Brillouin-Wigner Perturbation Theory

Every formula in the Rayleigh-Schrödinger chapters carries the same fine print: the energy denominators $E_n^{(0)}-E_k^{(0)}$ must stay large compared to the coupling. Non-Degenerate Time-Independent Perturbation Theory made that fine print exact — its series converges only inside the disk $\lambda<\Delta/2|v|$ — and Degenerate Perturbation Theory handles a gap of exactly zero. But nature is fond of the awkward middle ground: levels close without coinciding. A field-tuned pair of atomic sublevels, a vibrational Fermi resonance, two states near an avoided crossing — the gap is small, the naive series explodes term by term, yet the exact levels sit there perfectly finite. Brillouin-Wigner theory is the bookkeeping that refuses to panic. Instead of asking how far each virtual state lies from where you started ( $E_n^{(0)}$ ), it asks how far it lies from where you actually end up — the exact energy $E$ itself sits inside every denominator. The price is that the master equation becomes implicit: $E$ appears on both sides, and each truncation is an algebraic equation to solve rather than a number to read off. The reward is startling: for a two-level system the second-order truncation is not an approximation at all but the exact quadratic, level repulsion and avoided crossing included. This article derives the implicit pair of equations by projector surgery, proves the two-level exactness theorem, shows exactly how the Rayleigh-Schrödinger series re-emerges when the denominators are expanded, and then pays the theory's two honest debts: self-consistency (you must iterate) and the failure of size-extensivity (BW energies of independent systems refuse to add) — the defect that decides which perturbation theory runs modern many-body physics. Every number below is machine-verified.

Key takeaways

Brillouin-Wigner (BW) theory puts the exact energy $E$ in every denominator: $E = E_n^{(0)} + H'_{nn} + \sum_{k\neq n}|H'_{kn}|^2/(E-E_k^{(0)}) + \cdots$ — an implicit equation solved self-consistently, instead of Rayleigh-Schrödinger's (RS) explicit cascade.
Both BW equations are exact rearrangements of the Schrödinger equation, from projecting with $\hat{P}=|n^{(0)}\rangle\langle n^{(0)}|$ and $\hat{Q}=1-\hat{P}$ under intermediate normalization (verified on a random matrix to $10^{-11}$ ).
Near-degeneracy is BW's home turf: $E-E_k^{(0)}$ stays finite because the true level is repelled from its neighbours. For a gap-0.02 pair with coupling 0.05, RS second order lands 3.1× too deep; the BW fixed point is off by 0.003 — 29× better (computed).
Two-level exactness: BW truncated at second order gives $(E-E_a^{(0)})(E-E_b^{(0)})=|V|^2$ — the exact characteristic polynomial; avoided crossings come out exact.
Expanding the BW denominators about $E_n^{(0)}$ regenerates RS order by order — RS is BW Taylor-expanded, and that expansion is what dies near degeneracy.
The costs: every truncation must be solved for $E$ ; and BW is not size-extensive — two non-interacting copies give $E_{\mathrm{BW}}\neq 2E_A$ (gap $-0.0104$ in the worked model), which is why many-body physics belongs to RS-descended linked-cluster methods.
Key idea: put the exact energy $E$ in the denominators instead of the unperturbed $E_n^{(0)}$ — a divergent order-by-order expansion becomes a finite self-consistent equation that survives near-degeneracy.

The core insight

Put the exact energy $E$ in the denominators instead of the unperturbed $E_n^{(0)}$ , turning a divergent order-by-order expansion into a finite self-consistent (implicit) equation that survives near-degeneracy.

Where this lives today

The two-level BW quadratic is the formula experimentalists reach for whenever a control knob tunes two levels through an avoided crossing: field-mixed atomic sublevels between the weak-field Zeeman fan and the Paschen-Back limit obey exactly the $\bar{E}\pm\sqrt{(\Delta/2)^2+|V|^2}$ shape (the Breit-Rabi-type formulas of the magnetic-field serials), as do vibrational Fermi resonances and transmon avoided crossings. In quantum chemistry the story came full circle: multireference Brillouin-Wigner coupled-cluster methods re-adopt the implicit denominator to tame quasi-degenerate electronic states, correcting the extensivity defect perturbatively — while the $\hat{P}/\hat{Q}$ partition itself, as Feshbach projection, runs resonances and effective Hamiltonians in nuclear and scattering physics. One idea, three fields, same projector surgery.

