Regular Perturbation Theory

Regular perturbation theory is the first tool a physicist reaches for when a problem is hard at general coupling strength but trivial when the coupling is switched off. The idea is disarmingly simple: given $P(y;\varepsilon)=0$ that you cannot solve in closed form, but whose $\varepsilon=0$ version you can, posit that the answer depends smoothly on $\varepsilon$ and write it as a power series $y(\varepsilon)=y_0+\varepsilon y_1+\varepsilon^2 y_2+\cdots$ . Substitute, demand that each power of $\varepsilon$ vanish on its own, and a single hard problem dissolves into an infinite tower of easy ones — the zeroth order returning the unperturbed answer, every higher order a linear problem driven entirely by the orders already solved. The decisive adjective is regular. A perturbation is regular when the $\varepsilon\to 0$ limit preserves the character of the problem — a quadratic stays quadratic, a second-order equation stays second order, an integral keeps its endpoints, no boundary condition is orphaned — so the series is uniformly valid across the whole domain and often outright convergent. When instead $\varepsilon$ multiplies the highest derivative or leading coefficient, the limit is singular: roots run to infinity, thin boundary layers appear, or corrections grow secularly, and the naive series is worthless. Diagnosing which case you are in, before computing a single coefficient, is the central skill this article teaches. The same algebra recurs everywhere — perturbed roots of polynomials, small-parameter expansions of integrals, differential equations, and eigenvalue problems (the last being Rayleigh–Schrödinger theory in quantum mechanics). And although these series often converge, many of the most important ones do not: they are asymptotic in Poincaré's sense, divergent yet stunningly accurate when truncated well. The humble regular expansion is therefore the doorway to the entire asymptotic toolkit — WKB, Watson's Lemma, Borel summation, resurgence — that the rest of this chapter builds.

Key takeaways

It expands an unknown (a root, integral, eigenvalue, or solution of a differential equation) as a power series in a small $\varepsilon$ whose leading term is the exact solution at $\varepsilon=0$ .
It works when the perturbation is regular: the $\varepsilon\to 0$ limit is smooth, the number and type of solutions is preserved, and the series is uniformly valid on the whole domain.
You substitute $y=\sum_n\varepsilon^n y_n$ , collect equal powers of $\varepsilon$ , and solve a triangular hierarchy of linear problems, each driven only by lower orders — the same operator $L$ inverted at every step.
Lost uniform validity — a dropped derivative, escaped roots, or secular growth like $\varepsilon t\cos t$ — signals a singular perturbation needing boundary layers, multiple scales, or WKB.
Even a divergent regular series is powerful: truncated near its smallest term it beats what a zero radius of convergence suggests, linking this method to asymptotic and Borel-summable expansions.
Key idea: posit $y=\sum_{n\ge 0}\varepsilon^n y_n$ with $y_0$ the exact unperturbed solution, substitute, and equate equal powers of $\varepsilon$ to get a triangular hierarchy of linear problems — each order driven solely by lower orders — that is uniformly valid only if the $\varepsilon\to 0$ limit is non-singular.

The core insight

Posit $y=\sum_{n\ge 0}\varepsilon^n y_n$ with $y_0$ the exact unperturbed solution, substitute, and equate equal powers of $\varepsilon$ to obtain a triangular hierarchy of linear problems — each order driven solely by lower orders — that is valid uniformly only if the $\varepsilon\to 0$ limit is non-singular.

Where this lives today

Regular perturbation theory is the daily arithmetic of theoretical physics, not a museum piece. Every quantum-electrodynamics prediction — the electron's anomalous magnetic moment, computed to five loops and agreeing with experiment to better than one part in $10^{12}$ — is a regular expansion in the fine-structure constant $\alpha\approx1/137$ , and the fact that the QED series is ultimately asymptotic (Dyson, 1952) is Poincaré's insight at the frontier. In atomic physics the Stark and Zeeman shifts spectroscopists measure to exquisite precision are first- and second-order matrix elements of exactly the Rayleigh–Schrödinger form used above, underpinning modern frequency-comb metrology. In fluids the regular/singular dichotomy organises aerodynamics: low-Reynolds Stokes flow is a regular expansion, but the thin boundary layers on every wing are the singular case, resolved by the matched asymptotics whose cartoon is Derivation B. And where the series diverges too soon, the method's descendants take over — resurgent trans-series and Borel–Écalle summation now extract exact non-perturbative physics (instantons, renormalons) from the very divergence Poincaré first diagnosed, an active front in quantum field theory and string theory today.

