Winitzki - Perturbation Theory for An Harmonic Oscillations

8/3/2019 Winitzki - Perturbation Theory for An Harmonic Oscillations

1/8

Perturbation theory for anharmonic oscillations

Lecture notes by Sergei Winitzki

June 12, 2006

Contents

1 Introduction 1

2 What is perturbation theory 1

2.1 A first example . . . . . . . . . . . . . . . . . . 12.2 Limits of applicability of perturbation theory . 2

2.3 Higher orders . . . . . . . . . . . . . . . . . . . 22.4 Convergence and asymptotic expansions . . . . 32.5 How to guess the perturbative ansatz . . . . . . 3

3 Perturbation theory for differential equations 4

3.1 Precision and limits of applicability . . . . . . . 4

3.2 Convergence . . . . . . . . . . . . . . . . . . . . 43.3 Perturbative expansions with = 1 . . . . . . . 5

4 Anharmonic oscillations 5

4.1 Failure of a simple perturbation ansatz . . . . . 54.2 Lindstedt-Poincare method . . . . . . . . . . . 64.3 Precision of approximation . . . . . . . . . . . 74.4 Discussion . . . . . . . . . . . . . . . . . . . . . 7

5 Suggested literature 7

5.1 Remarks on Landau and Lifshitz 28 . . . . . 8

In these notes I explain the basics of perturbation theory andthen show how to apply this theory to describe anharmonicoscillations.

This document is in copyright, 2006-2010, by SergeiWinitzki. Permission is given to distribute this documentunder the GNU Free Documentation License version 1.2 orlater. The LATEX source code for this text is attached to thePDF file.

1 Introduction

Consider a one-dimensional system described by a coordinate

x(t). This corresponds to a point mass moving in a potentialV(x). When the system is near a stable equilibrium pointx = x0, we have V

(x0) = 0, and thus the equation of motion

x = 1m

V(x)

can be linearized as follows,

x = (x x0) V(x0)

m 1

2(x x0)2 V

(x0)

m ... (1)

If we only consider the linear term (harmonic approxima-tion), we obtain the equation of a harmonic oscillator withthe frequency

2 1m

V(x0).

Harmonic oscillations have a fixed period, T = 21, whichis independent of the amplitude. However, this is only an ap-proximation which is valid for small enough |x x0|. It is alsoimportant to study the effect of further terms in Eq. (1). Forinstance, we would like to know how the period of oscillationsdepends on the amplitude when the amplitude becomes large.This will give us an idea about the limits of applicability ofthe harmonic approximation.

Let us introduce the coordinate q(t) x(t) x0. We areinterested in describing an anharmonic oscillator,

q + 20q = f(q), (2)

where f(q) is a nonlinear function which represents a smallperturbation. The function f(q) can be expanded in a powerseries in q as follows,

f(q) = f2q2 + f3q

3 + ...,

where the parameters f2, f3, ..., are small in an appropriatesense. (Below we shall determine exactly how small they mustbe.)

In the general case, it is impossible to obtain an exact so-lution of Eq. (2). However, one can use the method calledperturbation theory and obtain an approximate solution asa function of the small parameters f2, f3, ...

I now review the basic principles of perturbation theory.(Please follow all the calculations with pen and paper if thisis your first experience with perturbation theory.)

2 What is perturbation theory

Often we have a problem whose exact solution cannot befound, but a slightly different problem is exactly solvable.Perturbation theory is a method for obtaining an approximatesolution of the first problem by modifying the exact solutionof the second problem.

2.1 A first example

Consider an algebraic (not differential) equation,

x5.23 0.000001 x 1 = 0. (3)

It is clear that this equation cannot be solved exactly. Butwe notice that Eq. (3) is almost the same as the equation

x5.23 1 = 0,

which we can easily solve: x = 1. In such situations, one canapply perturbation theory with great success.

The method of perturbation theory is applied to Eq. (3)as follows. For brevity, let us set a 5.23. The first step isto introduce a small parameter into the equation. One says

1

8/3/2019 Winitzki - Perturbation Theory for An Harmonic Oscillations

2/8

that Eq. (3) is a perturbation of the equation xa 1 = 0,and one describes this perturbation by introducing a smallparameter , so that Eq. (3) becomes

xa x 1 = 0. (4)In the particular example, we have = 106, but it is ac-tually easier to keep the value of arbitrary, as long as weremember that it is very small. The value = 0 describesthe unperturbed equation which we can solve exactly; theunperturbed solution is x0 = 1. Now we are ready for

the second step. It is reasonable to suppose that the solutionx of Eq. (4) is a small perturbation of the unperturbed so-lution, i.e. that x is only very slightly different from 1. Wemay imagine that we could solve Eq. (4) for every (sufficientlysmall) , so that the solution x is a function of . Since weknow that x( = 0) = x0 = 1, it seems reasonable to assumethat we may describe the solution x() as a series in ,

x() = 1 + A1 + A22 + ..., (5)

where A1, A2,... are some unknown constants. I would like toemphasize that the formula (5) is just a guess; at this point,we have no proof that the solution x is indeed of this form.

