Evolution Has a Hidden Speed Limit — and It's Weirder Than Diffusion

Evolution has a speed limit, and it turns out that speed limit does not behave the way everyone assumed. Scientists have long treated the search for new phenotypes — new body plans, new biochemical tricks, new adaptations — as a simple diffusion problem: the faster mutations accumulate, the faster a population stumbles onto something new. Tetsuhiro Hatakeyama, a biophysicist at the Earth-Life Science Institute in Tokyo, has now shown that this intuition is wrong in a deeply strange way. Make mutations rarer, and the waiting time for a new phenotype balloons far faster than a simple diffusion model would ever predict — not inversely with the mutation rate, but with a steeper, anomalous power law. The result, laid out in a 2026 preprint, packages a surprise about evolution inside a mathematical puzzle that reaches well beyond biology, into the physics of surfaces growing in turbulent ways.

The headline finding is a crooked number with an exponent hanging off it. For a population of reproducing organisms whose mutations are perfectly random and unbiased — the simplest possible case, with no natural selection at all — the average time for the first individual to reach a fixed phenotypic distance away scales approximately as $({\sigma }^2{)}^{-3/2}$, where ${\sigma }^2$ is the mutation variance per individual per generation. A naive diffusion argument predicts $({\sigma }^2{)}^{-1}$. The two differ by roughly half a power of ${\sigma }^2$, which may sound small, but it is not. Cut the mutation variance by a factor of ten, and a naive model says the wait grows tenfold; the observed law says it grows more than thirty-fold. Neutral evolution is slower than diffusion thinks it should be, over a broad and experimentally plausible range of mutation rates. And this anomalous regime is sandwiched between two ordinary ones, with the naive inverse-variance law recovered on either side.

This is the kind of result that makes a field stop and recheck its arithmetic. The simulations confirm it across seven population sizes, from one hundred individuals up to ten thousand. The anomaly survives even when natural selection is switched on, weakly favoring either extreme or intermediate phenotypes. And it raises a genuinely unsettling possibility: that the pace of evolutionary innovation, in the neutral or near-neutral regimes where new functions are often explored, has been systematically misestimated for decades.

The Science

Hatakeyama works with the Wright–Fisher model, the canonical mathematical description of neutral evolution. Picture a population of $N$ individuals, each carrying a one-dimensional phenotype, all starting at the same value. Each generation, every one of the $N$ organisms produces an offspring; in the neutral case, every individual is equally likely to be anyone's parent, with probability $1/N$. Each offspring inherits its parent's phenotype and then is nudged by a mutation — an independent, unbiased Gaussian step of variance ${\sigma }^2$.

This is the cleanest possible starting point: no selection, no structure, no spatial geography, a single number describing the phenotype. "If the anomaly shows up here," the reasoning runs, "it is fundamental, not an artifact of some biological complication."

The observable is the first-passage time: the generation at which at least one individual's phenotype crosses a threshold $|X_i(t)|>L$, escaping the interval $[-L,L]$. That crossing is the first emergence of an extreme phenotype — the moment a population, exploring neutral genetic space while preserving its current function, produces something new enough to serve as raw material for adaptation. Establishment and fixation, the steps where a new trait actually takes over, come later; first passage is the precondition.

The key theoretical trick is to stop thinking of the population as a bag of independent random walkers and start thinking of it as a single fluctuating continuum. Hatakeyama writes down a stochastic density equation for the evolving distribution of phenotypes, in the spirit of the Dean–Kawasaki description used for Brownian particles. The density $\rho (x,t)$ evolves by diffusion plus a noise term whose strength is set by $1/\sqrt{N}$. Crucially, the noise is not locally conserved: in Wright–Fisher reproduction, individuals can appear and vanish locally through random variation in offspring number, whereas true Brownian particles can only move across boundaries. That non-conserved noise turns out to be the mathematical seed of everything that follows.

What They Found

The simulations reveal a three-regime structure. On short timescales — small target distances — the population behaves like a blob of independent walkers: mean first-passage time scales quadratically with distance, and the mean waiting time follows diffusion's inverse-variance law. This is the regime where the population simply spreads out from its starting point like an ink drop.

