Time-Dependent Perturbation Theory & Fermi's Golden Rule
Strike a match, switch on a sodium lamp, watch a laser pointer glow: in every case atoms are
trading energy with light — soaking up a photon to jump upstairs, or dropping down and spitting
one back out. And they do it at astonishingly sharp colours. Sodium's flame is a
particular, unmistakable yellow; hydrogen glows a specific red; a helium-neon laser is locked to one
precise wavelength. Behind every one of these facts sits a single number you can actually calculate: a
transition rate — the probability per unit time that a quantum system, nudged by a
time-varying influence, hops from one state to another.
The time-independent Schrödinger
equation gave us the ladder of stationary energy levels, and
time-independent
perturbation theory told us how a static nudge shifts those levels. But a
stationary state, left alone, sits on its rung forever — it never jumps. To make an atom
change state, something has to shake it in time: an oscillating electromagnetic field, a
passing collision, a switched-on interaction. This page is about that shaking, and about the one
formula it culminates in — Fermi's Golden Rule — which turns "how strongly does the
perturbation connect these two states?" into "how many times per second does the jump happen?"
That rate is the quiet engine under an enormous amount of physics: the intensity of every spectral
line, the lifetime of every excited atom and unstable nucleus, the gain of every laser, and the cross
section of nearly every scattering and decay process. Learn to compute it once, here, and you have the
master tool for how fast quantum things happen.
The setup: a perturbation switched on in time
Start with a system whose stationary states we already know — the eigenstates
|n\rangle of an unperturbed Hamiltonian
\hat{H}_0, with \hat{H}_0|n\rangle = E_n|n\rangle.
At t=0 we turn on a small, time-dependent extra piece:
\hat{H}(t) = \hat{H}_0 + \hat{H}'(t),
with \hat{H}' weak enough that the old eigenstates are still a good language
to describe what happens. Because the |n\rangle form a complete basis, the
true evolving state can always be written as a superposition of them, but now with
time-dependent coefficients that carry the free phase explicitly:
|\psi(t)\rangle = \sum_n c_n(t)\,e^{-iE_n t/\hbar}\,|n\rangle.
All the interesting physics lives in how the c_n(t) drift away from their
starting values once \hat{H}' is on. Suppose the system starts cleanly in a
single initial state |i\rangle, so
c_i(0) = 1 and every other c_n(0) = 0. The
question of the whole page is: what is the probability of later finding it in some
different final state |f\rangle?
P_{i\to f}(t) = |c_f(t)|^2.
Feeding the ansatz into the time-dependent Schrödinger equation and projecting onto
\langle f| gives an exact but tangled set of coupled equations for the
c_n(t). Perturbation theory cuts the knot: since
\hat{H}' is small, we solve it in successive orders, keeping the leading
one.
First-order transition amplitude
To lowest order we set the coefficient on the right-hand side to its initial value
(c_i \approx 1, all others small) and simply integrate. The result is the
workhorse formula of the whole subject — the first-order transition amplitude:
c_f(t) = -\frac{i}{\hbar}\int_0^t \langle f|\hat{H}'(t')|i\rangle\;
e^{\,i\omega_{fi}t'}\,dt', \qquad
\omega_{fi} \equiv \frac{E_f - E_i}{\hbar}.
Read it slowly, because everything else on the page is a special case of it. The matrix element
\langle f|\hat{H}'|i\rangle asks how strongly the perturbation
connects the two states — if the perturbation cannot link them, this is zero and nothing happens.
The exponential e^{i\omega_{fi}t'} is a stopwatch ticking at the
Bohr frequency \omega_{fi}, the natural frequency of the
i\to f gap. The integral adds up the little kicks the perturbation
delivers over time, each weighted by that spinning phase.
The phase factor is the crux. If \hat{H}' pushes in step with
\omega_{fi}, the contributions pile up in phase and the amplitude
grows — a resonance. If it pushes at the wrong tempo, the spinning phase makes
successive kicks cancel, and almost nothing accumulates. That single tension — constructive build-up
at resonance, destructive cancellation off it — is the origin of sharp spectral lines.
P_{i\to f}(t) = \left|c_f(t)\right|^2
= \frac{1}{\hbar^2}\left|\int_0^t \langle f|\hat{H}'(t')|i\rangle\,e^{i\omega_{fi}t'}\,dt'\right|^2.
Everything now depends on the time profile of the perturbation. The single most important
profile — the one that describes light hitting an atom — is a sinusoid, and that is where we go next.
A sinusoidal perturbation: resonance and the birth of spectral lines
Light is an oscillating electromagnetic field, so shine it on an atom and the perturbation oscillates:
\hat{H}'(t) = 2\hat{V}\cos(\omega t)
= \hat{V}\big(e^{i\omega t} + e^{-i\omega t}\big).