We need to check that this guess is correct. Let us substituteEq. (5) into Eq. (4) and expand in powers of : the constantsA1, A2,... must be such that

1 + A1 + A22 + ...

a 1 + A1 + A22 + ... 1 = 0.(6)

For the moment, let us keep only linear terms in :

1 + aA1 1 + O(2) = 0.Since is arbitrary, the linear terms in must cancel sepa-rately from the quadratic terms, so we must have

A1 =

1

a .

Now let us look at the quadratic terms in Eq. (6) and disre-gard the cubic terms. We find

aA22 +

a(a 1)2

A212 A12 = 0, (7)

therefore

A2 = 12a

+3

2a2.

Let us stop at this point and summarize what we have done.Using perturbation theory, we have found an approximate so-lution of Eq. (4),

x 1 + A1 + A22 = 1 + 1a

a 32a2

2. (8)

For the values = 106 and a = 5.23, the term A22 is onlya tiny correction of order 1013, so the precision obtained byusing only the linear term, x 1 + /a, is already sufficientfor almost all practical purposes.

I would like to contrast the result obtained using perturba-tion theory, Eq. (8), with a numerical solution (i.e. a solutionfound by a numerical method, which essentially consists ofsubstituting different numbers x into Eq. (3) until a certainprecision is achieved). By using a computer program, one

could easily obtain an approximate solution of Eq. (3) withparticular numerical coefficients = 106, a = 5.23. How-ever, one would obtain a single number x; it would remain

unclear how this solution depends on a and . The numericalprocedure would have to be repeated if a different value ofa or were given. The perturbative result (8) is much moreuseful because it shows the behavior of x as a function of and a.

2.2 Limits of applicability of perturbation

theory

The solution (8) is not exact but approximate ; what is theprecision of this approximation? Here is how one can get aqualitative answer without knowing the exact solution. Thesecond term in Eq. (8) is linear in and represents a small cor-rection to the first term if || a. The third term representsa small correction to the second term if

|a 3|2a2

2 ||a

,

which is also satisfied if || a. Therefore, the condition|| a is the condition of applicability of perturbation theoryin our example. For very large such that || a, we shouldexpect that Eq. (8) becomes a bad approximation to the exactsolution. One says that perturbation theory breaks down

at || a.How can we estimate the error of the approximation (8)?

We can argue as follows. Perturbation theory works onlywhen is sufficiently small so that each term in the seriesis much smaller than the previous term. If we use only thefirst two terms, x 1 + /a, we are neglecting the thirdterm, which is a correction of order 2/a2. The third term, ifincluded, will certainly improve the precision. Therefore, theerror of the approximate solution 1 + /a can be estimatedas x C2/a2, where C is a constant roughly equal to|a 3| /2. Similarly, the error of the approximation (8) isapproximately equal to the fourth term, which is of order

|/a|3

.

2.3 Higher orders

It is clear that by using the method shown above, we cancalculate all the coefficients in the expansion (5), one by one.Let us describe the procedure in more general terms. Weassume the expansion

x() = 1 +Nn=1

Ann + O(N+1),

which contains N unknown constants A1

,...,AN

. This iscalled a perturbative expansion since it describes a smallperturbation of the unperturbed solution x = 1. We thensubstitute this expansion into Eq. (4), expand everything in, disregarding terms of order N+1 or higher, and obtain anequation of the form

1 1 + (nA1 1) +

aA2 +a(a 1)

2A21 A1

2

+... + (...)N + O(N+1) =0. (9)

Since is arbitrary, this equation can be satisfied only ifeachof the N coefficients at , 2, ..., N is equal to zero.

To establish this more rigorously, we divide Eq. (9) by and take the limit as 0; we find nA1 1 = 0. Thus thefirst nonvanishing term in Eq. (9) is of order 2. Then we can

2

8/3/2019 Winitzki - Perturbation Theory for An Harmonic Oscillations

3/8

divide by 2 and take the limit 0, which yields Eq. (7),etc. Thus, Eq. (9) means that the coefficient at each order in is equal to zero, independently of all other orders.

