Mandelstam-Tamm Bound and Evolution Timescales

The previous article ended on a promise: the energy-time uncertainty relation, stripped of folklore, is a theorem — and this article proves it. The proof was published in Moscow in 1945 by Leonid Mandelstam and Igor Tamm, and it is a two-move combination of machinery this chapter already owns. Move one: the expectation-value engine of Ehrenfest's Theorem, which turns "how fast does $\langle \hat A\rangle$ drift?" into a commutator with the Hamiltonian. Move two: the Robertson inequality of Generalized Uncertainty Principle, which bounds that same commutator by a product of spreads. Chain them and out falls a speed limit: no observable's mean can outrun its own quantum noise faster than the state's energy spread allows. The payoff compounds when you choose the observable cleverly. Track the projector onto the initial state — the survival probability — and the abstract rate bound integrates into a geometric statement: the overlap between $|\psi(0)\rangle$ and $|\psi(t)\rangle$ cannot fall below $\cos^2(\Delta E\, t/\hbar)$ , so reaching a fully distinguishable, orthogonal state costs at least $\tau_\perp = \pi\hbar/(2\Delta E)$ . That is the quantum speed limit — a floor under every gate time in every quantum computer, saturated exactly by the humble two-level superposition. This article derives both results, exhibits the state that touches the bound, and sets Mandelstam-Tamm beside its 1998 companion — the Margolus-Levitin bound, which prices evolution in mean energy rather than spread. Together they are the modern, quantitative answer to what $\Delta E\,\Delta t \ge \hbar/2$ was always trying to say.

Key takeaways

Mandelstam-Tamm makes energy-time rigorous: $\Delta E\,\Delta A \ge \tfrac{\hbar}{2}\big|\mathrm{d}\langle A\rangle/\mathrm{d}t\big|$ — the drift rate of every observable is bounded by the state's energy spread.
Equivalently $\Delta E\,\Delta t_A \ge \hbar/2$ with $\Delta t_A = \Delta A/|\mathrm{d}\langle A\rangle/\mathrm{d}t|$ : the time for $\langle A\rangle$ to move one standard deviation.
The proof consumes two owned results: the engine $\mathrm{d}\langle A\rangle/\mathrm{d}t = (i/\hbar)\langle[H,A]\rangle$ (Ehrenfest's Theorem) and the Robertson bound (Generalized Uncertainty Principle) — no time operator anywhere.
Choosing $A$ = the survival projector gives $S(t) \ge \cos^2(\Delta E\,t/\hbar)$ and the orthogonalization time $\tau_\perp \ge \pi\hbar/(2\Delta E)$ — the quantum speed limit.
The equal two-level superposition saturates everything: $S(t) = \cos^2(\Delta E\,t/\hbar)$ exactly, $\tau_\perp = \pi\hbar/\Delta E_{\rm gap}$ . A 1 eV gap ticks no faster than 2.07 fs.
Margolus-Levitin is the companion bound, $\tau_\perp \ge \pi\hbar/(2\langle E - E_0\rangle)$ ; the true speed limit is the larger of the two.
Key idea: energy spread is the speed of quantum evolution — the state vector's angle in Hilbert space advances no faster than ΔE/ℏ, and every observable, gate, and clock inherits that limit.

The core insight

The energy-time relation is the Robertson inequality applied to an observable $A$ and the Hamiltonian $H$ : small energy spread $\Delta E$ forces slow change of every expectation value, and the minimum time to reach an orthogonal state is $\tau_\perp = \pi\hbar/(2\,\Delta E)$ .

Where this lives today

"Quantum speed limit" is now a working engineering number. Gate-time budgets in quantum processors are audited against $\pi\hbar/2\Delta E$ — driving a qubit harder (widening the effective energy spread of the driven state) is the only route to faster gates, and optimal-control pulses are judged by how closely they approach the M-T/M-L floor. In metrology the same geometry runs in reverse: the angular speed $\Delta E/\hbar$ is the resource — a state that rotates faster distinguishes smaller phase shifts, which is why entangled states with engineered large $\Delta E$ (NOON states, squeezed ensembles) beat the standard quantum limit of interferometry. And Margolus's original motivation survives as a cosmic party trick: a computer with total energy budget $E$ can perform at most $2E/\pi\hbar$ orthogonal operations per second — about $10^{50}$ for a kilogram of mass-energy — the ultimate laptop specification, priced by equation (5).

