Time-Ordered Exponential and the Dyson Series

Every tool of the last four articles leaned on one luxury: a Hamiltonian that commutes with itself at all times, so the propagator is a single exponential — diagonalized once, a stack of spinning phases. Real driving destroys the luxury. Rotate a field over a spin, chirp a laser through resonance — now $\hat H(t_1)$ and $\hat H(t_2)$ are different operators that do not commute, and the innocent formula $e^{-\frac{i}{\hbar}\int \hat H\,\mathrm{d}t'}$ is simply wrong. Not approximately: for a spin kicked first about $z$ then about $x$ , it mispredicts a flip probability by twenty percentage points, and you will compute that failure exactly. The repair is one idea with heavy consequences: order matters, so keep the order. Chop the interval into frames, evolve each by its instantaneous exponential, multiply with later frames on the left — a film played strictly in sequence. The limit of that product is the time-ordered exponential, $\mathcal{T}\exp$ , the exact solution of the evolution operator's equation of motion for any Hamiltonian whatsoever. Expanded in powers of $\hat H$ it is the Dyson series: propagate, get kicked, propagate, get kicked — every history of kicks, summed. This is the chapter's mathematical flagship, and the payoff is proportional. The series' first term is time-dependent perturbation theory's entire engine; in the interaction picture it becomes the S-matrix of scattering theory and QFT; its combinatorial heart — a factorial cancelling a hypercube of orderings — is three lines of bookkeeping. Master this and the rest of driven quantum mechanics is applications.

Key takeaways

If $[\hat H(t_1), \hat H(t_2)] = 0$ for all times, $\hat U = \exp(-\frac{i}{\hbar}\int\hat H\,\mathrm{d}t')$ — the naive exponential survives.
If not, it fails: for $\sigma_z$ then $\sigma_x$ kicks it misses the exact propagator by Frobenius distance 0.76 and a flip probability by 0.50 vs 0.40.
The exact propagator is $\hat U(t,t_0) = \mathcal{T}\exp\!\big(-\frac{i}{\hbar}\int_{t_0}^t \hat H(t')\,\mathrm{d}t'\big)$ — frame-by-frame evolution, later frames leftmost.
Expanding gives the Dyson series: nested integrals of ordered products $\hat H(t_1)\cdots\hat H(t_n)$ , $t_1 \ge \cdots \ge t_n$ — a sum over kick histories.
The $1/n!$ in the symmetrized form is pure combinatorics: a hypercube holds $n!$ copies of the ordered simplex, and $\mathcal{T}$ makes them equal.
Truncated Dyson sums are not unitary; only the full series is. First order is the seed of time-dependent perturbation theory.
Key idea: when the Hamiltonian doesn't commute with itself at different times, evolution is a strictly ordered product — the time-ordered exponential, whose expansion is the Dyson series.

The core insight

Time evolution under a non-commuting time-dependent Hamiltonian is not $\exp(-\tfrac{i}{\hbar}\int H\,\mathrm{d}t)$ but its time-ordered version: chop $[t_0,t]$ into slices, evolve each by $e^{-iH(t_k)\Delta t/\hbar}$ , and multiply with later slices on the left. The infinite-product limit is $\mathcal{T}\exp$ , whose power-series expansion is the Dyson series.

Where this lives today

Every quantum computer runs this article: simulating evolution on gate hardware approximates an ordered exponential by a product of short frames — Trotterization is the film-strip definition of $\mathcal{T}\exp$ used as an algorithm, its error budget pure Dyson/Magnus bookkeeping. NMR and quantum-control labs design pulse sequences by engineering which terms of the ordered expansion survive (average-Hamiltonian theory). Floquet engineering — periodic driving that gives a material properties its static Hamiltonian lacks — is applied time-ordering, with Magnus supplying the effective Hamiltonian. And at the field-theory frontier the series remains the native language: every collider cross-section starts life as matrix elements of $\mathcal{T}\exp(-\frac{i}{\hbar}\int V_I)$ , diagram by diagram, exactly as the 1949 paper laid out.