In this way, a single condition (9) yields a system of Nequations for the N unknown constants. This system of equa-tions is easy to solve because the first equation contains onlyA1, the second equation contains only A1 and A2, and so on.Of course, the calculations will become more cumbersome forhigher orders, since we will have to retain many more terms,but there is no difficulty in principle. These calculations areso straightforward that one can program a computer to per-form these calculations symbolically. (A calculation is calledsymbolic if the computer is programmed to print the result-ing formula containing variables such as and a, as opposedto a numerical calculation where the computer prints onlya numerical value of x.)

2.4 Convergence and asymptotic expansions

In principle, one could calculate all the coefficients An, n =1, 2, 3,..., by the method shown above. Then one may askwhether the resulting infinite series

x() = 1 +

n=1

Ann (10)

will converge to the exact solution x. Unfortunately, thequestion of convergence is much more difficult to answer; toanswer it, one needs some information about the exact so-lution of the perturbed equation. In some cases (such asEq. (4) above, the series converges. However, in many cases,especially when applying perturbation theory to differentialequations, the perturbative series does not converge. Nev-ertheless, a truncated series (i.e. the series where we takeonly a fixed, finite number of first terms of the perturba-tive expansion) always provides excellent precision when isvery small, and hence can be successfully used in practice todetermine approximate solutions. For instance, Eq. (8) is avery accurate approximation to the solution of Eq. (4) when|/a| 1.

This is a somewhat paradoxical situation: A truncated se-ries is useful because it gives a good approximation to x() atsmall , but the full infinite series does not actually convergeto the exact value of x().

Let us consider this situation in some more detail. The sumof first N terms can be thought of as a function of and N,

x

(, N) 1 +

N

n=1An

n

.

This function may have very different behavior as N at fixed , and as 0 at fixed N. In particular, the func-tion x(, N) provides a very precise approximation to theexact solution x() at fixed N and small . The error of thisapproximation is of order N+1. What does it mean, in math-ematical terms, that the error is of order N+1? It meansthe following property,

lim0

x(, N) x()N

= 0. (11)

If this property holds, one says that the series x(, N) is anasymptotic expansion of the solution x(). Obviously,such a series is very useful in practice, since the precision of

a truncated series is very good at sufficiently small . Nev-ertheless, it often happens that the infinite series diverges,limN x(, N) = . There is no contradiction since theselimits are taken in very different ranges of the variables , N.An asymptotic expansion may be either convergent or diver-gent.

The conclusion is that perturbation theory will give anasymptotic expansion, but not always a convergent expan-sion. However, in practice it is almost always impossible toobtain all the coefficients An (i.e. to derive a formula for Anas a function ofn). Moreover, one usually computes only twoor three terms of the series, that is, one computes x(, N)with N = 2 or, at most, N = 3. So the possible divergenceat N is not important. It is much more important toobtain an estimate of the error and to determine the admis-sible values of for which the approximate solution x(, N)is still precise at a fixed, small value of N.

2.5 How to guess the perturbative ansatz

The starting point of perturbation theory is a perturbativeexpansion such as Eq. (5). In other words, we are first guess-ing the solution to the problem in a form of a certain formula,

or ansatz, containing in a certain way. We do not knowin advance that this ansatz is correct; the justification comeslater when we actually arrive at the solution (8). But it mighthappen that we have not guessed the ansatz correctly; thenthe procedure will not work.

For example, consider the equation

x2 2 (1 + ) x + 1 = 0. (12)

This equation looks like a perturbation of the equation x2 2x + 1 = 0, which has the solution x = 1. Let us assume theansatz

x() = 1 + A + O(2

) (13)and substitute it into Eq. (12), keeping only terms linear in. We find

2 = O(2).This condition cannot be satisfied because nothing can cancelthe linear term 2. Therefore, the ansatz (13) is incorrect.

The reason for this trouble will become clear if we look atthe exact solution of Eq. (12),

x() = 1 +

2 + 2. (14)

The leading correction is actually of order rather than oforder , and the expansion of x() is of the form

x = 1 + A1

+ A2 + A33/2 + ...

Therefore, the correct perturbative ansatz must be of thisform. After substituting such an ansatz into Eq. (12) andequating terms of equal order in , one gets A1 =

2, A2 =

1, etc.Now that we have the exact solution (14), it is easy to

see that the perturbative expansion is the Taylor series forthe analytic function x(

). Therefore, the perturbative

expansion actually converges for small enough . But it would

be difficult to judge the convergence of the expansion if we hadonly computed the first few coefficients A1, A2,...,An of theexpansion and did not have a formula for the exact solution.