The physical picture

Picture a knob $\varepsilon$ that smoothly deforms an easy problem into a hard one. If turning the knob a little only nudges the answer, the answer is a smooth function of $\varepsilon$ and a Taylor series captures it exactly; regular perturbation theory is just the bookkeeping that extracts that series without ever solving the hard problem in closed form. The hierarchy of orders is triangular — $y_0$ first, then $y_1$ using $y_0$ , then $y_2$ using $y_0,y_1$ — and at every step you invert the same unperturbed operator with a new source, so one well-understood inverse (a Green's function, or the unperturbed eigenbasis) generates the whole series mechanically. The danger sign is non-uniformity — a correction small at one place in the domain but not at another. The first figure makes the regular/singular divide visible with two concrete equations. A quadratic keeps its degree as $\varepsilon\to 0$ , so both roots stay finite and analytic, each seeding its own convergent branch. The cubic $\varepsilon x^3-x+1=0$ instead collapses to the linear $-x+1=0$ at $\varepsilon=0$ : it keeps one root and loses two, which have fled to infinity as $x\sim\pm\varepsilon^{-1/2}$ — the fingerprint of a singular perturbation no regular series can find.

Figure

Regular versus singular. The quadratic x² − (3+ε)x + 2 = 0 keeps degree 2 as ε → 0 so both roots stay finite and analytic (x₊ = 2 + 2ε − 2ε², x₋ = 1 − ε + 2ε²), while the cubic εx³ − x + 1 = 0 collapses to degree 1 and two of its three roots escape as x ≈ ±1/√ε

Even a divergent regular series can be the most accurate tool available, because its terms first shrink dramatically before they eventually grow. The second figure shows the property that earns the method its name: a genuinely convergent example, watched order by order. For the quadratic above at $\varepsilon=0.1$ the exact lower root is $0.91557\ldots$ , and the partial sums $S_0=1.000$ , $S_1=0.900$ , $S_2=0.920$ bracket it and close in, each order shaving the error by roughly a factor $\varepsilon$ . That geometric collapse is the payoff of regularity — and its breakdown, in the anharmonic oscillator below, is the birth of asymptotics.

Figure

Order-by-order convergence at ε = 0.1. The partial sums of x₋ = 1 − ε + 2ε² close in on the exact root 0.9156: S₀ = 1.000, S₁ = 0.900, S₂ = 0.920, with the absolute error falling 0.084 → 0.016 → 0.004, roughly one power of ε per order

The core equations

Everything begins with one act of faith made precise: the answer is a power series in the small parameter whose leading coefficient is the exact solution you already own. This ansatz is the method; the rest is consequences of substituting it.

y(\varepsilon)=y_0+\varepsilon\,y_1+\varepsilon^2 y_2+\cdots=\sum_{n=0}^{\infty}\varepsilon^n y_n \qquad (1)

where: $\varepsilon$ — small dimensionless parameter, $0<\varepsilon\ll 1$ ; $y_0$ — exact solution of the unperturbed ( $\varepsilon=0$ ) problem; $y_n$ — the $n$ -th order correction, found recursively. Reading: the leading term is exact, not approximate; the approximation lives entirely in truncating the sum. The engine that turns one ansatz into infinitely many solvable problems is the demand that each power of $\varepsilon$ balance separately: an identity holding for all small $\varepsilon$ forces every coefficient to vanish on its own.

P(y;\varepsilon)=P_0(y)+\varepsilon\,P_1(y)+\cdots=0 \;\Rightarrow\; \text{collect each }\varepsilon^n \qquad (2)

where: $P(y;\varepsilon)$ — the full problem (operator, equation, or functional); $P_0,P_1$ — the unperturbed operator and first perturbing piece. Reading: "equate powers of $\varepsilon$ " is the one move that converts a single equation into a solvable tower. The simplest incarnation is a perturbed root of a polynomial or transcendental equation, where the whole method reduces to one division — the cleanest place to see both mechanism and failure mode.

x(\varepsilon)=x_0+\varepsilon x_1+\varepsilon^2 x_2+\cdots,\qquad f(x_0)=0,\quad x_1=-\frac{g(x_0)}{f'(x_0)} \qquad (3)

where: $f(x)+\varepsilon g(x)=0$ — the perturbed equation; $x_0$ — a simple root of the unperturbed equation $f(x_0)=0$ ; $f'(x_0)$ — the derivative at the root, which must be non-zero for a regular expansion. Reading: the correction is small exactly when the root is simple; a double root ( $f'=0$ ) makes $x_1$ blow up and forces a fractional-power (singular) expansion. Linearising about $y_0$ shows why the same operator reappears at every order: the unknown $y_n$ always enters through the Fréchet derivative $L\equiv P_0'(y_0)$ , while all $\varepsilon$ -dependence and lower-order data are pushed into the source.

