Degenerate Perturbation Theory

Non-Degenerate Time-Independent Perturbation Theory built a beautiful machine and posted one warning on its side: no two unperturbed levels may share an energy. This article is what happens when the warning is ignored — and then the repair. The mixing formula of First-Order Energy and State Corrections divides every coupling by a gap; the borrowed-strength sum of Second-Order Energy Correction does it again, squared. Let two states share $E^{(0)}$ while the perturbation connects them, and both formulas divide finite numbers by zero. The series does not degrade politely; it detonates on page one. The detonation is a bluff, and calling it is the whole subject. Nothing physical diverges — the exact levels move smoothly as the perturbation switches on. What diverges is a choice: inside a degenerate subspace, $\hat{H}_0$ cannot tell its own eigenstates apart, so "the state we perturb around" was never well-defined; pick arbitrarily and the formulas punish you with infinities. The repair is to let the perturbation choose: restrict $\hat{H}'$ to the subspace and diagonalize the resulting $g\times g$ matrix $W$ — eigenvalues are the first-order splittings, eigenvectors the correct zeroth-order states, the combinations the exact eigenstates actually collapse to as $\lambda\to0$ . That one eigenvalue problem — its characteristic equation still called the secular equation, a word borrowed from planetary astronomy — gates this chapter's payoff run: fine structure, the Zeeman regimes, and the Stark effect are all degenerate problems in disguise. Every matrix element, splitting, and grid comparison below is machine-verified.

Key takeaways

The non-degenerate formulas divide by $E_n^{(0)}-E_m^{(0)}$ : a degenerate partner with $H'_{mn}\ne0$ makes the state correction and $E^{(2)}$ divergent — not physically, but because the zeroth-order basis was ambiguous and chosen wrong.
The cure: form $W_{ab}=\langle n_a^{(0)}|\hat{H}'|n_b^{(0)}\rangle$ on the $g$ -fold subspace and diagonalize. Eigenvalues = first-order shifts $E^{(1)}$ (the level generally splits); eigenvectors = the correct zeroth-order states.
Two-fold case in closed form: $E^{(1)}_\pm=\tfrac12(W_{aa}+W_{bb})\pm\tfrac12\sqrt{(W_{aa}-W_{bb})^2+4|W_{ab}|^2}$ , minimum separation $2|W_{ab}|$ — the avoided-crossing gap (verified symbolically).
In the eigenbasis of $W$ the dangerous terms have zero numerator: every divergence cancels and the non-degenerate machinery restarts, with the second-order sum running only over states outside the subspace.
Worked lab (all numbers verified against exact diagonalization): 2D square well pair $\psi_{12},\psi_{21}$ under $\hat{H}'=\lambda xy$ — good states $(\psi_{12}\pm\psi_{21})/\sqrt2$ , splitting $2\lambda\cdot256L^2/81\pi^4=6.49$ meV at $\lambda=0.1$ eV/nm², $L=1$ nm.
Degeneracy may lift only partially: a rank-deficient $W$ (worked: a point bump splits one state off a triplet, leaving a surviving doublet) hands the remainder to second order — and near-degeneracy to Brillouin-Wigner Perturbation Theory.
Key idea: inside a degenerate subspace, the perturbation picks its own basis — diagonalize $W$ there, read the splittings from its eigenvalues, and perturb around its eigenvectors.

The core insight

Inside a degenerate subspace, diagonalize the perturbation matrix $W_{ab}=\langle n_a^{(0)}|\hat{H}'|n_b^{(0)}\rangle$ : its eigenvalues are the first-order splittings and its eigenvectors are the correct zeroth-order states — the unique combinations toward which the exact eigenstates collapse as the perturbation switches off.

