Ehrenfest Theorem Revisited

Ehrenfest’s theorem gives exact equations for quantum expectation values. It becomes an approximate classical-trajectory statement only when those equations close on the means and the quantum state remains localized enough for its means to represent the observed motion.

The theorem and its operator derivation are developed in Ehrenfest Theorem. An introductory interpretation is given in Ehrenfest Theorem Overview. The dynamical question is:

$$\text{when does exact motion of means} \quad\Longrightarrow\quad \text{a useful classical trajectory?}$$

The answer requires more than the theorem itself. It requires control of packet width, higher moments, force variation, instability, and the time interval being modeled.

Exact Expectation-Value Equations

For one particle with

$$H = \frac{p^2}{2m}+V(x),$$

define

$$q(t)=\langle x\rangle_t, \qquad \bar p(t)=\langle p\rangle_t.$$

Ehrenfest’s theorem gives

$$\dot q = \frac{\bar p}{m}, \qquad \dot{\bar p} = -\langle V'(x)\rangle.$$

Therefore

$$m\ddot q = -\langle V'(x)\rangle.$$

These equations are exact under the usual domain and boundary assumptions. No large-action, narrow-packet, or small-$\hbar$ approximation has yet been made.

The classical equation for a point at $q(t)$ would instead be

$$m\ddot q_{\rm cl} = -V'(q_{\rm cl}).$$

The gap between the two equations is

$$\delta F_q = -\langle V'(x)\rangle + V'(q).$$

Classical center motion requires $\delta F_q$ to remain negligible at the accuracy and over the time interval of interest.

The Classical Approximation Is a Closure Problem

Write

$$\delta x=x-q, \qquad \langle\delta x\rangle=0,$$

and define the central moments

$$\mu_n = \langle(\delta x)^n\rangle.$$

In particular,

$$\mu_2=\sigma_x^2, \qquad \mu_3 = \langle(\delta x)^3\rangle.$$

A finite Taylor expansion with a controlled expected remainder gives the force near the mean. Smoothness alone does not guarantee convergence of an infinite moment series, and packet tails must be included:

$$\langle V'(x)\rangle = V'(q) + \frac{1}{2}V'''(q)\sigma_x^2 + \frac{1}{6}V^{(4)}(q)\mu_3 + \cdots.$$

The linear term in $\delta x$ vanishes because its expectation is zero. The mean equation becomes

$$m\ddot q = -V'(q) - \frac{1}{2}V'''(q)\sigma_x^2 - \frac{1}{6}V^{(4)}(q)\mu_3 - \cdots.$$

Thus the mean generally couples to the variance, skewness, and all higher moments. Their equations couple to still higher moments. Replacing this hierarchy by one Newtonian trajectory is a moment closure approximation.

For potentials at most quadratic,

$$V(x)=a+bx+cx^2,$$

the force $-V'(x)$ is affine in $x$. Then

$$\langle V'(x)\rangle = V'(\langle x\rangle)$$

for every state, and the mean equations close exactly. Even then, a broad, split, squeezed, or interference-rich state is not a classical point particle. Agreement of center motion with the classical trajectory is only one condition for a particle-like description.

Quantifying the Narrow-Packet Condition

The phrase “narrow packet” must name a comparison scale. Let $F_*\gt0$ be the force accuracy relevant to the problem. A direct condition is

$$\left\lvert \langle V'(x)\rangle - V'(q) \right\rvert \ll F_*.$$

Keeping only the leading width correction suggests

$$\epsilon_F = \frac{ \left\lvert V'''(q)\right\rvert \sigma_x^2 }{2F_*} \ll1.$$

This estimate is useful only if the Taylor expansion is controlled and higher moments are smaller. Near a point where the classical force vanishes, dividing by $\lvert V'(q)\rvert$ is unstable; use an externally chosen force scale, an absolute error, or a norm over the region sampled by the packet.

A more geometric statement is that the force should vary little over the packet’s resolved support. If $L_F$ is a local length scale for appreciable force variation, require

$$\frac{\sigma_x}{L_F} \ll1.$$

A useful absolute bound avoids an infinite Taylor series. If $|V'''(x)|\leq M$ throughout the region carrying the state (or globally), Taylor’s theorem and $\langle x-q\rangle=0$ give

$$|\langle V'(x)\rangle-V'(q)|\leq\frac{M}{2}\sigma_x^2.$$

For a local bound, the probability tails outside that region require a separate force estimate. All these conditions must hold throughout the evolution, not only at $t=0$.