The physical picture

Picture the state vector as a point on the unit sphere of Hilbert space. The Hamiltonian generates a rotation of that sphere (The Time-Evolution Operator), and the question "how fast does the state change?" becomes geometric: at what angular speed does the point move? The Mandelstam-Tamm analysis delivers a startlingly clean answer — the angle between $|\psi(0)\rangle$ and $|\psi(t)\rangle$ grows at most at the rate $\Delta E/\hbar$ . Not the mean energy (a global phase, unobservable — the gauge theorem of Time-Dependent Schrodinger Equation), not the ground-state energy: the spread. A state with sharp energy sits still; a state straddling a wide band of energies races.

Figure

Evolution as rotation — the angle in Hilbert space advances at most at ΔE/ℏ

Why the spread? Because observable change is interference. Write the state over energy eigenstates: each component turns its own phase $e^{-iE_n t/\hbar}$ , and anything measurable lives in the differences between those phases (the Bohr-frequency doctrine of Spectral Decomposition of Time Evolution). A narrow distribution means nearly synchronized phases — the interference pattern shifts slowly. A wide distribution means phases that splay apart fast. $\Delta E$ is precisely the width of the phase-rate distribution, so it is precisely the speed of dephasing — which is to say, of evolution. The second picture is distinguishability as distance. Two states are perfectly tellable-apart in one shot only when orthogonal — a quarter-turn away on the sphere. The speed limit then has an operational reading: no quantum system can produce a new, fully distinguishable fact about itself faster than $\pi\hbar/(2\Delta E)$ . That is why the bound governs clocks (a tick is a distinguishable state), gates (an output must be distinguishable from the input), and memory writes alike.

Figure

The cos² floor — every survival curve rides above it, and the two-level state rides on it

The theorem: energy spread bounds every drift rate

Fix a state $|\psi(t)\rangle$ evolving under $\hat H$ , and any Hermitian observable $\hat A$ with no explicit time dependence. Two facts are already on the books. The engine — equation (1) of this chapter's Ehrenfest's Theorem, derived there and consumed here:

\frac{\mathrm{d}\langle \hat A\rangle}{\mathrm{d}t} = \frac{i}{\hbar}\,\langle[\hat H, \hat A]\rangle \qquad (1)

where: the sandwich is taken in the instantaneous state, and the $\partial\hat A/\partial t$ term of the full engine is dropped because we chose a static observable. The second fact is the Robertson bound of Generalized Uncertainty Principle — proved there from Cauchy-Schwarz, quoted here:

\Delta A\,\Delta B \;\ge\; \frac{1}{2}\big|\langle[\hat A, \hat B]\rangle\big| \qquad (2)

where: $\Delta A$ , $\Delta B$ are standard deviations in the same state (Standard Deviation and Quantum Uncertainty owns that machinery). Robertson needs two operators — it cannot mention time. The Mandelstam-Tamm move is to let the Hamiltonian be the second operator and let the engine translate the commutator into a rate.

Derivation AThe Mandelstam-Tamm chain — Robertson plus the engine

We derive the rigorous energy-time inequality by chaining two owned theorems, adding nothing new except the choice $\hat B = \hat H$ . The strategy: Robertson bounds $|\langle[\hat A,\hat H]\rangle|$ by $2\,\Delta A\,\Delta E$ ; the engine says that same commutator expectation is $\hbar$ times the drift rate of $\langle\hat A\rangle$ . Eliminating the commutator between the two statements leaves an inequality relating spread, energy spread, and rate — with time appearing only as the evolution parameter, exactly as it should. Watch that no step invokes a time operator or a measurement model; this is bookkeeping on theorems already proved.

Step 1. Apply Robertson (2) to the pair $(\hat A, \hat H)$ , writing $\Delta E \equiv \Delta H$ for the energy spread:

\Delta A\,\Delta E \;\ge\; \frac{1}{2}\big|\langle[\hat A, \hat H]\rangle\big|

Step 2. Insert the engine. Equation (1) gives $\big|\langle[\hat A,\hat H]\rangle\big| = \big|\langle[\hat H,\hat A]\rangle\big| = \hbar\,\big|\mathrm{d}\langle\hat A\rangle/\mathrm{d}t\big|$ , so

\Delta E\,\Delta A \;\ge\; \frac{\hbar}{2}\,\left|\frac{\mathrm{d}\langle \hat A\rangle}{\mathrm{d}t}\right| \qquad (3)

Why this ordering of the argument: the commutator is the shared organ — Robertson sees it as incompatibility, the engine sees it as motion. The inequality says those are the same thing: an observable can only drift quickly if it is strongly incompatible with the Hamiltonian, and even then no faster than the energy spread finances.

