding to the upper branch M̄0
of the slow manifold remains unattainable, as it is separated from the x−axis by the unstable branch. As we
explained above (cf. (4.10)), xl is separated from x∗0 —where the x−axis reverses its normal stability type—by
8 The following is an indirect indication of the fact that, for ε > 0, no counterpart of M0 exists around the fold point to guide
the dynamics: first, recall that the part of the trajectory lying ε−close to M̄0 lies below it. Now, should M0 perturb close to the
fold point, in the sense of an invariant manifold joining the perturbed counterparts of M̄0 and M0 , the limit cycle would need
to cross it in order to approach the x−axis. This cannot occur, as it would violate uniqueness of solutions for our system, which
offers us the desired contradiction.

40
 y y
Y Y

y (0)
0
F > 0
y = y*(x) x*
0 0

F < 0
x x
X X
y y
Y Y
G < 0 G < 0

M0 G > 0
xf (x , y )
0 0

M0
x x
X X

Figure 7: Construction of the ε = 0 skeleton structure for the predator–prey model (4.1)–(4.2). Top-left panel :
Graph of the curve y = y0∗ (x) corresponding to the (ε = 0) F −nullcline. Top-right panel : The value x∗0 ,
where the normal stability type of the x−axis switches from stable (x > x∗0 ) to unstable (x < x∗0 ), and the
(ε = 0) carrying capacity (in the absence of predators) ȳ0 (0) found by solving the equation G(0, y; 0) = 0.
Bottom-left panel : Continuation of the carrying capacity ȳ0 (0) to fixed, nonzero predator populations to yield
the upper branch M̄0 of the manifold M0 , together with the fast dynamics in a neighborhood of that branch.
Additionally, the value xf a which this branch folds. Bottom-right panel : The lower branch M0 of the manifold
M0 , together with the full fast dynamics. Additionally, the second, also unstable fixed point (x0 , y0 ) lying on
that branch.

an O(1) distance which is necessary to allow the trajectory to ‘peel off’ after it spent the interval (x∗ε , xf )
getting ‘compressed’ towards the x−axis. Most interesting of all, the population remains close to the unstable
x−axis past x∗ε where the stability changes and although the only stable prey populations are those found on
the upper branch of M0 . In short, this means that the predator population must be reduced way below the
level xf at which the catastrophe occurred for the prey population to be built up again.

4.2 Relaxation oscillations in Rayleigh’s equation

As a second example of relaxation oscillations, we consider the well-known and studied Rayleigh’s equation,
 
1 2
εÿ + ẏ − 1 ẏ + y = 0, (4.11)
3
which can be thought of as modeling a linear spring with nonlinear friction/excitation (compare to our linear
model in Sect. 2.2) and is the prototypical
√ example of a Liénard equation.
√ In particular, the nonlinearity
corresponds to damping, for |ẏ| < 3, and to excitation, for |ẏ| > 3. As we will see in this section, this
equation possesses a solution corresponding to a relaxation limit cycle. In next week’s lecture, we will see that
this equation possesses a periodic orbit not only for arbitrarily small but also for arbitrarily large ε. In fact, it
can be proven that Rayleigh’s equation possesses a limit cycle for every positive value of ε. [In our setting, it
is instructive to try to show that (4.11) has a limit cycle by constructing a bounding box in the phase plane
and applying the Poincaré–Bendixson theorem.]
To construct the relaxation oscillation as in the last section, we write the singularly perturbed ODE (4.11)
as a system of two first-order ODEs,
ẏ = z,
(4.12)
εż = −y + z − 31 z 3 .
For ε = 0, this system possesses a manifold M0 of fixed points together with a vertical fibration over it, see
Fig. 9. The corresponding slow system yields the flow on M0 and reveals that the origin is a source. As

41
 y
Y

xl xf X x
Figure 8: The relaxation oscillation for the slow–fast, predator–prey model (4.1)–(4.2). The slow (blue) and
fast (red) components of the leading order relaxation oscillation are also sketched.