The physical picture

Think of evolution as a movie shown frame by frame. Each frame advances the state by $e^{-i\hat H(t_k)\Delta t/\hbar}$ — a small rotation of Hilbert space. If every frame rotated about the same axis, frame order would be irrelevant and the tiny exponents would pool into one integral. But a driven system rotates about an axis that tilts as time passes, and rotations about different axes do not commute: $x$ -then- $z$ is not $z$ -then- $x$ . The film must play in order.

Figure

Frames of a film — rotations about tilting axes compose in strict order, later frames applied on the left

A spin in a sweeping magnetic field makes this concrete: the sweep composes many small rotations about successively tilted axes, and the composite is not a rotation about the time-averaged axis — precisely why exponentiating the integrated Hamiltonian fails. And the failure is the working condition of every NMR pulse sequence, driven qubit gate, and chirped laser experiment. The second picture is the series expansion's. Each Dyson term is a history: coast, get kicked by $\hat H$ at $t_n$ , coast, get kicked at $t_{n-1}$ … with kick times integrated over all allowed orderings. Zero kicks, one kick, two — summing every history of every length reconstructs the exact evolution, the seed of the sum-over-histories view of quantum mechanics in its original operator clothing.

The equation to solve, and how the naive answer dies

The object and its law are the previous articles': the evolution operator obeys the operator Schrödinger equation, composition and unitarity built in (The Time-Evolution Operator owns the construction and the three-case taxonomy — this article is its hardest case). For general $\hat H(t)$ :

i\hbar\,\frac{\partial}{\partial t}\hat U(t,t_0) = \hat H(t)\,\hat U(t,t_0), \qquad \hat U(t_0,t_0) = \hat 1 \qquad (1)

where: $\hat U$ exists and is unitary for any driving — the question is only its form. Constant $\hat H$ : (1) integrates to $e^{-i\hat H(t-t_0)/\hbar}$ ; self-commuting $\hat H(t)$ : the same manipulation goes through with $\hat H(t-t_0) \to \int\hat H\,\mathrm{d}t'$ . The trap is assuming it always does.

Derivation AThe counterexample — twenty percentage points of wrong from one swapped order

We demonstrate, with the simplest possible driving, that $\exp(-\frac{i}{\hbar}\int\hat H\,\mathrm{d}t')$ is not the propagator. The strategy: a Hamiltonian so simple its exact evolution needs no series — constant $\hbar\omega\,\sigma_z$ for the first half, constant $\hbar\omega\,\sigma_x$ for the second — compared against the naive exponential of the integral. Everything is $2\times2$ ; the entire failure is the order of two rotations. Set $\hbar = 1$ , $\omega\tau = \pi/4$ per half.

Step 1. Exact propagator by composition. Each half has a constant Hamiltonian, so each half is an honest exponential; composition (later factor on the left) gives

\hat U_{\text{exact}} = e^{-i\sigma_x \tau}\,e^{-i\sigma_z \tau}

Why the order: the $\sigma_z$ half happens first, so it acts first — rightmost. Writing the factors the other way is already the mistake this article exists to prevent.

Step 2. The naive candidate. The integrated Hamiltonian is $\int_0^{2\tau}\hat H\,\mathrm{d}t' = (\sigma_z + \sigma_x)\tau$ , so the candidate is $\hat U_{\text{naive}} = e^{-i(\sigma_z + \sigma_x)\tau}$ — a single rotation about the averaged axis, the diagonal between $z$ and $x$ .

Step 3. Compare. Both matrices are perfectly unitary — the failure is not normalization but geometry:

\big\lVert \hat U_{\text{exact}} - \hat U_{\text{naive}} \big\rVert_F = 0.7600

where: the Frobenius norm of the difference is order one — the operators are not close in any sense. The physics: from spin-up, the exact sequence flips the spin with probability $0.5000$ ; the naive rotation predicts $0.4014$ . Twenty points of error, from pretending two rotations about different axes could be pooled.