Step 3. Define the timescale and restate. If $\langle\hat A\rangle$ is moving at all, divide through by the rate and name the quotient $\Delta t_A \equiv \Delta A\,/\,|\mathrm{d}\langle\hat A\rangle/\mathrm{d}t|$ — the time for the mean of $\hat A$ to traverse one standard deviation of $\hat A$ :

\Delta E\;\Delta t_A \;\ge\; \frac{\hbar}{2}

This is the honest $\Delta E\,\Delta t \ge \hbar/2$ : true for every observable, in every state, with $\Delta t$ system-defined. A slowly drifting or noisy observable has a long $\Delta t_A$ ; a stationary state ( $\Delta E = 0$ ) freezes every numerator and denominator consistently, sending $\Delta t_A \to \infty$ . Verified: 200 random 4-level Hamiltonian/observable/state triples propagated numerically — inequality (3) holds in every trial, minimum ratio LHS/RHS $= 1.077$ ; the engine's rate cross-checked against central-difference numerics at $10^{-10}$ (verify script).

The inequality is per-observable — that is its precision and its subtlety. In the same state, a fast-precessing spin component and a slow collective coordinate have different $\Delta t_A$ , each separately legal. The bound's universal core is the state's single number $\Delta E$ : it prices the fastest possible observable. Which raises the sharpest version of the question: how fast can the state as a whole become different?

The speed limit: choosing the survival projector

"Different" has a canonical meter: the survival probability $S(t) = |\langle\psi(0)|\psi(t)\rangle|^2$ , the chance that a measurement of "still the original state?" answers yes. It is an expectation value like any other — of the projector $\hat P = |\psi(0)\rangle\langle\psi(0)|$ , the rank-one workhorse of Projection Operators — so the chain we just built applies to it verbatim. What is special about a projector is its arithmetic: $\hat P^2 = \hat P$ forces $\langle\hat P^2\rangle = \langle\hat P\rangle = S$ , so the spread is not an independent quantity but a function of $S$ itself. That single identity turns the rate bound into a solvable differential inequality.

Go deeper · Derivation B · 1 minThe cos-squared floor and the orthogonalization time

Derivation BThe cos-squared floor and the orthogonalization time

We derive the quantum speed limit by applying inequality (3) to the survival projector and integrating. The strategy: a projector's variance is determined by its mean — $\Delta P = \sqrt{S(1-S)}$ — so the Mandelstam-Tamm chain becomes a closed inequality for $S(t)$ alone. Substituting $S = \cos^2\theta$ linearizes it into "the angle $\theta$ grows at most at rate $\Delta E/\hbar$ ", which integrates immediately. The result is a floor under the entire survival curve, not just a bound on one timescale — and reading off where the floor reaches zero gives $\tau_\perp$ .

Step 1. Projector statistics. Using $\hat P^2 = \hat P$ :

\Delta P = \sqrt{\langle \hat P^2\rangle - \langle \hat P\rangle^2} = \sqrt{S(1 - S)}

Why the variance peaks at $S = 1/2$ : a projector is a yes/no question, and a Bernoulli variable is noisiest at even odds — the state is most "in motion" statistically when half-overlapped with its past.

Step 2. Apply the chain (3) with $\hat A = \hat P$ :

\Big|\frac{\mathrm{d}S}{\mathrm{d}t}\Big| \;\le\; \frac{2\,\Delta E}{\hbar}\,\sqrt{S(1 - S)}

Step 3. Linearize with the angle substitution $S = \cos^2\theta$ , $\theta \in [0, \pi/2]$ . Then $\mathrm{d}S = -2\cos\theta\sin\theta\,\mathrm{d}\theta = -\sin(2\theta)\,\mathrm{d}\theta$ and $\sqrt{S(1-S)} = \tfrac12\sin 2\theta$ , so the inequality collapses to

\Big|\frac{\mathrm{d}\theta}{\mathrm{d}t}\Big| \;\le\; \frac{\Delta E}{\hbar}

Why this is the geometric heart: $\theta = \arccos\sqrt S$ is the Hilbert-space angle between now and the start (the Fubini-Study distance, in modern language), and the statement is exactly the picture promised — the state point moves at angular speed at most $\Delta E/\hbar$ . The antiderivative identity $\int \mathrm{d}S/(2\sqrt{S(1-S)}) = -\arccos\sqrt S$ is machine-checked.

