Adiabatic Approximation and Validity Conditions

The Adiabatic Theorem made a qualitative promise: change a Hamiltonian slowly enough and a system parked in an instantaneous eigenstate stays parked. Here "slowly enough" stops being a shrug and becomes a number you can compute before running the experiment — and you learn the three ways that number can lie to you. The centrepiece is a single dimensionless figure of merit, $\eta=\hbar\,|\langle m|\dot H|n\rangle|/(E_m-E_n)^2$ , derived here by asking how much amplitude actually leaks between instantaneous eigenstates when the drive is slow. The derivation explains the strangest feature of the criterion — why the energy gap enters squared — and hands you a closed form for a laboratory you can hold in your head: a spin riding a magnetic field around a cone, where $\eta=\omega\sin\alpha/2\omega_0$ . Then we stress-test the criterion by direct integration of the Schrödinger equation. A ramp with velocity kinks settles for an error falling only as $1/T^2$ — predicted analytically, measured to a fitted slope of $-1.99$ — while an infinitely smooth ramp plunges faster than any power of $T$ . Finally we exhibit the criterion's honest failure: a resonant micro-wobble that keeps $\eta\le 0.01$ at every instant yet destroys the state completely — the Marzlin–Sanders warning that $\eta\ll1$ is necessary, not sufficient. Where gaps close entirely, the swept avoided crossing is named and handed to Landau–Zener Transitions.

Key takeaways

The criterion is dimensionless: $\eta=\hbar\,|\langle m|\dot H|n\rangle|/(E_m-E_n)^2\ll1$ — one gap power suppresses the coupling, the second comes from fast phase averaging.
A two-clock statement: the drive must beat the internal phase clock $\hbar/\Delta E$ — always relative, never absolute; "slow" near a closing gap does not exist.
Spin on a swept cone (half-angle $\alpha$ , sweep $\omega$ , Larmor $\omega_0$ ): $\eta=\omega\sin\alpha/2\omega_0$ — a closed form worth memorizing.
The error's decay law is set by the schedule's smoothness: endpoint velocity kinks give a measured $1-P\propto1/T^2$ (fitted slope $-1.99$ ); a $C^\infty$ ramp falls faster than any power ( $5\times10^{-2}\to1.5\times10^{-8}$ over $T=10\to80$ ).
$\eta\ll1$ is necessary, not sufficient: a resonant modulation with $\eta\le0.01$ at all times still inverts the state completely (Marzlin–Sanders). Rigorous sufficient conditions exist (Kato 1950; Jansen–Ruskai–Seiler 2007).
Key idea: adiabaticity is a competition between the rate at which the Hamiltonian reshapes its eigenstates and the rate at which quantum phases spin; the criterion $\eta=\hbar|\langle m|\dot H|n\rangle|/(E_m-E_n)^2\ll1$ scores that competition — but only phase-coherent bookkeeping over the whole protocol, not the instantaneous score, decides the winner.

The core insight

Adiabaticity is a race between the rate at which the Hamiltonian reshapes its eigenstates and the rate at which quantum phases spin; the score $\eta=\hbar\,|\langle m|\dot H|n\rangle|/(E_m-E_n)^2\ll1$ decides it — one gap power for coupling, one for phase averaging — provided no drive resonance conspires with the internal clock.

Where this lives today

Adiabatic quantum computation stakes a hardware paradigm on this article's inequality: encode the answer in the ground state of $H_{\text{final}}$ , interpolate from an easy $H_{\text{initial}}$ , and the required runtime scales as the inverse square of the minimum spectral gap along the path — criterion (3) read as an engineering budget. Real annealers live by exactly our fine print: schedules smoothed at the endpoints (the $1/T^2$ kink law), the minimum-gap window crossed with Landau–Zener physics in mind, and control-line noise filtered at gap frequencies because resonant leakage defeats any slowness. The same bookkeeping runs STIRAP — stimulated Raman adiabatic passage — where population moves between atomic levels through a dark state steered slowly against a Rabi-frequency gap, its smooth pulse edges doing precisely the endpoint engineering derived above. And every shortcut-to-adiabaticity protocol begins by computing the very term $\langle m|\dot n\rangle$ this article bounds — then cancels it with an auxiliary field instead of suppressing it with patience.