Step 4. Name the obstruction. Differentiating $e^{A(t)}$ returns $\dot A e^{A}$ only when $[A(t), \dot A(t)] = 0$ ; in general the derivative smears $\dot A$ through the exponential (the exact formula integrates $e^{sA}\dot A e^{(1-s)A}$ over $s$ ). With $A = -\frac{i}{\hbar}\int\hat H$ , the obstruction is $[\int\hat H\,\mathrm{d}t', \hat H(t)] \neq 0$ — the self-commutator that vanished in every earlier article and refuses to vanish here. $\hat U_{\text{naive}}$ does not even satisfy equation (1): the failure is diagnosed before comparing answers. Verified: numerically — gap 0.7600, both matrices unitary to $10^{-12}$ , flip probabilities 0.5000 vs 0.4014; the frame-by-frame slice product converges to the exact propagator at machine precision; for the commuting family $H(t) = f(t)\sigma_z$ the naive exponential matches the slice product to $10^{-8}$ — the failure is ordering, nothing else (verify script, 16 checks green).

Iterating toward the answer — the Dyson series

No closed exponential solves (1) in general, so integrate it: both sides from $t_0$ to $t$ , initial condition absorbed, and the differential equation becomes a self-referential integral equation (Volterra form):

\hat U(t,t_0) = \hat 1 - \frac{i}{\hbar}\int_{t_0}^{t} \hat H(t')\,\hat U(t',t_0)\,\mathrm{d}t' \qquad (2)

where: the unknown $\hat U$ appears on both sides — circular-looking, actually a machine: feed any approximation into the right side and a better one comes out of the left. Iterating from $\hat U \approx \hat 1$ generates the expansion in powers of the Hamiltonian.

Go deeper · Derivation B · 1 minFeeding the equation to itself — the nested simplex integrals

Derivation BFeeding the equation to itself — the nested simplex integrals

We derive the Dyson series by iterating equation (2), watching two structural facts emerge on their own: the integration domains are ordered simplices ( $t_1 \ge t_2 \ge \cdots$ ), and the operator factors stand in that same order, later times leftmost. Nobody imposes the ordering — the iteration manufactures it, because each substitution reaches only into the past of the previous variable. That inevitability is the point.

Step 1. Zeroth and first order. Insert $\hat U \approx \hat 1$ into (2)'s right side:

\hat U^{(0)} = \hat 1, \qquad \hat U^{(1)} = -\frac{i}{\hbar}\int_{t_0}^{t}\hat H(t_1)\,\mathrm{d}t_1

Step 2. Second order. Substitute $\hat 1 + \hat U^{(1)}$ into (2) and keep the new term:

\hat U^{(2)} = \Big(\!-\frac{i}{\hbar}\Big)^{\!2}\int_{t_0}^{t}\!\mathrm{d}t_1\int_{t_0}^{t_1}\!\mathrm{d}t_2\;\hat H(t_1)\hat H(t_2)

Why the inner limit is $t_1$ , not $t$ : the substituted $\hat U$ was evaluated at $t_1$ , so its own integral reaches only $t_2 < t_1$ . The domain is the triangle $t \ge t_1 \ge t_2 \ge t_0$ — and the operator order matches it, later standing left of earlier. Order of factors = order of times, automatically.

Step 3. All orders. Each substitution appends one more past-reaching integral: term $n$ is $(-i/\hbar)^n$ times the $n$ -fold nested integral of $\hat H(t_1)\cdots\hat H(t_n)$ over the ordered simplex $t \ge t_1 \ge \cdots \ge t_n \ge t_0$ . The sum of all terms is the Dyson series — the exact solution of (1), order by order in the driving.

Step 4. Audit the truncations. Each partial sum solves (1) up to its order but is not unitary — unitarity is a conspiracy among all orders. For the counterexample above, the partial sums approach the exact propagator as $1.41 \to 1.15 \to 0.62 \to 0.25 \to 0.079$ through fourth order, while the first-order unitarity deficit is a whopping $1.75$ . Truncations are for small couplings or short times; never mistake them for evolution operators. Verified: numerically — the Volterra iteration on a 1600-slice grid reproduces the falling error ladder above and converges onto the exact propagator down to the quadrature floor ( $\sim10^{-3}$ , scaling as $1/N$ ); the first-order unitarity deficit is 1.75, decaying with order (verify script).