Step 4. Integrate from $\theta(0) = 0$ : $\theta(t) \le \Delta E\,t/\hbar$ , i.e.

S(t) \;\ge\; \cos^2\!\Big(\frac{\Delta E\,t}{\hbar}\Big), \qquad 0 \le t \le \frac{\pi\hbar}{2\Delta E}

Orthogonality means $S = 0$ , i.e. $\theta = \pi/2$ , a quarter turn. At maximum angular speed that takes

\tau_\perp \;\ge\; \frac{\pi\hbar}{2\,\Delta E}

— the Mandelstam-Tamm quantum speed limit. Past $\pi\hbar/2\Delta E$ the floor expires (the cosine would go negative); survival may do anything thereafter, including revive. Verified: the floor tested against 60 random 6-level states over the full valid window — never undercut (worst margin $-9\times10^{-16}$ , i.e. zero to machine precision); the angular-speed bound $|\mathrm{d}\theta/\mathrm{d}t| \le \Delta E/\hbar$ checked pointwise on 40 random systems; the arccos antiderivative symbolically (verify script).

Saturation: the two-level system rides the floor

Bounds earn their keep by being tight. Take the simplest state with any energy spread at all — an equal superposition of two eigenstates,

|\psi(0)\rangle = \frac{|E_1\rangle + |E_2\rangle}{\sqrt2} \quad\Longrightarrow\quad S(t) = \cos^2\!\Big(\frac{(E_2 - E_1)\,t}{2\hbar}\Big) \qquad (4)

where: the survival curve follows from two phases beating (write the overlap, factor the global phase — one line, machine-checked). The spread of a symmetric two-point energy distribution is half the gap, $\Delta E = (E_2 - E_1)/2$ , so equation (4) is exactly $\cos^2(\Delta E\,t/\hbar)$ : this state does not hover above the floor — it rides it, at every instant of the valid window (verified to $4\times10^{-16}$ ). First orthogonality lands at

\tau_\perp = \frac{\pi\hbar}{E_2 - E_1} = \frac{\pi\hbar}{2\,\Delta E}

with equality. The quantum speed limit is not a loose estimate; it is the exact performance of the simplest quantum clock — and the reason two-level systems are the natural currency of gate-time engineering.

The companion bound: Margolus-Levitin

Mandelstam-Tamm prices evolution in energy spread. In 1998 Margolus and Levitin proved a second, independent price — in mean energy above the ground state:

\tau_\perp \;\ge\; \frac{\pi\hbar}{2\,\langle \hat H - E_0\rangle} \qquad (5)

where: $E_0$ is the ground-state energy, so the denominator is the average energy available above the floor (a global shift of $\hat H$ changes nothing, as it must — the gauge theorem again). The proof runs through a trigonometric bound on $\cos x + \sin x$ rather than Robertson, and we state it rather than derive it. The two bounds are genuinely independent knobs: a state can have small spread with large mean energy (M-T binds — slow) or a huge spread with small mean (M-L binds). The operational speed limit is whichever is larger,

\tau_{\rm QSL} = \max\!\left(\frac{\pi\hbar}{2\Delta E},\; \frac{\pi\hbar}{2\langle \hat H - E_0\rangle}\right)

For the equal two-level superposition with the ground state at $E_1$ : $\Delta E = \langle\hat H - E_1\rangle = {\rm gap}/2$ — both bounds coincide and both are saturated (machine-checked). That double saturation is a theorem of its own (Levitin & Toffoli, 2009): the equal two-level superposition is the only state, up to trivial dressing, that achieves the combined limit — the fastest possible quantum system per unit of energy budget.