The physical picture

Carry a brimful cup across a room. Walk smoothly and the surface stays flat; jerk — even a small jerk, if it repeats at the slosh frequency — and coffee climbs the walls. The restoring frequency of the coffee is the energy gap: a stiff system forgives fast hands. The repeated small jerk is the villain of this article's final act — tiny, gentle, ruinous, because it is timed. The clean laboratory for everything here: a spin-½ whose field direction $\hat n(t)$ rides a cone of half-angle $\alpha$ , sweeping the azimuth at rate $\omega$ . The spin precesses at the Larmor rate $\omega_0$ — the internal clock, gap $\hbar\omega_0$ . The field direction moves at $\omega\sin\alpha$ — the external clock. When the internal clock is far faster, the spin hugs the moving field; the leakage bookkeeping below gives exactly $\eta=\omega\sin\alpha/2\omega_0$ .

Figure

The two clocks: a spin hugging a slowly swept field direction

Two subtleties hide in this picture. First, how the error decays with protocol time depends on the smoothness of the schedule — where the sweep velocity jumps, amplitude is knocked loose, and those boundary kicks dominate. Second, leakage amplitudes add coherently: a drive modulating in step with the internal clock can stack thousands of individually negligible kicks into complete inversion.

Figure

Measured adiabatic error vs protocol time: smoothness decides the law, resonance breaks it

The engine, recalled

The Adiabatic Theorem builds the machinery this article runs on: expand the exact state over instantaneous eigenstates $H(t)|n(t)\rangle=E_n(t)|n(t)\rangle$ and the Schrödinger equation becomes exact coupled amplitude equations driven by the off-diagonal $\langle m|\dot n\rangle$ . It also proves the identity used on every line here:

\langle m|\dot n\rangle \;=\; \frac{\langle m|\dot H|n\rangle}{E_n-E_m}\,,\qquad m\neq n \qquad (1)

where: $\langle m|\dot H|n\rangle$ — the matrix element of the rate of change of the Hamiltonian between two instantaneous eigenstates, the only way the drive talks across the spectrum; $E_n-E_m$ — the instantaneous gap. Equation (1) is the whole game in one line: eigenstates reshape at a rate set by $\dot H$ , demagnified by the gap. A perturbation that moves fast but connects states across a huge gap barely stirs them; a lazy drive near a small gap is violent. The amplitude in target state $m$ , starting from $n$ alone, obeys to first order

c_m(T) \;\approx\; -\int_0^T \frac{\langle m|\dot H|n\rangle}{E_n-E_m}\; e^{\,i\int_0^{t}\omega_{nm}(t')\,\mathrm{d}t'}\;\mathrm{d}t \qquad (2)

where: $\omega_{nm}=(E_n-E_m)/\hbar$ — the Bohr frequency at which the relative phase spins. The integrand is a slow envelope (the coupling) times a fast phasor; the criterion, the smoothness laws, and the resonant catastrophe are all statements about what that oscillating integral can accumulate.

Derivation AThe adiabaticity criterion from the leakage integral

The goal is the quantitative validity condition: a dimensionless number, computable from $H(t)$ alone, that bounds the population leaking out of an instantaneous eigenstate. The strategy is to evaluate the first-order amplitude (2) by exploiting the separation of time scales — integrate by parts against the fast phase, so the slow envelope is sampled only where it fails to be smooth. The result explains why the gap enters squared. Watch what the argument quietly assumes: an envelope with no Fourier weight at the Bohr frequency itself — the loophole exploited later in this article.

Step 1. Write the integrand of (2) as $F(t)\,e^{i\Phi(t)}$ with envelope $F=\langle m|\dot H|n\rangle/(E_n-E_m)$ and phase $\Phi(t)=\int_0^t \omega_{nm}\,\mathrm{d}t'$ , so $\dot\Phi=\omega_{nm}$ is the fast frequency.