Width and Covariance Dynamics

Introduce the centered momentum

$$\delta p=p-\bar p,$$

the momentum variance

$$\sigma_p^2 = \langle(\delta p)^2\rangle,$$

and the symmetrized covariance

$$C_{xp} = \frac{1}{2} \left\langle \delta x\,\delta p + \delta p\,\delta x \right\rangle.$$

Their exact equations begin with

$$\frac{d}{dt}\sigma_x^2 = \frac{2}{m}C_{xp},$$

and

$$\frac{d}{dt}C_{xp} = \frac{\sigma_p^2}{m} - \left\langle \delta x\,V'(x) \right\rangle.$$

The momentum variance obeys

$$\begin{aligned} \frac{d}{dt}\sigma_p^2 &= -\left\langle \left\{ \delta p, V'(x)-\langle V'(x)\rangle \right\} \right\rangle. \end{aligned}$$

The anticommutator is

$$\{A,B\}=AB+BA.$$

These equations show why checking only $q(t)$ is incomplete. Position–momentum covariance drives spreading; force variation couples the second moments to higher ones; and nonlinear evolution can make an initially Gaussian packet non-Gaussian.

For a quadratic Hamiltonian, the first and second moments form a closed system. For a generic nonlinear potential, no finite collection of moments closes exactly without an additional approximation.

Free Gaussian: Exact Center, Spreading State

For an initially unchirped free Gaussian with $\sigma_x(0)=\sigma_0$, $C_{xp}(0)=0$, and $\sigma_p(0)=\hbar/(2\sigma_0)$,

$$q(t) = q_0+\frac{p_0}{m}t$$

is exactly the classical free trajectory. The width is

$$\sigma_x(t) = \sigma_0 \sqrt{ 1+ \left( \frac{t}{t_{\rm sp}} \right)^2 },$$

with spreading time

$$t_{\rm sp} = \frac{2m\sigma_0^2}{\hbar}.$$

A particle-like description over a measurement scale $L_{\rm obs}$ requires

$$\sigma_x(t) \ll L_{\rm obs}.$$

The center can remain perfectly classical after that localization condition has failed. Larger mass and wider initial packets increase $t_{\rm sp}$, but narrowing a packet indefinitely does not make it indefinitely classical: the uncertainty relation increases its momentum spread and accelerates dispersion.

The exact wavefunction and convention for $\sigma_0$ are given in Gaussian Wave Packets. The scale analysis is developed in Dispersion and Classical Limit.

Quadratic Potentials: Exact Mean Motion

For a harmonic oscillator,

$$V(x) = \frac{1}{2}m\omega^2x^2,$$

the mean satisfies

$$\ddot q+\omega^2q=0$$

for every state. A coherent state retains a Gaussian shape and constant variance while its center follows the classical phase-space ellipse. A squeezed state has the same exact mean equation while its width oscillates. A superposition of separated packets can also have a classically moving mean.

This exact quadratic exception is powerful but easy to overread. It proves closure of the first moments, not classicality of the full probability distribution or the absence of interference.

Nonlinear Example: Quartic Correction

Consider

$$V(x) = \frac{1}{2}m\omega^2x^2 + \lambda x^4.$$

Since

$$V'(x) = m\omega^2x+4\lambda x^3,$$

and

$$\langle x^3\rangle = q^3+3q\sigma_x^2+\mu_3,$$

the exact mean equation is

$$m\ddot q = -m\omega^2q - 4\lambda q^3 - 12\lambda q\sigma_x^2 - 4\lambda\mu_3.$$

The terms $-m\omega^2q-4\lambda q^3$ are the classical force evaluated at $q$. The remaining terms are the exact contribution of packet width and skewness to the mean force. Even if $\mu_3=0$ initially, nonlinear evolution can generate nonzero higher moments.