The physical picture

Picture the reference state $|n^{(0)}\rangle$ dressed by virtual excursions into the other unperturbed states. Every perturbation theory prices those excursions by an energy denominator — the question is measured from where. Rayleigh-Schrödinger asks: how far is $|k^{(0)}\rangle$ from where I started, $E_n^{(0)}$ ? Brillouin-Wigner asks: how far is it from where I actually end up, the exact $E$ ? For well-separated levels the two questions have nearly the same answer; for a close neighbour they differ catastrophically — and only the second stays sane, because the exact level moves. Level repulsion — exact in the two-level laboratory of Non-Degenerate Time-Independent Perturbation Theory, generalized by the pairwise-push theorem of Second-Order Energy Correction — pushes $E$ away from its neighbour, so the BW denominator is held open by the very physics that slams the RS one shut.

Figure

Two ways to price a virtual excursion: distance from the start vs distance from the destination

The implicitness is not a bug to be apologized for — it is the entire mechanism. RS converts the exact eigenvalue problem into an explicit cascade by Taylor-expanding $1/(E-E_k^{(0)})$ about $E_n^{(0)}$ , and that expansion is what has a finite radius of convergence. BW simply declines to expand: it keeps the denominator whole and pays with a self-consistency condition — truncate, then solve. One virtual channel gives a quadratic done by hand; many channels give a fixed-point iteration (eighteen passes for twelve digits in the worked example below). There is a second, quieter consequence of that global $E$ : because one number threads through every denominator, the theory cannot factorize over independent subsystems. Describe two non-interacting molecules together and molecule B's energy contaminates molecule A's denominators — the energies refuse to add. RS terms, with their fixed unperturbed denominators, do factorize. That single structural fact, made quantitative in the third derivation, is why BW rules isolated few-level problems and RS-descended machinery rules everything extended.

The implicit pair

The Hamiltonian splits as $\hat{H}=\hat{H}_0+\lambda\hat{H}'$ with known spectrum $\hat{H}_0|k^{(0)}\rangle=E_k^{(0)}|k^{(0)}\rangle$ ; we hunt the exact eigenpair $(E,|\psi\rangle)$ that grows out of $|n^{(0)}\rangle$ as $\lambda$ turns on. Two tools come with credentials — the projectors onto the reference line and its complement (the algebra of Projection Operators),

\hat{P}=|n^{(0)}\rangle\langle n^{(0)}|,\qquad \hat{Q}=1-\hat{P}=\sum_{k\neq n}|k^{(0)}\rangle\langle k^{(0)}| \qquad (1)

where: $\hat{P}$ picks out the model state, $\hat{Q}$ spans the virtual space, and both commute with $\hat{H}_0$ because they are built from its eigenstates — the one fact the whole derivation leans on. Normalization is the intermediate convention legitimized in Non-Degenerate Time-Independent Perturbation Theory: $\langle n^{(0)}|\psi\rangle=1$ , so corrections live entirely in the $\hat{Q}$ space.

Derivation AThe implicit pair by projector surgery

We derive both Brillouin-Wigner master equations — the implicit energy condition and the self-consistent state equation — as exact rearrangements of the Schrödinger equation, using nothing but the $\hat{P}/\hat{Q}$ split and intermediate normalization. No expansion happens anywhere in this card; that is the point. The exact $E$ enters the denominator forced, not chosen.

Step 1. Rearrange the eigenvalue problem. $(\hat{H}_0+\lambda\hat{H}')|\psi\rangle=E|\psi\rangle$ becomes $(E-\hat{H}_0)|\psi\rangle=\lambda\hat{H}'|\psi\rangle$ . Everything unknown is on display: $E$ on the left, the exact state on both sides.

Step 2. Project with $\hat{P}$ . Acting with $\langle n^{(0)}|$ and using $\langle n^{(0)}|\hat{H}_0=E_n^{(0)}\langle n^{(0)}|$ plus $\langle n^{(0)}|\psi\rangle=1$ :

E=E_n^{(0)}+\lambda\langle n^{(0)}|\hat{H}'|\psi\rangle \qquad (2)

Why exact yet useless alone: the shift is one matrix element into the exact state — which is unknown. The $\hat{Q}$ projection closes the loop.