Step 2. Use $e^{i\Phi}=\frac{1}{i\dot\Phi}\frac{\mathrm{d}}{\mathrm{d}t}e^{i\Phi}$ and integrate by parts:

c_m(T)= -\left[\frac{F(t)}{i\,\omega_{nm}(t)}\,e^{i\Phi(t)}\right]_0^T \;+\;\int_0^T \frac{\mathrm{d}}{\mathrm{d}t}\!\left(\frac{F}{i\,\omega_{nm}}\right) e^{i\Phi}\,\mathrm{d}t

Why this move: the fast phasor is nearly a total derivative, so the integral is controlled not by the envelope's size but by how much it changes per phase revolution — "fast phases average slow couplings away," as mathematics.

Step 3. The boundary terms are the leading contribution. Their magnitude is

\left|\frac{F}{\omega_{nm}}\right| \;=\; \frac{\hbar\,|\langle m|\dot H|n\rangle|}{(E_n-E_m)^2}\;\equiv\;\eta

One power of the gap came from the coupling (1); the second came from the phase integration in Step 2. This is the promised dimensionless score: $[\hbar]\,[\mathrm{energy/time}]/[\mathrm{energy}]^2=1$ .

Step 4. The remaining integral in Step 2 is smaller by another factor of (slow rate)/(gap frequency) and can be integrated by parts again, and again: each schedule derivative that exists and vanishes at the endpoints buys another power of $1/T$ . The hierarchy terminates only at the drive's smoothness — the error's decay law is a statement about the schedule's differentiability class, quantified in the next derivation.

Step 5. Adiabatic following therefore demands, level pair by level pair,

\eta \;=\; \frac{\hbar\,\bigl|\langle m|\dot H|n\rangle\bigr|}{(E_m-E_n)^2} \;\ll\; 1 \qquad (3)

read as a two-clock inequality: with $\dot H\sim\Delta H/T$ , condition (3) says the drive time must exceed the internal phase time $\hbar/\Delta E$ by the further ratio of coupling to gap. Verified: identity (1) checked symbolically on a two-level model; $\eta$ confirmed dimensionless; the spin-cone closed form of the next section and the boundary-term amplitude of Step 3 reproduced by direct TDSE integration (verify script, checks A1, C1–C3, E2).

The spin-cone laboratory

Every claim so far can be tested on a system small enough to integrate exactly: $H(t)=\tfrac{\hbar\omega_0}{2}\,\hat n(t)\cdot\boldsymbol\sigma$ , $\hat n=(\sin\alpha\cos\omega t,\ \sin\alpha\sin\omega t,\ \cos\alpha)$ . The instantaneous eigenstates are the half-angle eigenspinors constructed in Spinors and Two-Component Wavefunctions; the gap is rigidly $\hbar\omega_0$ . Feeding this into definition (3) — two lines once $\dot H=\tfrac{\hbar\omega_0}{2}\,\omega\,(\partial_\phi\hat n)\cdot\boldsymbol\sigma$ is written — collapses the criterion to a formula you can carry in your head:

\eta_{\text{cone}} \;=\; \frac{\omega\,\sin\alpha}{2\,\omega_0} \qquad (4)

where: $\omega\sin\alpha$ — the angular speed of the field direction; $\omega_0$ — the Larmor frequency, the gap over $\hbar$ . The criterion is exactly the cartoon: direction-speed over precession-speed. Lap count drops out — only the instantaneous rate matters to $\eta$ . The same laboratory now delivers the article's owned result: what actually controls how good the protocol is at finite $T$ .

Go deeper · Derivation B · 2 minThe quadratic error law for kinked ramps — predicted, then measured

Derivation BThe quadratic error law for kinked ramps — predicted, then measured

The goal is the sharpest practical statement about finite-time protocols: for a schedule whose sweep velocity jumps at the endpoints (the ubiquitous "linear ramp"), the leakage is dominated by the two boundary kicks of Step 3 above, giving a $1/T^2$ population error with a computable coefficient — while a $C^\infty$ schedule, all of whose derivatives vanish at the endpoints, has no boundary terms at any order and must fall faster than any power of $T$ . The strategy: turn the integration-by-parts hierarchy into predictions, then check both laws by brute-force integration of the Schrödinger equation. The subtlety: the two kicks interfere, so the honest observable is the phase-averaged error.

