The one good idea: solve the problem next door
Almost no equation that matters can be solved exactly. The quadratic x^2 + x - 1 = 0 has a tidy formula; the equation x^2 + x - 1 = epsilon x^5 does not, and never will. But look closely: when epsilon is a small number, that second equation is barely different from the first. So here is the whole philosophy of regular perturbation theory in one breath — see regular perturbation — when you can't solve the problem you've got, find a nearby problem you CAN solve, measure the gap between them with a small parameter, and patch up the easy answer with corrections that grow more refined as you account for more of the gap.
The small parameter is almost always written epsilon (sometimes lambda or a physical ratio like v/c). Setting epsilon = 0 hands you the unperturbed problem — the clean, solvable one. Turning epsilon up a little switches on the small effect you actually care about: a faint nonlinearity, a slight damping, a weak coupling, a tiny mass. The bet is that the true answer depends smoothly on epsilon, so that near epsilon = 0 it can be described by its values and its rate of change there — exactly the spirit of the linear approximation you met in Volume I, except now we go far beyond the first correction.
The machinery: expand, substitute, match powers
Here is the move that makes the whole method tick. You ASSUME the answer is itself a power series in epsilon: x = x_0 + epsilon x_1 + epsilon^2 x_2 + epsilon^3 x_3 + ... — a series not in a space or time variable but in the small parameter. Each x_n is a fixed number (or, for differential equations, a fixed function) waiting to be found. The leading term x_0 is the unperturbed answer; x_1 is the first correction, x_2 the correction to the correction, and so on. The magic is that you can find them one at a time, never solving anything harder than the easy problem.
The reason it works is a single, sturdy principle: if two power series in epsilon are equal for ALL small epsilon, then their coefficients must match power by power. The constant terms agree, the coefficients of epsilon agree, the coefficients of epsilon^2 agree, and so on — each gives you a separate equation. So you substitute the assumed series into the hard problem, expand everything, collect like powers of epsilon, and set each coefficient to zero. One messy problem fractures into an orderly tower of easy ones, indexed by the power of epsilon. That fracturing is the entire trick.
- Identify the knob. Write the problem so the hard part carries an explicit small factor epsilon, and check that epsilon = 0 leaves a problem you can actually solve. If it doesn't, stop — you have a singular problem, not a regular one.
- Posit the series. Write the unknown as x = x_0 + epsilon x_1 + epsilon^2 x_2 + ... and substitute it everywhere the unknown appears.
- Collect powers. Expand all products and powers, group every term by its power of epsilon, and set the coefficient of each power separately to zero. This is where the principle of matching coefficients does its work.
- Solve up the ladder. Solve the epsilon^0 equation for x_0; feed it into the epsilon^1 equation and solve for x_1; feed both into the epsilon^2 equation for x_2. Each rung is a fresh easy problem whose only new ingredient is the answers you already have.
A worked picture: a slightly bent quadratic
Let's chase one root of x^2 - 1 = epsilon x in pictures. With epsilon = 0 the equation is just x^2 = 1, so the unperturbed root is x_0 = 1 — the problem next door, solved on sight. Now write x = 1 + epsilon x_1 + epsilon^2 x_2 + ... and substitute. Squaring the series, x^2 = 1 + 2 epsilon x_1 + epsilon^2 (x_1^2 + 2 x_2) + ..., and the right-hand side epsilon x = epsilon + epsilon^2 x_1 + .... Now demand equality power by power.
Solve x^2 - 1 = epsilon x near the root x = 1. Let x = 1 + e*x1 + e^2*x2 + ... (e means epsilon) x^2 = 1 + 2e*x1 + e^2*(x1^2 + 2*x2) + ... e*x = e + e^2*x1 + ... Match powers of e in (x^2 - 1) - e*x = 0 : order e^0 : 1 - 1 = 0 -> automatically true (this fixed x0 = 1) order e^1 : 2*x1 - 1 = 0 -> x1 = 1/2 order e^2 : (x1^2 + 2*x2) - x1 = 0 -> x2 = (x1 - x1^2)/2 = 1/8 So x = 1 + (1/2) e + (1/8) e^2 + ... Check against the exact root x = (e + sqrt(e^2 + 4)) / 2 : its Taylor series in e is 1 + e/2 + e^2/8 - ... -- the terms agree.