The time-ordering symbol — one factorial, all the bookkeeping

The nested simplex integrals are exact but clumsy — every integral has a different upper limit. The repair is a notational machine with real combinatorial content: the time-ordering symbol $\mathcal{T}$ , which rearranges a product of time-labeled operators so time decreases left to right,

\mathcal{T}\big[\hat H(t_1)\hat H(t_2)\big] = \theta(t_1 - t_2)\,\hat H(t_1)\hat H(t_2) + \theta(t_2 - t_1)\,\hat H(t_2)\hat H(t_1) \qquad (3)

where: $\theta$ is the unit step. Under $\mathcal{T}$ , operators at different times behave as if they commuted — integrate over any domain in any order; the symbol restores chronology afterward. That freedom converts simplices into hypercubes:

\hat U(t,t_0) = \sum_{n=0}^{\infty}\frac{1}{n!}\Big(\!-\frac{i}{\hbar}\Big)^{\!n}\int_{t_0}^{t}\!\mathrm{d}t_1\cdots\int_{t_0}^{t}\!\mathrm{d}t_n\;\mathcal{T}\big[\hat H(t_1)\cdots\hat H(t_n)\big] \equiv \mathcal{T}\exp\!\left(-\frac{i}{\hbar}\int_{t_0}^{t}\hat H(t')\,\mathrm{d}t'\right) \qquad (4)

where: all $n$ integrals run over the full interval, $1/n!$ compensates the overcounting, and the result earns the name "exponential": same coefficients, same shape, ordering included. Equation (4) is the standard form.

Figure

The square cut by its diagonal — two triangles, one ordered integrand, and the 1/2! that generalizes

Go deeper · Derivation C · 1 minWhere the factorial comes from — a square cut along its diagonal

Derivation CWhere the factorial comes from — a square cut along its diagonal

We prove the simplex and hypercube forms equivalent at second order, where every feature of the general argument already appears. The claim: the ordered simplex integral equals $\frac{1}{2!}$ times the full-square integral of the time-ordered product. The strategy is pure geometry — count the copies of the ordered region tiling the cube, and check that $\mathcal{T}$ makes every copy contribute equally.

Step 1. Cut the square. The domain $[t_0,t]^2$ of the pair $(t_1,t_2)$ splits along the diagonal $t_1 = t_2$ into two triangles: $t_1 > t_2$ (the simplex of Derivation B) and $t_2 > t_1$ (its mirror). The diagonal itself has zero area.

Step 2. Apply the definition (3) on each triangle. On the first, $\mathcal{T}$ leaves $\hat H(t_1)\hat H(t_2)$ alone; on the mirror it swaps to $\hat H(t_2)\hat H(t_1)$ — which is the same integrand as the first triangle after renaming the dummy variables $t_1 \leftrightarrow t_2$ . The two triangles contribute identically:

\int_{t_0}^{t}\!\mathrm{d}t_1\int_{t_0}^{t}\!\mathrm{d}t_2\,\mathcal{T}\big[\hat H(t_1)\hat H(t_2)\big] = 2\int_{t_0}^{t}\!\mathrm{d}t_1\int_{t_0}^{t_1}\!\mathrm{d}t_2\,\hat H(t_1)\hat H(t_2)

Why $\mathcal{T}$ is essential: without it the mirror triangle holds the differently-ordered product $\hat H(t_2)\hat H(t_1)$ with $t_2$ later — a genuinely different operator when the self-commutator is nonzero — and the two triangles would NOT be equal. Time-ordering is what makes the dummy-relabeling argument legal.

