Fermi's Golden Rule

Fermi's Golden Rule answers one of the most consequential questions in quantum mechanics: when a weak perturbation couples a discrete initial state to a dense continuum of final states, how fast does the system leave the initial state? The answer is a constant rate — a well-defined number of transitions per unit time — equal to $(2\pi/\hbar)|V_{fi}|^2\rho(E_f)$ , the squared coupling matrix element times the density of final states, evaluated where energy is conserved. That the answer is a rate at all, rather than the oscillation of an isolated two-level system, is the whole point: a rate means a lifetime, a linewidth, an exponential decay law — the irreversibility that lets excited atoms actually decay. This article derives the rule and then does something the standard textbook treatment usually skips: it proves, on a concrete model, why coupling one state to a continuum turns the perturbative linear growth $P=\Gamma t$ into genuine exponential decay $e^{-\Gamma t}$ , and pins down the exact time window in which the rate is meaningful. The derivation consumes two neighbours by name and re-derives nothing they own: the sharpening resonance function $\sin^2(\omega_{fi}t/2)/(\omega_{fi}/2)^2\to 2\pi t\,\delta(\omega_{fi})$ is First-Order Transition Amplitudes and Probabilities', and the counting factor $\rho(E_f)$ is Density of Final States'. What is owned here is the assembly — δ-limit times density gives a rate — and the self-consistency that upgrades that rate to a decay law. A word on the name, because it is a genuine trap. The rule was first written down by Dirac in 1927, in the paper that founded time-dependent perturbation theory. Fermi never derived it; he named it — "Golden Rule No. 2" — in his 1949–50 Chicago lecture notes, and the name stuck to the wrong person. State this accurately: the rule is Dirac's, the name is Fermi's. It underlies photoabsorption, the Einstein $A$ coefficient for spontaneous emission, beta-decay rates, and scattering rates in solids.

Key takeaways

Fermi's Golden Rule gives the constant transition rate from a discrete state into a continuum under a weak perturbation: $\Gamma_{i\to f}=\dfrac{2\pi}{\hbar}|\langle f|\hat V|i\rangle|^2\,\rho(E_f)$ , evaluated on the energy shell.
It is assembled, not derived from scratch: the resonance function's δ-limit (owned by first-order transition probability) times the density of states (owned by density of final states) produces the rate — this article owns the assembly and its consequences.
A single final state gives Rabi oscillation, never a rate. Only a genuine continuum of final states makes a constant rate exist.
The rate is valid in a window $1/\Delta\omega_\text{band}\ll t\ll 1/\Gamma$ : long enough for the resonance to sharpen into a δ-function, short enough that the initial state is barely depleted.
Iterating the rate self-consistently upgrades the linear growth $P=\Gamma t$ into exponential decay $P_i(t)=e^{-\Gamma t}$ , with lifetime $\tau=1/\Gamma$ and natural linewidth $\Gamma_E=\hbar\Gamma$ .
The name is a historical accident: the rule is Dirac's (1927); Fermi merely called it "Golden Rule No. 2" (1950).
Key idea: sum the sharpening first-order resonance over a continuum, and $\sin^2(\omega_{fi}t/2)/(\omega_{fi}/2)^2\to 2\pi t\,\delta(\omega_{fi})$ turns a $t^2$ blip into a probability linear in $t$ ; its slope is the rate $\Gamma=(2\pi/\hbar)|V_{fi}|^2\rho(E_f)$ .

The core insight

Sum the first-order transition probability over a dense band of final states. The resonance function $\sin^2(\omega_{fi}t/2)/(\omega_{fi}/2)^2$ sharpens with time (height $\propto t^2$ , width $\propto 1/t$ , area $\propto t$ ) into $2\pi t\,\delta(\omega_{fi})$ . Integrating against the density of states leaves a probability that grows linearly in time — and a linear probability has a constant slope, the rate $\Gamma=(2\pi/\hbar)|V_{fi}|^2\rho(E_f)$ .