we discussed in last week’s lecture, the skeleton structure of the phase space for ε > 0 may be obtained by
concatenating the ε = 0 fast and slow dynamics. This idea leads, here again, to the leading order construction
of the limit cycle sketched in that same figure. Here also, a local analysis must be carried out around either
fold point where the slow manifold is not expected to perturb to establish that trajectories near Mε enter a
neighborhood of either of these points, get captured by the fast dynamics, and exit it.9 Our sketch also suggests
the amplitude (to leading order) of the oscillation, as well as that the limit cycle is (globally) stable. The period
of the oscillation can be also obtained to leading order in the following way: first, the fast transients last an
o(1) amount of time and thus can be omitted to leading order—hence, the period of the limit cycle equals,
to leading order, the sum of the amounts of time the cycle spends on either stable branch of M0 . By the
antisymmetry of the vector field, these two amounts are equal and thus the leading order term for the period
equals the time it takes the trajectory to move from (−2/3, 2) to (2/3, 1). Now, since M0 = graph(z − z 3 /3)
and ẏ = z, we find
z = ẏ = 1 − z 2 ż,

whence we obtain the period to leading order,

ˆ T /2 ˆ 1
1 − z2
T =2 dt = 2 dz + o(1) = 3 − 2 ln 2 + o(1).
0 2 z

9 In fact, an analysis around either one of these two fixed points is enough, as the vector field is antisymmetric with respect to

reflections about the origin.

42
 z

−2/3 2/3 y

−1

−2

Figure 9: Concatenation of the ε = 0 fast and slow dynamics in the phase space for Rayleigh’s equation (4.11).
The slow (blue) and fast (red) components of the leading order relaxation oscillation are also sketched.

Lecture no.5 March 05, 2013

5.1 The method of strained coordinates

[See also [Kevorkian & Cole, Sect. 4.1-2] for a detailed discussion of the method of strained coordinates, a.k.a.,
Poincaré–Lindstedt method.] At this point, it might seem that we have all we need to deal with perturbed
ODEs: a procedure to approximate inner solutions (if any exist) and another one for outer solutions (both to
arbitrary asymptotic accuracy), and even a way to combine (match) them into a uniformly valid asymptotic
expansion. Additionally, in the last two lectures, we have examined a geometric framework which, among other
things, allows us to go as far as predicting where layers will turn up and how many (if any) solutions satisfying
our initial/boundary value problem may be expected to exist.
Nevertheless, a large part of the machinery we developed in the lectures up to this one can only deal with
problems in which the independent variable takes values in a bounded set. Consider, for example, the regularly
perturbed problem we treated in Section 2.1: both the two-term asymptotic expansion we obtained there (cf.
(2.6) and (2.6)) and the exact solution (2.2) grow unboundedly (in fact, exponentially) as x → ∞, albeit the
difference between them also grows exponentially, so that it remains O(ε2 ) only over a finite time horizon.
Even worse, consider the problem in exercise 01(b) of Homework #02, which deals with a regularly perturbed
harmonic oscillator,
ÿ(t) + (1 + ε) y(t) = 0, with y(0) = 0 and ẏ(0) = 1. (5.1)
As we saw, the two-term asymptotic expansion we derived only remains uniformly valid over a finite time,
although both it and the actual solution remain bounded. Indeed, postulating the asymptotic expansion
X
y(t) = εn yn (t),
n≥0

substituting into (5.1), and collecting terms of equal order in ε, we obtain, up to and including terms of O(ε),
ÿ0 + y0 = 0, with y0 (0) = 0 and ẏ0 (0) = 1, (5.2)
ÿ1 + y1 = −y0 , with y1 (0) = 0 and ẏ1 (0) = 0. (5.3)
The first of these problems yields y0 (t) = sin t. Substituting into the second ODE and using variation of
constants [Bender & Orszag, Section 1.5], we obtain for y1 the general solution
 ˆ t   ˆ t 
y1 (t) = C+ + sin2 (s) ds cos(t) + C− − cos(s) sin(s) ds sin(t).
0 0

43
Here, C± are arbitrary constants. Evaluating the integrals and imposing the initial conditions, we finally find
1 1
y1 (t) = t cos t − sin t.
2 2
It now becomes evident that the term εy1 becomes commensurate with the leading order term y0 in the regime
t = O(1/ε); in other words, the asymptotic expansion stops being well-ordered in that regime, see also Fig. 10.
(It can be easily shown that all subsequent terms in the asymptotic expansion become commensurate with y0 ;
the problem is not limited to εy1 .)
In this particular case, the result can be anticipated. Indeed, the forcing term in (5.3) is oscillatory with
frequency equal to the system’s eigenfrequency; hence, one should expect resonance. (The way this manifests
´t
itself mathematically is through the term 0 sin2 (s) ds which becomes unbounded as t → ∞—note that the
integrand is periodic and non-negative!) From yet another point of view, our linear model can be solved to
yield the exact solution
1 √

y(t) = √ sin 1 + ε t . (5.4)
1+ε
Expanding the ε−dependent terms in powers of ε, we obtain
 