Where this lives today

Avoided crossings — the $2|W_{ab}|$ gap of Derivation B — are now a calibration tool: superconducting-qubit spectroscopy sweeps a flux bias through a degeneracy and reads the coupling strength directly off the minimum gap, the two-level closed form serving as the fitting function of a thousand cooldown logs. Conical intersections in photochemistry are the secular problem's multi-dimensional descendants (two available knobs loosen the no-crossing rule), and valley/orbital degeneracies in quantum-dot devices are lifted by strain and gates — a designed $W$ in device clothing. In quantum chemistry, CASSCF and its relatives are industrial-scale degenerate perturbation theory: diagonalize exactly inside the near-degenerate subspace, perturb the rest — this article's Step 3 at $g\sim10^6$ .

The physical picture

Picture the degenerate subspace as the bottom of a perfectly round bowl. A marble there has no preferred direction — exactly a $g$ -fold level under $\hat{H}_0$ , where any orthonormal frame is an equally legitimate "eigenbasis." Tilt the bowl by a hair: suddenly there is a lowest direction and a highest, and they are properties of the tilt, not the bowl. Degenerate perturbation theory is the act of finding the tilt's principal axes before doing anything else.

Figure

One degenerate rung, tilted: the perturbation picks the axes and fans the level

The divergence of the naive formulas is a coordinate singularity, not a physical one — longitude becoming meaningless at the north pole. The exact eigenstates move smoothly with $\lambda$ , but as $\lambda\to0$ they converge to one specific frame; if your starting vectors differ from it by a finite rotation, "first order" is asked to undo a finite rotation with an infinitesimal parameter — which is what an infinite coefficient is trying to say. Non-Degenerate Time-Independent Perturbation Theory saw this coming from the other side: its solvability audit noted that a degenerate partner enlarges the null space of $(\hat{H}_0-E_n^{(0)})$ and imposes extra conditions the naive ansatz cannot satisfy. Those extra conditions are the secular equation below. There is also a bluff-free preview of the outcome. Exact levels repel — Second-Order Energy Correction proved the pairwise push — so states that start at one energy get shoved apart into a fan, and first order in the right basis reads the fan's spacings off a small hermitian matrix. Everything in this chapter's application run — fine structure's $j$ -ladders, the Zeeman fans, the Stark doublet — is this picture wearing different operators.

Figure

The λxy laboratory: the tilt selects diagonal combinations of the degenerate pair

Where exactly the machinery breaks

Write the failure before the fix. For an isolated level, First-Order Energy and State Corrections delivered the mixing coefficients and Second-Order Energy Correction the energy bill:

|n^{(1)}\rangle=\sum_{m\neq n}\frac{H'_{mn}}{E_n^{(0)}-E_m^{(0)}}\,|m^{(0)}\rangle,\qquad E_n^{(2)}=\sum_{m\neq n}\frac{|H'_{mn}|^2}{E_n^{(0)}-E_m^{(0)}} \qquad (1)

where: $H'_{mn}=\langle m^{(0)}|\hat{H}'|n^{(0)}\rangle$ , and both sums are owned by the two serials just named — quoted here only to read their common failure mode off their faces. Let state $|m^{(0)}\rangle$ be a degenerate partner: $E_m^{(0)}=E_n^{(0)}$ with $H'_{mn}\neq0$ . One term in each sum acquires a zero denominator under a finite numerator. The first-order state correction and the second-order energy are both infinite — for every basis choice in the subspace except the special ones this article constructs. The diagnosis is sharper than "divide by zero." Project the order- $\lambda$ equation onto a degenerate partner $\langle m^{(0)}|$ : the left side dies identically (same eigenvalue), leaving $H'_{mn}=0$ as a requirement — a condition on the perturbation that is generically false. Unless your basis already makes $\hat{H}'$ diagonal inside the subspace, the expansion was inconsistent from its first line; the divergence is the inconsistency's invoice arriving.