Where this lives today

The golden rule is the workhorse rate formula across physics. In atomic and molecular physics it gives photoabsorption and photoionisation cross sections and the radiative lifetimes of excited states. In condensed matter it sets electron–phonon and impurity scattering rates, hence electrical resistivity and carrier mobility, and the recombination rates in LEDs and lasers. In nuclear and particle physics, dressed with the correct phase-space density, it gives decay widths and the shapes of beta spectra (the Sargent rule, rate $\propto Q^5$ , is a golden-rule phase-space statement). Wherever a quantum system leaks into a continuum — a qubit into its electromagnetic environment (the $T_1$ relaxation time), a molecule into a solvent bath — the leading estimate of the rate is Dirac's 1927 formula. Its breakdown at short times (the Zeno regime) and long times (non-exponential tails) is itself an active experimental subject in cavity QED and open quantum systems.

The physical picture

Think of the first-order probability of reaching a final state of energy $E_f$ as a diffraction pattern in energy: a central peak at $E_f=E_i$ whose height grows like $t^2$ and whose width shrinks like $1/t$ . A single final state just rides that peak — its probability oscillates or, exactly on resonance, grows quadratically, and never settles into a steady rate. Now surround that peak with a dense band of final states. As time grows the peak narrows, so fewer states stay inside it, but each of those states carries a taller probability. The two effects — height $\uparrow t^2$ , width $\downarrow 1/t$ — combine so that the total probability collected across the band grows only linearly in $t$ . That linear slope is the rate. This is the entire trick, and the first figure makes it visible. As $t$ increases the resonance sharpens toward a delta function centred on $E_f=E_i$ ; the continuum "catches" exactly the area under the peak, which grows as $t$ . The rate factorises cleanly into two independent questions: how strongly are the states coupled (the matrix element $|V_{fi}|^2$ ) and how many states are there to decay into (the density $\rho$ ). A weak coupling into a dense band can be as fast as a strong coupling into a sparse one.

Figure

The resonance function sharpens with time (height ∝ t², width ∝ 1/t, area ∝ t) toward the energy-conserving δ-function

Because the rate $\Gamma$ is constant, the survival probability of the initial state obeys $\dot P_i=-\Gamma P_i$ , which integrates to exponential decay $P_i=e^{-\Gamma t}$ , lifetime $\tau=1/\Gamma$ , and — by Fourier-transforming the decaying amplitude — a Lorentzian spectral line of full width $\Gamma_E=\hbar\Gamma$ . This is the last figure's message: the golden rule secretly contains the exponential decay law, but only in a specific time window. Too early (shorter than the inverse bandwidth), the probability is quadratic and no rate has formed yet — the quantum Zeno regime. Too late (once $\Gamma t\sim 1$ ), the naive linear growth would exceed one, and only re-summing the depletion restores the physical exponential.

Figure

The three eras of the survival probability: Zeno transient, golden-rule plateau, exponential depletion — with the plateau window flagged

The rate and its ingredients

The golden rule is not a fresh calculation; it is the marriage of two results this article does not own. It is worth stating both parents cleanly before assembling them, so the reader sees exactly which piece comes from where. The first parent is the first-order transition probability into a single final state, which oscillates as a diffraction-in-energy function. The second is the counting factor that turns a sum over discrete states into an energy integral. We write the master rate, then trace each factor to its owner.

\Gamma_{i\to f}=\frac{2\pi}{\hbar}\,\big|\langle f|\hat V|i\rangle\big|^2\,\rho(E_f)\Big|_{E_f=E_i} \qquad (1)

where: $\Gamma_{i\to f}$ — transition rate (probability per unit time) from $|i\rangle$ into the continuum; $\hat V$ — the (constant-amplitude) perturbation operator; $\langle f|\hat V|i\rangle$ — the matrix element coupling initial and final states; $\rho(E_f)$ — the density of final states per unit energy, evaluated on the energy shell $E_f=E_i$ . Physically: strength of coupling ( $|V_{fi}|^2$ ) times number of destinations per unit energy ( $\rho$ ), with $2\pi/\hbar$ the universal conversion factor from the δ-limit. For a harmonic perturbation $\hat V(t)=\hat V_0\cos\omega t$ the same assembly holds, but the energy shell shifts by one quantum $\hbar\omega$ — the system absorbs ( $E_f=E_i+\hbar\omega$ ) or is stimulated to emit ( $E_f=E_i-\hbar\omega$ ). Harmonic (Sinusoidal) Perturbations and Resonance owns that shifted resonance; here we only note where the shell lands.