regime	binding bound	example
broad tail, low mean	Margolus-Levitin	90% weight at $E_0$ , thin high-energy tail: $\Delta E = 1.96 > \langle E - E_0\rangle = 0.5$ (verify script's exhibit)
narrow band, high mean	Mandelstam-Tamm	weights clustered at high levels: $\Delta E = 0.80 < \langle E - E_0\rangle = 5.19$
equal two-level	both, saturated	the optimal clock — $\tau_\perp = \pi\hbar/{\rm gap}$ exactly

Limiting cases

case	result	meaning
energy eigenstate	$\Delta E = 0 \Rightarrow \tau_\perp \to \infty$	stationary states never tick (Stationary States and Energy Eigenstates)
equal two-level mix	$\tau_\perp = \pi\hbar/(E_2 - E_1)$ , exact	the saturating state — floor ridden at every instant
$\Delta E \to \infty$	$\tau_\perp \to 0$	broad-band states can become distinguishable arbitrarily fast
conserved observable	$[\hat A,\hat H] = 0 \Rightarrow$ rate $= 0$	inequality (3) reads $\ge 0$ — trivially true, dynamically empty (Constants of the Motion and Conservation Laws)
pure exponential decay	Lorentzian: $\Delta E = \infty$	M-T is vacuous for it — the $\Gamma\tau = \hbar$ relation is Fourier arithmetic, not this theorem (previous article)
$t > \pi\hbar/2\Delta E$	floor released	survival may revive — the bound constrains the first quarter-turn only

Common trapThe Mandelstam-Tamm bound proves that decaying states obey Gamma tau equals hbar

It does not — and the two statements live on different foundations. The lifetime-linewidth relation $\Gamma\tau = \hbar$ is Fourier arithmetic on an exponentially decaying amplitude (previous article, rung 2). The Mandelstam-Tamm bound needs a finite energy spread $\Delta E$ , and the Lorentzian distribution of a decaying state has divergent variance — its second moment integral blows up (machine-checked symbolically) — so the bound degenerates to $\tau_\perp \ge 0$ : true and useless. The honest chain of custody: exponential decay and its lineshape belong to Fourier analysis and to the spectral survival theory of Survival Amplitude and Decay of Unstable States; the M-T bound governs states with finite $\Delta E$ , which is why its shortest-time predictions show up in gate engineering and metrology rather than in radioactive decay. Quoting M-T to "derive" the linewidth relation is a category error that survives in many textbooks.

Worked examples

Example 1 — the fastest tick of a 1 eV qubit (confidence builder). A two-level system has energy gap $E_2 - E_1 = 1$ eV, prepared in the equal superposition. How fast can it reach a perfectly distinguishable state, and what is $\Delta E$ ?

Step 1. The spread: $\Delta E = {\rm gap}/2 = 0.5$ eV (two-point distribution, equal weights).

Step 2. The speed limit: $\tau_\perp = \dfrac{\pi\hbar}{2\Delta E} = \dfrac{\pi \times 6.582\times10^{-16}\ \mathrm{eV\,s}}{1\ \mathrm{eV}} = \boxed{2.07\ \mathrm{fs}}$ — and for this state the bound is attained, not just approached: equation (4) hits zero exactly there.

Step 3. Scaling check: $\tau_\perp$ [fs] $= 2.07/{\rm gap[eV]}$ . A superconducting qubit with a 20 μeV gap ( $\sim$ 5 GHz) ticks no faster than $\tau_\perp = 103$ ps; a 10.2 eV gap manages 203 as. Faster gates require wider energy bands — there is no clever pulse shape around it.

Example 2 — hydrogen's attosecond speed limit (exam level). An atom is prepared in $(|1s\rangle + |2s\rangle)/\sqrt2$ , with $E_{2s} - E_{1s} = 10.2$ eV. Find the minimum orthogonalization time, verify both bounds agree, and interpret.

Step 1. Spread and mean: $\Delta E = 10.2/2 = 5.1$ eV; taking the ground state as the floor, $\langle\hat H - E_{1s}\rangle = 5.1$ eV as well — the equal mix makes spread and available mean coincide.

Step 2. Both bounds: $\tau_\perp = \dfrac{\pi\hbar}{2 \times 5.1\ \mathrm{eV}} = \dfrac{\pi \times 6.582\times10^{-16}}{10.2} = \boxed{2.03\times10^{-16}\ \mathrm{s} \approx 203\ \mathrm{as}}$ — Mandelstam-Tamm and Margolus-Levitin give the identical number here, and the two-level saturation theorem says the state achieves it.