L\,y_n = R_n[y_0,\dots,y_{n-1}],\qquad L\equiv P_0'(y_0) \qquad (4)

where: $L$ — the unperturbed linear operator, common to every order; $R_n$ — the source at order $n$ , built entirely from already-known lower orders. Reading: invert $L$ once (a Green's function or an eigenbasis) and you can grind out the whole series, because only the right-hand side changes. In quantum mechanics this machinery is Rayleigh–Schrödinger theory. A solvability (Fredholm) condition — orthogonality of the source to the null vector of $H_0-E_0$ — fixes the first-order energy without ever solving for the first-order state.

\big(H_0-E_0\big)|\psi_0\rangle=0,\qquad E_1=\langle\psi_0|V|\psi_0\rangle \qquad (5)

where: $H_0$ — the unperturbed Hamiltonian with eigenpair $(E_0,\psi_0)$ ; $V$ — the perturbation, $H=H_0+\varepsilon V$ ; $E_1$ — the first-order energy correction. Reading: the first-order shift is just the expectation value of the perturbation in the unperturbed state — a diagonal matrix element. Pushing to the next order forces the intermediate states to appear, weighted by energy denominators — which is where near-degeneracy becomes dangerous, since a small gap can make a "correction" enormous.

E_2=\sum_{m\neq 0}\frac{|\langle\psi_m|V|\psi_0\rangle|^2}{E_0-E_m} \qquad (6)

where: $E_m,\psi_m$ — the other unperturbed eigenpairs; $E_0-E_m$ — the energy denominator. Reading: for the ground state every denominator is negative, so $E_2<0$ always lowers it; small gaps ( $E_0\approx E_m$ ) break the regular series and demand degenerate perturbation theory. The same expand-and-collect logic works on definite integrals: expand the integrand and integrate term by term, converting one hard integral into a series of moments.

I(\varepsilon)=\int_a^b f(x)\,\mathrm{e}^{-\varepsilon h(x)}\,\mathrm{d}x=\sum_{n=0}^{\infty}\frac{(-\varepsilon)^n}{n!}\int_a^b f(x)\,h(x)^n\,\mathrm{d}x \qquad (7)

where: $f,h$ — smooth integrand factors with all moments finite; $\varepsilon$ — the small parameter. Reading: the integral simply inherits the regular series of its integrand; when the parameter sits in an exponent over an infinite range, this dovetails into Watson's Lemma and Laplace's method. Finally, what makes even divergent series respectable is Poincaré's definition of an asymptotic expansion: for fixed truncation $N$ , the remainder is controlled as $\varepsilon\to 0$ , whether or not the sum converges.

\Big|\,y(\varepsilon)-\sum_{n=0}^{N}\varepsilon^n y_n\,\Big|=O(\varepsilon^{N+1})\quad(\varepsilon\to 0) \qquad (8)

where: $N$ — the (fixed) truncation order; $O(\varepsilon^{N+1})$ — the remainder bound. Reading: convergence takes $N\to\infty$ at fixed $\varepsilon$ ; asymptoticity takes $\varepsilon\to 0$ at fixed $N$ — a divergent series can still be superbly asymptotic.

Derivation APerturbed roots of a quadratic in the regular case

The goal is to find both roots of the perturbed quadratic $x^2-(3+\varepsilon)x+2=0$ as convergent power series in $\varepsilon$ , to show that order-by-order matching and the exact quadratic formula agree coefficient for coefficient, and to see explicitly why a preserved degree is what guarantees a regular expansion in which every root stays finite and analytic. We run the order-by-order algebra for the branch seeded by $x_0=1$ , quote the twin branch, and confirm both against the closed form.

Step 1. Set $\varepsilon=0$ : the equation becomes $x^2-3x+2=(x-1)(x-2)=0$ , so the unperturbed roots are $x_0=1$ and $x_0=2$ , and the degree is still two. Why: two distinct simple roots survive the limit, so we expect two independent regular series, each seeded by one exact root.

Step 2. Expand the branch about $x_0=1$ as $x=1+\varepsilon x_1+\varepsilon^2 x_2+\cdots$ and substitute into $x^2-(3+\varepsilon)x+2=0$ . The pieces are $x^2=1+2\varepsilon x_1+\varepsilon^2(2x_2+x_1^2)+\cdots$ and $-(3+\varepsilon)x=-3-3\varepsilon x_1-\varepsilon-\varepsilon^2 x_1-3\varepsilon^2 x_2+\cdots$ , plus the constant $+2$ . Why: writing every term to the order we want keeps the collection honest and prevents dropping a cross term.