Step 3. Project with $\hat{Q}$ . Since $[\hat{Q},\hat{H}_0]=0$ : $(E-\hat{H}_0)\hat{Q}|\psi\rangle=\lambda\hat{Q}\hat{H}'|\psi\rangle$ . On the $\hat{Q}$ subspace $E-\hat{H}_0$ has eigenvalues $E-E_k^{(0)}$ , $k\neq n$ — all nonzero unless the exact level lands on an unperturbed other level (generically it does not; repulsion actively helps). So it inverts there:

\hat{Q}|\psi\rangle=\frac{\hat{Q}}{E-\hat{H}_0}\,\lambda\hat{H}'|\psi\rangle \qquad (3)

Why the inverse is legitimate: we invert on the complement only — never on the full space, where $E-\hat{H}$ is singular by construction. The object $\hat{Q}(E-\hat{H}_0)^{-1}\hat{Q}$ is a resolvent restricted to the virtual space — the analytic machine catalogued in Green Functions in Quantum Mechanics.

Step 4. Reassemble. $|\psi\rangle=\hat{P}|\psi\rangle+\hat{Q}|\psi\rangle$ with $\hat{P}|\psi\rangle=|n^{(0)}\rangle$ :

|\psi\rangle=|n^{(0)}\rangle+\frac{\hat{Q}}{E-\hat{H}_0}\,\lambda\hat{H}'|\psi\rangle \qquad (4)

Equations (2) and (4) are the Brillouin-Wigner pair — exact, closed on each other. Iterating (4) and feeding (2) generates the BW series with $E$ inside every denominator:

E=E_n^{(0)}+\lambda H'_{nn}+\lambda^2\sum_{k\neq n}\frac{|H'_{kn}|^2}{E-E_k^{(0)}}+\lambda^3\sum_{k,l\neq n}\frac{H'_{nk}H'_{kl}H'_{ln}}{(E-E_k^{(0)})(E-E_l^{(0)})}+\cdots \qquad (5)

Verified: on a random Hermitian 6×6 with a diagonal $\hat{H}_0$ , the exact eigenpair under intermediate normalization satisfies (4) and (2) to $10^{-11}$ — both relations confirmed as identities, not approximations (article verifier, section 1).

Equation (5) has exactly the shape of the Rayleigh-Schrödinger series of Second-Order Energy Correction — same matrix elements, same virtual sum — with every $E_n^{(0)}$ in a denominator replaced by $E$ . That substitution makes the right side depend on the left, so a truncation of (5) is an equation in $E$ , not a formula. Truncate after the $\lambda^2$ sum and you hold the BW second-order equation, the workhorse of everything below:

E=E_n^{(0)}+\lambda H'_{nn}+\lambda^2\sum_{k\neq n}\frac{|H'_{kn}|^2}{E-E_k^{(0)}} \qquad (6)

where: the unknown appears on both sides; one virtual channel clears to a quadratic (next derivation), many channels get iterated: seed $E\approx E_n^{(0)}$ , evaluate, feed back, repeat. The map's derivative at the fixed point sets the speed — $0.219$ in the worked example, a factor of five per pass.

The two-level theorem

The cleanest advertisement for the implicit denominator: for two levels it is not an approximation at all. Take the reference $|a\rangle$ coupled to a single partner $|b\rangle$ by $\lambda v$ , no diagonal perturbation — the laboratory Hamiltonian whose exact solution and convergence disk ( $\lambda_c=\Delta/2|v|$ ) were established in Non-Degenerate Time-Independent Perturbation Theory. That article proved when RS fails; here is what replaces it.

Go deeper · Derivation B · 1 minTwo levels: where the second-order truncation is exact

Derivation BTwo levels: where the second-order truncation is exact

We show that equation (6), truncated at second order, reproduces the exact eigenvalues of the two-level system — avoided crossing included, and the degenerate limit where the RS series has zero radius. The strategy: clear the one denominator and recognize the characteristic polynomial.

Step 1. Write (6) for the pair. With $E_a^{(0)}$ , $E_b^{(0)}$ , coupling $H'_{ba}=v$ , and $H'_{aa}=H'_{bb}=0$ , the virtual sum has exactly one term:

E=E_a^{(0)}+\frac{\lambda^2v^2}{E-E_b^{(0)}} \qquad (7)

Why the truncation is not a truncation: higher BW terms need a path $a\to k\to l\to a$ through the virtual space — which is the single state $|b\rangle$ , and $H'_{bb}=0$ kills every longer path. The series terminates identically at second order.