1 1
y(t) = sin(t) + ε t cos t − sin t + O(ε2 );
2 2

this matches the asymptotic expansions we derived above (the same applies to all orders in ε), and which is
only well-ordered for finite times. It is interesting to note here that, although the asymptotic expansion is not
well-ordered past t = O(1/ε), the expansion derived above remains, in fact, convergent for all t ∈ R. This is
far from surprising: indeed, the only t−dependent term in (5.4) is
√
 
1
1 + ε t = 1 + ε + O ε2


t;
2

the (asymptotically) largest term in this asymptotic expansion is t, while εt is the (asymptotically) largest
perturbative term. For t 6 1/ε, this last term stops being perturbative and becomes commensurate—or even
asymptotically larger—than t, the value around which we expand. Hence, it stands to reason that the Taylor
expansion we produce is not an asymptotic expansion of the actual solution; this, nevertheless, is unrelated to
whether it is convergent or not.
Since the asymptotic series we produced is convergent, its terms must progressively get smaller. Nevertheless,
the bigger the time interval we wish to approximate the solution over, the more terms we have to include in the
asymptotic expansion. If we allow the time interval to become unbounded, every truncated series will diverge
from the actual solution, as the series is not uniformly convergent over any unbounded
√ time interval. From a
phenomenological point of view, the exact solution has a period equal to 2π/ 1 + ε, whereas the asymptotic
expansion involves 2π−periodic functions; it is only to be expected that the phase difference accumulated at
each period grows with time. This consideration also suggests the remedy, which lies at the heart ofP the method
of strained coordinates: first, we change the timescale (rescale time) via τ = ω(ε) t, where ω(ε) = 1+ n≥1 εn ωn
is the new timescale. (The leading order term in this expansion is chosen equal to one for definiteness.) [Make
sure you understand why we have the freedom to do this without compromising the applicability of the method
and what are the significance and consequences of this choice.] Then, we express the solution in terms of the
new temporal variable τ and write y(t) = ψ(τ ). The initial value problem that ψ has to satisfy is

ω 2 (ε)ψ 00 (τ ) + (1 + ε) ψ(τ ) = 0,
with ψ(0) = 0 and ψ 0 (0) = ω −1 (ε).

Expanding the new dependent variable as ψ(τ ) = n≥0 εn ψn (τ ) and following the same procedure as before,1
P
we obtain

ψ000 + ψ0 = 0, with ψ0 (0) = 0 and ψ00 (0) = 1,

ψ100 + ψ1 = −2ω1 ψ000 − ψ0 , with ψ1 (0) = 0 and ψ10 (0) = −ω1 .
1 Naturally, this problem can be solved explicitly, just like our original problem could; we derive the asymptotic expansion of

the solution rather than the solution itself solely to demonstrate the method.

44
 y

0
t

Figure 10: The exact solution—y, in blue—and the two-term expansion—y0 +√εy1 , in red—corresponding
to regularly perturbed harmonic oscillator (5.1). The amplitude lines y = ±1/ 1 + ε corresponding to the
maxima and minima of the exact solution are also shown.

The first of these problems is identical to that for y0 , and hence ψ0 (τ ) = sin(τ ). Substituting into the problem
for ψ1 , we find
ψ100 (τ ) + ψ1 (τ ) = (2ω1 − 1) sin(τ ), with ψ1 (0) = 0 and ψ10 (0) = −ω1 . (5.5)
Here also, the forcing is resonant and hence it will lead to an unbounded solution if not eliminated. To avoid
this, we do eliminate it by making the choice ω1 = 1/2. Substituting into (5.5) and solving, subsequently, the
problem, we find ψ1 (τ ) = − sin(τ )/2, so that
h ε i h ε i
ψ(τ ) = 1 − + O(ε2 ) sin(τ ), or equivalently y(t) = 1 − + O(ε2 ) sin 1 + ε/2 + O(ε2 ) t .


2 2
Note that the asymptotic expansion we obtain in this way corresponds to asymptotically expanding, in powers
of ε, the solution’s amplitude and eigenfrequency but not the sinusoidal term itself.