Step 3. Generalize. The cube $[t_0,t]^n$ tiles into $n!$ simplices, one per ordering; $\mathcal{T}$ maps each tile's integrand onto the ordered one, so the cube integral is $n!$ times the simplex integral. Dividing by $n!$ recovers Derivation B's terms — equation (4) is proved, and the exponential's familiar $1/n!$ is revealed as a counting of orderings.

Step 4. Collect the commuting limit. If $[\hat H(t_1), \hat H(t_2)] = 0$ always, $\mathcal{T}$ does nothing, the hypercube integrals factor into powers of $\int\hat H\,\mathrm{d}t'$ , and (4) collapses to the naive exponential — exactly when it was legal. One symbol, both regimes. Verified: symbolically — the $n = 2$ simplex integral equals $\frac{1}{2}(\int f)^2$ for exponential and polynomial $f$ (sympy, exact); numerically — for a genuinely non-commuting matrix family, the square's time-ordered integral equals twice the simplex integral to 0.5% (grid tolerance), and the commuting-family propagator matches the naive exponential to $10^{-8}$ (verify script).

The first-order bridge — where perturbation theory is born

Truncate (4) at one kick and sandwich between eigenstates $|i\rangle \to |f\rangle$ of a split Hamiltonian — the workhorse of driven quantum mechanics appears:

c_f^{(1)}(t) = -\frac{i}{\hbar}\int_{t_0}^{t}\langle f|\hat V(t')|i\rangle\,e^{i\omega_{fi}t'}\,\mathrm{d}t' \qquad (5)

where: $\hat V$ is the driving part of $\hat H = \hat H_0 + \hat V(t)$ and $\omega_{fi} = (E_f - E_i)/\hbar$ is the transition's Bohr frequency — the phase factor is the fingerprint of unperturbed evolution between kicks. Equation (5) is stated as the bridge: its systematic development — rates, resonance, Fermi's golden rule — belongs to Time-Dependent Perturbation Theory and its chapter, which runs on it. One structural remark completes the picture. The Dyson series' natural habitat is the interaction picture, where free evolution is absorbed into the operators and the state moves only under $\hat V_I(t) = e^{i\hat H_0t/\hbar}\hat V(t)e^{-i\hat H_0t/\hbar}$ — a Hamiltonian that is always effectively time-dependent and non-self-commuting, so the $\mathcal{T}$ -exponential is constitutive there, not optional. Taking $t_0 \to -\infty$ , $t \to +\infty$ of that object defines the S-matrix. The picture itself — construction, equation of motion, why it splits the work — is owned by The Interaction (Dirac) Picture, two articles ahead; here you need only know where the series will live.

Limiting cases

case	result	meaning
$[\hat H(t_1), \hat H(t_2)] = 0$ for all pairs	$\hat U = \exp\big(-\frac{i}{\hbar}\int\hat H\,\mathrm{d}t'\big)$	$\mathcal{T}$ is inert; includes $\hat H(t) = f(t)\hat A$ with one fixed operator
time-independent $\hat H$	$\hat U = e^{-i\hat H(t-t_0)/\hbar}$	the whole preceding chapter; spectral decomposition applies
weak/short driving	$\hat U \approx \hat 1 - \frac{i}{\hbar}\int\hat H\,\mathrm{d}t'$	first-order truncation → equation (5); not unitary
interaction picture, $t_0 \to -\infty$ , $t \to +\infty$	$\hat S = \mathcal{T}\exp\big(-\frac{i}{\hbar}\int V_I\,\mathrm{d}t\big)$	the S-matrix (statement; owned by the scattering chapter)
slow driving	dynamical phase + ordered corrections	adiabatic/geometric-phase territory (approximation-methods chapter)

Common trapFor a time-dependent H, just exponentiate the integral of H

The single most common error in driven quantum mechanics, wrong whenever $[\hat H(t_1), \hat H(t_2)] \neq 0$ — Derivation A prices it at twenty percentage points of flip probability for two orthogonal kicks. The naive exponential does not even satisfy the equation of motion: differentiating $e^{A(t)}$ gives $\dot Ae^A$ only when $A$ and $\dot A$ commute; the first correction to the exponent is the commutator term $\frac{1}{2}\int\int[\hat H(t_1), \hat H(t_2)]$ — the start of the Magnus expansion. Two companion traps: a truncated Dyson sum is not unitary — probabilities from it can exceed 1 (first-order deficit 1.75 in the worked counterexample) — and the ordering is later-times-left, which the iteration produces automatically but hand-written expressions scramble with depressing regularity.