\Gamma_{i\to f}=\frac{2\pi}{\hbar}\,\big|\langle f|\hat V_0|i\rangle\big|^2\,\rho(E_f)\Big|_{E_f=E_i\pm\hbar\omega} \qquad (2)

where: $\hat V_0$ — the amplitude operator of the harmonic drive; $E_i+\hbar\omega$ — absorption resonance; $E_i-\hbar\omega$ — stimulated-emission resonance. The constant-perturbation case (1) is the $\omega\to0$ limit of (2). The single-final-state probability that seeds everything is owned by first-order transition probability; we import it verbatim as the input to the sum, not as something to re-derive.

P_{i\to f}^{(1)}(t)=\frac{|V_{fi}|^2}{\hbar^2}\,\frac{\sin^2(\omega_{fi}t/2)}{(\omega_{fi}/2)^2},\qquad \omega_{fi}=\frac{E_f-E_i}{\hbar} \qquad (3)

where: $P^{(1)}_{i\to f}$ — first-order probability to a specific state $|f\rangle$ ; $\omega_{fi}$ — the Bohr angular frequency of the transition. This is the resonance function that the sum sharpens into a δ-function.

Derivations

Derivation AAssembling the golden rule from the δ-limit and the density of states

The rule is not derived from the Schrödinger equation here — it is assembled from two owned results. The goal of this card is to make the assembly airtight: to show precisely how the sharpening resonance function of first-order transition probability, weighted by the density of final states, collapses to a probability linear in time whose slope is the rate. Each imported piece is flagged with its owner; the only new content is how they combine and where the $2\pi/\hbar$ comes from.

Step 1. Import the single-state probability from first-order transition probability (its equation, not re-derived here):

P_{i\to f}^{(1)}(t)=\frac{|V_{fi}|^2}{\hbar^2}\,\frac{\sin^2(\omega_{fi}t/2)}{(\omega_{fi}/2)^2}.

Why: this is the exact leading-order probability into one final state; the golden rule is what happens when many such states are available.

Step 2. Replace the sum over discrete final states by an energy integral weighted by the density of states — the definition owned by Density of Final States:

P(t)=\sum_f P_{i\to f}^{(1)}(t)\;\longrightarrow\;\int \mathrm{d}E_f\,\rho(E_f)\,P_{i\to f}^{(1)}(t).

Why: a continuum has no discrete labels to sum; $\rho(E_f)\,\mathrm{d}E_f$ counts the states in each energy slice.

Step 3. Use the sharpening limit (owned by first-order transition probability, cashed here). Since $\int_{-\infty}^{\infty}\sin^2(\omega t/2)/(\omega/2)^2\,\mathrm{d}\omega=2\pi t$ and the function concentrates at $\omega=0$ , it behaves as $2\pi t\,\delta(\omega_{fi})$ . Convert to an energy delta via $\delta(\omega_{fi})=\hbar\,\delta(E_f-E_i)$ :

\frac{\sin^2(\omega_{fi}t/2)}{(\omega_{fi}/2)^2}\;\longrightarrow\;2\pi t\,\hbar\,\delta(E_f-E_i).

Why: the area under the peak grows as $t$ while the width shrinks — the defining behaviour of a nascent delta of weight $2\pi t$ . The extra $\hbar$ is the Jacobian $\delta(ax)=\delta(x)/|a|$ with $a=1/\hbar$ .

Step 4. Insert and integrate. With $|V_{fi}|^2$ and $\rho$ slowly varying across the sharp peak, pull them out at $E_f=E_i$ :

P(t)=\frac{|V_{fi}|^2}{\hbar^2}\,\rho(E_i)\,(2\pi t\,\hbar)=\frac{2\pi t}{\hbar}\,|V_{fi}|^2\,\rho(E_i).