Step 3. Collect powers. At order $\varepsilon^0$ : $1-3+2=0$ , confirming the seed. At order $\varepsilon^1$ : $2x_1-3x_1-1=0$ , i.e. $-x_1-1=0$ , so $x_1=-1$ . Why: the same linear coefficient $2x_0-3=-1$ multiplies $x_1$ ; it is non-zero precisely because $x_0$ is a simple root, which is what makes the division legal.

Step 4. At order $\varepsilon^2$ : $2x_2+x_1^2-3x_2-x_1=0$ , i.e. $-x_2+x_1^2-x_1=0$ . With $x_1=-1$ this is $-x_2+1+1=0$ , so $x_2=2$ . Hence $x_-(\varepsilon)=1-\varepsilon+2\varepsilon^2+O(\varepsilon^3)$ , and the identical procedure about $x_0=2$ gives $x_+(\varepsilon)=2+2\varepsilon-2\varepsilon^2+O(\varepsilon^3)$ . Why: each new order is a linear equation for one unknown coefficient with the same $-1$ multiplying it — the triangular hierarchy in miniature.

Step 5. Cross-check with the quadratic formula: $x_\pm=\tfrac12\big[(3+\varepsilon)\pm\sqrt{1+6\varepsilon+\varepsilon^2}\big]$ , and expanding the surd, $\sqrt{1+6\varepsilon+\varepsilon^2}=1+3\varepsilon-4\varepsilon^2+O(\varepsilon^3)$ , reproduces $x_+=2+2\varepsilon-2\varepsilon^2$ and $x_-=1-\varepsilon+2\varepsilon^2$ exactly. The discriminant vanishes at $\varepsilon=-3+2\sqrt2\approx-0.172$ , where the two roots collide into a double root — that collision sets the radius of convergence of both series. Why: the closed form both verifies the recursion and reveals the convergence disk, which the coefficients alone would only hint at through their growth. *Verified: order-by-order matching and the surd expansion agree to $O(\varepsilon^2)$ (SymPy series); the nearest branch-point at $\varepsilon=-3+2\sqrt2$ gives radius $\approx 0.172$ .*

Go deeper · Derivation B · 2 minFailure of regularity in the boundary-layer equation

Derivation BFailure of regularity in the boundary-layer equation

The goal is to show that the innocuous-looking two-point problem $\varepsilon y''+y'+y=0$ with $y(0)=0$ , $y(1)=1$ cannot be captured by any single regular power series in $\varepsilon$ , and to see exactly how the structure breaks: because $\varepsilon$ multiplies the highest derivative, the $\varepsilon\to 0$ equation drops from second order to first, one boundary condition can no longer be enforced, and a thin boundary layer must reinstate it. This is the archetype every physicist must recognise on sight — a singular perturbation masquerading as a small one.

Step 1. Try the naive regular ansatz. At leading order the $\varepsilon y''$ term drops and $y_0'+y_0=0$ , so $y_0=C\mathrm{e}^{-x}$ — a first-order equation with a single free constant. Why: one constant cannot satisfy two boundary conditions, so the regular series is dead on arrival; the very act of setting $\varepsilon=0$ has destroyed a solution.

Step 2. Decide which condition to keep. The physically correct choice keeps the right boundary $y(1)=1$ , giving $C\mathrm{e}^{-1}=1$ and the outer solution $y_{\text{outer}}(x)=\mathrm{e}^{\,1-x}$ ; the left condition $y(0)=0$ is left violated for now. Why: away from $x=0$ the neglected term $\varepsilon y''$ really is small, so the outer solution is trustworthy there — the failure is localised near one end.

Step 3. Resolve the missing end by rescaling. Set $X=x/\varepsilon$ ; then $\varepsilon y''=\varepsilon^{-1}y_{XX}$ and $y'=\varepsilon^{-1}y_X$ , and the equation becomes $y_{XX}+y_X+\varepsilon y=0$ , whose leading order $y_{XX}+y_X=0$ is again second order. Why: stretching the coordinate by $1/\varepsilon$ magnifies the thin region so the highest derivative reasserts itself — the lost solution reappears inside the layer.

Step 4. Solve the inner problem $y_{\text{inner}}=A+B\mathrm{e}^{-X}$ . Impose $y(0)=0$ : $A+B=0$ . Match to the outer solution as $X\to\infty$ , where $y_{\text{outer}}(0)=\mathrm{e}$ : this forces $A=\mathrm{e}$ , hence $B=-\mathrm{e}$ and $y_{\text{inner}}=\mathrm{e}\big(1-\mathrm{e}^{-X}\big)$ . Why: matching is the statement that the inner and outer descriptions agree in the overlap region, which glues the two incomplete solutions into one uniform answer.