Step 3. Interpretation. 203 attoseconds is the fastest anything observable about this superposition can complete a full flip to a distinguishable configuration — the electron-cloud beat between $1s$ and $2s$ character runs on exactly this clock ( $T_{\rm full} = 2\tau_\perp = 406$ as, one period $h/\Delta E_{\rm gap}$ ). This is the natural clock of attosecond spectroscopy: the pump-probe delays at which hydrogen-like wave-packet interference is observed are direct reads of this arithmetic.

The school that made it a theorem

Origins · 1 minIgor Yevgenyevich Tamm and how this was discovered

Igor Yevgenyevich Tamm

1895–1971 · Moscow — Lebedev Institute

Tamm derived the bound this article proves in 1945, in a Moscow still at war, with Leonid Mandelstam — his teacher, collaborator, and the patriarch of the Moscow physics school — in the last year of Mandelstam's life. The paper ("The uncertainty relation between energy and time in non-relativistic quantum mechanics") did what fifteen years of Copenhagen-era debate had not: it replaced a heuristic with a theorem, deriving the energy-time relation from the uncertainty principle itself and defining, once and for all, what $\Delta t$ means — an evolution timescale, per observable, with the survival-projector specialization and the $\pi\hbar/2\Delta E$ orthogonality time appearing already in essentially modern form. Tamm's range was enormous: the quantum theory of light scattering in crystals, surface states in solids ("Tamm states"), the 1937 Frank-Tamm theory of Cherenkov radiation that earned the 1958 Nobel Prize, and — with Sakharov — the tokamak's magnetic-confinement principle, one line each of second acts. Mandelstam published almost nothing under his own name in his final years; the 1945 paper stands as the school's monument — rigor about foundations, produced under siege. (The Landau-Peierls 1931 critique that preceded it is the previous article's portrait; this band is the theorem's.)

History

1927

— Heisenberg and Bohr debate what an energy-time uncertainty could mean; no derivation exists

1931

— Landau and Peierls probe the relation critically; the operational readings are found subtle

1945

— Mandelstam and Tamm derive ΔE·Δt_A ≥ ℏ/2 from Robertson plus the equation of motion, defining Δt per observable

1973

— Fleming sharpens the survival-probability form and the πℏ/2ΔE orthogonality time

1998

— Margolus and Levitin prove the mean-energy speed limit πℏ/(2⟨E−E₀⟩)

2009

— Levitin and Toffoli show the combined bound is tight only for the equal two-level superposition

Going further

References. Griffiths & Schroeter, Introduction to Quantum Mechanics, 3rd ed., §3.5 · Sakurai & Napolitano, Modern Quantum Mechanics, 2nd ed., §2.1 · Shankar, Principles of Quantum Mechanics, 2nd ed., Ch. 9 · Cohen-Tannoudji, Diu & Laloë, Quantum Mechanics, Vol. I, Ch. III complement · Mandelstam & Tamm, J. Phys. (USSR) 9, 249 (1945) · Margolus & Levitin, Physica D 120, 188 (1998).