Why: the delta collapses the integral to the value of the smooth integrand exactly on the energy shell.

Step 5. Read off the rate as the slope of a probability linear in $t$ :

\Gamma=\frac{\mathrm{d}P}{\mathrm{d}t}=\frac{2\pi}{\hbar}\,|V_{fi}|^2\,\rho(E_i).

Why: a linear $P(t)$ has a constant derivative — this is what "a rate exists" means. Verified: symbolically in verify_aq-qm-fermi-golden-rule.py — the δ-collapse assembly reproduces $\Gamma=(2\pi/\hbar)|V_{fi}|^2\rho$ exactly; dimensions $[\Gamma]=[\text{J}]^2\cdot[\text{J}]^{-1}\cdot[\text{J\,s}]^{-1}=\text{s}^{-1}$ ; the nascent-delta area $\int\sin^2(\omega t/2)/(\omega/2)^2\,\mathrm{d}\omega=2\pi t$ recomputed numerically to 0.02%.

Go deeper · Derivation B · 2 minWhy a continuum turns the rate into exponential decay

Derivation BWhy a continuum turns the rate into exponential decay

This is the article's owned structural result, and it repairs a defect that first-order theory alone cannot: taken literally, $P=\Gamma t$ would exceed one at long times. The goal here is to prove — not merely assert — that a discrete state coupled to a genuine continuum decays exponentially at exactly the golden-rule rate, and to find the time window in which the rate is meaningful. The proof is by explicit construction: couple one state to a flat band and watch it decay.

Step 1. Set up the self-consistency. First-order theory drains the initial state at rate $\Gamma$ , but only while it is still occupied. Allowing the initial amplitude to deplete gives the rate equation

\frac{\mathrm{d}P_i}{\mathrm{d}t}=-\Gamma\,P_i.

Why: the probability of leaving per unit time is $\Gamma$ times the probability of still being present — the Markov (memoryless) assumption, exact in the golden-rule window.

Step 2. Integrate:

P_i(t)=e^{-\Gamma t},\qquad \tau=\frac1\Gamma,\qquad \Gamma_E=\hbar\Gamma.

Why: exponential survival, with the naive linear law $1-\Gamma t$ recovered as its short-time tangent, $1-e^{-\Gamma t}\approx\Gamma t$ .

Step 3. Prove it on a model. Couple a single state (energy $0$ ) with amplitude $v$ to a flat band of $N$ levels spanning $[-W,W]$ , density $\rho=N/(2W)$ . Diagonalise the $(N{+}1)\times(N{+}1)$ Hamiltonian numerically and compute the true survival amplitude $\langle i|e^{-i\hat Ht/\hbar}|i\rangle$ . Why: this is the honest test — no perturbation theory, just exact evolution of a discrete-but-dense model that mimics a continuum until the recurrence time $2\pi/\Delta E$ .

Step 4. Read the decay. Over the window $1/W\ll t\ll 1/\Gamma$ (and well short of the recurrence), $\ln P_i(t)$ is a straight line of slope $-\Gamma$ . Fitting the model gives $\Gamma_\text{fit}=0.0040$ against the golden-rule prediction $\Gamma_\text{GR}=2\pi v^2\rho=0.0040$ — agreement to $0.13\%$ — and the instantaneous rate $-\dot P_i/P_i$ is constant across the window to within a few percent. Why: this is direct numerical proof that the golden rule is the exponential decay constant of a genuine continuum, valid precisely in the window the figure marks. Verified: verify_aq-qm-fermi-golden-rule.py — the toy discrete-to-band model decays at the golden-rule rate to 0.13%; the plateau (constant instantaneous rate) holds with coefficient of variation below 15% across the window; $P_i=e^{-\Gamma t}$ solves the rate equation symbolically and its tangent is $\Gamma t$ .

Limiting cases

The golden rule lives in a specific window; its edges are physically instructive, and each is a named regime elsewhere in the chapter.