Step 1. For the cone swept through one full turn in total time $T$ at constant rate, the envelope $F/\omega_{nm}$ of the criterion derivation is constant in the interior: the interior integral vanishes identically and only the endpoint kicks survive. Each kick has amplitude $\eta_T=\pi\sin\alpha/(\omega_0 T)$ — formula (4) with $\omega=2\pi/T$ .

Step 2. Two kicks separated by the full accumulated phase $\Phi(T)$ add coherently:

1-P_{\text{stay}} \;=\; \bigl|\,\eta_T\,e^{i\Phi(T)}-\eta_T\,\bigr|^2\!\sim\; 4\,\eta_T^{\,2}\sin^2\!\tfrac{\Phi(T)}{2}

Why the averaging: $\Phi(T)\approx\omega_0 T$ sweeps many cycles as $T$ varies, so individual protocols scatter anywhere inside the envelope $4\eta_T^2$ ; the phase-averaged error is $2\eta_T^{\,2}\propto 1/T^2$ — that, not any single run, is the law.

Step 3. Measure it. Integrating the TDSE (piecewise-exponential stepping, convergence-checked) for $\omega_0=1$ , $\alpha=60^\circ$ , averaging each nominal $T$ over a cluster of nearby values:

\begin{array}{r|ccccc} T & 40 & 80 & 160 & 320 & 640\\ \hline \langle 1-P\rangle & 8.1\times10^{-3} & 2.5\times10^{-3} & 4.8\times10^{-4} & 1.3\times10^{-4} & 3.5\times10^{-5} \end{array}

Log–log fit: slope $-1.99$ . Measured-to-predicted ( $2\eta_T^2$ ) ratios: $0.87,\ 1.08,\ 0.84,\ 0.93,\ 0.98$ — coefficient right, not just exponent.

Step 4. Replace the schedule by a $C^\infty$ ramp (a bump-function smoothstep: every derivative zero at both endpoints) traversing the same loop. Now no integration by parts ever terminates on a boundary term, so the error must fall faster than any fixed power. Measured: $1-P = 5.2\times10^{-2},\ 5.2\times10^{-2},\ 6.9\times10^{-5},\ 1.5\times10^{-8}$ at $T=10,20,40,80$ — the local log–log slope steepens from $-0.01$ to $-9.6$ to $-12.2$ , the numerical signature of super-polynomial decay. At $T=80$ the smooth schedule beats the kinked one by five orders of magnitude for the same total time.

Step 5. The practical law: in the adiabatic regime the error is set not by how slow you go but by how you start and stop — endpoint engineering beats protocol lengthening, a fact annealing and STIRAP pulse designers exploit deliberately. Verified: slope $-1.991$ , coefficient ratios $0.87$ – $1.08$ against the $2\eta_T^2$ prediction, smooth-ramp slopes $-0.01\to-9.6\to-12.2$ , and integrator convergence at the $10^{-8}$ floor (doubled step count agrees), verify script checks E0–E4, G2.

When the criterion lies: the resonant counterexample

The derivation of $\eta$ contained a confession: integration by parts presumes the envelope has no Fourier weight at the Bohr frequency. Break that deliberately. Freeze the cone sweep and instead wobble the cone angle by a whisker, $\alpha(t)=\alpha_0+\varepsilon\sin(\omega_d t)$ , $\varepsilon=0.02$ rad ( $1.1^\circ$ ) — tuned to the gap, $\omega_d=\omega_0$ . Every instantaneous scorecard calls this drive innocent: $\eta(t)\le\varepsilon/2=0.01$ at every moment. But the envelope of integral (2) now oscillates in step with the phasor, so the phasor never averages the coupling away: each cycle deposits the same signed sliver of amplitude, and the slivers stack. This is the driven two-level physics solved exactly in Transition Probability for a Periodic Two-Level System — on resonance its generalized Rabi frequency $\Omega_R=\sqrt{\Omega_0^2+\Delta^2}$ reduces to the bare drive strength, and population cycles completely no matter how weak the drive; in the rotating wave approximation the co-rotating half of the wobble is a static transverse field that nothing suppresses. The measured verdict: with $\eta\le0.01$ throughout, $P_{\text{stay}}$ collapses to $0.000$ by $t=\pi/(\varepsilon\omega_0/2)\approx314/\omega_0$ . This is the Marzlin–Sanders inconsistency (2004) in its simplest clothing, and it recalibrates $\eta$ : necessary, not sufficient. The rigorous theorems — Kato's 1950 gap-condition proof, the Jansen–Ruskai–Seiler bounds (2007) — close precisely this loophole by constraining $\dot H$ and $\ddot H$ over the whole protocol, not instant by instant.