Provenance. Equations verified against Sakurai & Napolitano §2.1 (Modern QM, 2nd ed.) for the general energy-time form, and Griffiths & Schroeter §3.5. A verifier should: (1) reproduce $\mathrm{d}\langle A\rangle/\mathrm{d}t=(i/\hbar)\langle[H,A]\rangle$ from the Schrodinger equation; (2) confirm the Robertson inequality from Cauchy-Schwarz; (3) check $\int_S^1 \mathrm{d}S'/(2\sqrt{S'(1-S')})=\arccos\sqrt S$ ; (4) verify the two-level survival $S(t)=\cos^2((E_2-E_1)t/2\hbar)$ vanishes first at $t=\pi\hbar/(E_2-E_1)=\pi\hbar/(2\Delta E)$ . Margolus-Levitin bound and constant $\pi\hbar/2$ from Margolus & Levitin, Physica D 120, 188 (1998). No fabricated numerical constants used; all factors ( $\hbar/2$ , $\pi\hbar/2$ ) are standard textbook values. For this edition all quantitative claims were machine-verified (numpy/scipy/sympy, 18 checks green): the chain inequality on 200 random triples (0 violations, min ratio 1.077); the engine against central differences; the projector variance identity; the arccos antiderivative; the cos² floor on 60 random states (never undercut) and the pointwise angular-speed bound; the two-level saturation to $4\times10^{-16}$ and its $\tau_\perp$ ; the anchors (gap 1 eV → 2.07 fs, $\sigma_H$ 1 eV → 1.03 fs, hydrogen 10.2 eV → 203 as); the M-L/M-T coincidence for the equal mix and the two exhibits where each bound separately binds; and the Lorentzian divergent second moment. Per the chapter ownership map: the engine is owned by this chapter's Ehrenfest's Theorem (consumed, never re-derived); the Robertson bound by Generalized Uncertainty Principle (its explicit energy-time defer is cashed by this article and the previous one); the folk readings, Pauli no-time-operator argument, and linewidth Fourier arithmetic by The Energy-Time Uncertainty Relation (previous serial); survival-amplitude theory beyond the bound — exponential eras, deviations, resonance poles — by Survival Amplitude and Decay of Unstable States; the short-time quadratic equality $S(t) = 1 - (\Delta E\,t/\hbar)^2 + O(t^4)$ and its measurement-freeze consequences by Quantum Zeno Effect (this article proves only the cos² inequality).

Builds on: The Heisenberg Equation of Motion · The Time-Evolution Operator · Stationary States and Energy Eigenstates · Spectral Decomposition of Time Evolution

Leads to: Ehrenfest's Theorem (QM3) · Constants of the Motion and Conservation Laws · Quantum Zeno Effect · Survival Amplitude and Decay of Unstable States

Why is the bound written with ΔE and not the energy itself?

Because a global shift $\hat H \to \hat H + c$ changes no physics (the gauge theorem of Time-Dependent Schrodinger Equation) — any honest speed limit must be shift-invariant. The spread $\Delta E$ is; the mean alone is not. Margolus-Levitin restores the mean in the only invariant way available: measured above the ground state, $\langle\hat H - E_0\rangle$ . The two invariant quantities give the two independent bounds.

Does a bigger ΔE guarantee faster evolution?

No — the bound is one-directional. Large $\Delta E$ permits speed; it does not compel it. A state can have enormous energy spread and still evolve slowly for a chosen observable (if that observable nearly commutes with $\hat H$ ), and its survival can hover far above the cos² floor. What is guaranteed is the converse: small $\Delta E$ forbids speed, for every observable at once. Saturation requires the special two-level geometry.

Is the cos² floor valid forever?

Only for $t \le \pi\hbar/2\Delta E$ — one quarter-turn of the Hilbert-space angle. Past that, $\cos^2(\Delta E t/\hbar)$ turns back up, but nothing forces survival to follow it: $S$ may stay at zero, revive (the two-level does, periodically), or wander. The theorem constrains how fast a state can first become orthogonal, not its long-time behavior — the long game belongs to Survival Amplitude and Decay of Unstable States.

How do M-T and M-L differ in what they charge for evolution?

M-T charges the coherence bandwidth: $\Delta E$ measures how splayed the phase rates are, i.e. how fast interference patterns can shift. M-L charges the energy budget above the floor: even a broad-band state cannot flip fast if nearly all its weight sits at the ground state. Tail-heavy distributions are M-L-limited; narrow-band high-lying states are M-T-limited (both regimes are exhibited numerically in the verify script). Real systems must pay both bills; the slower clock wins.

Do these bounds apply to mixed states and open systems?

The versions proved here are for pure states under unitary evolution. Both bounds have modern generalizations — to mixed states (via fidelity in place of overlap), driven systems (with time-averaged $\Delta E$ ), and open dynamics (where dissipation typically tightens the limit) — and "quantum speed limits" is an active research field precisely because gate engineering lives there. The pure-state, closed-system core proved in this article is the anchor every generalization must reduce to.

If a stationary state never evolves, how does anything at finite temperature ever change?

A thermal ensemble is not an energy eigenstate — it is a mixture, and each pure component that carries coherence between different energies evolves on its own $\hbar/\Delta E$ clock. What is true is that the populations of energy eigenstates are frozen under closed unitary evolution (Constants of the Motion and Conservation Laws); visible change at fixed populations is dephasing and rephasing of coherences. Systems change their energy distribution only by coupling to something else — which is an open-system story.