Step 5. Conclude. A boundary layer of thickness $O(\varepsilon)$ at $x=0$ carries the solution from $0$ up to the outer value $\mathrm{e}$ ; no single power series in $\varepsilon$ is uniformly valid, which is the defining signature of a singular perturbation and the entry point to matched asymptotic expansions. Why: the exact solution $y=(\mathrm{e}^{r_+x}-\mathrm{e}^{r_-x})/(\mathrm{e}^{r_+}-\mathrm{e}^{r_-})$ with $r_\pm=(-1\pm\sqrt{1-4\varepsilon})/2\varepsilon$ has $r_-\sim-1/\varepsilon$ , the fast mode the regular series could never produce. *Verified: matched inner+outer construction reproduces the exact two-exponential solution in the limit $\varepsilon\to 0$ ; $r_+\to-1$ gives the outer decay, $r_-\to-\infty$ the boundary layer.*

Go deeper · Derivation C · 1 minAnharmonic oscillator ground-state energy to first order

Derivation CAnharmonic oscillator ground-state energy to first order

The goal is to compute the first-order shift of the harmonic-oscillator ground-state energy under a quartic perturbation $\varepsilon x^4$ using Rayleigh–Schrödinger theory, to obtain the clean result $E_0(\varepsilon)=\tfrac12+\tfrac34\varepsilon+O(\varepsilon^2)$ , and to flag the surprise behind it: although every coefficient is finite and elementary, the full series diverges — making this the canonical bridge from regular perturbation theory to the asymptotic and Borel-summation methods of the rest of the chapter. We work in units $\hbar=m=\omega=1$ with $H=\tfrac12 p^2+\tfrac12 x^2+\varepsilon x^4$ .

Step 1. Take the unperturbed data: $E_0^{(0)}=\tfrac12$ and the normalised ground state $\psi_0(x)=\pi^{-1/4}\mathrm{e}^{-x^2/2}$ . The perturbation is $V=x^4$ . Why: Rayleigh–Schrödinger builds every correction from the exactly known unperturbed eigenstate, so this Gaussian is all the input the first order needs.

Step 2. Apply the first-order formula (5): $E_1=\langle\psi_0|x^4|\psi_0\rangle$ , the expectation of $x^4$ in the ground state. Why: the first-order shift is always the diagonal matrix element of the perturbation — no knowledge of the corrected state is required, thanks to the solvability condition.

Step 3. Evaluate the Gaussian moment: $\langle x^4\rangle=\dfrac{1}{\sqrt\pi}\displaystyle\int_{-\infty}^{\infty}x^4\mathrm{e}^{-x^2}\,\mathrm{d}x=\dfrac{1}{\sqrt\pi}\cdot\dfrac{3\sqrt\pi}{4}=\dfrac34$ , using the standard result $\int x^4\mathrm{e}^{-x^2}\mathrm{d}x=\tfrac34\sqrt\pi$ . Why: the fourth moment of a normalised Gaussian is $3\sigma^4$ with $\sigma^2=\tfrac12$ here, giving $3\cdot\tfrac14=\tfrac34$ — a one-line integral that fixes the entire first-order physics.

Step 4. Assemble: $E_0(\varepsilon)=\tfrac12+\tfrac34\varepsilon+O(\varepsilon^2)$ . The next coefficients, from the Bender–Wu recursion, are $-\tfrac{21}{8}$ , $+\tfrac{333}{16}$ , $-\tfrac{30885}{128},\ldots$ and grow factorially: the series has zero radius of convergence. Why: the quartic well has no bound states for $\varepsilon<0$ , so the energy is singular at $\varepsilon=0$ and its Taylor series can only be asymptotic — yet truncating near the smallest term still gives excellent numbers, which is the whole justification for perturbation theory in quantum field theory. *Verified: $\langle x^4\rangle=3/4$ exactly (SymPy Gaussian integral); leading shift $E_1=3/4$ ; factorial growth of higher coefficients is the classic Bender–Wu result.*

Limiting cases

Whether a small term is a harmless regular perturbation or a treacherous singular one is decided entirely by what the $\varepsilon\to 0$ limit does to the structure of the problem — derivative order, polynomial degree, domain finiteness. The table collects the four canonical outcomes.