At intermediate distances, the behavior changes. The population has developed a "dilute edge" — a thin, barely-populated tail of pioneering individuals far ahead of the crowded bulk. And at this edge something extraordinary happens. The mean first-passage time scales as the cube of the target distance, $⟨T_L⟩\propto L^3$, which translates into the anomalous $({\sigma }^2{)}^{-3/2}$ dependence on mutation rate.

Mean first-passage time scales as (σ²)⁻³/² in the anomalous regime

Label	Value
Inverse-variance regime (low sigma²)	-1
Anomalous regime (intermediate sigma²)	-1.5
Inverse-variance regime (high sigma²)	-1

Scaling exponent on mutation variance, schematic from Figure 1(a).

The physical mechanism, as Hatakeyama unravels it, is elegant. First, shared ancestry gives the population's tail an exponential shape, not the Gaussian shape a single random walker would have. Two individuals in a current population trace back to a common ancestor about $N$ generations in the past on average; the mutations separating them accumulate along two independent branches, and averaging over the random coalescence time converts Gaussian mutation steps into an exponential separation distribution with decay length $\sigma \sqrt{N}$. Shared history — the fact that reproducing organisms are correlated through their genealogies — is what produces this heavy tail. Independent walkers could never do this.

Second, the growing density fluctuations at that dilute edge follow the Kardar–Parisi–Zhang equation, a famous piece of statistical physics originally developed to describe how vapor deposits onto a growing surface. KPZ fluctuations are famous for their anomalous $t^{1/3}$ growth law: the height of a growing interface wanders upward not as the square root of time, as ordinary diffusion would dictate, but as the cube root. Apply the Cole–Hopf transformation $h=\log \rho$ to the density equation, and the dilute edge of an evolving population becomes a one-dimensional KPZ surface. The log-density at the target must climb a height gap to reach the one-individual cutoff, and since KPZ heights climb as $t^{1/3}$, inverting that gives a cubic time-dependence.

The mathematics coheres. An exponential tail means the required log-density rise is linear in distance; KPZ fluctuations mean that rise accumulates as the cube root of time; combine the two and you get passage time $\propto L^3$. The amplitude of the cubic law, and both crossover positions, can be predicted self-consistently. The lower crossover marks where the tail transitions from Gaussian to exponential; the upper crossover marks the distance beyond which the whole population cloud — drifting as a diffusing center of mass — reaches the target faster than an edge excursion can.

A quadratic–cubic–quadratic law governs scaled first-passage time

Label	Value
L~ < L~-	2
L~- < L~ < L~+	3
L~ > L~+	2

Scaling exponent of <T~_L> on L~, from Figure 2(b). The cubic regime is bounded by the lower crossover L~- and upper crossover L~+.

The scaling collapse in Figure 2 is striking: rescale time by the population size $N$ and phenotypic distance by $\sigma \sqrt{N}$, and all seven population sizes fall onto a single master curve following the quadratic–cubic–quadratic sequence. The anomaly is not a property of some particular population size or mutation rate; it is a universal feature of the reproduction process itself.

The mean first-passage time, rescaled as $⟨{\tilde{T}}_L⟩=⟨T_L⟩/N$, obeys $⟨{\tilde{T}}_L⟩\simeq A{\tilde{L}}^3$ in the intermediate regime, where $\tilde{L}=L/(\sigma \sqrt{N})$. Unwrapping the units gives the headline result:

$$⟨T_L⟩\simeq A\frac{L^3}{{\sigma }^3\sqrt{N}}.$$

Holding population size and target distance fixed, this is the $({\sigma }^2{)}^{-3/2}$ law. Note the $\sqrt{N}$ in the denominator — larger populations find new phenotypes faster, but only as the square root of population size, another subtle departure from naive expectations.