5.2 The multiscale method

[See [Holmes, Sect. 3.1-2] for a detailed discussion of the multiscale method.] Looking back at the idea un-
derpinning the method of strained coordinates, it becomes intuitively clear that the method should do well
whenever a sole timescale is involved. This raises the question of how can one treat problems involving multiple
timescales. To elucidate the problem, we deal with the example presented in [Holmes, Sect. 3.1-2],
ÿ(t) + εẏ(t) + y(t) = 0, with y(0) = 0 and ẏ(0) = 1, (5.6)
which models a weakly damped harmonic oscillator. A standard application of the Liouville transform, Y (t) =
eεt/2 y(t), puts the problem in the form
ε2
 
Ÿ (t) + 1 − Y (t) = 0, with Y (0) = 0 and Ẏ (0) = 1.
4
This is the model problem (5.1) in disguise: the asymptotically small parameter ε appearing in that model has
been replaced by −ε2 /4 in this one. Consequently, the method of strained coordinates can be employed to deal
with (5.6), at least after the Liouville transform has been employed. Can it deal equally well with the problem
before the application of that transform? To find out, we introduce the new temporal variable τ = ω(ε)t and
dependent variable y(t) = ψ(τ ) to recast the problem in the form
ω 2 (ε) ψ 00 (τ ) + ε ω(ε) ψ 0 (τ ) + ψ(τ ) = 0, with ψ(0) = 0 and ψ 0 (0) = ω −1 (ε).

45
 εn ωn and ψ(τ ) = εn ψn (τ ) and working as
P P
Postulating the asymptotic expansions ω(ε) = 1 + n≥1 n≥0
before, we obtain

ψ000 + ψ0 = 0, with ψ0 (0) = 0 and ψ00 (0) = 1,

ψ100 + ψ1 = −2ω1 ψ000 − ψ00 , with ψ1 (0) = 0 and ψ10 (0) = −ω1 .

The zeroth-order initial value problem has been solved before, ψ(τ ) = sin(τ ). Substituting into the first-order
problem yields
ψ100 + ψ1 = 2ω1 sin(τ ) − cos(τ ), with ψ1 (0) = 0 and ψ10 (0) = −ω1 .
A straightforward application of the variation of constants formula yields the solution
 
1
ψ1 (τ ) = − sin(τ ) + ω1 cos(τ ) τ.
2

The parenthesized term is a linear combination of sines and cosines with the same frequency, and hence it is a
sinusoidal function itself. In particular,
r
1 1
sin(τ ) + ω1 cos(τ ) = A sin(τ + θ), with A = + ω12 and θ = arctan(2ω1 ),
2 4
and hence, finally,
ψ1 (τ ) = −A τ sin(τ + θ).
In other words, ψ1 is comprised of a sole,psecular term. To suppress secularities, one would need to set A = 0;
nevertheless, that is impossible as A = 1/4 + ω12 . [In fact, this formula suggests that the choice ω1 = i/2
will make A vanish. You should research this possibility, in particular as far as higher order terms in the
expansion(s) are concerned. Keep in mind that sin(iετ /2) = −i sinh(ετ ) and that the damping contributes an
exponential term to the exact solution; also, that the damping-induced frequency shift is O(ε2 ). These two
observations point to the idea of using ω1 to deal with damping and the even coefficients ω2n to deal with the
frequency shift. Does this work out to all orders?] It appears, therefore, that the method of strained coordinates
cannot handle this problem.
The method in question fails because the problem involves p not one but two timescales: an O(ε) one,
τ1 = εt/2 (associated with damping), and an O(1) one, τ2 = 1 − ε2 /4t (associated with oscillatory motion).
Having chosen ω0 = 1, we have made the implicit choice to focus on the latter. In that regime, the exponent
of the exponential term e−εt/2 contributed by damping to the (exact) solution will be perceived as O(ε) and
thus expanded asymptotically (here, as a McLaurin series). Recalling the McLaurin series of the exponential,
we plainly see that we must anticipate the presence of secular terms in our asymptotic expansion. The most
straightforward way to avoid such secularities would be to avoid expanding the exponential, and this is precisely
what the method of multiple scales does.
The method works as follows: first, next to the timescale t1 = t suggested by the formulation of the problem,
we introduce a second timescale t2 = εα t (where the parameter α is free, for the moment). We will attempt to
derive a solution to the problem of the form y(t; ε) = ψ(t1 , t2 ; ε). Since both t1 and t2 are defined in terms of
the temporal variable t, chain rules yields
d ∂ ∂
= + εα = ∂1 + εα ∂2 ,
dt ∂t1 ∂t2
where the last equation merely introduces notation. Our initial value problem becomes accordingly
2
(∂1 + εα ∂2 ) ψ + ε (∂1 + εα ∂2 ) ψ + ψ = 0, with ψ(0, 0; ε) = 0 and (∂1 + εα ∂2 ) ψ(0, 0; ε) = 1. (5.7)