For a state symmetric about $q=0$, the mean can remain at zero by symmetry while the distribution changes width and develops non-Gaussian structure. In that case the mean trajectory contains very little of the physically relevant dynamics.

Approximate Newtonian Motion

A useful trajectory approximation requires several conditions at once:

1. The state is localized relative to the force-variation scale.
2. The momentum spread is small relative to the required momentum resolution.
3. Packet spreading, shearing, and splitting remain negligible over the chosen time interval.
4. Tunneling and interference between distinct classical branches are irrelevant at the chosen resolution.
5. The mean is representative of the distribution rather than lying in a low-probability region between peaks.
6. Environmental noise and measurement backaction are negligible or included in the target classical stochastic model.

The packet-level synthesis, including group velocity, covariance tubes, external-force distortion, and branch splitting, continues in Wave Packets and Classical Trajectories.

When these conditions hold,

$$q(t)\approx q_{\rm cl}(t), \qquad \bar p(t)\approx p_{\rm cl}(t)$$

can be a controlled approximation. The approximation is state-, observable-, time-, and resolution-dependent.

From Force Error to Position Error

A small force correction must be integrated in time before it becomes a trajectory-accuracy statement. Compare $q(t)$ with the classical solution having the same initial position and velocity. On a fixed interval $0\leq t\leq T$, suppose

$$|\delta F_q(t)|\leq R, \qquad |V'(x)-V'(y)|\leq L|x-y|,$$

where $R\geq0$ has units of force and $L\geq0$ has units of force per length. The second bound must hold between the two centers throughout that interval; a bound $|V''|\leq L$ there suffices. The first must include the full quantum distribution, including its tails. For example, the earlier global bound $|V'''|\leq M$ and a verified width bound $\sigma_x(t)\leq\sigma_*$ give $R=M\sigma_*^2/2$.

Let $e(t)=q(t)-q_{\rm cl}(t)$. Subtract the equations of motion and integrate twice, using $e(0)=\dot e(0)=0$:

$$|e(t)|\leq\frac{1}{m}\int_0^t(t-s)\left[L|e(s)|+R\right]ds.$$

Iterating this positive integral inequality gives

$$|e(t)|\leq \begin{cases} \displaystyle\frac{R}{L}\left[\cosh\!\left(\sqrt{L/m}\,t\right)-1\right],&L>0,\\[6pt] \displaystyle\frac{Rt^2}{2m},&L=0. \end{cases}$$

To see the bound directly, successive iterations produce the series $\frac{R}{m}\sum_{j\geq0}(L/m)^jt^{2j+2}/(2j+2)!$. The remaining iterated term vanishes on a finite interval because $e$ is bounded there and the factorial dominates. Equivalently, the bound is the solution of $m\ddot b=Lb+R$ with $b(0)=\dot b(0)=0$. This is a second-order comparison estimate; Teschl (2012, §2.4) develops the related continuous-dependence and Gronwall estimates for first-order systems.

Compare the resulting length with the allowed position error. The small-time limit is $Rt^2/(2m)$; later, force sensitivity can amplify the accumulated error. The bound is sufficient and conservative: its exponential envelope is not a measured Lyapunov exponent, and it can overestimate errors in stable motion. Nor does it prove the assumed width bound or show that the center represents a single localized branch. Those remain separate checks.

Ehrenfest Time and Unstable Motion

Classical instability can shorten the interval over which a localized packet shadows one trajectory. If an initial phase-space uncertainty $\delta z_0$ grows approximately as

$$\delta z(t) \sim \delta z_0e^{\lambda t},$$

with positive Lyapunov exponent $\lambda$, then reaching a classical variation scale $L$ takes roughly

$$t_E \sim \frac{1}{\lambda} \log\left( \frac{L}{\delta z_0} \right).$$

In semiclassical families where the initial packet scale shrinks with $\hbar$, this produces a logarithmic dependence on an action ratio such as $S/\hbar$.

This estimate is not one universal clock. Its coefficient and even its scaling depend on the system, observable, state family, and error norm. Integrable systems can have algebraic spreading times, and some semiclassical observables remain accurate beyond a naive packet-shadowing time. The robust lesson is narrower: long-time and small-$\hbar$ limits need not commute in unstable dynamics.