Notice three things this tiny example reveals. First, every equation after the leading one is LINEAR in the new unknown, even though the original equation was nonlinear — squaring x_0 = 1 against a small correction only ever produces one copy of the new term. Perturbation theory routinely linearizes the nonlinear, which is precisely why it is tractable. Second, this case has an exact answer to check against, and the series reproduces its Taylor expansion term for term — reassurance that the bookkeeping is sound. Third, you stop whenever you like: two terms already give a root accurate to within a few percent at epsilon = 0.1, and you never paid the price of the quadratic formula.
Lifting it to differential equations
The same engine runs on differential equations, where the unknowns x_0, x_1, x_2 become unknown FUNCTIONS rather than numbers. Take a weakly nonlinear spring, d^2x/dt^2 + x + epsilon x^3 = 0, released from rest at amplitude 1. With epsilon = 0 it is the plain harmonic oscillator whose answer you know cold: x_0(t) = cos t. Substitute x = x_0 + epsilon x_1 + ... and collect powers of epsilon. The order-epsilon equation comes out as d^2x_1/dt^2 + x_1 = -x_0^3 = -cos^3(t), with x_1 starting from rest at zero.
Look at what just happened, because it is the heart of why the method is so powerful for differential equations: every correction x_n solves the SAME differential operator on the left — here d^2/dt^2 + 1, the simple oscillator — with a right-hand side built only from the lower-order solutions you already have. You solve the unperturbed equation once and then reuse its machinery forever, each time with a known forcing term. The hard nonlinear coupling has been demoted to a sequence of standard forced-oscillator problems, the kind Volume I and the earlier rungs taught you to solve.
When it works, when it fails, and how to tell
Regular perturbation succeeds when three conditions hold. The unperturbed problem is genuinely solvable; the small parameter really is small in the regime you care about; and the answer depends smoothly on epsilon, with no sudden change of character at epsilon = 0. When those hold, the series gives accurate answers cheaply, and the error after keeping terms up to epsilon^n is of order epsilon^{n+1} — shrink epsilon by ten and the leading error drops by a factor of ten, a hundred, a thousand as you keep one, two, three corrections.
Now the honest part — three ways the method betrays you, each a doorway to the rest of this rung. The first is the secular term you just saw: the spring example's naive series predicts an amplitude that grows forever, when the real spring oscillates with steady amplitude at a slightly shifted frequency. The series is fine for a few cycles and absurd over long times; the cure is to let the frequency itself depend on epsilon, the Poincaré-Lindstedt method of guide 5. The second is the singular case: if epsilon multiplies the HIGHEST derivative, setting epsilon = 0 lowers the order of the equation and you cannot satisfy all the boundary conditions — a thin boundary layer forms where your regular series is simply wrong.
The third surprise is the deepest, and it should change how you think about series forever: the perturbation series may DIVERGE for every nonzero epsilon and still be magnificent. This is the realm of the asymptotic series — the running theme of this whole rung. A divergent asymptotic series does not converge as you add terms, yet truncated at the right place its partial sum can match the true answer to astonishing precision. The terms first shrink, reach a smallest term, then grow again; you stop at the smallest term and the error is roughly that smallest term's size. "Divergent" sounds like failure but here it is just a different, more honest, kind of promise — exactly as you saw with the asymptotic expansions earlier in this rung.
So how do you tell, before investing hours, whether your problem is regular? Run a quick sanity scan in the spirit of dominant balance: set epsilon = 0 and ask whether the leftover problem is the same TYPE, with the same number of solutions and the same boundary conditions, as the full one. If yes, expand with confidence. If the order drops, a boundary disappears, or a bounded solution wants to blow up, you are in singular territory and a plain power series in epsilon will quietly lie to you. The whole art is knowing which world you are in before you start turning the crank.