Note that this initial value problem involves a PDE and not an ODE; in other words, the notation we use above
suggests that t1 and t2 are treated as variables independent of each other although they P are both functions
of the same variable t. Introducing, here also, the asymptotic expansion ψ(t1 , t2 ; ε) = n≥0 εn ψn (t1 , t2 ) and
substituting in the initial value problem (5.7), we obtain at O(1)

∂12 ψ0 + ψ0 = 0, with ψ0 (0, 0) = 0 and ∂1 ψ0 (0, 0) = 1. (5.8)

46
The solution to this linear, second order, parabolic PDE is readily found to be

ψ0 (t1 , t2 ) = C+ (t2 ) cos(t1 ) + C− (t2 ) sin(t1 ), with C+ (0) = 0 and C− (0) = 1. (5.9)

Collecting the next order terms from each term entering the left member of (5.7), we obtain

2εα ∂1,2 ψ0 + ε∂12 ψ1 + ε∂1 ψ0 + εψ1 = 0, with ψ1 (0, 0) = 0 and ε∂1 ψ1 (0, 0) + εα ∂2 ψ0 (0, 0) = 0. (5.10)

If α > 1, then the first term in the left member of the PDE above, as well as the second term in the left member
of the second initial condition, are higher order. The remaining terms comprise a problem that is resonant
because of the inhomogeneity, i.e., the advective term ∂1 ψ0 . Hence, a ≤ 1. Next, if α < 1, then the first term
in the left member of the PDE is leading order (and similarly for the second term in the left member of the
second initial condition); hence,

∂1,2 ψ0 = 0, with ψ1 (0, 0) = 0 and ∂2 ψ0 (0, 0) = 0.

Substituting from (5.9) into this problem, we obtain

0 0 0
−C+ (t2 ) sin(t1 ) + C− (t2 ) cos(t1 ) = 0, with C+ (0) = 0.

As
p discussed above, the left member of the first equation is a sinusoidal term with amplitude equal to
[C+0 (t )]2 + [C 0 (t )]2 ; hence, it vanishes iff it holds identically that C 0 (t ) = C 0 (t ) = 0. Recalling the
2 − 2 + 2 − 2
initial conditions for C± reported in (5.9), we conclude that C+ (t2 ) = 0 and C− (t2 ) = 1. Equivalently,
ψ0 (t1 , t2 ) = sin(t1 ), and resonance will once again appear in the next-order problem involving ψ1 . Hence, we
select α = 1 (as the principle of least degeneracy also dictates) to recast (5.10) in the form

∂12 ψ1 + ψ1 = −2∂1,2 ψ0 − ∂1 ψ0 , with ψ1 (0, 0) = 0 and ∂1 ψ1 (0, 0) + ∂2 ψ0 (0, 0) = 0.

Substituting from (5.9) for ψ0 , we rewrite the PDE above in the form
 0  0
∂12 ψ1 + ψ1 = 2C+
 
(t2 ) + C+ (t2 ) sin(t1 ) − 2C− (t2 ) + C− (t2 ) cos(t1 ).

The only way to avoid resonance is to set the coefficients of the sinusoidal terms in the inhomogeneity equal
0
to zero, 2C± (t2 ) + C± (t2 ) = 0. Recalling also the initial conditions for C± reported in (5.9), we conclude that
C+ (t2 ) = 0 and C− (t2 ) = e−t2 /2 . Hence,

ψ0 (t1 , t2 ) = e−t2 /2 sin(t1 ).

Resubstituting for t1 and t2 from their definitions—t1 = t and t2 = εt (recall that α = 1)—we obtain the
leafing order term in the multiscale asymptotic expansion of y,

y0 (t) = ψ0 (t, εt) = e−εt/2 sin(t).

This clearly matches the exact solution to leading order, where by ‘leading order’ we understand the process
of expanding asymptotically the sinusoidal term while keeping the exponential intact. The same process can
be repeated to all orders to yield higher order terms in the multiscale asymptotic expansion of the solution.

47
Homework set no.3 due March 12, 2013

01. Consider the various statements in Section 3.4.

(a) Prove statement #01, under the normal hyperbolicity assumption in that same section.

(b) Prove the second half of Statement #02; namely, that Tz M0 and Tz F 0 (z) intersect transversally in Rn+` .