Worked examples

Example 1 — two kicks, done exactly and naively (confidence builder). A spin- $\tfrac12$ evolves under $\hat H = \hbar\omega\sigma_z$ for $0 \le t < \tau$ , then $\hat H = \hbar\omega\sigma_x$ for $\tau \le t < 2\tau$ , with $\omega\tau = \pi/4$ and $\hbar = 1$ . Compare the exact propagator with the naive exponential, starting from spin-up.

Step 1. Exact: each segment is constant — compose two honest exponentials, later on the left: $\hat U_{\text{exact}} = e^{-i\sigma_x\pi/4}\,e^{-i\sigma_z\pi/4}$ .

Step 2. Naive: $\int\hat H\,\mathrm{d}t = (\sigma_z + \sigma_x)\pi/4$ , giving one rotation about the diagonal axis.

Step 3. Numbers. Frobenius gap $= \mathbf{0.7600}$ . Spin-flip probability from $|{\uparrow}\rangle$ : exact $\mathbf{0.5000}$ , naive $\mathbf{0.4014}$ — a 20-point miss on a two-frame film.

Step 4. Dyson repair ladder: partial sums (Volterra iteration, numeric) close the gap as $1.41 \to 1.15 \to 0.62 \to 0.25 \to 0.079$ through fourth order — converging, but slowly at this coupling, and non-unitary at every truncation. The moral is double: the series is exact summed, and strong driving wants many orders or the exact frame product.

Example 2 — a Gaussian pulse on a two-level system (exam level). A two-level system with splitting $\hbar\omega_{fi} = \Delta$ ( $\hbar = 1$ , $\Delta = 1$ ) is driven by $\hat V(t) = V_0\,e^{-t^2/2\tau_p^2}\,\sigma_x$ with $V_0 = 0.02$ , $\tau_p = 1$ . Find the transition probability to first order in the Dyson series and compare with exact integration.

Step 1. Equation (5) with $\langle f|\sigma_x|i\rangle = 1$ : $c_f^{(1)} = -iV_0\int_{-\infty}^{\infty}e^{-t^2/2\tau_p^2}e^{i\omega_{fi}t}\,\mathrm{d}t$ .

Step 2. The integral is a Gaussian's Fourier transform, exact: $\sqrt{2\pi}\,\tau_p\,e^{-\omega^2\tau_p^2/2}$ . So $|c_f^{(1)}| = V_0\sqrt{2\pi}\,\tau_p\,e^{-\omega_{fi}^2\tau_p^2/2} = 0.02 \times 2.5066 \times e^{-0.5} = \mathbf{0.03041}$ .

Step 3. Probability: $P^{(1)} = |c_f^{(1)}|^2 = \mathbf{9.25\times10^{-4}}$ . Exact numerical integration of the Schrödinger equation across the pulse gives $P = 9.24\times10^{-4}$ — agreement to $\mathbf{0.09\%}$ . One kick is the whole story at this drive strength.

Step 4. Read the physics: the transition probability is the pulse's spectral content at the Bohr frequency — a slow pulse ( $\omega_{fi}\tau_p \gg 1$ ) is exponentially unable to flip the system (the adiabatic seed). And respect the regime: rerun at $V_0 = 0.5$ and first order misses by 87% — the higher kicks have arrived.