The secular problem

Set the degenerate stage. A level $E_n^{(0)}$ of $\hat{H}_0$ carries $g$ orthonormal eigenstates:

\hat{H}_0|n_a^{(0)}\rangle=E_n^{(0)}|n_a^{(0)}\rangle,\qquad a=1,\dots,g \qquad (2)

where: $g$ is the degeneracy and the frame $\{|n_a^{(0)}\rangle\}$ is any orthonormal basis of the subspace — $\hat{H}_0$ cannot prefer one frame; the perturbation can and does. Degeneracy is not exotic: symmetry manufactures it wholesale (Symmetries and Conservation in Quantum Dynamics proved non-abelian symmetry makes it compulsory), so this article's situation is the rule in three dimensions, not the exception. The unknown is which combination inside the subspace to perturb around:

|n^{(0)}\rangle=\sum_{a=1}^{g}c_a\,|n_a^{(0)}\rangle \qquad (3)

where: the coefficients $c_a$ are not free to choose — they are determined by demanding the perturbation series be consistent at first order. The object that decides everything is the perturbation restricted to the subspace,

W_{ab}=\langle n_a^{(0)}|\hat{H}'|n_b^{(0)}\rangle \qquad (4)

where: $W$ is a $g\times g$ hermitian matrix ( $W_{ba}=W_{ab}^*$ ) — small: $g$ is 2 for this article's lab, 4 for hydrogen's $n=2$ shell. Its eigenvalue problem is the entire content of first-order degenerate theory, and its characteristic equation $\det(W-E^{(1)}\mathbb{1})=0$ still carries its astronomical name: the secular equation (the portrait band tells the planetary story).

Derivation AThe W matrix from the order-lambda equation

We derive the central eigenvalue problem: consistency of the perturbation series at first order forces the coefficients $c_a$ of equation (3) to be eigenvectors of $W$ , with the first-order energies as eigenvalues. The route is the same projection that produced the solvability conditions in Non-Degenerate Time-Independent Perturbation Theory — but now the null space of $(\hat{H}_0-E_n^{(0)})$ is $g$ -dimensional, so projection yields $g$ conditions instead of one, and they organize into a matrix equation. Watch for where hermiticity is used: it is what guarantees real splittings and an orthonormal good basis.

Step 1. Expand as usual: $|n\rangle=|n^{(0)}\rangle+\lambda|n^{(1)}\rangle+\cdots$ and $E_n=E_n^{(0)}+\lambda E^{(1)}+\cdots$ , with $|n^{(0)}\rangle$ the still-unknown combination (3). Collecting order $\lambda$ in $\hat{H}|n\rangle=E_n|n\rangle$ gives the familiar first-order equation

(\hat{H}_0-E_n^{(0)})\,|n^{(1)}\rangle=(E^{(1)}-\hat{H}')\,|n^{(0)}\rangle

identical in form to the non-degenerate case; the difference is entirely in what "solvable" costs.

Step 2. Project onto each member of the degenerate frame. For every $a$ , the bra $\langle n_a^{(0)}|$ kills the left side, because $\langle n_a^{(0)}|(\hat{H}_0-E_n^{(0)})=0$ — the operator annihilates the whole subspace from the left, not just one state. What remains is

\langle n_a^{(0)}|(E^{(1)}-\hat{H}')|n^{(0)}\rangle=0 \qquad\Longrightarrow\qquad \sum_{b=1}^{g}W_{ab}\,c_b=E^{(1)}c_a

one equation per frame vector: $g$ solvability conditions, assembled into the eigenvalue problem of the hermitian matrix $W$ . Why this settles the basis ambiguity: a hermitian $g\times g$ matrix has $g$ real eigenvalues and an orthonormal eigenbasis. Generic $\hat{H}'$ gives $g$ distinct eigenvalues — then the good frame is unique (up to phases), the level splits into $g$ sublevels $E_n^{(0)}+\lambda E^{(1)}_k$ , and the marble-in-the-bowl arbitrariness is gone. The perturbation has named its own axes.