Drop this into the first-order integral. The cosine splits into two counter-rotating exponentials, and
combined with the stopwatch e^{i\omega_{fi}t'} they produce two terms
oscillating at (\omega_{fi}+\omega) and
(\omega_{fi}-\omega). One of them can go resonant:
-
Absorption (E_f > E_i): the term resonates when
\omega \approx +\omega_{fi}, i.e.
\hbar\omega \approx E_f - E_i. The atom climbs a rung by
swallowing a photon of just the right energy.
-
Stimulated emission (E_f < E_i): the other term
resonates when \omega \approx -\omega_{fi}, i.e.
\hbar\omega \approx E_i - E_f. The passing field coaxes the atom
down a rung, and it emits a photon in lockstep with the driving wave — the principle behind the
laser (Light Amplification by Stimulated Emission of Radiation).
Keeping the resonant term, the transition probability works out to the celebrated lineshape:
P_{i\to f}(t) = \frac{|V_{fi}|^2}{\hbar^2}\;
\frac{\sin^2\!\big[(\omega - \omega_{fi})\,t/2\big]}{\big[(\omega - \omega_{fi})/2\big]^2},
\qquad V_{fi} = \langle f|\hat{V}|i\rangle.
This is a sinc² curve in the detuning \omega - \omega_{fi}:
a tall central peak sitting exactly at resonance, flanked by small ripples. Two features matter enormously.
As the interaction time t grows, the peak height at exact
resonance climbs as t^2 (take the limit
\omega \to \omega_{fi} and the fraction becomes
t^2), while the peak width shrinks as
1/t. The peak becomes taller and narrower — the system grows ever pickier
about matching the exact resonant frequency the longer you watch. Drag the time slider below and see it
happen.
That sharpening is the whole reason spectral lines are sharp. A long-lived interaction admits only
frequencies extremely close to \omega_{fi}; everything else cancels away.
Push t to infinity and the peak becomes infinitely tall and infinitely
thin — a Dirac delta function enforcing
\hbar\omega = |E_f - E_i| exactly. Energy conservation for the
atom-plus-photon is not put in by hand; it emerges from letting the perturbation act for a
long time.
Into a continuum: Fermi's Golden Rule
A jump between two isolated levels at exact resonance actually oscillates — probability sloshes
up to |f\rangle and back down again (we meet that as
Rabi flopping in the Watch-out below). To get a steady rate, the
final state must not be a lone level but one of a continuum — a dense band of nearly
degenerate states, described by a density of states
\rho(E_f) (the number of available final states per unit energy). A decaying
atom emits into the continuum of photon modes; a scattered particle flies off into a continuum of
momenta. Sum the sinc² probability over that band, take the long-time limit so the delta function bites,
and the sharp lineshape integrates to a probability that grows linearly in time. A
probability linear in time is a constant rate — and that rate is Fermi's Golden Rule.
-
For transitions from an initial state |i\rangle into a continuum of final
states around |f\rangle, the transition rate
(probability per unit time) is
\Gamma_{i\to f} = \frac{2\pi}{\hbar}\,
\big|\langle f|\hat{H}'|i\rangle\big|^2\,\rho(E_f).
-
The rate is constant in time: the transition probability grows
linearly, P_{i\to f}(t) = \Gamma\, t, so a fixed number of jumps
happen per second.
-
It is proportional to the matrix element squared,
|\langle f|\hat{H}'|i\rangle|^2 — the strength of the coupling — and to
the density of final states \rho(E_f) — how many places
there are to land.
-
The long-time limit enforces energy conservation through a delta function,
E_f = E_i \pm \hbar\omega (the - for
absorption of energy \hbar\omega, the + for
emission); it is this delta that \rho(E_f) is evaluated against.
Dirac derived it in 1927; Fermi called it "golden" because he used it so often it deserved a nickname —
and the name stuck to the wrong person, in the grand tradition of physics. Its power is that it reduces
an entire dynamical calculation to two ingredients you can often compute by hand: a matrix element and a
density of states.
What the rule buys you: selection rules, lifetimes, lasers
Selection rules. If the matrix element
\langle f|\hat{H}'|i\rangle = 0, the rate is exactly zero: the transition is
forbidden, no matter how much light you shine. Very often the matrix element vanishes
by symmetry — a parity or angular-momentum mismatch between
|i\rangle, \hat{H}' and
|f\rangle makes the integral cancel. This is why only certain atomic
transitions light up: the periodic table's spectra are a giant catalogue of which matrix elements
happen to be nonzero.
Lifetimes. An excited atom left alone still decays — spontaneous emission — at a rate
\Gamma given by (a fully quantised-field version of) the golden rule. If
each atom jumps with constant probability per unit time \Gamma, a population
decays exponentially, N(t) = N_0 e^{-\Gamma t}, and the characteristic
lifetime is simply
\tau = \frac{1}{\Gamma}.