3

8/3/2019 Winitzki - Perturbation Theory for An Harmonic Oscillations

4/8

The conclusion is that one will run into trouble if one triesto use an incorrect perturbative expansion. When this hap-pens, another expansion ansatz must be chosen, e.g. contain-ing

or ln instead of . One can often use physical intu-

ition or qualitative considerations about the exact solution toguess the correct ansatz.

3 Perturbation theory for differen-

tial equations

Let us apply the method of perturbation theory to a differ-ential equation. For example, consider the equation

x + 2.33x = 0.000001x3, x(0) = 15. (15)

It is clear that this equation can be thought of as a smallperturbation of

x + 2.33x = 0, x(0) = 15,

whose exact solution can be easily found,

x(t) = 15e2.33t. (16)

Therefore, we rewrite Eq. (15) as

x + ax = x3, x(0) = A, (17)

where we introduced a small parameter 106 and (non-small) parameters a 2.33, A 15. Let us now apply themethod of perturbation theory to this problem.

As the next step, we assume that the solution of Eq. (17)is a small perturbation of the solution (16), i.e.,

x(t) = Aeat + f1(t) + O(2), (18)

where the function f1(t) needs to be determined. The initialcondition for f1(t) is found by setting t = 0 in Eq. (18), whichyields

A + f1(0) + O(2) = A.

Since the term linear in must vanish, we obtain f1(0) = 0.We now substitute the ansatz (18) into Eq. (17) and expandto first order in :

f1 + af1 = A3e3at + O(2).

All the terms linear in must cancel separately from termsquadratic in . Therefore, we obtain the equation

f1 + af1 = A3e3at, f1(0) = 0.

The solution is

f1(t) = A3 eat e3at

2a.

Therefore, the approximate solution of Eq. (17) is

x(t) Aeat + A3 eat e3at

2a.

Let us now find a second-order correction. We start with

the ansatz

x(t) = Aeat + f1(t) + 2f2(t) + O(

3), (19)

where f1(t) is already known and f2(t) is to be determined.The initial condition for f2(t) is f2(0) = 0. Substituting thisinto Eq. (17), we find

f2+af2 = 3A2e2atf1(t) =

3A5

2a

e3at e6at , f2(0) = 0.

The solution is

f2(t) = A5 9e

at 15e3at + 6e6at20a2

.

Therefore, the second-order approximate solution is

x(t) = Aeat + A3eat e3at

2a

+ 2A59eat 15e3at + 6e6at

20a2. (20)

3.1 Precision and limits of applicability

As a rule, the perturbative expansion provides good precisionif its successive terms rapidly diminish in magnitude. There-fore, the condition of applicability of perturbation theory toEq. (17) is

A3 eat e3at

2a

Aeat, for all t.It straightforwardly follows that || 2aA2. The conditionfor the quadratic term in Eq. (20) is

2A59eat 15e3at + 6e6at

20a2A3 e

at e3at2a

,which gives || 9aA2/10. It follows that perturbationtheory is well within its limit of applicability for the values = 106, A = 15, a = 2.33.

The precision of the approximate solution (20) can be esti-

mated as follows. The first term is x0(t) = Aeat, the secondterm is of order

A3a1eat A2

ax0 104x0,

and the third term is of order

2A5a2eat

A2

a

2x0 108x0.

If we computed the next term of the perturbative series, itwould be of order 1012x0, but we did not include that termin Eq. (20). Therefore, the expected error is of order 1012x0.

3.2 Convergence

In principle, we can compute all the unknown functions f1(t),f2(t), ..., in the perturbative expansion

x(t) = eat +n=1

nfn(t). (21)

To compute each fn(t), we would need to solve a simple linearequation of the form

fn + afn = (...) ,

where the omitted expression in the right-hand side will de-pend on the functions f1, ..., fn1. Therefore, we can easily

4

8/3/2019 Winitzki - Perturbation Theory for An Harmonic Oscillations

5/8

determine fn(t) if the previous functions f1, ..., fn1 are al-ready found. As before, the calculations become more cum-bersome at higher orders, but there is no difficulty in princi-ple. A computer can be programmed to compute the resultsymbolically to any order.

The expansion (21) is asymptotic as 0, i.e. the partialsum

x(n)(t) = eat +

Nn=1

nfn(t)

satisfies the equation up to terms of order N+1 and is there-fore a good approximation to the exact solution when issmall. We can show that this expansion converges. Indeed,in this case we can actually find the exact solution of Eq. (17)by separating the variables,

xA

dx

x3 ax =t0

dt = t.