Step 2. Clear the denominator: $(E-E_a^{(0)})(E-E_b^{(0)})=\lambda^2v^2$ . Expand and compare with $\det(H-E\mathbb{1})=0$ for the matrix $\begin{pmatrix}E_a^{(0)}&\lambda v\\\lambda v&E_b^{(0)}\end{pmatrix}$ : they are the same polynomial, term by term (verified symbolically).

Step 3. Solve. The two roots are

E_\pm=\frac{E_a^{(0)}+E_b^{(0)}}{2}\pm\sqrt{\left(\frac{E_a^{(0)}-E_b^{(0)}}{2}\right)^2+\lambda^2v^2} \qquad (8)

— exactly the diagonalization result, avoided crossing and all. The implicit equation knew about both levels: seeding the iteration near $E_a^{(0)}$ finds one root, near $E_b^{(0)}$ the other.

Step 4. Push to the degenerate limit $E_a^{(0)}=E_b^{(0)}$ , where the RS radius collapses to zero and every RS term beyond first order is infinite. Equation (8) stays finite: $E_\pm=E_a^{(0)}\pm\lambda|v|$ — the symmetric split Degenerate Perturbation Theory obtains from its W-matrix. One formula covers both regimes. Verified: quadratic == characteristic polynomial and roots == exact eigenvalues, symbolically; degenerate limit $E_\pm=\pm\lambda v$ , symbolically; all numeric instances below to machine precision (article verifier, section 2).

From Brillouin-Wigner back to Rayleigh-Schrödinger

If both theories are exact rearrangements of the same equation, they must contain each other. The dictionary runs through the denominator. Write the exact shift as $\delta=E-E_n^{(0)}$ and expand each BW denominator about the unperturbed energy — the step BW refused, performed now deliberately, in a regime where it is safe:

\frac{1}{E-E_k^{(0)}}=\frac{1}{E_n^{(0)}-E_k^{(0)}}\left[1-\frac{\delta}{E_n^{(0)}-E_k^{(0)}}+\frac{\delta^2}{(E_n^{(0)}-E_k^{(0)})^2}-\cdots\right] \qquad (9)

where: the leading term restores exactly the RS denominator, and since $\delta$ is itself $\mathcal{O}(\lambda^2)$ for an off-diagonal perturbation, each further term is pushed two powers of $\lambda$ up. Substituting (9) into (6) and collecting powers regenerates Rayleigh-Schrödinger order by order: the $\lambda^2$ coefficient is the familiar $\sum|H'_{kn}|^2/(E_n^{(0)}-E_k^{(0)})$ of Second-Order Energy Correction — recovered, not re-derived — and the $\delta$ -correction supplies (part of) RS's fourth order. For the two-level model the exact root's Taylor series has $c_2=-v^2/\Delta$ , vanishing odd coefficients, and $c_4=+v^4/\Delta^3$ ; the expansion of (7) reproduces each one (verified symbolically through $\lambda^6$ ). Read (9) as a diagnosis. RS is BW with every denominator Taylor-expanded — and that series converges only while the shift is small compared to the gap. Near degeneracy the ratio passes one, (9) diverges, and the RS cascade inherits the divergence term by term. Nothing is wrong with the physics; something is wrong with the expansion. BW's entire content is the refusal to perform (9) prematurely.

The price: self-consistency and the extensivity defect

Two honest costs. The first is mild: every truncation is an implicit equation — a quadratic for one channel, a fixed-point iteration $E\mapsto f(E)$ (convergent when $|f'(E^*)|<1$ ) for several. Cheap in practice, but the iteration finds one root per seed, so you must know which level you are hunting. The second cost is structural, and it decided the history of many-body physics.

Go deeper · Derivation C · 1 minTwo non-interacting copies: the size-extensivity defect made quantitative

Derivation CTwo non-interacting copies: the size-extensivity defect made quantitative

We take the most innocent compound system imaginable — two identical, strictly non-interacting two-level subsystems — and show that composite BW-2 is not twice the single copy, though the exact ground energy trivially is. The global $E$ in the denominators is the culprit. Model: each copy has levels $0$ and $\Delta=1$ , internal coupling $v=0.3$ , no interaction between copies.

Step 1. The single copy is the two-level theorem's system: BW-2 is exact, $E_A=\tfrac12-\tfrac12\sqrt{1+4v^2}=-0.08310$ . The exact composite ground energy is $2E_A=-0.16619$ — tensor additivity, confirmed on the full $4\times4$ (verifier, $10^{-12}$ ).