Step 3. Confirm the divergences die in this frame. Take $|n^{(0)}\rangle$ = an eigenvector of $W$ and re-examine the dangerous terms of (1): the numerator between two different $W$ -eigenvectors in the same subspace is $\langle n_{k'}^{(0)}|\hat{H}'|n_k^{(0)}\rangle$ , which vanishes because $W$ is diagonal in its own eigenbasis. Zero numerators retire the zero denominators; intra-subspace terms drop out of both sums in (1), and the second-order energy becomes

E_{n,k}^{(2)}=\sum_{m\;:\;E_m^{(0)}\neq E_n^{(0)}}\frac{|\langle m^{(0)}|\hat{H}'|n_k^{(0)}\rangle|^2}{E_n^{(0)}-E_m^{(0)}}

a sum over outside states only — finite, and back on the non-degenerate rails. Verified: the projection algebra and the intra-block vanishing are checked symbolically, and the healed second order is confirmed against exact diagonalization in the laboratory below (article verifier, 35 checks).

Read the result physically. Eigenvalues of $W$ = the first-order splittings (the hero figure's fan spacings); eigenvectors = the correct zeroth-order states, the $\lambda\to0$ limits of the true eigenstates. And the machinery is self-deactivating: for $g=1$ , $W$ 's lone entry is the ordinary first-order shift — the apparatus collapses to the formula it generalizes.

E_{n,k}=E_n^{(0)}+\lambda\,E^{(1)}_k,\qquad k=1,\dots,g \qquad (5)

where: $E^{(1)}_k$ runs over the spectrum of $W$ . If all $g$ eigenvalues differ, the degeneracy is completely lifted and each sublevel proceeds through the higher orders as an ordinary non-degenerate level. If some coincide — a rank-deficient or symmetry-protected $W$ — the degeneracy is partially lifted: the surviving block keeps its ambiguity, and the contest is settled at second order within that block (the worked examples exhibit exactly this). When the survivors are merely close rather than equal, the honest tool is the resummed machinery of Brillouin-Wigner Perturbation Theory, which keeps the exact energy in its denominators and never divides by a small gap.

Go deeper · Derivation B · 2 minTwo-fold degeneracy in closed form

Derivation BTwo-fold degeneracy in closed form

We solve the $g=2$ case completely — the workhorse instance, and the one whose closed form every later application quotes. The goal: eigenvalues and eigenvectors of a general $2\times2$ hermitian $W$ , the size of the minimum splitting, and the geometry of the good states as a rotation angle. Everything follows from the quadratic formula, but the reading of each piece — especially the incompressible gap $2|W_{ab}|$ — is the part worth owning.

Step 1. The secular determinant, written out for the frame $\{|a\rangle,|b\rangle\}$ with $W_{ba}=W_{ab}^*$ :

\det\begin{pmatrix}W_{aa}-E^{(1)}&W_{ab}\\W_{ab}^{*}&W_{bb}-E^{(1)}\end{pmatrix}=0

expands to the quadratic $(E^{(1)})^2-(W_{aa}+W_{bb})E^{(1)}+(W_{aa}W_{bb}-|W_{ab}|^2)=0$ .

Step 2. Its roots are the two first-order shifts:

E^{(1)}_{\pm}=\frac{W_{aa}+W_{bb}}{2}\;\pm\;\sqrt{\left(\frac{W_{aa}-W_{bb}}{2}\right)^{2}+|W_{ab}|^{2}}

mean of the diagonals, plus-or-minus a hypotenuse. The separation between the two sublevels is $\sqrt{(W_{aa}-W_{bb})^2+4|W_{ab}|^2}$ . Why the levels cannot cross: the separation is a sum of squares under a square root. Tuning any single external knob can zero the diagonal mismatch $W_{aa}-W_{bb}$ , but at that point the separation bottoms out at $2|W_{ab}|$ — strictly positive whenever the perturbation couples the pair. This is the avoided-crossing gap, the first-order face of the repulsion theorem that Second-Order Energy Correction proved for distant levels and its chapter opener proved exactly for two.