After some calculations, this gives the solution

x(t) = Aeat

1 A2

a

1 e2at

12

.

This is an analytic function of near = 0, therefore a Tay-lor expansion in will converge for small enough (moreprecisely, for || < aA2). The expansion (21) is equivalentto a Taylor expansion of the above function in . Hence, weconclude that the perturbative expansion will converge forsmall enough .

3.3 Perturbative expansions with = 1

In the example considered in this section, we saw that per-turbation theory works well as long as || A2 a. If thiscondition holds, the perturbative expansion is useful even if

the parameter is itself not small. Even if = 10, the per-turbative expansion will be precise for sufficiently small A.For instance, we may set a = = 1 and conclude that theequation

x + x = x3 (22)

may be solved using perturbation theory if the initial valuex(0) = A is small, |A| 1.

This may appear paradoxical at first sight: we constructedthe perturbative expansion under the assumption that issmall, disregarded higher powers of , and now we want touse the result with = 1. The reason for the success of theperturbative expansion with = 1 can be seen by examin-

ing the formula (20). In that formula, every power of ismultiplied also by A2/a. In other words, the perturbative ex-pansion is a series not in powers of but in powers ofA2/a.The value of does not actually need to be small.

The lesson is that one may apply perturbation theory suc-cessfully even to an equation such as Eq. (22), where no smallparameters are apparently present. To apply perturbationtheory, we may introduce a parameter = 1 and rewrite theequation as

x + x = x3, x(0) = A. (23)

The perturbative ansatz can still be formulated as in Eq. (19),as if is small, even though in reality = 1. Corrections of

order 2 in the final result (20) will be actually smaller thanterms of order , not because itself is small, but becauseevery is multiplied by a small number A2. One says that the

parameter is formal; by this one means that serves merelyas a bookkeeping parameter that measures the smallness ofcorrections. Of course, it takes a certain amount of intuitionto realize that the parameter should be attached to thex3 term rather than to some other term in Eq. (22). Thisintuition usually comes from physical considerations; in thecase of Eq. (22), one expects that the x3 term is insignificantwhen the initial value x(0) is very small, especially since thesolution of x + x = 0 exponentially decays with time.

However, I would like to stress that perturbation theorydoes involve a certain amount of guessing, because one needsto start from a certain perturbative ansatz, and there areno absolute rules about choosing that ansatz. How does oneknow that the ansatz is guessed correctly? Proving the con-vergence of a perturbative expansion is difficult unless oneactually knows something about the exact solution. In theabsence of a proof of convergence, one can justify the pertur-bative expansion by verifying that higher-order terms of theexpansion are smaller in magnitude than lower-order terms.This will show the limits of applicability of perturbation the-ory and also provide an estimate of the resulting precision.

4 Anharmonic oscillationsIn the preceding sections, we have considered toy pertur-bation problems where the exact solution is in fact knownand there is no need to use perturbation theory. The anal-ysis of these examples helps understand possible pitfalls ofguessing the perturbative expansion in various cases. We arenow ready to consider the problem of anharmonic oscillations,where exact solutions are not known.

As a first example of an anharmonic oscillator, consider apoint mass m moving in the potential

V(x) =1

2kx2 +

1

4mx4,

where is a small parameter, while k is not small. This po-tential has a minimum at x = 0 (for all ), so we expect oscil-latory behavior near this stable equilibrium point. Of course,oscillations will not proceed according to a sine curve such asx(t) = sin 0t. One says that in this case the oscillations areanharmonic (i.e. not harmonic).

It is clear that the potential V(x) is a small perturbationof the harmonic potential V0(x) =

12kx

2. Suppose that theoscillator was released from rest at a position x = A = 0.Then the equation of motion is

x + 20x + x3 = 0, x(0) = A, x(0) = 0, (24)

where 20 k/m is the frequency of oscillations in the unper-turbed potential V0(x).

4.1 Failure of a simple perturbation ansatz

Let us now apply perturbation theory to Eq. (24). The un-perturbed solution is

x0(t) = A cos 0t.

As a first attempt, let us assume a perturbative expansion of

the form

x(t) = x0(t) + x1(t) + 2x2(t) + ..., (25)

5

8/3/2019 Winitzki - Perturbation Theory for An Harmonic Oscillations

6/8

where x1(t), x2(t), ... are unknown functions. Substitutingthis ansatz into Eq. (24), we find

x1 + 20x1 = x30,

x2 + 20x2 = 3x20x1.