Step 2. Run BW on the composite. The reference $|a\rangle\otimes|a\rangle$ sits at $E^{(0)}=0$ ; the perturbation reaches only the two singly-excited states, each at $\Delta=1$ with element $v$ . The composite BW-2 equation is

E_{AB}=\frac{2v^2}{E_{AB}-\Delta} \qquad (10)

where: the denominator is $E_{AB}-E_{k_A}^{(0)}-E_B^{(0)}$ — copy A's gap contaminated by copy B's energy. Isolated copy A would write $E_A-E_{k_A}^{(0)}$ . That mismatch is the defect, visible before any algebra.

Step 3. Solve (10): $E_{AB}=\tfrac12(1-\sqrt{1+8v^2})=-0.15574$ versus $2E_A=-0.16619$ : a gap of $-0.01045$ — six percent of the answer, from no interaction whatsoever — and it compounds with system size.

Step 4. Run RS on the same composite: $E^{(2)}_{AB}=2v^2/(0-\Delta)=-0.18=2\times(-0.09)$ — exactly twice the single-copy value, at every size. RS denominators contain only unperturbed differences, so copy B cancels out of copy A's denominators identically: $E_a^{(0)}+E_b^{(0)}-E_{k_A}^{(0)}-E_b^{(0)}=E_a^{(0)}-E_{k_A}^{(0)}$ . The RS sum factorizes (only linked terms survive), and additivity holds order by order. Why this exiled BW from many-body theory: a per-particle energy that depends on how much matter you describe is unusable for solids or long molecules. The linked-cluster theorem (Brueckner, Goldstone, 1955) proved RS's additivity to all orders and settled the choice; coupled-cluster methods are its descendants — a prose pointer to the many-body serials, nothing more here. *Verified: $E_A=-0.0830952$ ; composite BW root $\tfrac12(1-\sqrt{1.72})=-0.155744$ ; gap $-0.010447$ ; exact composite $=2E_A$ to $10^{-12}$ ; RS additivity exact (article verifier, section 6).*

Limiting cases

case	result	meaning
well-separated levels	$E\approx E_n^{(0)}$ : BW $\to$ RS term by term	expansion (9) converges fast; RS is less work
near-degenerate pair, $\Delta\to$ small	RS radius $\lambda_c=\Delta/2\lvert v\rvert\to0$ ; BW quadratic stays finite	level repulsion holds the BW denominator open
exact degeneracy $\Delta=0$	$E_\pm=E^{(0)}\pm\lambda\lvert v\rvert$	the W-matrix split of Degenerate Perturbation Theory recovered
two levels, any coupling strength	BW second order is EXACT	the virtual space is one-dimensional; the series terminates
$\lambda\to0$	$E\to E_n^{(0)}$ , iteration converges in one step	both theories reduce to nothing, as they must
$N$ non-interacting copies	$E_{\mathrm{BW}}\neq N\,E_A$	size-inextensivity — RS/linked-cluster wins for extended systems

Worked examples

Example 1 — beyond the convergence radius: RS drifts, BW lands. The two-level laboratory of Non-Degenerate Time-Independent Perturbation Theory with $\Delta=1$ , $v=0.3$ has RS radius $\lambda_c=1.667$ (its theorem, quoted). Work at $\lambda=2.0$ — outside the disk.

Step 1. Exact answer first (equation 8): $E_\pm=0.5\pm0.5\sqrt{1+4(2.0)^2(0.09)}$ , so $E_-=-0.28102$ , $E_+=+1.28102$ .

Step 2. Try RS anyway. The partial sums do something subtler than explode: the error shrinks to an optimal $0.041$ at order $K=8$ , then turns around — $0.098$ at $K=20$ , $0.334$ at $K=30$ , $1.81$ at $K=42$ (computed). Textbook asymptotic-series behaviour: information at low orders, then a quiet walk away from the answer. Summing "more accurately" makes you wronger.

Step 3. Now BW at second order — which for two levels is exact. Clearing the denominator gives the quadratic with root $E_-=-0.28102$ : error zero, at the same order where RS's best effort is $0.041$ .

Step 4. Solve it the many-channel way to see the machinery: $E\mapsto\lambda^2v^2/(E-1)$ seeded at $E=0$ gives $E_1=-0.360$ , $E_2=-0.265$ , $E_3=-0.285$ , contraction $|f'(E_-)|=0.219$ — a factor of ~5 per pass, $10^{-12}$ in 18 iterations. Seeding near the upper level converges to $E_+$ instead. Answer: $E_-=-0.28102$ exactly from BW-2; RS's best truncation misses by $0.041$ .