The man who unified the theories on a bus

Origins · 1 minFreeman Dyson and how this was discovered

Freeman Dyson

1923–2020 · Cornell and the Institute for Advanced Study

In 1948 the new quantum electrodynamics existed as three mutually unintelligible dialects: Schwinger's operator calculus, Tomonaga's parallel formulation, and Feynman's diagrams, which computed right answers by rules nobody could derive. Dyson — a 24-year-old English ex-number-theorist who had spent the year listening to Feynman at Cornell and studying Schwinger at Michigan — saw the bridge during a Greyhound ride across Nebraska in September 1948: expand the evolution operator as a time-ordered series, and Feynman's diagrams are the terms, one picture per history of kicks. The 1949 paper, "The Radiation Theories of Tomonaga, Schwinger, and Feynman," proved the dialects identical and handed physics its working language — the series bearing his name, the $\mathcal{T}$ -ordering bookkeeping, the demonstration that renormalization tames it order by order. Characteristically, he never took a PhD; Oppenheimer hired him to the Institute for Advanced Study anyway. Six further decades of heretic-generalism followed — spheres around stars, tree-planting on comets — but his monument is the equation on this page: the exact evolution of anything driven, written as every possible sequence of kicks, summed.

History

1926

— Dirac's time-dependent perturbation theory computes transition amplitudes one kick at a time

1948

— Schwinger, Tomonaga, and Feynman hold three working but incompatible formulations of QED

1949

— Dyson's paper unifies them: the time-ordered series makes Feynman's diagrams theorems

1950

— Wick's theorem turns time-ordered products into systematic diagram rules

1954

— Magnus recasts the ordered exponential as a single exponent of nested commutators

1968

— Average-Hamiltonian theory brings the Magnus/Dyson machinery into NMR pulse design

Going further

References. Sakurai & Napolitano, Modern Quantum Mechanics, 2nd ed., Ch. 2 & 5 · Shankar, Principles of Quantum Mechanics, 2nd ed., Ch. 18 · Cohen-Tannoudji, Diu & Laloë, Quantum Mechanics, Vol. II, Ch. XIII · Griffiths, Introduction to Quantum Mechanics, 3rd ed., Ch. 11 · Dyson, F.J., The Radiation Theories of Tomonaga, Schwinger, and Feynman, Phys. Rev. 75, 486 (1949).

Provenance. Equations are standard and can be verified against Sakurai & Napolitano Modern QM 2nd ed. Sec. 2.2 and Ch. 5 (Dyson series, interaction picture), Shankar 2nd ed. Ch. 18, and Cohen-Tannoudji Vol. II Ch. XIII. The Dyson resummation and S-matrix follow Dyson, Phys. Rev. 75, 486 (1949). A verifier should recompute: (i) the derivative-of-exponential obstruction $[A, \dot A]$ ; (ii) the iteration producing nested simplex integrals; (iii) the $n!$ counting of orderings in the cube; (iv) the first-order amplitude reproducing Fermi's golden rule. For this edition, verification was machine-executed (sympy/numpy/scipy, 16 checks green): Frobenius gap 0.7600 with flip probabilities 0.5000 vs 0.4014, both operators unitary to $10^{-12}$ ; the Volterra error ladder $1.41/1.15/0.62/0.25/0.079$ with first-order unitarity deficit 1.75; the symbolic $n = 2$ simplex identity (two function families) and the numeric non-commuting square-equals-twice-simplex check (0.5%); the commuting-family reduction ( $10^{-8}$ ); the Gaussian-pulse integral with $|c^{(1)}| = 0.03041$ , $P^{(1)} = 9.25\times10^{-4}$ vs exact ODE $9.24\times10^{-4}$ (0.09%); the strong-drive breakdown (87% at $V_0 = 0.5$ ). Ownership: the operator equation of motion, composition, and three-case taxonomy belong to The Time-Evolution Operator; commutator algebra to The Commutator and Compatible Observables; the interaction picture's construction to The Interaction (Dirac) Picture; the development of equation (5) — rates, resonance, golden rule — to the time-dependent-perturbation chapter; the S-matrix to the scattering chapter. Stated here as bridges only.

