The Shape of Selection: How Clipping a Population's Upper or Lower End Predicts Its Evolutionary Future

Fishermen have been unknowingly running a massive evolution experiment for centuries. By hauling in the biggest fish, they set up a selection pressure that favors smaller body sizes—one of the starkest examples of what biologists call "unnatural selection." The catch: nobody really understood how such repeated harvesting events reshape a population's traits over time, or whether we could deliberately use this principle to steer biological systems toward desired outcomes.

A paper published in July 2026 by physicists R. K. Singh, Michael Assaf, Jason R. Green, and Erez Aghion cracks open this question in a surprising way. Their work shows that the mathematics of probability—the same equations that describe how pollen grains jitter in water or how stock prices fluctuate—predict something remarkable about any system you periodically cull: depending on whether you remove individuals from the top or bottom of a trait distribution, the population will settle into one of two fundamentally different futures. Clip the larger end, and the population converges to a stable equilibrium that barely changes no matter how long you keep harvesting. Clip the smaller end, and the population locks into a fixed shape even as it drifts steadily toward ever-larger trait values, accelerating away from where it started.

Neither outcome requires any genetic modification. Neither requires changing the underlying rules that govern how individuals grow, diffuse, or compete. You simply remove some fraction of the population at regular intervals—and the mathematical structure of randomness does the rest.

The implications stretch well beyond fish populations. Any system with intrinsic variability and stochastic dynamics—cancer cell populations, antibiotic-resistant bacteria, investment portfolios, opinion distributions in a population—can in principle be harvested. And if these researchers are right, the same clean mathematics that governs a pollen grain's random walk might eventually help us design interventions that steer complex systems toward stability, or toward controlled growth, or away from catastrophic failure.

The Science

To understand what Singh and colleagues discovered, it helps to start with a question that sounds almost too simple to matter: what happens to a population of randomly moving individuals when you periodically remove some of them?

The researchers formalized this problem using the language of probability densities—the mathematics that describes how a population is distributed across different trait values at any given moment. If you're tracking body size in a fish population, for instance, the probability density tells you what fraction of fish fall into each size class. A population with most individuals clustered around medium sizes would have a bell-shaped density curve; one dominated by small individuals would be skewed left.

Without any intervention, such a density evolves according to the underlying stochastic process—the random rules that govern how individuals gain or lose the trait in question. For a fisherman measuring body size, this might involve growth rates, death rates, and the random fluctuations in nutrition and environment that affect individual fish differently. For a physicist watching a pollen grain jitter in water, it's the ceaseless bombardment of water molecules. The mathematics are the same.

Singh and colleagues focused on what they call "harvesting": the periodic removal of individuals above or below a fixed threshold trait value. They distinguished between two complementary protocols. Backward reshaping removes the upper portion of the population—everything above a threshold such as the median—and renormalizes the remaining density to sum to one. Forward reshaping does the opposite, clipping the lower portion and keeping only the individuals above the threshold.

The researchers modeled a range of stochastic processes to test whether their findings would generalize. They started with the simplest case: Brownian motion, the random walk that describes particles diffusing through a fluid. But they also examined more exotic dynamics—processes that spread faster or slower than ordinary diffusion, a phenomenon physicists call anomalous diffusion. Specifically, they looked at continuous-time random walks, where particles pause for random intervals between steps; run-and-tumble dynamics, which capture the dart-and-pause movement of bacteria; and Lévy walks, which feature long ballistic flights interspersed with pauses. All of these occur in biological systems.

To make the mathematics tractable, the researchers assumed a reflecting boundary at zero—a constraint that keeps trait values from going negative, as makes sense when tracking body sizes or population counts. They set up harvesting events at regular intervals of duration $\tau$, then tracked how the probability density evolved between cuts and transformed instantaneously at each cut.

The key insight, mathematically, was that the consecutive harvesting events break the usual associativity of the stochastic evolution. Without harvesting, a system evolving for time $t$ looks the same whether you calculate it all at once or in pieces—the order doesn't matter. But once you start clipping the density at intermediate intervals, the calculation becomes path-dependent. The system "remembers" the harvests. And this memory, the researchers found, drives it toward one of two attractors depending on which direction you reshape.

For numerical verification, the team ran extensive simulations—tracking millions of independent "random walkers" through hundreds of harvesting cycles and comparing the resulting density distributions against their analytical predictions. The match was precise.

What They Found

The results split cleanly into two cases, each with its own mathematical signature.