Case	Result	Meaning
Regular (uniform)	$y=\sum_n \varepsilon^n y_n$ , e.g. $\ddot y+y=\varepsilon\,y^3$	Same order/degree and same number of solutions as $\varepsilon=0$ ; the series is uniformly valid on the whole (bounded) domain.
Singular — lost order	$\varepsilon\,y''+y'+y=0$	$\varepsilon$ multiplies the highest derivative; at $\varepsilon=0$ the equation drops to first order and one boundary condition cannot be met — a layer of width $O(\varepsilon)$ appears.
Singular — lost roots	$\varepsilon x^3-x+1=0$	Degree falls $3\to 1$ as $\varepsilon\to 0$ ; two roots escape as $x\sim\pm\varepsilon^{-1/2}$ , a fractional-power expansion no integer series can reach.
Singular — secular growth	$\ddot y+y=-2\varepsilon\,\dot y$	The naive series throws up terms like $\varepsilon t\cos t$ that grow on the slow scale $t\sim1/\varepsilon$ ; valid only for $t=O(1)$ , cured by multiple scales or Poincaré–Lindstedt.

Common trapAssuming a regular power series always exists

The most common and costly mistake is to write $y=\sum_n\varepsilon^n y_n$ reflexively and start grinding, without checking that the $\varepsilon\to 0$ limit is non-singular. Three warning signs doom the integer-power ansatz. First, if $\varepsilon$ multiplies the highest derivative (or a polynomial's leading coefficient), the limiting problem loses order — you need boundary layers or WKB. Second, if the honest answer contains $\varepsilon^{1/2}$ , $\varepsilon^{-1/2}$ , or $\ln\varepsilon$ , no integer-power series can represent it; you must guess the correct gauge functions first (the job of the order-symbols article preceding this one). Third, if $\varepsilon y_1/y_0$ grows without bound anywhere (secular terms), the series is non-uniform even though every coefficient is finite. Push the naive series in any of these cases and you get confident nonsense — recognise the singular structure and switch methods.

Worked examples

1 · Convergence of a perturbed root, to the digit (fully worked)

For $x^2-(3+\varepsilon)x+2=0$ at $\varepsilon=0.1$ , compute the lower root exactly and compare with the zeroth-, first-, and second-order truncations of its perturbation series. Quantify how fast the error falls.

Step 1. Exact value. The quadratic formula gives $x_-=\tfrac12\big[(3.1)-\sqrt{3.1^2-8}\big]=\tfrac12\big[3.1-\sqrt{1.61}\big]=\tfrac12(3.1-1.268858)=\mathbf{0.915571}$ .

Step 2. Perturbative truncations, from $x_-(\varepsilon)=1-\varepsilon+2\varepsilon^2$ : the zeroth order is $S_0=1.000000$ , the first order is $S_1=1-0.1=0.900000$ , and the second order is $S_2=0.900+2(0.01)=\mathbf{0.920000}$ .

Step 3. Errors. $|S_0-x_-|=0.084429$ , $|S_1-x_-|=0.015571$ , $|S_2-x_-|=0.004429$ . The ratios $0.015571/0.084429\approx0.18$ and $0.004429/0.015571\approx0.28$ show each order shrinking the error by roughly one power of $\varepsilon$ — the hallmark of a convergent regular series inside its disk $|\varepsilon|<0.172$ .

Step 4. The truncations bracket the exact root ( $S_0$ above, $S_1$ below, $S_2$ above) and spiral inward, as the second figure shows. At $\varepsilon=0.1$ two terms already pin the root to better than $0.5\%$ : $x_-\approx 0.920$ versus the exact $0.9156$ .

2 · Second-order shift in a two-level system (energy denominators with numbers)

Take $H=\begin{pmatrix}0&\varepsilon\\\varepsilon&1\end{pmatrix}$ (so $E_a=0$ , $E_b=1$ , off-diagonal coupling $\varepsilon\Delta$ with $\Delta=1$ ). Find the perturbative shifts of both levels and test them against the exact eigenvalues at $\varepsilon=0.1$ .

Step 1. First order. Since $V=\begin{pmatrix}0&1\\1&0\end{pmatrix}$ is purely off-diagonal, $E_1=\langle a|V|a\rangle=0$ for the lower state and likewise $0$ for the upper — no linear shift. The levels do not move at first order because the perturbation has no diagonal part.

Step 2. Second order, from (6). For the lower state, the only other state is $|b\rangle$ , so $E_2^{(a)}=\dfrac{|\langle b|V|a\rangle|^2}{E_a-E_b}=\dfrac{1}{0-1}=\mathbf{-1}$ . For the upper state, $E_2^{(b)}=\dfrac{|\langle a|V|b\rangle|^2}{E_b-E_a}=\dfrac{1}{1-0}=\mathbf{+1}$ . The levels repel: the lower one is pushed down, the upper one up.