Builds on: The Time-Evolution Operator · Time-Dependent Schrodinger Equation · Spectral Decomposition of Time Evolution

Leads to: The Interaction (Dirac) Picture · Symmetries and Conservation in Quantum Dynamics · Equivalence of the Dynamical Pictures

How do I quickly tell whether I'm allowed to drop the time-ordering symbol?

Compute (or eyeball) $[\hat H(t_1), \hat H(t_2)]$ for two generic times. Safe cases: $\hat H$ constant; $\hat H(t) = f(t)\hat A$ with one fixed operator; any $\hat H(t)$ built from a single commuting family (always diagonal in one fixed basis). Danger signs: a field's direction changes, not just its magnitude; non-commuting operators carry different time profiles ( $f(t)\sigma_z + g(t)\sigma_x$ ); rotating frames not yet transformed away. When in doubt, keep $\mathcal{T}$ — it is inert precisely when it was unnecessary.

Is the Dyson series guaranteed to converge?

For bounded Hamiltonians on finite intervals, yes — the $n$ -th term is bounded by $(\lVert H\rVert\,\Delta t/\hbar)^n/n!$ and the factorial wins. The caveats: convergence can be slow when $\lVert H\rVert\Delta t/\hbar$ is not small (Example 1's ladder), unbounded operators need domain care, and in quantum field theory the corresponding expansion is famously asymptotic rather than convergent — that subtlety belongs to QFT, not to finite-dimensional driven systems.

What's the difference between the Dyson series and the Magnus expansion?

Two repairs of the same failure. Dyson expands the propagator as a sum — easy to generate, but truncations are not unitary. Magnus keeps the propagator an honest exponential, $\hat U = e^{\Omega}$ , and expands the exponent: $\Omega = -\frac{i}{\hbar}\int\hat H + \frac{1}{2}(-\frac{i}{\hbar})^2\int\int[\hat H(t_1), \hat H(t_2)] + \cdots$ , anti-hermitian at every order — so every truncation is exactly unitary. Perturbation theory and field theory run on Dyson; pulse design and numerical integrators, where losing unitarity is fatal, run on Magnus. The first Magnus correction is the very commutator that killed the naive exponential.

Why do later times stand on the left — it feels backwards?

Because operators act on states from the left, so the rightmost factor acts first. Reading a product right-to-left is reading time forward: $\hat U = \hat U_{\text{late}}\hat U_{\text{early}}$ applies the early evolution, then the late. The film strip keeps it straight — newest frames on top, pile applied bottom-up. Writing the factors the other way is not a convention change; it is a different (wrong) operator whenever the self-commutator is nonzero.

Does the first-order formula already contain Fermi's golden rule?

It contains its seed. Equation (5) says the first-order amplitude is the drive's Fourier component at the Bohr frequency. Specialize to a long harmonic perturbation and $|c^{(1)}|^2$ develops a sharp sinc² peak at resonance; the long-time limit against a continuum of final states turns the peak into a delta function and the probability into a constant rate — the golden rule. The specialization and its validity window are the time-dependent-perturbation chapter's; the input is exactly this article's one-kick term.

Is Trotterization on a quantum computer the same thing as time-ordering?

It is the definition of the time-ordered exponential used as an algorithm. The frame product $\prod_k e^{-i\hat H(t_k)\Delta t/\hbar}$ (later factors left) is $\mathcal{T}\exp$ in the $\Delta t \to 0$ limit; Trotter circuits implement finite- $\Delta t$ versions, their per-step error the commutator terms Magnus catalogs. The further Trotter idea — splitting a time-independent $\hat H = \sum_j\hat H_j$ into non-commuting pieces applied in sequence — is the same mathematics with "time slices" played by "terms of the Hamiltonian."

Each Dyson term looks like particles getting kicked at definite times — is that a real picture?

It is a real decomposition, not a real sequence of events. The system does not choose kick times; the amplitude integrates over all possible kick histories, and histories interfere — exactly as path amplitudes interfere in the sum-over-histories formulation this structure seeds. The picture becomes literal bookkeeping in Feynman diagrams: each vertex one kick, each internal line one stretch of free propagation. What is physical is always the summed, squared amplitude.