Backward Reshaping: Convergence to a Quasi-Steady State

When researchers removed the upper portion of the probability density—clipping above the median, for instance—the system behaved as follows: after enough harvesting cycles, the population stopped evolving in any meaningful sense. Not because it had died out, but because it had settled into a quasi-steady state: a probability distribution that looks identical at every harvesting tick, provided you observe it at those precise moments (the "harvesting clock").

This convergence was robust. Starting from a narrow peak or a broad distribution, with different initial conditions and different parameter values, the sequence of probability densities after each harvest converged to the same shape. The researchers proved, through a monotonicity argument, that this was inevitable: under backward reshaping, the sequence of densities must eventually intersect at every point in the domain, and the only admissible intersection is identical overlap.

For Brownian motion specifically, they derived the exact form of this quasi-steady state. It turns out to be a half-Gaussian—a Gaussian distribution folded at zero—characterized by a width that depends only on the diffusion coefficient $D$ and the harvesting interval $\tau$:

$${\sigma }_{\ast }\simeq 1.53\sqrt{D\tau }$$

The median of this steady-state distribution settles at approximately $0.731{\sigma }_{\ast }$. Notably, this value matches the 75th percentile of the original Gaussian, a constraint that emerges from requiring the distribution to be self-consistent after each cut-and-diffusion cycle.

Convergence to quasi-steady state: median approach

Label	Value
n = 1	1
n = 2	0.82
n = 3	0.76
n = 4	0.74
n = 5	0.735
n = 6	0.732

Normalized median m(nτ)/√(2Dτ) approaches ~0.731 after repeated backward reshaping events

Figure 2 from the paper illustrates this convergence beautifully. Panel (a) shows the median position of Brownian particles over time, normalized by $\sqrt{2D\tau }$. Different simulation parameters collapse onto the same curve, approaching a constant value after roughly six harvesting events. Panel (c) shows the actual probability densities at successive harvesting times, plotted one after another—starting from a broad initial distribution in purple and converging toward the predicted half-Gaussian steady state shown in black. The match between simulation and theory is exact.

The researchers emphasized that this quasi-steady state depends only on the harvesting threshold and frequency, not on any feature of the initial population. Two populations that start wildly different—say, one dominated by small individuals and one dominated by large individuals—will converge to identical quasi-steady states after enough backward harvests. This independence from initial conditions is a hallmark of attractor dynamics, and it means that a manager applying backward reshaping doesn't need to know the starting state to predict the long-term outcome.

Forward Reshaping: Fixed Shape, Constant Drift

The complementary protocol yielded a strikingly different result. When the researchers clipped the lower portion of the density—removing smaller individuals and keeping only those above the median—the population locked into a fixed shape even as it drifted steadily toward larger trait values.

The mechanism creates a feedback loop. Without harvesting, a diffusing population naturally expands outward from zero. Forward reshaping removes the individuals closer to zero, compressing the remaining density into the upper half of the distribution. After the population diffuses again, it has spread farther from zero than it would have without the harvest. The next forward reshaping removes even more of the lower portion, pushing the median rightward again. Each cycle accelerates the drift away from the origin.

Eventually, this feedback loop saturates. The variance of the distribution converges to a constant value, while the median advances at a constant rate. The shape stabilizes—even as the entire distribution marches toward infinity.

For Brownian motion, the researchers derived the asymptotic behavior:

$$\Delta m\equiv m((n+1)\tau )-m(n\tau )\stackrel{n\to \infty }{\to }\sqrt{2D\tau }$$

The median advances by a fixed increment $\sqrt{2D\tau }$ at each harvesting cycle, regardless of how many cycles have already occurred. Meanwhile, the variance converges to:

$${\sigma }^2((n+1)\tau )\stackrel{n\to \infty }{\to }\pi D\tau$$

Convergence to fixed shape: variance evolution

Label	Value
n = 1	0
n = 2	0.414
n = 3	0.665
n = 4	0.796
n = 5	0.864
n = 6	0.897

Variance σ²(nτ)/Dτ converges to π ≈ 3.14 after repeated forward reshaping

Figure 2(b) shows this constant-rate drift in action. The normalized median versus normalized time approaches a straight line—the signature of linear drift—after about five or six harvesting events, regardless of whether the harvesting interval $\tau$ is 1 or 2. Panel (d) shows the probability densities at successive harvesting times, shifted by the instantaneous mean to remove the drift. When viewed this way, the shapes collapse perfectly onto a single curve: the fixed-shape attractor of forward reshaping.