A strongly-coupled transition (big matrix element) is fast and short-lived; a nearly-forbidden one can
linger for seconds — the "metastable" states that make phosphorescence glow in the dark.
Spontaneous vs stimulated. Our sinusoidal treatment gave stimulated emission,
driven by an applied field; a full theory also has spontaneous emission into empty modes.
Einstein tied the two together before quantum electrodynamics existed with his
A and B coefficients — the fixed ratio between
spontaneous decay and stimulated up/down rates that any working laser exploits by engineering a
population inversion so stimulated emission wins.
And it does not stop at atoms. Replace \hat{H}' by a scattering potential and
the same rule gives collision cross sections; replace it by the weak interaction and it gives nuclear
beta-decay rates. Fermi's Golden Rule is the universal answer to "how fast?"
Worked examples
Example 1 — a rate straight from the rule. A perturbation connects an initial state
to a continuum with matrix element |M| = 2\times10^{-3}\,\text{eV} and the
final-state density is \rho = 5\times10^{18}\,\text{J}^{-1}. Fermi's rule gives
\Gamma = (2\pi/\hbar)\,|M|^2\,\rho. The lesson is structural: the rate is
linear in \rho and quadratic in the coupling. Double the
density of final states and the rate doubles; double the matrix element and the rate goes up
four-fold. Everything else is just plugging in
\hbar = 1.055\times10^{-34}\,\text{J·s} with the units made consistent.
Example 2 — a forbidden transition. Suppose |i\rangle and
|f\rangle are both even functions of x and the
perturbation is \hat{H}' \propto x (an odd function — the usual dipole coupling
to a field). Then the integrand
\langle f|x|i\rangle = \int \psi_f^*(x)\,x\,\psi_i(x)\,dx is
even × odd × even = odd, so it integrates to zero over a symmetric range. The
matrix element vanishes, the rate is zero, and the transition is forbidden by parity —
exactly why, for example, the 2s \to 1s transition in hydrogen is so
spectacularly slow. You did not need to compute a single number; symmetry killed it.
Example 3 — lifetime from a rate. An excited state decays with rate
\Gamma = 2.5\times10^{8}\,\text{s}^{-1} (a typical allowed optical transition).
Its lifetime is
\tau = \frac{1}{\Gamma} = \frac{1}{2.5\times10^{8}\,\text{s}^{-1}}
= 4.0\times10^{-9}\,\text{s} = 4\,\text{ns}.
A few nanoseconds — the ballpark for ordinary allowed atomic transitions, and why fluorescent things
flash off almost the instant you kill the excitation. A metastable state with
\Gamma a billion times smaller would glow for seconds instead.
A fair worry: the first-order formula gave a probability
P_{i\to f}(t), yet the golden rule is a rate. The bridge is that,
once you sum the sinc² lineshape over a continuum of final states, all the wiggly time-dependence
collapses and the surviving probability is simply proportional to t:
P(t) = \Gamma t. Differentiate and the rate
\Gamma = dP/dt is a constant — no time dependence left. That linear
growth is what makes a constant rate meaningful at all: a fixed slice of probability accrues each
second, so a fixed fraction of a population jumps per second, which is exactly what "rate" means.
There is a catch worth naming: because P = \Gamma t grows without bound, it
would eventually exceed 1, which is nonsense. First-order theory is only
valid in the window \Gamma t \ll 1 — long enough for the delta function to
sharpen, short enough that the initial state is not appreciably depleted. Inside that window, the
golden rule reigns.
The classic trap. Fermi's Golden Rule gives a steady, constant rate — but that only
happens when the system decays into a dense set of final states (that is what the density of
states \rho(E_f) is doing in the formula). If instead you drive a transition
between two isolated discrete levels on exact resonance, there is no rate at all in the
golden-rule sense. The probability does not grow linearly forever; it oscillates
sinusoidally — sloshing fully up to the excited state and all the way back down again. These are
Rabi oscillations, and the population cycles at the Rabi frequency
\Omega = |V_{fi}|/\hbar:
P_{i\to f}(t) = \sin^2\!\big(\Omega t\big)
\quad\text{(two levels, on resonance)} \qquad\text{vs.}\qquad
P_{i\to f}(t) = \Gamma\,t \quad\text{(into a continuum).}
So before you reach for the golden rule, ask: is there somewhere for the system to decay into a
continuum? If yes — a photon field, a band of momenta, a dense spectrum — the rule applies and you
get a lifetime. If the final states are just a couple of sharp levels, you get coherent flopping
instead, and you must solve the two-state problem directly. Same physics, utterly different behaviour.