Figure

Beyond the radius at λ = 2.0: the RS error is U-shaped in truncation order; BW-2 sits at zero

Example 2 — a near-degenerate ladder with a spectator. Three levels at $E^{(0)}=\{0,\ 0.02,\ 5\}$ : a whisper-thin pair (gap $0.02$ ) plus a distant spectator. Couplings $H'_{01}=0.05$ , $H'_{02}=0.2$ , $H'_{12}=0.1$ , diagonal zero. Find the ground level.

Step 1. Diagnose first: the close pair's RS radius is $\lambda_c=\text{gap}/2|v|=0.02/0.10=0.2$ , and we work at $\lambda=1$ — five times beyond it. RS is doomed in advance; the spectator at $5$ is harmless.

Step 2. RS second order: $E^{(2)}=(0.05)^2/(0-0.02)+(0.2)^2/(0-5)=-0.133$ . The exact ground level is $-0.042785$ : RS lands 3.1× too deep — worse than computing nothing.

Step 3. BW second order: $E=(0.05)^2/(E-0.02)+(0.2)^2/(E-5)$ ; iterating from $E=0$ lands at $E=-0.045877$ — error $0.0031$ , 29× better at identical cost. The near-degenerate denominator $E-0.02\approx-0.066$ was held open by the level's own repulsion; RS's $-0.02$ was three times smaller and inflated the term by the same factor.

Step 4. Read the residual honestly: BW-2 is not exact here (spectator paths generate genuine third-order terms), but the near-degenerate physics is captured non-perturbatively. Answer: exact $-0.0428$ ; RS-2 $-0.133$ (unusable); BW-2 $-0.0459$ .

Origins · 1 minLéon Brillouin and how this was discovered

Léon Brillouin

1889–1969 · Sorbonne, Collège de France, then Columbia

Brillouin was born into the problem: his father Marcel held the Collège de France physics chair he would later occupy himself, and Léon grew up around the Solvay circle whose scientific secretary he became. His name is stitched across twentieth-century physics — zones in crystals, a function in paramagnetism, the B in a semiclassical method — but the piece that lives here arrived in 1932-33, in a run of Comptes Rendus notes on perturbation and self-consistent-field problems. His observation was disarmingly simple: the divergences plaguing perturbation expansions came from expanding energy denominators about the wrong point; keep the true energy inside them and the equations close on themselves — implicit, but finite. Wigner tidied the scheme in 1935 and the pairing stuck (Lennard-Jones had written implicit denominators in 1930 — a priority honestly shared). The war broke his life in two: director of Radio France in 1939, he fled after the fall of France and rebuilt his career in America — Wisconsin, Columbia, Harvard, then IBM — where his late obsession became the physics of information, negentropy and Maxwell's demon. He liked machinery that admitted what it did not know; an equation with its own answer inside, solved by successive approximation, was exactly his taste.

History

1930

— Lennard-Jones writes implicit-denominator corrections for molecular problems: the idea's first appearance

1932

— Brillouin's Comptes Rendus notes systematize the expansion with the exact energy in every denominator

1935

— Wigner reorganizes and extends the scheme; the Brillouin-Wigner pairing enters the textbooks

1955

— Brueckner and Goldstone prove the linked-cluster theorem: Rayleigh-Schrödinger is size-extensive to all orders, and many-body theory chooses sides

1998

— multireference Brillouin-Wigner coupled-cluster methods revive the implicit denominator for quasi-degenerate molecules

Going further

Sakurai & Napolitano, Modern Quantum Mechanics, 2nd ed., §5.1-5.2 — the projector route at the source register.
Shankar, Principles of Quantum Mechanics, 2nd ed., Ch. 17 — RS and BW side by side.
Cohen-Tannoudji, Diu & Laloë, Vol. II, Ch. XI — the two-level problem as organizing principle.
Lindgren & Morrison, Atomic Many-Body Theory, 2nd ed., Ch. 9 — BW vs RS and linked-cluster size-extensivity.