The initial conditions for the functions x1(t) and x2(t) arex1(0) = x1(0) = x2(0) = x2(0) = 0 since the initial condi-tion x(0) = A, x(0) = 0 is already exactly satisfied by theunperturbed solution x0(t). Note that

x30 = A3 cos3 0t = A3 cos30t 3cos 0t

4,

therefore we obtain

x1(t) =A3

3220[cos30t cos 0t 120t sin 0t] .

Thus the approximate first-order solution is

x(t) A cos 0t + A3

3220[cos30t cos 0t 120t sin 0t] .

We omit the calculation of x2(t) because there is already a

problem with the above expression for x1(t). Namely, it in-cludes the term t sin 0t which describes oscillations with agrowing amplitude. Clearly, the term t sin 0t will becomearbitrarily large at late times, thus breaking down the pertur-bation theory, even if is very small. Such terms that growat late times are called secular terms (from the Latin saecu-lum = century; the mental image is, perhaps, that a smallcontribution that persists over centuries will eventually growinto a large contribution). The reason for the appearance ofthe secular term is, technically, the presence of cos 0t in theright-hand side of the equation for x1(t), which is analogousto the presence of a periodic force whose frequency resonates

with the frequency of the oscillator. The resonance leads toa linear growth of the oscillation amplitude.The problem of secular terms is generic to any situation

with anharmonic oscillations. Secular terms may not appearin all orders in ; for instance, if we apply perturbation theoryto the equation

x + 20x = x2,

secular terms will first appear only at the second order in .This, of course, merely postpones the problem.

Secular terms appearing in a perturbative expansion areunsatisfactory for two reasons. First, they limit the applica-bility of perturbation theory to small values of time t. Per-

turbation theory breaks down at |t| 0A2

. Higher-ordercoefficients in the perturbative expansion will also containsecular terms; thus, we cannot improve the precision at latetimes, no matter how many orders of we compute. Sec-ond, an approximate solution containing secular terms doesnot reproduce the correct behavior of anharmonic oscillations:Since the total energy of the system is conserved, the ampli-tude of oscillations cannot actually increase without bound.We conclude that the perturbative ansatz (25) is incorrect.Let us now come up with an improved ansatz.

4.2 Lindstedt-Poincare method

We shall begin with some qualitative considerations. Whenthe potential is unperturbed ( = 0), the oscillations are har-monic and their frequency 0 is independent of the amplitude

A. In the presence of the perturbation (i.e. with = 0), wewould expect that the frequency of oscillations will depend onthe amplitude A as well as on .

Since we expect that the perturbed solution is still oscilla-tory but with a slighly different frequency, it makes sense toincorporate a change of frequency in the perturbative ansatz:

x(t) = A cos t + x1(t) + 2x2(t) + ..., (26)

where = 0 + 1 +

22 + ... (27)

is the perturbed frequency of the oscillations, and the con-stants 1, 2, ... are to be determined together with the un-known functions x1(t), x2(t), ...

The perturbation method we shall use is known as theLindstedt-Poincare method. This method uses the ex-pansions (26)-(27) with a further mathematical trick: onechanges the time variable from t to the dimensionless phasevariable t, where is the perturbed frequency. Afterthis change of variable, the perturbative ansatz becomes

x() = A cos + x1() + 2x2() + ..., (28)

while Eq. (24) needs to be rewritten through derivativesdx/d and d2x/d2. Let us denote these derivatives by aprime, then we have

x =d2x

dt2=

d2x

d22 2x,

and finally Eq. (24) becomes

0 + 1 +

22 + ...2

x + 20x + x3 = 0. (29)

The initial conditions for this equation are x(0) = A andx(0) = 0.

Now we can apply the perturbation ansatz (28) straight-forwardly to Eq. (29). Keeping only the linear terms in , weobtain the following equation for x1():

x1 +x1 = 2A10

cos A3

20cos3 , x1(0) = x

1(0) = 0. (30)

The solution of this equation will have secular terms if theright-hand side contains cos or sin . Since

cos3 =cos3 3cos

4,

the right-hand side of Eq. (30) is

2A

10

+3

4

A3

20

cos A

3

420cos3.

A secular term will appear if the coefficient at cos is nonzero.However, the value of 1 is still undetermined; so we canchoose the value of 1 to avoid the appearance of secu-lar terms. This is the key point of the Lindstedt-Poincaremethod. It is clear that we need to choose

1 = 38

A2

0.