The failure catalogue

Three ways adiabaticity dies, three remedies:

Closing gaps. Where two levels approach, $\eta\propto1/\Delta E^2$ blows up and no finite slowness survives a true crossing. The generic case is the avoided crossing swept at finite rate — the two-level Hamiltonian with eigenvalues $\pm\tfrac12\sqrt{\alpha^2t^2+\Delta^2}$ pinching to minimum gap $\Delta$ at $t=0$ . The fraction that jumps the gap is an exactly solvable problem — Landau, Zener, Stückelberg and Majorana all solved it in the same year, 1932 — and it belongs to Landau–Zener Transitions, which derives it from scratch. Here we keep only the criterion's verdict: the fate of the state is sealed in the brief window near the pinch, where $\Delta^2$ competes with the sweep rate — the same two ingredients $\eta$ says matter.

Exact degeneracies. If two levels are degenerate while the drive acts, $\eta$ is meaningless — but symmetry can rescue the protocol. When an operator commutes with the full Hamiltonian at all times and distinguishes the degenerate partners, the drive cannot connect them — the mechanism Good Quantum Numbers and Symmetry-Adapted Bases proves for perturbation theory ("a good quantum number is a label the perturbation cannot change"), running here in continuous time. Engineer the labels, and passage sails through a naively fatal $\Delta E=0$ .

Resonant drives. The counterexample above: no gap closes, $\eta$ stays tiny, the state dies anyway by coherent accumulation. Remedy: keep the drive's spectral weight away from all Bohr frequencies — a spectral condition, invisible to any instantaneous criterion.

Limiting cases

case	result	meaning
deep adiabatic, smooth schedule	error falls faster than any power of $T$	no boundary terms, no resonances; the practical ideal
deep adiabatic, kinked schedule	$\langle 1-P\rangle \approx 2\eta_T^2 \propto 1/T^2$	endpoint velocity jumps dominate; measured slope $-1.99$ in the cone laboratory
sweep through avoided crossing	jump fraction set by $\Delta^2$ vs $\hbar\alpha$	criterion diverges at the pinch; exact treatment in Landau–Zener Transitions
sudden limit $T\ll\hbar/\Delta E$	state frozen; populations $=\vert \langle n_{\text{new}}\vert \psi\rangle\vert ^2$	the opposite extreme of the two-clock race — the drive wins outright
resonant modulation, any $\eta$	complete inversion in time $\pi/(\varepsilon\omega_0/2)$	coherent accumulation defeats the instantaneous criterion (Marzlin–Sanders)
degenerate levels + symmetry label	leakage forbidden exactly	good-quantum-number rescue: the drive cannot change the label

Worked examples

Example 1 — score the spin sweep (confidence builder). An electron spin in a $\sim0.1$ T field: Larmor (gap) frequency $f_0=2.8$ GHz. The field direction rides a cone of half-angle $\alpha=60^\circ$ . Score sweep rates $f_{\text{sweep}}=1$ MHz and $1$ GHz.

Step 1. The two clocks. Internal: $\omega_0=2\pi f_0=1.76\times10^{10}\ \text{s}^{-1}$ . External: $\omega=2\pi f_{\text{sweep}}$ , with $\sin60^\circ=0.866$ .