Case	Result	Meaning
Short-time (Zeno) regime, $t\ll 1/\Delta\omega_\text{band}$	$P_{i\to f}(t)\propto t^2$	The resonance has not yet sharpened; probability is quadratic and no rate exists. Repeated measurement here freezes decay — the quantum Zeno effect.
Golden-rule (rate) window, $1/\Delta\omega_\text{band}\ll t\ll 1/\Gamma$	$P(t)\approx\Gamma t,\;\Gamma=\tfrac{2\pi}{\hbar}\|V_{fi}\|^2\rho(E_f)$	The δ-function has formed but the initial state is barely depleted. A well-defined constant rate holds — the regime the rule describes.
Long-time (depletion) regime, $\Gamma t\gtrsim 1$	$P_i(t)=e^{-\Gamma t}\neq 1-\Gamma t$	Naive first order predicts $P>1$ ; re-summing depletion restores exponential decay (Wigner–Weisskopf). Eventually even that fails — non-exponential tails.
Discrete final state (no continuum)	$\rho\to\sum_n\delta(E-E_n)$ ; Rabi flopping	With a single destination the system oscillates coherently; the exact two-level result is owned by Transition Probability for a Periodic Two-Level System. The rule strictly needs a smooth $\rho$ .

Common trapThe golden rule is not valid at all times, and not for a single final state

Two failure modes are routinely forgotten. First, timing: at very short times the probability is quadratic (no rate yet), and at very long times the initial state depletes so first-order theory would predict probability exceeding one. The rule holds only in the intermediate window $1/\Delta\omega_\text{band}\ll t\ll 1/\Gamma$ . Second, the continuum requirement: applying $\Gamma=(2\pi/\hbar)|V_{fi}|^2\rho$ to a discrete two-level system is nonsense — there you get Rabi oscillations, not a rate, and $\rho$ is not even defined. Finally, $\rho(E_f)$ must be evaluated on the energy shell ( $E_f=E_i$ , or $E_i\pm\hbar\omega$ ), and $V_{fi}$ is the matrix element of the perturbation amplitude, not the full time-dependent $V(t)$ .

Worked examples

Example 1 — a discrete level coupled to a flat band: compute $\Gamma$ , the lifetime, and the window. A discrete state sits at $E=0$ , coupled with matrix element $v=8\times10^{-4}$ (in units where $\hbar=1$ ) to each of $N=2001$ band states evenly spread over $[-1,1]$ in energy.

Step 1. Density of final states: the band has $N=2001$ states over width $2W=2$ , so $\rho=N/(2W)=2001/2=1000.5$ states per energy unit (equivalently $\rho=1/\Delta E$ with level spacing $\Delta E=2/2000=10^{-3}$ ).

Step 2. Apply the golden rule (constant perturbation): $\Gamma=(2\pi/\hbar)|v|^2\rho=2\pi\,(8\times10^{-4})^2\,(1000.5)=\mathbf{4.02\times10^{-3}}$ per unit time.

Step 3. Lifetime: $\tau=1/\Gamma=\mathbf{249}$ time units.

Step 4. The window. The band recurrence time is $2\pi/\Delta E=2\pi/10^{-3}\approx\mathbf{6.3\times10^{3}}$ time units. Since $\tau=249\ll 6.3\times10^3$ , the initial state decays fully before the finite band can echo it back — the golden-rule window $1/W\;(=1)\ll t\ll\tau\;(=249)$ is wide and clean.

Step 5. Verify by exact diagonalisation. Diagonalising the $2002\times2002$ Hamiltonian and fitting $\ln P_i(t)$ over $[0.1\tau,2\tau]$ gives $\Gamma_\text{fit}=4.02\times10^{-3}$ — matching the golden-rule prediction to $0.13\%$ , with a genuinely exponential decay. The rate is real, not an artefact.

Example 2 — the hydrogen 2p radiative lifetime. The golden rule, summed over the photon continuum (the field-quantisation step belongs to Spontaneous Emission and the Einstein A Coefficient), gives the Einstein $A$ coefficient $A=\omega^3|\mathbf d_{fi}|^2/(3\pi\varepsilon_0\hbar c^3)$ . We compute the 2p→1s lifetime as the golden-rule skeleton.