Step 3. The good states, parametrized as a rotation (real $W_{ab}$ case for the geometry; a phase on $|b\rangle$ restores full generality):

\tan2\theta=\frac{2W_{ab}}{W_{aa}-W_{bb}},\qquad |+\rangle=\cos\theta\,|a\rangle+\sin\theta\,|b\rangle,\quad |-\rangle=-\sin\theta\,|a\rangle+\cos\theta\,|b\rangle

Two limits calibrate the dial. $W_{ab}\to0$ : $\theta\to0$ , the original frame was already good — this is what a symmetry-adapted basis buys, the subject of Good Quantum Numbers and Symmetry-Adapted Bases, next. $W_{aa}=W_{bb}$ : $\theta=45°$ exactly, maximal mixing — the frame is chosen entirely by the coupling, as in the laboratory below where symmetry forces equal diagonals.

Step 4. Sanity-check the extremes. Nondegenerate limit: decouple the pair ( $W_{ab}=0$ , $W_{aa}\neq W_{bb}$ ) and $E^{(1)}_\pm$ reduce to the two diagonal elements — two independent ordinary levels. Strong-mixing limit: equal diagonals give $E^{(1)}_\pm=W_{aa}\pm|W_{ab}|$ and the fifty-fifty combinations — the pattern of the $\lambda xy$ laboratory, and (in other clothes) of every two-level avoided crossing in the chapter ahead. Verified: closed form == exact eigenvalues over random hermitian draws (complex $W_{ab}$ ); determinant-quadratic identity symbolic; gap minimum $2|W_{ab}|$ at equal diagonals symbolic; $\tan2\theta$ eigenvector confirmed numerically; both limits checked (article verifier).

Choosing the basis without diagonalizing

One shortcut deserves its statement here, because the base of every application uses it: if a hermitian operator $\hat{A}$ commutes with both $\hat{H}_0$ and $\hat{H}'$ , then

\langle a'|[\hat{A},\hat{H}']|a\rangle=(\alpha_{a'}-\alpha_a)\langle a'|\hat{H}'|a\rangle=0 \qquad (6)

where: $\alpha_a$ are $\hat{A}$ -eigenvalues labeling the degenerate frame: states with different labels cannot be connected by $\hat{H}'$ , so in an $\hat{A}$ -adapted frame $W$ arrives block-diagonal — often fully diagonal — and the secular equation is pre-solved. This two-line observation is the germ of an entire doctrine — which operators, what "good quantum number" precisely means, when the shortcut fails — owned by the next serial, Good Quantum Numbers and Symmetry-Adapted Bases, which re-solves this article's laboratory by symmetry alone. Here it appears only as the closed form's $W_{ab}=0\Rightarrow\theta=0$ limit. The marquee applications are owned downstream by design: the fine-structure perturbations run this shortcut in the coupled basis — executed across Relativistic Kinetic Energy Correction, Spin-Orbit Coupling, Darwin Term, assembled in Fine Structure of Hydrogen — the magnetic fans belong to Zeeman Effect and Weak-Field (Anomalous) Zeeman Effect, and the classic $4\times4$ showpiece, hydrogen's $n=2$ shell in an electric field with its $\pm3e\mathcal{E}a_0$ doublet, is the property of Linear Stark Effect in Hydrogen — named here once, computed there.

Limiting cases

case	result	meaning
$g=1$	$W=(H'_{nn})$ : $E^{(1)}=\langle n^{(0)}\|\hat{H}'\|n^{(0)}\rangle$	machinery self-deactivates to the ordinary first-order formula
$W_{ab}=0$ chosen frame	$\theta=0$ : frame already good	symmetry-adapted basis; the next serial's whole subject
all $E^{(1)}_k$ distinct	complete lifting into $g$ sublevels	each sublevel continues as an ordinary non-degenerate level
repeated $E^{(1)}_k$	partial lifting; block survives	re-run the contest at second order inside the surviving block
$W\propto\mathbb{1}$	rigid shift, no splitting at first order	perturbation blind to the subspace's structure (equal diagonals and zero off-diagonals)
survivors close, not equal	small denominators return	resummed denominators: Brillouin-Wigner Perturbation Theory