The solution of Eq. (30) is then

x1() =A3

3220(cos3 cos ) ,

6

8/3/2019 Winitzki - Perturbation Theory for An Harmonic Oscillations

7/8

and thus the approximate first-order solution of the originalequation is

x(t) A cos t + A3

3220(cos3t cos t) , (31)

0 38

A2

0.

The solution (31) is well-behaved for all t and describes os-cillations with a slightly perturbed frequency and with slightdeviations from pure cosine shape.

4.3 Precision of approximation

Of course, the solution (31) is still approximate. The majorimprovement over the initial ansatz (25) is that there are nosecular terms, i.e. the amplitude of oscillations does not growwith time. However, the next correction to the frequency is oforder 2, so we expect that the first-order solution will devi-ate significantly from the exact solution after times t 2.Therefore, the solution (31) can be expected to be preciseonly for t 2. Moreover, the standard condition of appli-cability of perturbation theory is that the higher-order terms

are smaller, which gives|| A33220

A, 38

|| A20

0.

This condition is satisfied if || A2 20, which means thateither the perturbation in the potential is small, or the am-plitude A of oscillations is small. Physically, this conditionmeans that the change in the energy due to the perturbationof the potential, 1

4A4, is much smaller than the typical en-

ergy of the unperturbed oscillator, 12

20A2. Note also that

negative values of such that < 20A2 are physicallysignificant. For these values, the solution x(t) has the initial

accelerationx(0) = 20A + || A3 > 0,

which means that the oscillator does not move towards thestable equilibrium point x = 0 from its the initial posi-tion, x(0) = A. In this case the motion cannot be picturedas merely a small perturbation of the harmonic oscillation.These arguments show that we should not expect perturba-tion theory to yield any good results if

A220 1.The precision of the approximation (31) can be estimated

as follows. The first correction is of order || A220 times theunperturbed solution. Therefore, one expects that the nextcorrection will be of order

|

|

2 A440 times the unperturbed

solution. So the relative error of the approximation (31) is oforder ||2 A440 .

In reality, the error of the approximation will be some con-stant, say C1, times ||2 A440 . It is not easy to compute C1rigorously, but our estimates show that C1 might be a numberof order 1.

Using the Lindstedt-Poincare method, one can find furtherterms in the perturbative expansion (28)-(29). At the n-thstep, the function xn(t) will be determined from an equationsuch as

xn + xn = (...) ,

where the terms in the right-hand side will depend on pre-

viously found constants 1, ..., n1 and functions x1(t), ...,xn1(t). The value of n will be determined from the con-dition that no secular terms should arise in this equation.

The n-th order solution will be valid until times t n1.For times t n1, the error of the n-th order solution isCn

A220 n+1 times the unperturbed solution, where Cn isa proportionality constant that may be difficult to compute.

Note that the perturbation series may or may not converge(for fixed ) depending on how quickly the constant Cn growswith n. If Cn grows too quickly for large n, say as n!, wewill obtain a divergent series. Nevertheless, a fixed numberof terms of this series will be a useful approximation to thesolution.

4.4 Discussion

Let us summarize the results we obtained. We have ap-plied perturbation theory to Eq. (24) with a straightforwardansatz (25), but the attempt failed because of the appearanceof secular terms such as t cos 0t. The second attempt was touse an improved ansatz (26)-(27), which takes into account achange in the oscillation frequency,

0 = 0 + 1 + 22 + ...,due to the perturbation of the potential. However, perturba-tion theory based on Eq. (24) and the ansatz (26)-(27) would

have failed, had we not used the Lindstedt-Poincare method,which mandates a change of variable, t = t, involv-ing the perturbed frequency . Equation (24) is rewritten inthe new variables x(), yielding Eq. (29), and straightforwardperturbation theory can now be applied. At each order in ,one can choose the constants 1, 2, ..., so that no secularterms appear. The result is a well-behaved approximate so-lution for x(t). The n-th order solution is valid until timest n1, as long as the amplitude of oscillations A and theparameter are sufficiently small so that || A2 0.

What is the reason for the failure of perturbation theorywithout the Lindstedt-Poincare method? We can easily un-

derstand this by looking at the solution (31). The straight-forward perturbative ansatz (25) is a Taylor expansion ofthat solution in . For instance, the term A cos t will beexpanded as

A cos[(0 + 1) t] = A cos 0t A1t sin 0t + O(2).This expansion is only meaningful for small enough suchthat || 1t 1: The second term grows without bound atlate times, so the expansion breaks down at t 1. Nowit becomes clear why the secular term t sin 0t appears whenone uses a simple perturbative expansion in .