Quantum Chaos Preview places this packet-shadowing time beside spectral statistics, periodic-orbit sums, quantum maps, transport, and operator-growth diagnostics.

Semiclassical Limits and Correspondence places the same finite-time condition beside the distinct correspondence statements for spectra, coarse-grained eigenstate densities, and trajectory-sum amplitudes.

Relation to Decoherence

Environmental decoherence can suppress observable interference between spatially separated packets and can favor localized pointer-state families. It can therefore help sustain an effective ensemble of classical-looking alternatives.

It does not make the closed Ehrenfest equations sufficient by itself. Open-system dynamics generally adds friction, diffusion, noise, or measurement-conditioned terms to the equations for moments. Decoherence can suppress coherence while simultaneously broadening a momentum distribution. It also does not select one outcome in an interpretation-neutral unitary account.

Use Decoherence as a Classical-Limit Bridge for the relation to closed-system trajectory approximations, What Is Decoherence? for the mechanism, and What Decoherence Does Not Solve for the outcome boundary.

Failure Modes

• Broad packet: force variation across the state makes higher moments important.
• Split packet: the mean can lie where the probability density is nearly zero.
• Anharmonic shearing: an initially Gaussian packet develops skewness and fine structure.
• Tunneling: motion crosses a classically forbidden region without being captured by one Newtonian trajectory.
• Interference: branch phases affect later probabilities even if the branches separately look localized.
• Chaotic amplification: small initial quantum widths become macroscopic on an Ehrenfest timescale.
• Singular potentials or boundaries: formal commutator manipulations may require careful operator domains and boundary terms.
• Open dynamics: damping, diffusion, and measurement backaction change the mean and covariance equations.

The phase-space route to the same issue is developed in Classical Limit of the Moyal Bracket. It compares the full distributional generator with classical Liouville flow rather than tracking only first moments.

Common Mistakes

• Saying that Ehrenfest’s theorem is approximate; the theorem is exact, while the Newtonian closure is approximate.
• Replacing $\langle V'(x)\rangle$ by $V'(\langle x\rangle)$ without an error estimate.
• Checking that a packet is narrow only at the initial time.
• Assuming exact harmonic mean motion makes every harmonic-oscillator state classical.
• Ignoring skewness and higher moments in an anharmonic potential.
• Treating the mean of a bimodal state as a likely particle position.
• Calling $t_E$ a universal sharp breakdown time.
• Assuming decoherence removes spreading or supplies a unique outcome.

References

• L. E. Ballentine, Quantum Mechanics: A Modern Development, 2nd ed., World Scientific, 2014.
• P. Ehrenfest, “Bemerkung über die angenäherte Gültigkeit der klassischen Mechanik innerhalb der Quantenmechanik,” Zeitschrift für Physik 45, 455–457, 1927.
• M. C. Gutzwiller, Chaos in Classical and Quantum Mechanics, Springer, 1990.
• G. A. Hagedorn, “Semiclassical quantum mechanics. I. The $\hbar\to0$ limit for coherent states,” Communications in Mathematical Physics 71, 77–93, 1980, doi:10.1007/BF01230088.
• E. J. Heller, “Time-dependent approach to semiclassical dynamics,” Journal of Chemical Physics 62, 1544–1555, 1975, doi:10.1063/1.430620.
• R. G. Littlejohn, “The semiclassical evolution of wave packets,” Physics Reports 138, 193–291, 1986, doi:10.1016/0370-1573(86)90103-1.
• G. Teschl, Ordinary Differential Equations and Dynamical Systems, Graduate Studies in Mathematics 140, American Mathematical Society, 2012, §2.4, author’s text.
• W. H. Zurek, “Decoherence, einselection, and the quantum origins of the classical,” Reviews of Modern Physics 75, 715–775, 2003.