Common trapPlugging a degenerate state into the non-degenerate formula

The single most common exam disaster in this chapter: computing $E^{(1)}=\langle\psi|\hat{H}'|\psi\rangle$ with whatever degenerate basis state was handed to you. The number you get is basis-dependent — someone who wrote the same subspace in a rotated frame gets a different "answer" — and both of you are generally wrong, as the diverging state correction is trying to announce. The diagonal elements of $W$ are honest first-order shifts only if the off-diagonals vanish. Check $W_{ab}$ first, always: if it is nonzero, diagonalize (or find the symmetry that does it for you). The only situation where you may skip the check is $g=1$ — an actually non-degenerate level.

Worked examples

Example 1 — the $\lambda xy$ laboratory, end to end. An electron in a 2D square well of side $L=1$ nm ( $0\le x,y\le L$ ) feels the perturbation $\hat{H}'=\lambda xy$ with $\lambda=0.1$ eV/nm². Split the first excited level.

Step 1. The stage. Unperturbed states are products $\psi_{n_xn_y}$ of Infinite Square Well modes with $E=(n_x^2+n_y^2)E_1$ , $E_1=\hbar^2\pi^2/2mL^2=0.3760$ eV. The first excited level is the two-fold pair $\psi_{12},\psi_{21}$ at $5E_1=1.8801$ eV — degenerate because the square's $x\leftrightarrow y$ symmetry cannot tell "node in $x$ " from "node in $y$ ."

Step 2. Build $W$ from 1D matrix elements. With $\langle1|x|1\rangle=\langle2|x|2\rangle=L/2$ and $\langle1|x|2\rangle=-16L/9\pi^2$ (both computed symbolically): the diagonals factor as $W_{aa}=W_{bb}=\lambda\langle x\rangle_1\langle y\rangle_2=\lambda L^2/4$ , and the off-diagonal as $W_{ab}=\lambda\langle1|x|2\rangle\langle2|y|1\rangle=\lambda\left(16L/9\pi^2\right)^2=\lambda\cdot256L^2/81\pi^4$ — the two minus signs cancel: $W_{ab}>0$ .

Step 3. Diagonalize by inspection (equal diagonals ⇒ $\theta=45°$ ): good states $(\psi_{12}\pm\psi_{21})/\sqrt2$ , shifts $E^{(1)}_\pm=\lambda\left(L^2/4\pm256L^2/81\pi^4\right)$ . Numerically: common shift $25.0$ meV, half-splitting $\lambda\cdot256L^2/81\pi^4=3.24$ meV, splitting $6.49$ meV, the $+$ (symmetric) combination on top — it piles amplitude along the $y=x$ diagonal where $xy$ is largest, and pays for it.

Step 4. Audit against the exact answer. Diagonalizing $\hat{H}_0+\lambda xy$ in a 196-state product basis: splitting and common shift match the first-order values within 2% and 3% respectively (the residue is honest $O(\lambda^2)$ ), and the upper exact eigenvector overlaps $(\psi_{12}+\psi_{21})/\sqrt2$ at $|\langle\cdot|\cdot\rangle|^2>0.95$ . Validity was never in doubt: $|W|\sim25$ meV against a $\sim1.1$ eV gap to the nearest outside level. Answer: the level fans into $1.8801\,\text{eV}+25.0\pm3.24$ meV, good states the $\pm45°$ diagonal combinations — all six numbers machine-verified.

Example 2 — partial lifting: a point bump under a triplet. The same well's level $E=50E_1$ is three-fold degenerate: $(n_x,n_y)=(1,7),(7,1),(5,5)$ . Perturb with a point bump at the quarter-point, $\hat{H}'=\lambda\,\delta(x-L/4)\,\delta(y-L/4)$ , $\lambda=0.002$ eV·nm².