Step 3. Perturbative energies: $E_a(\varepsilon)\approx 0-\varepsilon^2=-\varepsilon^2$ and $E_b(\varepsilon)\approx 1+\varepsilon^2$ . At $\varepsilon=0.1$ these are $-0.010000$ and $1.010000$ .

Step 4. Exact check. Diagonalising gives $\lambda_\pm=\tfrac12\pm\tfrac12\sqrt{1+4\varepsilon^2}$ ; at $\varepsilon=0.1$ , $\sqrt{1.04}=1.019804$ , so $\lambda_-=\mathbf{-0.009902}$ and $\lambda_+=\mathbf{1.009902}$ . The second-order formula matches the exact eigenvalues to four decimals (error $\sim10^{-4}=O(\varepsilon^4)$ ), and the sign of every energy denominator correctly predicts level repulsion.

Physics of the portrait

Regular perturbation theory has no single inventor — Newton expanded, Laplace and Lagrange perturbed the planets, Rayleigh perturbed vibrations — but one figure turned the folk-craft of "expand in a small parameter" into a rigorous theory and, crucially, understood why the resulting series so often diverge yet still work. That figure is Henri Poincaré, who in the same decade both founded the modern theory of asymptotic series and, in his three-volume treatise on celestial mechanics, mapped the exact places where naive perturbation series break down.

Origins · 1 minHenri Poincaré and how this was discovered

Henri Poincaré

1854–1912 · asymptotic series and celestial-mechanics perturbation theory

Henri Poincaré was the last universalist of mathematics, and perturbation theory sits near the centre of his legacy. Working on the three-body problem for King Oscar II's prize (1887–1890), he confronted the perturbation series astronomers had used for a century to predict planetary motion — and found that they generically diverge, undone by small denominators and orbital resonances. Rather than discard them, he asked how a divergent series can still be useful. His 1886 memoir gave the answer that still stands, defining an asymptotic expansion by the property that a fixed truncation has error $O(\varepsilon^{N+1})$ as $\varepsilon\to 0$ , convergent or not — legitimising the entire enterprise, since the perturbation series of physics need not converge to be predictive. In Les Méthodes Nouvelles de la Mécanique Céleste (1892–1899) he catalogued the failures of regular expansions — secular terms growing like $\varepsilon t$ , resonant denominators, homoclinic tangles — opening both dynamical-systems theory and the modern regular/singular distinction. The tools this article ends with (optimal truncation, Stokes phenomenon, Borel summation, resurgence) are all downstream of his insistence that a divergent series is not a failure but a structure to be understood.

History

1877

— Rayleigh perturbs the eigenfrequencies of vibrating systems in Theory of Sound

1886

— Poincaré defines the asymptotic series, legitimising divergent-but-useful expansions

1892

— Poincaré's Méthodes Nouvelles de la Mécanique Céleste anatomises secular and resonant breakdown

1926

— Schrödinger casts stationary-state perturbation theory in wave mechanics

1969

— Bender and Wu prove the quartic-oscillator series diverges factorially and analyse its Borel structure

Going further

Texts. Bender & Orszag, Advanced Mathematical Methods for Scientists and Engineers, Ch. 7 (the definitive perturbation-theory reference) · Hinch, Perturbation Methods, Ch. 1 · Stone & Goldbart, Mathematics for Physics, chapter on perturbation methods · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., perturbation sections · Hassani, Mathematical Physics, and Byron & Fuller, Mathematics of Classical and Quantum Physics, for the linear-operator and Green's-function viewpoint · Nayfeh, Perturbation Methods, for the singular cases.

Provenance. A verifier should recompute: the quadratic-root series via both order-by-order matching and the quadratic formula, confirming $x_+=2+2\varepsilon-2\varepsilon^2$ , $x_-=1-\varepsilon+2\varepsilon^2$ and the convergence radius $|\varepsilon|<3-2\sqrt2\approx0.172$ (note this corrects a transcription in the thin migrated draft, whose box read $2+\varepsilon+\varepsilon^2$ ); the Gaussian moment $\langle x^4\rangle=3/4$ giving $E_1=3\varepsilon/4$ (Bender–Orszag §7); the two-level second-order shifts $E_2^{(a)}=-1$ , $E_2^{(b)}=+1$ against the exact $\tfrac12\pm\tfrac12\sqrt{1+4\varepsilon^2}$ ; and the boundary-layer matched solution against the exact two-exponential form with $r_\pm=(-1\pm\sqrt{1-4\varepsilon})/2\varepsilon$ . Standard references for all four: Bender & Orszag, Hinch, Stone & Goldbart.