Universal convergence across anomalous diffusion types

Label	Value
Brownian	1.53
Run-and-tumble	1.71
CTRW	2.05
Lévy walk	1.38

Width parameter A = σ/√(Dτ) varies by dynamics but remains finite and predictable

Anomalous Dynamics: The Principle Holds

One of the paper's most striking claims is that these results aren't limited to ordinary Brownian motion. The researchers ran simulations of three other stochastic processes—run-and-tumble particles, continuous-time random walks, and Lévy walks—and found that the same qualitative behavior emerged.

In all cases, after sufficient forward reshaping cycles, the probability densities collapsed onto a single curve when shifted by the instantaneous mean. The shape was fixed; only the location varied. The details differed (the width of the steady distribution depends on the underlying dynamics), but the qualitative outcome was universal.

The researchers noted an interesting subtlety: for processes where the median grows subdiffusively—slower than $\sqrt{t}$—the convergence to a fixed shape requires longer diffusion steps between harvests. For subdiffusive continuous-time random walks with a power-law waiting time distribution, the time interval between harvests must be large enough that the distribution's bulk is well-described by a Gaussian. But under those conditions, the attractor behavior appears reliably.

Backward reshaping showed similar generality. The appendix demonstrates convergence to a quasi-steady state for run-and-tumble particles, CTRW, and Lévy walks, using the same shifting technique to remove any residual drift and show that the shapes collapse.

A Predator-Prey Application

To ground these abstract results in a concrete ecological context, the researchers extended their analysis to a stylized predator-prey model. Here, the trait $x$ represented predator body size, and larger predators were assumed to be more effective hunters. Prey abundance $y$ grew logistically but was consumed at a rate proportional to mean predator size $⟨x(t)⟩$.

Without intervention, predator size diffuses randomly, causing the mean size to drift upward as $⟨x(t)⟩\sim \sqrt{\sigma t}$. As predators grow larger, they eventually overexploit the prey, driving the system to collapse.

But the researchers showed that periodic backward harvesting—removing predators above the median size—could stabilize this dynamics. By repeatedly clipping the upper tail of the predator size distribution, the system was driven back toward the quasi-steady state. The prey population stabilized, surviving indefinitely despite the stochastic fluctuations in predator size.

Conversely, forward reshaping would accelerate the drift toward larger predator sizes, hastening the prey collapse. The two harvesting protocols yield opposite ecosystem outcomes from the same underlying mathematics.

Why This Changes Things

The paper's core insight is that external selection pressure—harvesting—can fundamentally reshape a population's trajectory without altering the underlying stochastic dynamics. This is a powerful idea because it decouples two things that are usually conflated: the rules of individual-level variation, and the aggregate behavior of the population.

For evolutionary biology, this reframes a long-standing puzzle. Evolutionary change is often treated as something that "emerges" from random mutation and selection operating on individuals. But Singh and colleagues show that the aggregate pattern of trait distributions under sustained selection pressure follows deterministic mathematical attractors—regardless of the details of mutation, recombination, or fitness landscapes. The attractor depends only on the harvesting protocol: which end of the distribution you trim, and how often.

This is both a simplification and a revelation. It means that predicting long-term evolutionary outcomes may be easier than we thought: you don't need to know the full genetic architecture or the precise fitness effects of each variant. You just need to know the harvesting regime and the basic statistics of the stochastic dynamics. The attractor will take over.

For conservation biology and fisheries management, the implications are immediate. Commercial fishing typically removes larger individuals—exactly the forward reshaping protocol. The paper predicts that under such a regime, fish populations should experience constant upward drift in body size distributions toward larger values, even after accounting for the removal itself. This is a known phenomenon: fishermen have documented evolutionary responses to size-selective harvesting for decades. But the new framework provides a clean mathematical prediction for the rate of this drift: $\sqrt{2D\tau }$ per harvesting cycle, where $D$ captures the intrinsic body-size dynamics and $\tau$ is the time between fishing seasons.

More importantly, the paper suggests that a shift to backward reshaping—protecting large individuals and selectively harvesting smaller ones—could drive populations toward a stable equilibrium rather than an ever-accelerating growth trajectory. The quasi-steady state for a given harvesting threshold and frequency is predictable and reproducible. Managers could in principle compute the target distribution and design harvesting quotas to approach it.

The predator-prey model makes the stakes concrete. In the unharvested system, stochastic drift in predator size eventually drives prey to extinction. Backward reshaping stabilizes the predator distribution, preventing this catastrophe. Forward reshaping accelerates it. The same intervention applied in opposite directions yields opposite ecosystem outcomes—a clean demonstration that harvesting is not merely a way to reduce population size, but a way to steer the entire system.