Step 1. Transition energy $E=10.2$ eV, so $\omega=E/\hbar=1.55\times10^{16}$ rad/s.

Step 2. Dipole matrix element: the radial-plus-angular integral gives $|\langle 1s|\hat z|2p\rangle|=\tfrac{2^7\sqrt2}{3^5}a_0=0.745\,a_0=3.94\times10^{-11}\,$ m, so $|\mathbf d_{fi}|=e\cdot0.745\,a_0=6.31\times10^{-30}$ C·m.

Step 3. $A=\dfrac{\omega^3|\mathbf d_{fi}|^2}{3\pi\varepsilon_0\hbar c^3}=\mathbf{6.26\times10^{8}}$ s $^{-1}$ .

Step 4. Lifetime $\tau=1/A=\mathbf{1.60\,ns}$ — the textbook value, and the linewidth $\Gamma_E=\hbar/\tau=4.1\times10^{-7}$ eV. The $\omega^3$ factor (two powers from the photon density of states, one from the coupling) is why atomic transitions decay so fast compared with the 21-cm line.

The historians' trap

Origins · 1 minFermi Golden Rule and how this was discovered

Fermi Golden Rule

1927 (Dirac) / 1950 (name) · Cambridge → Chicago

The result is one of the most-used formulas in physics, and it is named after the wrong person. Paul Dirac wrote it down in 1927, in "The quantum theory of the emission and absorption of radiation" (Proc. R. Soc. Lond. A 114, 243) — the paper that founded time-dependent perturbation theory and first quantised the electromagnetic field. Enrico Fermi never derived it; two decades later, in his 1949–50 University of Chicago lecture notes (published as Nuclear Physics, 1950), he tabulated it as "Golden Rule No. 2" — a memorable label for a standard tool — and the nickname migrated onto his name. Fermi did apply it decisively, most famously to beta decay in 1934, where the final-state phase space is the electron–antineutrino continuum. So the honest statement is: the rule is Dirac's, the name is Fermi's, and the beta-decay application is genuinely Fermi's. It is a clean example of Stigler's law of eponymy — no scientific discovery is named after its original discoverer.

History

1927

— Dirac derives the transition rate in his radiation paper (the true origin)

1934

— Fermi applies it to beta decay, with the electron–antineutrino continuum as final states

1949

— Fermi's Chicago lectures label it "Golden Rule No. 2"

1950

— the name is fixed in print in Fermi's Nuclear Physics

1930

— Weisskopf and Wigner derive the exponential decay and Lorentzian linewidth

Going further

References. Griffiths & Schroeter, Introduction to Quantum Mechanics, 3rd ed., §11.2–11.3 · Sakurai & Napolitano, Modern Quantum Mechanics, 2nd ed., §5.7–5.8 · Shankar, Principles of Quantum Mechanics, 2nd ed., §18.2 · Cohen-Tannoudji, Diu & Laloë, Quantum Mechanics, Vol. II, Ch. XIII · Dirac, Proc. R. Soc. Lond. A 114, 243 (1927) · Fermi, Nuclear Physics (Chicago, 1950).

Provenance. Equations cross-checked against Griffiths & Schroeter (3rd ed.) §11.2–11.3, Sakurai & Napolitano (2nd ed.) §5.7–5.8, Shankar §18.2, and Cohen-Tannoudji Vol. II Ch. XIII. A verifier should recompute: (i) the nascent-delta normalisation $\int \sin^2(\omega t/2)/(\omega/2)^2\,\mathrm{d}\omega=2\pi t$ ; (ii) dimensional consistency $[w]=\text{s}^{-1}$ from $|V_{fi}|^2\rho/\hbar$ ; (iii) the free-particle density of states $\rho=\mathcal V m k/[(2\pi)^3\hbar^2]\,\mathrm{d}\Omega$ ; (iv) the Einstein A coefficient $A=\omega^3|\mathbf d|^2/(3\pi\varepsilon_0\hbar c^3)$ and the hydrogen 2p lifetime ~1.6 ns. Historical attribution to Dirac (1927) with Fermi's 'Golden Rule No. 2' naming is standard.