Why This Changes Things

The result matters first because it overturns a default assumption in evolutionary theory. Neutral exploration has been treated as ordinary diffusion, in part because each individual lineage is an ordinary random walk. But a population is not a list of independent lineages; it is a web of shared ancestry, and that web changes the statistics of the collective. Reproduction introduces correlations that a single-walker picture cannot capture, and those correlations funnel through the dilute edge to produce anomalous first-passage scaling. The assumption that more identical random walkers just speed up arrival while preserving the inverse-variance law is simply wrong.

The practical consequence is that the pace of evolutionary innovation is more sensitive to mutation rate than anyone thought. In the intermediate regime, a population with weaker mutation waits not merely longer, but disproportionately longer, to produce its first extreme phenotype. This has implications for how we think about the evolution of new functions from neutral standing variation, the exploration of genotype space that precedes adaptation, and even the timescales over which antibiotic resistance or novel metabolic capabilities might appear. Estimates of phenotype discovery times that average over local mutational neighborhoods, as some network-based models do, could systematically miss the cost of searching distant parts of genotype space.

The anomaly also persists under weak stabilizing and disruptive selection — selection that gently favors either intermediate or extreme phenotypes. Empirically this means the three-regime structure is not a fragile artifact of perfect neutrality; it survives real biological conditions. Stabilizing selection lengthens the waiting time; disruptive selection shortens it; both preserve the anomalous intermediate law. The effect is largest where the whole population cloud drifts toward the target, and smallest during initial Gaussian spreading.

There is also a broader intellectual payoff. KPZ fluctuations are a recurring theme in modern statistical physics, appearing in growing interfaces, directed polymers in random media, and now — unexpectedly — at the edge of an evolving population. Mathematicians have previously predicted KPZ-type fluctuations in first-arrival times for particles correlated by a shared, time-dependent random environment. But in those models the mean first-passage time does not become anomalous; only the fluctuations do. Hatakeyama's framework suggests a fresh principle: whenever a dilute tail with correlated trajectories experiences KPZ growth, the mean waiting time can become anomalous. The edge-KPZ mechanism may be a generic route to nontrivial first-passage statistics across a whole class of correlated diffusive systems, not just evolving populations.

What's Next

The most immediate open question is experimental. The edge-KPZ theory makes concrete, testable predictions: measure first-passage times in a population with tunable mutation rates, and the anomalous $({\sigma }^2{)}^{-3/2}$ scaling should appear, with crossover positions determined by population size and target distance. Microbial evolution experiments, with their fast generations and controllable mutation rates, seem the natural arena.

A deeper challenge is generalizing to genotype networks. Real populations do not explore a smooth one-dimensional phenotype axis; they move through discrete networks of genotypes, where a given phenotype may be reachable through many nearby genetic routes. Hatakeyama notes that averaging over local mutational neighborhoods can miss the time required to search distant parts of a network — a population may need to cross a different region entirely before a target phenotype becomes accessible. Earlier simulations of neutral populations in finite, discrete genotype spaces did find an anomalous dependence of discovery time on mutation rate, and the edge-KPZ framework may be the right language to explain them. The difficulty is identifying the analog of a "dilute edge" on a genotype network and determining whether its fluctuations admit an effective KPZ description.

There is also unfinished business within the theory itself. The coefficient $A$ in the cubic law is not derived from first principles; it is inferred from the lower crossover and verified against the upper one. A fully closed theory would predict the amplitude of KPZ height excursions directly. The edge-KPZ equation is also a heuristic reduction — the noise amplitude is approximated by its value near the one-individual cutoff, and the Itô corrections are absorbed into a constant drift. Deriving the edge theory rigorously from the microscopic Wright–Fisher process would put the anomaly on firmer footing.

The most exciting implication is also the most speculative. If correlated dilute tails produce anomalous first passage generally, then the search for something new — a phenotype, a mutation, a polymer configuration, a new state in a complex system — is governed by physics that collective behavior imposes, not just by the underlying single-particle dynamics. Evolution, it seems, has a collective speed that no individual random walk can predict. The anomaly is not a correction to diffusion; it is a different kind of motion altogether, one where the history shared by a population, and the wild fluctuations at its leading edge, set the real pace of discovery.