(c) Prove formula (3.22) for DĜ0 (x, ŷ).

02∗ . Consider the initial-value problem

ẏ(t) = 1 − y(t) y 2 (t),

with y(0) = ε.

Here, 0 < ε  1 as usual. Perform a full analysis of this model for t ∈ R̄+ . In particular: locate the boundary
layer, derive inner and outer expansions for the solution (to leading order, at least), and match them over prop-
erly chosen intermediate timescales. [Note that the approximate formula(s) for the outer solution(s) can—and
therefore should!—yield the solution y as an explicit function of t. The inner solution, on the other hand, can
only be obtained implicitly.]

03. Consider the following initial-value problem for a harmonic oscillator with weak cubic damping,

ÿ(t) + εẏ 3 (t) + y(t) = 0, with y(0) = 0 and ẏ(0) = 1.

Use the method of multiple scales (two-timing) to derive a leading order solution which remains uniformly valid
over the entire positive real axis.

04. Consider the following Mathieu equation describing a weakly parametrically forced harmonic oscillator,

ÿ(t) + (δ(ε) + ε cos(ωt)) y(t) = 0,

with y(0) = yI and ẏ(0) = 0.

Here, δ(ε) is assumed to possess the asymptotic expansion 1 + n≥1 εn δn .

P

(a) Obtain the leading order behavior of the solution to the above initial-value problem for ω > 2, by using a
two-timescale expansion (in the original time variable t and the slow time variable τ = εt).
(b∗ ) Do the same for the case ω = 2 and show that, if |δ(ε) − 1| is small enough, then the origin is destabilized.
Is the origin destabilized for other values of ω as well?

48
Lecture no.6 March 12, 2013

6.1 Metastable patterns in a reaction–diffusion system

A summary of the theory of metastable patterns in a simple reaction–diffusion system is coming up; until then,
your primary sources of information are [Carr & Pego] and [Fusco & Hale].

49
Lecture no.7 March 19, 2013

50
References
[Bender & Orszag] C. M. Bender and S. A. Orszag, Advanced Mathematical Methods for Scientists and Engi-
neers, Springer-Verlag, New York, 1999 (originally Mc-Graw Hill, 1978) [available online in google books]
[Carr & Pego] J. Carr and R. L. Pego, Metastable patterns in solutions of ut = ε2 uxx − f (u), Comm. Pure
Appl. Math. XLII(3), 523–576, 1989
[Fusco & Hale] G. Fusco and J. K. Hale, Slow-Motion Manifolds, Dormant Instability, and Singular Perturba-
tions, J. Dyn. Diff. Eqs. 1(1), 75–94, 1989
[Hek] G. Hek, Geometric singular perturbation theory in biological practice, J. Math. Biol. 60, 347–386, 2010
[Holmes] M. Holmes, Introduction to Perturbation Methods, Springer-Verlag, New York, 1995 [available online
in google books]

[Jones] C. K. R. T. Jones, Geometric singular perturbation theory, in: [ Arnold]Dynamical Systems, Monteca-
tini Terme, L. Arnold (ed.), Lecture Notes in Mathematics, 1609, Springer-Verlag, Berlin, 1994, pp. 44–118
[Kaper] T. J. Kaper, An introduction to geometric methods and dynamical systems theory for singular per-
turbation problems, Proc. Sympos. Appl. Math. 56, 85-131, 1999

[Kevorkian & Cole] J. Kevorkian and J. D. Cole, Multiple Scale and Singular Perturbation Methods, Springer-
Verlag, New York, 1996 (updated version of the authors’ Perturbation Methods in Applied Mathematics,
Springer-Verlag, 1981) [available online in google books]
[O’Malley] R. E. O’Malley, Jr., Singular Perturbation Methods for Ordinary Differential Equations, Springer-
Verlag, New York, 1991

[Segel & Slemrod] L. A. Segel and M. Slemrod, The Quasi–Steady–State Assumption: a study case in pertur-
bation, SIAM Rev. 31(3), 446–477, 1989
[Stewart & Sun] G. W. Stewart and J. Sun, Matrix Perturbation Theory, Academic Press, New York, 1990
[Verhulst] F. Verhulst, Methods and Applications of Singular Perturbations: Boundary Layers and Multiple
Timescale Dynamics, Springer, New York, 2005 [available online in google books]
[Wong] R. Wong, Asymptotic Approximation of Integrals, SIAM, Philadelphia, 2005 (originally Academic Press,
1989) [available online in google books]

51