The Elephant in Evolution's Boa: Why a Trait's 'Optimum' Is Really a Moving Compromise

A boa constrictor has swallowed an elephant, and its body bulges into the unmistakable outline of the prey inside. When Antoine de Saint-Exupéry drew that image in The Little Prince, he meant it as a test: could you see the elephant inside the hat, or did you just see a hat? The authors of this paper — evolutionary biologists across the University of Florida, Wake Forest, Kentucky, and Santo Tomás in Chile — borrow that exact image to make an argument about evolution. The "boa" is our assumption that natural selection pulls a trait toward a single tidy peak, like a ball rolling into a bowl. The "elephant" is what actually happens in nature: the peak isn't where we think it is, and sometimes there isn't even one peak at all.

The strange, wonderful claim at the heart of this paper is that the optimum a population evolves toward is not fixed — it can move, split, and reshape itself even when the underlying targets of selection never change at all. The optimum is a weighted average of the optima favored by survival and by reproduction, and the weights are simply the respective strengths of selection on each. Change those strengths — across a landscape, from one season to the next — and the whole adaptive surface contorts. The paper shows this can produce fitness landscapes that are lopsided, or double-peaked, or completely flat in the middle: shapes we thought were exotic anomalies but which may actually be the default in a fluctuating world.

This matters for something far bigger than a mathematical footnote. It reshapes how we interpret phylogenetic trees, how we forecast whether species can rescue themselves from climate change, and how we understand why so much of nature appears to sit stubbornly away from its supposed optima. It is the kind of paper that quietly reframes a field's null hypothesis.

The Science

To grasp what's new here, you need the half-century-old framework it builds on. Russell Lande's 1976 paper — the beating heart of modern macroevolution — showed something elegant: if you model fitness as a smooth Gaussian curve with a single optimum, then selection acts like a spring. The steeper and narrower the curve around the optimum, the harder the population is pulled toward the peak. This "spring" logic gave rise to the Ornstein-Uhlenbeck (OU) model, which has become the workhorse of evolutionary inference. Nearly every study that fits a trait's history to a phylogeny — whether tracing beak size in Darwin's finches or body mass across mammals — uses a descendant of Lande's model. It assumes a trait is pulled toward a fixed optimum with a constant pull strength, buffeted by random fluctuations.

The OU model is elegant, and it fits data well. But it rests on a hidden assumption the authors wanted to crack open: that there is a single Gaussian fitness function with a single optimum and a single width. Lande himself defended the Gaussian as a local approximation — "any smooth fitness function can be closely approximated by a Gaussian function in the vicinity of the optimum." The new paper's provocation is that real organisms don't experience one fitness function. They survive, and then they reproduce, and selection acts on each separately. An individual's lifetime fitness is the product of its probability of surviving and its expected reproduction. If the trait that maximizes survival differs from the trait that maximizes reproduction, something has to give.

The authors start with the deceptively simple case where both components are Gaussian. The product of two Gaussians is itself a Gaussian — but a shifted one. Here is the first revelation. The composite optimum is a weighted average of the survival optimum and the reproduction optimum:

$$z^{\ast }=\mathcal{P}{z}_F^{\star }+(1-\mathcal{P})z_S^{\star }$$

where the weight $\mathcal{P}$ is the fraction of total selection strength coming from reproduction. If reproduction imposes strong stabilizing selection and survival imposes weak selection, the effective optimum slides toward the reproduction optimum. If the ratio flips, it slides the other way. The weights aren't arbitrary — they're the relative strengths of stabilizing selection. This is a mechanistically grounded, testable statement about where a population's trait should sit, and it comes purely from the arithmetic of multiplying fitness components (Ponciano et al., 2026).

There's also a subtler, almost poignant consequence hiding in that product. When the survival and reproduction optima disagree, the maximal fitness itself is depressed by a factor between 0 and 1. Discordant optima cost the organism: no individual can be simultaneously perfect at surviving and perfect at reproducing, so the best anyone can do is a compromise that is strictly worse than either pure optimum.

What They Found

The real novelty arrives when the authors let the selection strengths vary — across space from habitat to habitat, or across time from generation to generation. This is where the Little Prince's boa makes its dramatic entrance.

Model the strength of stabilizing selection on reproduction, $\delta$, and on survival, $\gamma$, as random variables drawn from Gamma distributions, representing patches of habitat with different selection intensities. The composite fitness function is then an average over all those patches. And here's the payoff: even though every single component is Gaussian, the average of the resulting non-Gaussian composites is itself generally non-Gaussian. Multiply two Gaussians per patch, then average the products across patches, and the total is no longer a clean bell curve. This is not a mathematical trick — it's a biological statement. A population spread across a heterogeneous landscape experiences a fitness surface that is asymmetric, or bimodal, or flat-topped, even when every local patch is perfectly Gaussian.

The figure the authors lean on is their reproduction of the Saint-Exupéry drawing

. On the left, the boa with the elephant inside. On the right, a sweep of fitness functions produced by varying how much heterogeneity lives in reproduction selection versus survival selection. When one component dominates, the surface is a pronounced peak — the elephant standing tall. When heterogeneity is balanced across both components, the surface flattens, and the elephant is "lying down." The topology of the fitness landscape — the thing that governs the rate and direction of evolution — depends not just on where the optima are, but on how variable the selection strengths are across the environment.