Exercises

1. Derive the leading correction to Newton’s force law for a narrow packet.

Solution

Let $x=q+\delta x$ with $q=\langle x\rangle$ and $\langle\delta x\rangle=0$. Expand

$$V'(q+\delta x) = V'(q) + V''(q)\delta x + \frac{1}{2}V'''(q)(\delta x)^2 + \cdots.$$

Taking the expectation value gives

$$\langle V'(x)\rangle = V'(q) + \frac{1}{2}V'''(q)\sigma_x^2 + \cdots.$$

Therefore

$$m\ddot q = -V'(q) - \frac{1}{2}V'''(q)\sigma_x^2 + \cdots.$$

2. Why does exact classical mean motion in a harmonic oscillator not imply that every state is classical?

Solution

The harmonic force is linear, so the first moments close exactly for every state. But the theorem says nothing by itself about the packet width, number of peaks, interference fringes, squeezing, or measurement statistics. A Schrödinger-cat superposition can have the same mean equation as a coherent state while having a radically different probability distribution and coherence structure.

3. Derive the width correction for the quartic oscillator.

Solution

For

$$V(x) = \frac{1}{2}m\omega^2x^2 + \lambda x^4,$$

the force expectation is

$$-\langle V'(x)\rangle = -m\omega^2q - 4\lambda\langle x^3\rangle.$$

Writing $x=q+\delta x$ gives

$$\langle x^3\rangle = q^3+3q\sigma_x^2+\mu_3.$$

Hence

$$m\ddot q = -m\omega^2q - 4\lambda q^3 - 12\lambda q\sigma_x^2 - 4\lambda\mu_3.$$

4. A free Gaussian has initial width $\sigma_0$. At what time does its width reach $\sqrt{2}\sigma_0$?

Solution

The width is

$$\sigma_x(t) = \sigma_0 \sqrt{1+(t/t_{\rm sp})^2}.$$

Setting $\sigma_x(t)=\sqrt{2}\sigma_0$ gives

$$1+ \left( \frac{t}{t_{\rm sp}} \right)^2 =2.$$

Therefore

$$t=t_{\rm sp} = \frac{2m\sigma_0^2}{\hbar}.$$

5. Explain why the mean position of an equal superposition of two well-separated packets can be a poor trajectory variable.

Solution

If the packets have equal weight and centers at $+a$ and $-a$, symmetry can give $\langle x\rangle=0$. The probability density may nevertheless be concentrated near the two packet centers and nearly vanish around zero. The mean then lies between the likely outcomes and does not represent either branch. One must track the distribution, branch weights, and possible interference rather than one mean coordinate.

6. Derive the exact equation for $d\sigma_x^2/dt$.

Solution

Since

$$\sigma_x^2 = \langle x^2\rangle-q^2,$$

Ehrenfest’s identity gives

$$\frac{d}{dt}\langle x^2\rangle = \frac{1}{m} \langle xp+px\rangle.$$

Also,

$$\frac{d}{dt}q^2 = \frac{2q\bar p}{m}.$$

Subtracting yields

$$\begin{aligned} \frac{d}{dt}\sigma_x^2 &= \frac{1}{m} \left( \langle xp+px\rangle -2q\bar p \right) \\ &= \frac{2}{m}C_{xp}. \end{aligned}$$

7. For $V(x)=V_0\cos(x/\ell)$ with $V_0>0$ and $\ell>0$, suppose an independent calculation establishes $\sigma_x(t)\leq\sigma_*$ throughout the interval of interest. Bound the separation of the quantum and classical centers with matching initial position and velocity. If $\sigma_*=0.1\ell$, how long does this estimate guarantee a position error at most $0.01\ell$?

Solution

The derivative bounds are global:

$$|V''|\leq\frac{V_0}{\ell^2}=L, \qquad |V'''|\leq\frac{V_0}{\ell^3}=M.$$

Thus the force residual satisfies $R=M\sigma_*^2/2$. With $\gamma=\sqrt{V_0/(m\ell^2)}$, the position estimate becomes

$$|q(t)-q_{\rm cl}(t)|\leq \frac{\sigma_*^2}{2\ell}\left[\cosh(\gamma t)-1\right].$$

For the stated width this is $0.005\ell[\cosh(\gamma t)-1]$. It is at most $0.01\ell$ when

$$\gamma t\leq\operatorname{arcosh}3\simeq1.76275.$$

The guarantee applies only on the intersection of this time interval with the interval on which the width bound has actually been established. It controls the center error, not the full wavefunction or the probability of every position outcome. The global derivative bounds ensure that no additional tail truncation was used in the force estimate.