Step 2. Apply the closed form (4), $\eta=\omega\sin\alpha/2\omega_0$ . For 1 MHz: $\eta=(2\pi\times10^{6})(0.866)/(2\times1.76\times10^{10})$ $=1.5\times10^{-4}$ . For 1 GHz: $\eta=0.155$ .

Step 3. Read the boundary numbers. At 1 MHz the protocol is deeply adiabatic — a kinked one-turn ramp leaves a population error of order $2\eta^2\approx5\times10^{-8}$ . At 1 GHz, $\eta$ sits within a factor $\sim6$ of unity: percent-level leakage per kick, the adiabatic language itself failing. $\eta=1.5\times10^{-4}$ (safe) versus $\eta=0.155$ (marginal).

Example 2 — the innocent drive that kills the state (exam level). Same spin, same $2.8$ GHz gap; cone angle held near $60^\circ$ but wobbling by $\varepsilon=0.02$ rad at exactly the gap frequency. How long until complete inversion?

Step 1. Score it naively: the criterion gives $\eta(t)\le\varepsilon/2=0.01$ at every instant — a hundred times smaller than the marginal case above. Naive verdict: perfectly safe.

Step 2. Score it honestly: the wobble is a weak driven two-level problem transferring population at $\Omega_w=\varepsilon\omega_0/2$ . Complete inversion at $t_{\text{flip}}=\pi/\Omega_w=2\pi/(\varepsilon\omega_0)$ , i.e. $\omega_0t\approx314$ — about 50 precession periods.

Step 3. Put in the clock: $t_{\text{flip}}=314/(2\pi\times2.8\times10^{9}\ \text{s}^{-1})=1.8\times10^{-8}$ s. Total inversion in $\approx18$ ns, by a $1.1^\circ$ wobble the criterion certifies as safe — TDSE-verified ( $P_{\text{stay}}=0.000$ ). Audit the drive spectrum, not just $\eta$ .

Origins · 1 minEttore Majorana and how this was discovered

Ettore Majorana

1906–1938 · Rome, Naples

Majorana was the prodigy of Fermi's Via Panisperna group in Rome. In 1932 — the same year Landau, Zener and Stückelberg each attacked the swept level crossing — Majorana solved his version: a spin in a magnetic field whose direction varies in time, passing near a field zero ("Atomi orientati in campo magnetico variabile"). His nonadiabatic spin-flip formula, derived for real beam experiments rather than a toy model, is why the modern literature says Landau–Zener–Stückelberg–Majorana. Fermi ranked him with Galileo and Newton in raw ability and despaired of his indifference to publishing: fewer than ten papers, at least two of which name whole fields (this one, and the Majorana fermion). Appointed professor in Naples in 1937, he boarded the Palermo–Naples mail steamer in March 1938 and was never seen again. The physics he clarified runs daily: every avoided-crossing sweep in an annealer traverses his 1932 problem.

History

1916

— Ehrenfest's adiabatic hypothesis: the quantities that survive slow change are the ones the old quantum theory should quantize

1928

— Born and Fock prove the quantum adiabatic theorem for discrete non-degenerate spectra

1932

— Landau, Zener, Stückelberg and Majorana independently solve the swept two-level crossing, quantifying the theorem's failure

1950

— Kato re-proves the adiabatic theorem with operator methods, honest gap conditions, and degeneracies handled

2004

— Marzlin and Sanders exhibit resonant counterexamples to the naive criterion, forcing the modern rigor program (Jansen–Ruskai–Seiler bounds, 2007)

Going further

References. Griffiths & Schroeter, Introduction to Quantum Mechanics, 3rd ed., Ch. 11 (Adiabatic Approximation, Berry's Phase) · Sakurai & Napolitano, Modern Quantum Mechanics, 2nd ed., §5.6 · Messiah, Quantum Mechanics, Vol. II, Ch. XVII (the rigorous adiabatic theorem) · Born & Fock, Z. Phys. 51, 165 (1928) · Kato, J. Phys. Soc. Jpn. 5, 435 (1950) · Marzlin & Sanders, Phys. Rev. Lett. 93, 160408 (2004) · Jansen, Ruskai & Seiler, J. Math. Phys. 48, 102111 (2007).

