Scrambling and the Fast-Scrambling Scale

Scrambling is the loss of access to initially localized information by simple subsystems, not the destruction of that information. In an all-to-all chaotic system, each infected degree of freedom can influence a finite fraction of the remaining system, so operator support grows exponentially until saturation. Starting from one effective degree of freedom, reaching an order-one fraction of a system with entropy $S$ then takes a time proportional to $\log S$. This derivation explains the fast-scrambling scale while leaving its coefficient, diagnostic, and gravitational regime explicit.

Required background. Quantum Information supplies mutual information, reduced states, and channel language; Gauge Theory and String-Theory Preparation supplies the large-system and holographic setting used to interpret horizon degrees of freedom.

Helpful background. Semiclassical Gravity supplies the Hawking temperature and horizon timescale; Quantum Fields in Curved Spacetime supplies the exterior field description; and Black-Hole Information supplies the recovery problem for which scrambling is one intermediate step.

Scrambling as information spreading

Consider $N$ effective two-state units. Their unconstrained maximum entropy is $S_{\mathrm{sys}}=N\log 2$, and more generally $S_{\mathrm{sys}}=O(N)$ in the large-entropy regime used below. Insert a small message in one unit. A system is scrambled relative to a chosen family of simple subsystems when no such subsystem retains appreciable mutual information with a reference that purifies the message. A related diagnostic measures the support developed by initially simple Heisenberg operators. Growth of one operator alone is insufficient to establish reference decoupling; the relevant operator family and state must also be controlled.

This definition depends on the subsystem algebra and accuracy threshold. It is stronger than local relaxation and weaker than sampling a Haar-random unitary. A few-point observable may thermalize before the message is globally spread; conversely, a state can scramble without its full unitary resembling Haar measure at high moments.

For a concrete counterexample, start with a Bell pair of reference $R$ and message qubit $1$, append $N-1$ zero qubits, and apply controlled-NOT gates from qubit $1$ to every other qubit. The result is

$$\frac{|0\rangle_R|0\cdots0\rangle+|1\rangle_R|1\cdots1\rangle}{\sqrt2}.$$

The circuit spreads $X_1$ into $X_1X_2\cdots X_N$, but every individual output qubit still has mutual information $\log2$ with $R$ for $N\ge2$. It has spread one operator and copied a classical correlation without scrambling the full quantum message in the decoupling sense.

The epidemic growth derivation

Let $s(t)$ be the effective number of units reached by an operator. In an all-to-all model, an infected unit meets an uninfected fraction $1-s/N$. With interaction time $\tau$, the mean-field equation is

$$\frac{ds}{dt}=\frac{1}{\tau}s\left(1-\frac{s}{N}\right), \qquad s(0)=1.$$

Separating variables gives

$$\int_1^{s(t)}\frac{N\,ds'}{s'(N-s')} =\frac{t}{\tau},$$

and therefore

$$s(t)=\frac{N}{1+(N-1)e^{-t/\tau}}.$$

Define the scrambling time by $s(t_*)=fN$ for a fixed $0\lt f\lt1$. Solving yields

$$t_*=\tau\log\left[\frac{f(N-1)}{1-f}\right] =\tau\log N+O(\tau) =\tau\log S_{\mathrm{sys}}+O(\tau).$$

The logarithm follows from two inputs: exponential growth while $s\ll N$ and a single-unit seed. Changing the threshold $f$ alters only the additive $O(\tau)$ term. Starting with $s_0$ infected units gives $t_*\simeq\tau\log(N/s_0)$, which is the first independent limit check.

Sekino and Susskind formulate the fast-scrambler conjecture and its horizon motivation in Sekino and Susskind 2008, § 2, equation (2.4). Their argument motivates a lower scale of order inverse temperature times $\log S$ for systems with sufficiently nonlocal interactions; it is not a theorem that every Hamiltonian with all-to-all couplings scrambles efficiently.

From the interaction time to a black-hole scale

In the strongly coupled thermal models motivating this estimate, the dissipative time is of order $\beta=1/T$; this is a regime assumption, not a universal bound on every relaxation process. If the relevant operator growth is governed by a Lyapunov rate $\lambda_L$, then a small connected effect of size $S^{-1}$ becomes order one when

$$S^{-1}e^{\lambda_L t_*}\sim1, \qquad t_*\sim\frac{1}{\lambda_L}\log S.$$

In a system saturating the thermal chaos bound, $\lambda_L=2\pi/\beta$, this becomes

$$t_*\simeq\frac{\beta}{2\pi}\log S.$$

The factor $2\pi$ belongs only to that regulated OTOC convention and saturation regime. A broader fast-scrambling estimate is $t_*=O(\beta\log S)$. For a stationary black hole, $T_{\rm H}=\kappa/(2\pi)$ after fixing the horizon and normalization of surface gravity, so the scale can also be written $t_*=O(\kappa^{-1}\log S_{\rm BH})$. This restores dimensions: entropy is dimensionless and both $\beta$ and $\kappa^{-1}$ have units of time.

The near-horizon shock-wave calculation gives a gravitational realization of exponential sensitivity in an appropriate large-AdS-black-hole regime; see Shenker and Stanford 2014, § 2, equations (9)–(14). That controlled example does not establish the same coefficient for an arbitrary evaporating, asymptotically flat hole.

Checks and failure regimes

Two limits discriminate the mechanism. First, if $S$ is not large, there is no parametric separation between the interaction time and $t_*$; the logarithmic formula is then only an interpolation. Second, for a spatially local system in $d$ dimensions, a finite propagation velocity makes the reached volume grow at most like $t^d$ before finite-size saturation. Crossing a linear size $L\sim S^{1/d}$ then takes a power of $S$, not $\log S$. The epidemic equation has deliberately removed that locality constraint.

An integrable all-to-all model can also fail: nonlocal connectivity alone does not guarantee chaotic operator growth. Symmetries, conserved charges, weak couplings, and sparse interaction graphs can create slower sectors. A robust calculation should compare an operator-size diagnostic, a mutual-information diagnostic, and the relevant OTOC window rather than infer scrambling from one decaying two-point function.

The scale says when information is delocalized, not when it is decoded. It does not supply a recovery circuit or prove a smooth interior. Generic quantum information, stable thermodynamics, and explicit string constructions remain outside this bounded derivation; they enter only through the stated entropy, temperature, and interaction assumptions.

Exercises

In the epidemic model, take $N=10^6$, $s(0)=10$, and define scrambling by $s(t_*)=N/2$. Express $t_*/\tau$ and compare it with the one-unit seed.

Solution

For a general seed $s_0$, the logistic solution gives $t_*/\tau=\log[(N-s_0)/s_0]$ at $f=1/2$. Thus $t_*/\tau=\log(99999)\simeq11.513$. A one-unit seed gives $\log(999999)\simeq13.816$. Increasing the seed by a factor of ten shortens the time by approximately $\log 10$, confirming that the leading logarithm counts the exponential amplification needed to reach macroscopic support.

References

• Sekino, Yasuhiro, and Leonard Susskind. 2008. “Fast Scramblers.” Journal of High Energy Physics 10, 065. arXiv:0808.2096.
• Shenker, Stephen H., and Douglas Stanford. 2014. “Black Holes and the Butterfly Effect.” Journal of High Energy Physics 03, 067. arXiv:1306.0622.