Provenance note: Implicit BW equations and projector derivation follow Sakurai & Napolitano §5.1-5.2 and Shankar Ch. 17; the exact two-level reduction $E_\pm=\bar E\pm\sqrt{(\Delta/2)^2+|V|^2}$ is standard (verify by diagonalizing the 2×2 Hamiltonian). Size-inextensivity of BW and the contrast with RS linked-cluster expansions follow Lindgren & Morrison, Atomic Many-Body Theory, Ch. 9. A verifier should: (1) confirm the projected equations reproduce the Schrödinger equation; (2) check the BW second-order quadratic against direct 2×2 diagonalization; (3) Taylor-expand to recover the RS second-order denominator $E_n^{(0)}-E_k^{(0)}$ . No fabricated numerical constants appear. — All three verifier items are executed in this article's verify script, together with every shipped number ( $\lambda=2.0$ error table and optimal $K=8$ ; the 18-pass iteration, contraction $0.219$ ; the three-level exhibit $-0.133$ / $-0.0459$ / exact $-0.0428$ ; the extensivity gap $-0.0104$ ). Edition note: the base's three derivations are retained as this article's three cards, recast into the current grammar with computed exhibits. The two-level laboratory's exact solution and radius belong to Non-Degenerate Time-Independent Perturbation Theory, quoted not re-derived; the recovered RS second-order formula to Second-Order Energy Correction; the exactly-degenerate case to Degenerate Perturbation Theory — touched here only as the $\Delta\to0$ limit of the quadratic.

Is it not circular to write the unknown energy inside its own formula?

It is implicit, not circular — the same status as $x=\cos x$ : a constraint that pins $E$ down without handing it to you as a finished expression. Circularity would mean the equation is empty; instead the BW pair is an exact rearrangement of the Schrödinger equation, so its solutions are precisely the true eigenvalues. The resolution is root-finding: clear denominators for one channel, iterate for many.

How do I actually solve a BW equation on an exam?

One virtual channel: clear the denominator and solve the quadratic — two roots, two levels, pick by continuity with $\lambda\to0$ . Several channels: fixed-point iteration — seed with $E_n^{(0)}$ (or the RS estimate), evaluate the right side, feed back, repeat; each pass gained a factor of five in this article's examples. If the iteration bounces instead of settling, the map's slope exceeds one there: rearrange the equation about the level you want and iterate again.

When do I use BW versus the W-matrix of degenerate perturbation theory?

By the gap. Exactly degenerate: the W-matrix of Degenerate Perturbation Theory — cheaper and cleaner. Nearly degenerate (gap comparable to or smaller than the coupling): BW, whose two-level truncation covers the whole crossover in one quadratic and reduces to the W-matrix split as the gap goes to zero. Well-separated: ordinary RS — less work, explicit answers, size-extensive.

Is BW at a given order always more accurate than RS at the same order?

No. Near degeneracy BW wins dramatically (29× in the three-level exhibit). For well-separated levels the two agree to leading order, differences buried at $\mathcal{O}(\lambda^4)$ — neither is systematically better there, and RS's explicitness and extensivity make it the default; a truncated BW root can even sit farther from the truth than the RS sum. The honest rule: BW's advantage is structural (finite denominators), not a uniform accuracy theorem.

Why exactly does the two-level case terminate at second order?

Count the paths. The $m$ -th BW term threads $m-1$ intermediates through the $\hat{Q}$ space. With one virtual state $|b\rangle$ and $H'_{bb}=0$ , any path longer than $a\to b\to a$ must hop $b\to b$ — matrix element zero — so third order and beyond vanish identically; the "truncation" discards nothing. With $H'_{bb}\neq0$ the same closure happens after absorbing it into $E_b^{(0)}$ .

What does size-extensive mean, and why should a single-atom problem care?

Size-extensive: $N$ non-interacting copies cost $N$ times one copy, order by order. BW fails this (6% for two copies in the worked model) because the total energy sits in every denominator. For one isolated atom or one avoided crossing you may not care — BW is at its best there. You care the moment you compare sizes, compute per-particle binding energies, or scale toward a solid: an error that grows with $N$ corrupts exactly the differences you want.

Does the BW fixed-point iteration always converge, and to which level?

Neither is guaranteed. Convergence needs the map's slope at the root below one ( $0.219$ in Example 1 — comfortable); strong coupling or an unlucky rearrangement pushes it past one and the iteration walks away. The cure: rearrange which term is solved for — seeding about $E_b^{(0)}$ found the upper root in Example 1. Each seed basin belongs to one root; continuity with the $\lambda\to0$ spectrum is the compass.