Provenance. Equations cross-checked against Griffiths & Schroeter Ch. 11 (Eqs. 11.25–11.62), Berry (1984), and Messiah Vol. II Ch. XVII. A verifier should confirm: (1) the non-adiabatic coupling identity $\langle m|\dot n\rangle=\langle m|\dot H|n\rangle/(E_n-E_m)$ by differentiating the eigen-equation; (2) the validity condition's $1/(\Delta E)^2$ scaling via integration by parts of the first-order amplitude; (3) the spin-1/2 Berry phase $\gamma_\pm=\mp\Omega/2$ from monopole curvature; (4) the Landau–Zener exponent $\pi\Delta^2/(2\hbar\alpha)$ against Zener (1932). No nonstandard numerical constants are asserted. Ownership note: provenance items (3) and (4) belong to Geometric (Berry) Phase and Landau–Zener Transitions respectively; items (1) and (2) are verified in this article's script.

Builds on: The Interaction Picture and Time-Dependent Hamiltonians · Time-Dependent Perturbation Theory: General Formalism · First-Order Transition Amplitudes and Probabilities Leads to: Transition Probability for a Periodic Two-Level System · Dyson Series and the Time-Ordered Evolution Operator · Constant Perturbation Switched On

Is the adiabatic criterion a theorem or a rule of thumb?

A rule of thumb with a theorem behind it — the gap between the two is this article's point. The first-order estimate producing $\eta\ll1$ is honest perturbative bookkeeping, and rigorous versions (Born–Fock 1928, Kato 1950) prove genuine theorems under gap and smoothness conditions. But the instantaneous criterion is only necessary: the Marzlin–Sanders construction — here, a resonant wobble with $\eta\le0.01$ that fully inverts the state — shows sufficiency needs the drive's whole time dependence controlled, which the rigorous bounds do by constraining $\dot H$ and $\ddot H$ over the protocol.

Why do I keep seeing the condition written with the Bohr frequency instead of the gap squared?

Same statement, different clothes: $\eta\ll1$ reads as $|\langle m|\dot H|n\rangle|/\Delta E\ll\Delta E/\hbar=\omega_{mn}$ — "eigenstates must reshape slowly compared to the Bohr frequency." One form emphasizes the dimensionless score, the other the two clocks; they carry identical content.

What actually breaks at a velocity kink? The Hamiltonian is still continuous there.

Continuity of $H$ is not the currency — smoothness of the schedule is. Each integration by parts trades a power of $1/T$ for one envelope derivative evaluated at the endpoints; a kink in $\dot H$ leaves the first boundary term standing, hence amplitude $\propto1/T$ and population $\propto1/T^2$ . Physically: an abrupt start is a small quench, and quenches project — gently here, but dominantly at long $T$ .

Is the sudden approximation just the adiabatic criterion reversed?

The other end of the same race. Adiabatic: drive time $\gg\hbar/\Delta E$ , populations of instantaneous eigenstates frozen. Sudden: drive time $\ll\hbar/\Delta E$ , the state vector itself frozen, populations redistributed as overlaps with the new eigenbasis. Between them sits the swept avoided crossing, solved exactly in Landau–Zener Transitions.

My system has a degeneracy along the path. Is adiabatic passage automatically impossible?

No — it depends on whether the drive can use it. If a symmetry operator commutes with $H(t)$ throughout and labels the partners differently, the coupling vanishes identically and passage is protected — the continuous-time good-quantum-numbers theorem. Danger arises when the degeneracy is approached by levels the drive does connect: then $\eta$ genuinely diverges and crossing physics takes over.

Where does the Berry phase fit into this article?

It doesn't — deliberately. The same instantaneous-eigenbasis expansion also produces a geometric phase riding on the surviving amplitude, but that phase is a closed-loop phenomenon with its own gauge structure and its own article: Geometric (Berry) Phase derives and measures it in the same spin-cone laboratory used here. This article polices whether you stay in the eigenstate; that one asks what phase you carry home if you do.