The Lindstedt-Poincare method can be also applied to any

equation of the formx + 20x = f(x, x),

where is the expansion parameter and f(x, x) is some ex-pression involving x and x, or to systems of such equations.

5 Suggested literature

E. J. Hinch, Perturbation methods (Cambridge UniversityPress, 1995). This is a short introductory book with manyexamples.

A. H. Nayfeh, Perturbation methods (Wiley, 1973). This is

a more advanced book showing a great multitude of tricks oneneeds to apply in order to develop a successful perturbationtheory in different cases.

7

8/3/2019 Winitzki - Perturbation Theory for An Harmonic Oscillations

8/8

5.1 Remarks on Landau and Lifshitz 28I warn against reading Landau and Lifshitzs Mechanics(third edition, Butterworth - Heinemann, 1976, which is thelast English edition) if it is your first source of information onthis topic. They discuss anharmonic oscillations in 28 with-out using the Lindstedt-Poincare method and obtain correctresults through a slightly incorrect mathematical reasoning.

Since Landau and Lifshitz is an extremely well respectedtextbook that has very few errors, I give full details of thecalculation. The problem is that their Eq. (28.12) cannot be

derived from the preceding Eq. (28.11) merely, as they say,by omitting terms of above the second order of smallness:one needs to suppress certain terms without explanation.

Their Eq. (28.11) is fully equivalent to Eq. (28.9), which wemay rewrite (explicitly introducing the parameter = 1) as

x + 20x =

x2 + x3

. (32)

They use the ansatz

x(t) = a cos t + x(2)(t), = 0 + (1) + (2) + ...,

where x(2), (1), etc., are quantities of higher order. Let us

make this explicit by rewriting the ansatz as

x(t) = a cos t + x1(t), = 0 + 1 + ... (33)

Now if we follow Landau and Lifshitz and substitute theansatz (33) into Eq. (32), we will find

2a01 cos t +

x1 + 20x1

= a2 cos2 t + O(2).

We equate the first-order terms and obtain an equation forx1(t),

x1 + 20x1 = 2a01 cos t a2 cos2 t. (34)

At this point, we observe that we have not yet fully expandedthe right-hand side of Eq. (34) in because the in cos tstill contains the term 1. The crucial point is that thecorrect result for x1(t), which is their Eq. (28.12),

x1(t) = a2

220+

a2

620cos2t,

will be obtained only if we expand in the first term in theright-hand side of Eq. (34) but do not expand in the secondterm (even though we expanded it just a second ago)!

Indeed, our Eq. (34) is identical to the unnumbered equa-tion in Landau and Lifshitz just after Eq. (28.11). If we doexpand the cos t everywhere (as we should, since we did ex-pand x2 + x3 in ), we will obtain

cos t = cos 0t 1t sin 0t + ...

and then Eq. (34) becomes

x1 + 20x1 = 2a01 cos 0t a2 cos2 0t + O().

The terms O() need to be neglected; in other words, wemerely replace by 0 in the right-hand side of Eq. (34).Then we obtain the no-resonance condition, 1 = 0, and theequation for x1(t) whose solution is

x1(t) = a2

220+

a2

620cos20t.

In other words, we will not obtain Eq. (28.12) of Landau andLifshitz because our x1(t) will contain cos 20t while theirscontains cos2t.

In order to obtain cos 2t in the expression for x1(t), wemust not replace the term cos2 t by cos2 0t in the right-hand side of Eq. (34). However, in order to derive the no-resonance condition, 1 = 0, we must replace the term cos tby cos 0t in the same equation. If we do not replace costby cos 0t, we will not get any resonance since the left-handside of Eq. (34) contains the unmodified frequency 0, whilethe right-hand side contains a different frequency, . Thenour solution x1(t) would have been

x1(t) = a

cos t + a10

cos t a2

220+

a2

620cos2t + O(),

which is of course also incorrect because of the presence ofthe first two terms. We have no grounds to demand thatthese terms be absent since they are not secular terms (nota resonance). Note especially the peculiar term of order 1/:No choice of 1 will make this term disappear.

The only way to derive Eq. (28.12) from Eq. (28.11) inLandau and Lifshitz is to selectively replace by 0 in

some terms but not in other terms. This is mathematicallyincorrect, since we are not allowed to throw away some termsof order while keeping other terms of the same order. Sincewe have performed a mathematically incorrect calculation,we cannot be sure that the result is correct. In contrast,the Lindstedt-Poincare method is conceptually simpler andperforms the derivation without any mathematically incorrectoperations.

8