Prerequisites. Order Symbols and Gauge Functions · Taylor's theorem and power series · Linear ordinary differential equations. Leads to. Asymptotic Series and Poincaré Expansions · WKB Approximation · Watson's Lemma · Borel Summation and Resurgence.

Ask the tutor: Bring me any equation with a small parameter and I will show you in one line whether a regular series will work — and if it won't, which singular method to reach for instead.

How do I decide, before computing anything, whether to use a regular expansion?

Set $\varepsilon=0$ and stare at what is left. Ask: does a differential equation keep its order, a polynomial its degree, and the domain stay finite with all boundary conditions still enforceable? If yes to all three, the perturbation is regular and the integer-power ansatz $y=\sum_n\varepsilon^n y_n$ is safe. If the order drops, roots escape, a limit runs to infinity, or a boundary condition becomes unenforceable, it is singular — reach for matched asymptotics, multiple scales, or WKB. This thirty-second check saves hours, because every failure mode is already visible in the $\varepsilon=0$ problem before you compute a single coefficient.

Why does the same linear operator show up at every order?

Because perturbation theory is Newton's method in disguise. Expanding $P(y;\varepsilon)=0$ about $y_0$ , the unknown at each order enters only through the linearisation $L\equiv P_0'(y_0)$ , while all $\varepsilon$ -dependence and lower-order data collect into the source $R_n$ . So the structure is always $L\,y_n=R_n$ with the same $L$ and only the source changing. One inverse of $L$ — a Green's function, or the unperturbed eigenbasis — then generates the entire series, which is why the second-order energy is just a weighted sum over unperturbed states.

What exactly is a secular term and why is it fatal?

A secular term grows without bound in an independent variable — the classic case is $\varepsilon t\cos t$ in a weakly damped or nonlinear oscillator. Its coefficient $\varepsilon$ is small, so at short times it looks harmless, but growing with $t$ it eventually overtakes $y_0$ : once $\varepsilon t\sim 1$ , i.e. $t\sim1/\varepsilon$ , the "correction" is no longer a correction. The series is then non-uniform — accurate for $t=O(1)$ but useless on the long timescale where the interesting physics (slow amplitude decay, frequency shift) lives. The cure is a better ansatz, not more terms: Poincaré–Lindstedt rescales the frequency, and multiple scales introduces a slow time $T=\varepsilon t$ so the growth is absorbed into a smoothly evolving amplitude.

If the QED perturbation series diverges, why do we trust its predictions?

Because divergent does not mean wrong — it means asymptotic. Dyson argued in 1952 that the QED series must diverge (flipping the coupling's sign destabilises the vacuum, so the function is not analytic at zero coupling), yet the first terms of the electron's anomalous moment agree with experiment to twelve figures. Poincaré's definition explains it: a fixed truncation has error $O(\varepsilon^{N+1})$ , and for $\alpha\approx1/137$ the terms shrink for a very long time before factorial growth wins. Truncating near the smallest term leaves an error exponentially small in $1/\alpha$ . The divergence only matters beyond that floor — precisely where Borel summation and resurgence take over.

Why must f′(x₀) ≠ 0 for a regular root expansion?

Because the first-order correction is $x_1=-g(x_0)/f'(x_0)$ , and a vanishing derivative makes it blow up. Geometrically $f'(x_0)=0$ means $x_0$ is a double root where two roots have merged; under a perturbation they split apart as a fractional power, $x_0\pm c\sqrt\varepsilon$ , so the correct ansatz is a series in $\varepsilon^{1/2}$ , not $\varepsilon$ . Forcing an integer-power series there divides by zero at first order — the algebra telling you the gauge is wrong. This is the polynomial analogue of degenerate perturbation theory, where a vanishing energy gap plays the role of the vanishing derivative.

How is regular perturbation of an integral related to Watson's Lemma?

They are two ends of the same idea. If the small parameter multiplies a bounded smooth factor over a finite interval, just expand the integrand in $\varepsilon$ and integrate term by term (equation 7); the integral inherits the regular series of its integrand. But when the parameter sits in an exponent over a semi-infinite range — $\int_0^\infty f(t)\mathrm{e}^{-t/\varepsilon}\,\mathrm{d}t$ — naive expansion fails, because the exponential concentrates all the weight near one endpoint. There the tool is Watson's Lemma, which expands only the slow factor near that endpoint, yielding an asymptotic series. The finite regular case is the gentle cousin; Watson's Lemma and Laplace's method are what it becomes once an endpoint runs to infinity.