Step 1. For a point perturbation every matrix element is a product of wavefunction values at the bump: $W_{ab}=\lambda\,\varphi_a(P)\,\varphi_b(P)$ with $P=(L/4,L/4)$ — $W$ is rank one by construction. The amplitudes: $\sin(n\pi/4)$ at $n=1,5,7$ gives $+\tfrac{\sqrt2}2,-\tfrac{\sqrt2}2,-\tfrac{\sqrt2}2$ , so $\varphi_{17}(P)=\varphi_{71}(P)=-1/L$ and $\varphi_{55}(P)=+1/L$ (each verified).

Step 2. A rank-one hermitian matrix $\lambda\,vv^{\mathsf T}$ has one nonzero eigenvalue $\lambda|v|^2$ and a zero eigenvalue of multiplicity $g-1$ . Here: one sublevel shifts by $3\lambda/L^2=6.0$ meV — its state $\propto(-\varphi_{17}-\varphi_{71}+\varphi_{55})/\sqrt3$ , the combination that maximizes amplitude at the bump — while a doublet survives unshifted: the two independent combinations whose wavefunctions vanish at the bump point simply do not know the perturbation exists.

Step 3. Exact-diagonalization audit (400-state basis, exact point-bump elements): the singlet lands within 12% of $+6.0$ meV (the residue is the point perturbation's notoriously slow second order); both doublet members stay pinned within $0.4$ meV of zero. Answer: first order lifts one state and leaves a two-fold ambiguity for second order. Partial lifting is not failure — it is the machinery reporting that the bump has one number's worth of opinion about a three-dimensional subspace (rank of $W$ = 1).

Origins · 1 minSecular Equation and how this was discovered

Secular Equation

1774–1929 · Paris, Berlin, Göttingen

The oldest word in this article is not "quantum" — it is secular, from the Latin saeculum, an age or century. In 1774 Lagrange, and after him Laplace, attacked the grand anxiety of Enlightenment astronomy: do Jupiter's and Saturn's slow orbital drifts — their secular variations, visible only across centuries — ever destroy the solar system? Linearizing the planets' mutual pulls produced exactly this article's object: a small symmetric matrix of couplings whose eigenvalues are slow precession frequencies, extracted from a characteristic determinant astronomers named the secular equation. Laplace's celebrated stability conclusion rested on those eigenvalues being real — hermiticity doing then what it does in Step 2 above — and Le Verrier ground out the eight-planet secular matrix by hand in 1856. The quantum inheritance was direct: Schrödinger's 1926 perturbation paper, built as he acknowledged on the "method familiar from celestial mechanics," put $\det(W-E^{(1)}\mathbb{1})=0$ at the center of degenerate problems, and von Neumann and Wigner's 1929 no-crossing analysis sharpened what its roots may do. Every Stark-doublet homework runs Lagrange's planetary algorithm with couplings in electron-volts instead of arcseconds per century.

History

1774

— Lagrange's memoir on the secular variations of planetary nodes and inclinations sets up the coupled linear problem

1784

— Laplace argues solar-system stability from the reality of the secular equation's roots

1856

— Le Verrier hand-computes the secular matrix of the eight-planet system

1926

— Schrödinger's perturbation paper imports the secular method into quantum mechanics

1929

— von Neumann and Wigner prove the no-crossing rule for the secular roots of hermitian families

Going further

Griffiths & Schroeter, Introduction to Quantum Mechanics, 3rd ed., §7.2 — the source register for the W-matrix method and the "good states" language.
Sakurai & Napolitano, Modern Quantum Mechanics, 2nd ed., §5.2 — the two-level closed form and projector viewpoint.
Cohen-Tannoudji, Diu & Laloë, Vol. II, Ch. XI — degenerate theory with full rigor and the $\lambda\to0$ limit made precise.
Shankar, Principles of Quantum Mechanics, 2nd ed., Ch. 17 — careful on normalization and the degenerate danger.