For population genetics, the framework connects to a rich literature on branching processes and fitness waves. The researchers noted that forward reshaping resembles the "cutoff-front mechanism" studied by Brunet and Derrida in the context of genetic traveling waves, and the "fitness waves" of evolutionary dynamics in growing populations. But here, the branching component is stripped away: diffusion supplies the spread, and truncation converts it into an effective drift. This simplification makes the mathematics tractable while preserving the essential dynamics.

For cancer biology, the implications are speculative but tantalizing. Tumor populations are stochastic, with individual cells varying in traits like proliferation rate, drug resistance, and metastatic potential. If clinicians could periodically remove a portion of the tumor—say, through surgery or a treatment that targets a specific subpopulation—the remaining cells might converge to one of the two attractors described in the paper. A backward reshaping protocol might drive the residual tumor toward a stable, less aggressive equilibrium. A forward reshaping protocol might accelerate it toward a more aggressive state. Either way, the mathematical framework could inform treatment sequencing and timing.

For economics and finance, the connection is less direct but conceptually rich. Stock returns, asset prices, and portfolio returns all exhibit stochastic dynamics with variance that grows over time. Periodic "harvesting"—rebalancing portfolios by selling winners or losers—might similarly steer these distributions toward predictable attractors. The framework might help explain why certain investment strategies exhibit long-term convergence to specific patterns, or why market interventions have the distributional consequences they do.

What's Next

The paper opens more questions than it closes, which is the mark of good science.

Most immediately, the researchers worked primarily with median-based thresholds. What happens when the harvesting threshold is set at a fixed percentile other than 50%, or at an absolute trait value? The paper includes some analysis of fixed-threshold reshaping, showing that backward reshaping at a fixed location also converges to a quasi-steady state. But the quantitative predictions—how the width of the steady distribution depends on the threshold—remain to be worked out for non-median cases.

The predator-prey model is deliberately toy. Real ecological systems involve spatial structure, age structure, density-dependent feedbacks, and non-Gaussian fluctuations. Whether the attractor dynamics persist in more realistic settings is an empirical question. Field experiments in controlled systems—microcosm studies with bacteria or rotifers, for instance—could test whether backward harvesting truly stabilizes predator-prey cycles as the model predicts.

The generalization to anomalous dynamics also warrants deeper analysis. The paper demonstrates convergence numerically for CTRW, run-and-tumble, and Lévy walks, but the analytical framework is developed primarily for Brownian motion. For processes where the variance grows as $t^{\alpha }$ for some $\alpha \ne 1$, the scaling of the attractor width with $D\tau$ may take different forms. A unified theory covering the full anomalous diffusion spectrum would be valuable.

Perhaps most ambitiously, the framework assumes that harvesting acts on a continuous probability density—that the population is large enough that removing individuals merely perturbs the distribution smoothly, without creating discrete effects. In small populations, stochastic extinction events and Allee effects might disrupt the attractor dynamics. Extending the theory to finite or sparse populations—where the "density" is literally a small number of individuals—would bridge the gap to conservation applications in endangered species.

The researchers themselves note several directions for future work: testing whether the results hold for processes with non-diffusive dynamics (ballistic motion, subdiffusion slower than any power law), incorporating density-dependent feedback into the stochastic dynamics, and extending the framework to multi-species communities where multiple traits interact.

There is also the broader question of what "harvesting" means beyond the biological contexts the paper emphasizes. The mathematical structure applies to any stochastic system with intrinsic noise and a lower bound. Opinion dynamics, cultural evolution, linguistic change, technological adoption—all of these involve distributions over trait space that evolve stochastically, and all could in principle be "harvested" by interventions that selectively remove individuals at one end of the distribution. Whether the framework helps understand such phenomena remains to be seen.

---

The deepest implication of the paper may be philosophical. Evolution is often framed as a tension between random variation and deterministic selection. But Singh and colleagues reveal an additional layer: the statistics of the distribution itself—the shape, the spread, the position relative to a threshold—can determine the long-term trajectory, independent of the genetic details. This is a form of "dynamical" selection that acts on probability distributions rather than genes.

It suggests that populations are more malleable than we might suppose. A manager who understands the harvesting attractors—who knows whether backward or forward reshaping is being applied, and at what frequency—can predict the long-term evolution even without detailed knowledge of the underlying genetics. This is a powerful tool for intervention, and a humbling reminder that the consequences of our actions on biological systems may be more predictable, and more consequential, than we appreciate.

Fishermen have been running this experiment for centuries, reshaping fish populations by pulling up the largest individuals. Now, thanks to this paper, we can see the mathematics that underlie their unwitting intervention. Whether we use that knowledge to sustain fisheries, stabilize ecosystems, or engineer biological systems toward desired outcomes is a question for the future. The mathematics, at least, is now clear.