Builds on: First-Order Transition Amplitudes and Probabilities · Time-Dependent Perturbation Theory: General Formalism · Density of Final States · The Interaction Picture and Time-Dependent Hamiltonians Leads to: Absorption and Stimulated Emission · Spontaneous Emission and the Einstein A Coefficient · Semiclassical Theory of Atom–Radiation Interaction · The Dipole Approximation

If the rule is really Dirac's, why does everyone call it Fermi's?

Because Fermi's 1949–50 Chicago lecture notes labelled it "Golden Rule No. 2," and that memorable name — attached to one of the century's most famous physicists — stuck. Dirac derived it in 1927 without giving it a catchy name. This is a textbook instance of Stigler's law of eponymy: results are rarely named after their original discoverer. The honest citation credits Dirac for the formula and Fermi for the name (and for the beta-decay application, which genuinely is his).

Why does the transition probability grow as t² at first but linearly later?

For a single final state exactly on resonance, $P^{(1)}=(|V_{fi}|^2/\hbar^2)(t^2/4)$ — genuinely quadratic. That quadratic growth is what you see before the resonance function has sharpened enough for a continuum to matter (the Zeno regime). Once $t$ exceeds the inverse bandwidth, the band samples a peak whose area grows only as $t$ , so the summed probability is linear. The crossover between $t^2$ and $t$ is exactly where "a rate begins to exist."

Where does the factor 2π come from?

From the area under the resonance function: $\int_{-\infty}^{\infty}\sin^2(\omega t/2)/(\omega/2)^2\,\mathrm{d}\omega=2\pi t$ . Equivalently, $\sin^2(\omega t/2)/(\omega/2)^2\to 2\pi t\,\delta(\omega)$ . The $2\pi$ is the weight of the nascent delta; it is a pure piece of Fourier analysis, not physics input. The verify script recomputes this integral numerically to better than 0.02%.

Why must ρ be the density of final states, evaluated on the energy shell?

The long-time limit replaces the resonance function by $\delta(E_f-E_i)$ (or $\delta(E_f-E_i\mp\hbar\omega)$ ). When integrated against $\rho(E_f)|V_{fi}|^2$ , the delta picks out the value of the smoothly varying integrand exactly at resonance. So only the density of states where energy is conserved contributes — and it must be the final states, since those are the destinations being summed over.

What happens if there is no continuum — say a single final state?

Then $\rho$ is not defined and the rule does not apply. A single final state gives Rabi oscillations: the probability sloshes coherently back and forth between the two levels, never settling into a decay. This is the exact two-level problem owned by Transition Probability for a Periodic Two-Level System. The golden rule strictly requires a smooth density of states — a genuine continuum or a dense enough band that recurrences are pushed beyond the timescale of interest.

Why doesn't the artificial box volume V appear in the physical rate?

It cancels. The density of final states scales as $\rho\propto V$ (more volume, more modes), but plane-wave final states are normalised as $\psi=V^{-1/2}e^{i\mathbf k\cdot\mathbf r}$ , so $|V_{fi}|^2\propto 1/V$ . The product $|V_{fi}|^2\rho$ is $V$ -independent, as it must be — the box is a fictitious scaffold for counting. Density of Final States proves this cancellation explicitly.

Does the golden rule violate energy conservation at short times?

Not in the end, but it looks fuzzy at finite time. The resonance function has width $\Delta\omega\sim 2\pi/t$ , so transitions to states within $\hbar/t$ of the energy shell are allowed — the energy–time uncertainty relation $\Delta E\,\Delta t\sim\hbar$ (owned by The Energy-Time Uncertainty Relation). As $t\to\infty$ the width vanishes and energy conservation becomes strict. Real, permanent transitions only occur to energy-conserving states; the near-shell "virtual" reach is transient.