Provenance note: Equations and structure follow Griffiths §7.2 (degenerate theory, Stark effect), Sakurai & Napolitano §5.2 (two-level formula, good quantum numbers), Cohen-Tannoudji Ch. XI, and Townsend Ch. 11. Verifiers should recompute: the secular-equation derivation; the two-level eigenvalue formula and gap $2|W_{12}|$ ; the hydrogen matrix element $\langle200|z|210\rangle=-3a_0$ and resulting $\pm3eE_za_0$ Stark splitting; and the selection-rule argument from $[A,H']=0$ . No non-standard numbers are introduced. — The secular derivation, two-level closed form, and gap are executed in this article's verify script (35 checks, including the laboratory's exact-diagonalization audits). Edition note: the base article's worked derivation of the hydrogen $n=2$ linear Stark splitting ( $\langle200|z|210\rangle=-3a_0$ , $\pm3e\mathcal{E}a_0$ ) is RELOCATED by chapter ownership to Linear Stark Effect in Hydrogen (its verifier items travel with it); the good-quantum-number selection-rule argument is stated here as equation (6) and owned in full by Good Quantum Numbers and Symmetry-Adapted Bases.

Nothing physical diverges, so why does the math insist on infinities?

Because the naive ansatz makes a false promise: that the exact eigenstate tends to your chosen basis vector as $\lambda\to0$ . When the truth is a finitely rotated combination, first order is asked to produce a finite rotation with an infinitesimal coefficient — impossible, and the divergent coefficient is the contradiction surfacing. It is precisely the failure mode of expanding around the wrong point, not of expanding at all.

Do I diagonalize the perturbation over the whole Hilbert space or just the subspace?

Just the $g$ -dimensional degenerate subspace. Diagonalizing $\hat{H}'$ everywhere would be solving the full problem (and is generally impossible); the theory's entire economy is that only the ambiguous directions need exact treatment. Couplings to outside levels are handled perturbatively afterwards, through the healed second-order sum — they are safe because their denominators are honest finite gaps.

What if W comes out proportional to the identity?

Then first order shifts the whole level rigidly and lifts nothing — equal diagonals and vanishing off-diagonals mean the perturbation is blind to the subspace's internal structure at this order. The basis ambiguity survives intact, and you must go to second order within the subspace (an effective $W^{(2)}$ built from outside couplings) or wait for a different perturbation. Example 1's perturbation with the well's full symmetry group would do this; the $xy$ tilt specifically breaks the right symmetry to act at first order.

How do I know the eigenvectors of W are really the λ→0 limits of the exact states?

The two-level case shows it directly: diagonalize exactly and take $\lambda\to0$ ; the limit vectors are $W$ 's eigenvectors whenever its eigenvalues are distinct. The laboratory audit makes it concrete — at small finite $\lambda$ the exact eigenvector already overlaps the predicted combination at better than 95%. With repeated eigenvalues the limit is genuinely ambiguous within the surviving block — the honest content of "partial lifting."

Why is it called the secular equation when nothing here is slow or centuries-long?

Pure inheritance. The determinant was invented for planetary perturbations that accumulate over centuries — saeculum — where the eigenvalues were precession rates. Quantum mechanics borrowed the machinery whole (Schrödinger cited celestial mechanics explicitly in 1926) and the name rode along. The portrait band tells the story; the mathematics is identical, which is the real point: eigenvalue perturbation theory predates quantum mechanics by 150 years.

In Example 2, could I have predicted the surviving doublet without computing anything?

Almost. A point bump is a rank-one perturbation — it measures each state by a single number, its amplitude at the bump. Any $g$ -dimensional subspace has a $(g-1)$ -dimensional sub-subspace of combinations vanishing at that point, and those states cannot feel the bump at first order. So "rank of $W$ = 1 ⇒ $g-1$ survivors" is general. What needs the actual numbers is the surviving states' identity and the singlet's shift $\lambda\sum_a|\varphi_a(P)|^2$ .