The paper quantifies this with a concrete numeric example

: reproduction optimum at $z_F^{\star }=3$, survival optimum at $z_S^{\star }=10$, with selection-strength heterogeneities set by Gamma parameters $a_F=1,b_F=4,a_S=2,b_S=7$. Neither the survival nor the reproduction optimum is where the population actually ends up. The joint optimum lands at 8.969 — nearly two-thirds of the way from the reproduction optimum toward the survival optimum, pulled there by the relative strengths of the two selection regimes. Neither target alone explains where the trait sits (Ponciano et al., 2026).

The paper then plugs this machinery into a real dataset — the classic long-term study of Chamaecrista fasciculata (partridge pea) by Etterson and Shaw, in which leaf traits were tracked against survival and fecundity across environments

. Fitting a maximum-likelihood model under the heterogeneous-dual-fitness framework, the estimated fecundity optimum sits at log leaf thickness $-0.7301$, while the survival optimum sits at $-0.01167$, and the composite optimum lands between them at $-0.7237$. The reproduction and survival optima genuinely disagree by a wide margin, and the overall optimum is a compromise weighted by selection strengths — exactly the pattern the model predicts (Ponciano et al., 2026). The empirical world, it turns out, carries the elephant inside the boa.

Where the composite optimum lands vs. each component's target

Label	Value
Reproduction optimum	-0.7301
Composite optimum	-0.7237
Survival optimum	-0.01167

Data: Ponciano et al. (2026), fitted from Etterson & Shaw (2001).

The deeper structural claim can be seen in how the two components stack up. Reproduction's optimum and survival's optimum diverge sharply, yet the composite fitness at its peak is dragged well away from both — a direct demonstration that the relative strengths of selection, not just the targets, determine where a trait evolves.

This framework also licenses the authors to construct, for the first time, a family of stochastic differential equation models of trait evolution that go beyond the classic OU process. Lande's derivation shows that a Gaussian fitness surface yields OU dynamics — a linear spring pulling toward a fixed optimum. Insert the non-Gaussian heterogeneous fitness surface instead, and the resulting diffusion process is no longer a simple spring. Different kinds of heterogeneity produce different families of stochastic models — a genuinely new toolbox for macroevolutionary inference. Rather than assuming the classic form and fitting constants, researchers can now ask which kind of heterogeneity in selection best explains an observed trait trajectory.

Why This Changes Things

The shift here is a change in what we treat as the null model. The field has long leaned on the OU process because it's tractable and because Lande justified it from first principles. But as the authors note, empirical syntheses routinely find that the strength of selection varies across climatic gradients, predator regimes, resource levels, and years. A growing catalog of field studies reports inter-annual shifts in both the strength and form of selection. If selection strength wanders, this paper shows, the entire geometry of the adaptive surface — its shape, its peak location, even the number of peaks — wanders with it, even when the targets stay put (Ponciano et al., 2026).

That insight has uncomfortable implications for how we've been reading the fossil record and phylogenetic trees. Comparative methods routinely fit OU models to trait data and interpret the fitted optimum as "where selection is pulling this lineage." If the optimum we estimate is actually a width-weighted compromise that has been shifting because selection strengths fluctuate, then estimates of past optima may be systematically misleading. The authors flag the statistical identifiability problem explicitly: with finite data, it can be hard to tell apart a model with a fixed optimum and lots of noise from a model with a fluctuating optimum and less noise. This is a real, honest caveat — the kind of intellectual honesty that makes the paper's claims more trustworthy rather than less (Ponciano et al., 2026).

There is also a profound upside to the non-Gaussian surfaces this framework generates. Evolutionary biologists have long fretted about the "paradox" that organisms so rarely sit at their apparent optima, attributing the gap to maladaptive constraint or bad luck. This paper offers a cleaner answer: what looks like suboptimality may be the correct optimum of a composite, heterogeneous fitness function. Species distributed across heterogeneous environments aren't failing to reach a single peak; they're averaging over many peaks carved by variable selection strengths.

The most forward-looking implication is for evolutionary rescue — the hope that a population confronted with rapid environmental change can evolve fast enough to avoid extinction. The authors point out that current rescue theory is built on fixed fitness-landscape geometry. But if the landscape itself reshapes along the climate gradient — as the strength of thermal selection tightens or loosens — then the rate at which a population can track a moving optimum is governed not just by heritability and population size but by how sharply selection strengthens across the new conditions. A landscape that flattens as the climate shifts buys a threatened population more time; one that sharpens and pulls away can collapse it. The framework opens the door to modeling rescue under explicitly dynamic landscape geometry (Ponciano et al., 2026).

Estimated selection-strength heterogeneity for each fitness component

Label	Value
Reproduction selection strength	0.4986
Survival selection strength	0.4746

Shape parameters a_F and a_S estimated via maximum likelihood (Ponciano et al., 2026).

What's Next

The authors are careful to flag what remains open, and the list is genuinely rich. First is the identifiability problem: disentangling a fixed-optimum, noisy-history model from a fluctuating-width, cleaner-history model with limited data is genuinely hard, and the paper treats this as a frontier rather than a solved problem (Ponciano et al., 2026). Second, the current construction assumes fecundity is a constant given survival and reproduction. Relaxing that — letting $\beta$ itself depend on the trait — is an obvious and promising next step that the authors flag explicitly.

The paper also carves out space for testing alternative mechanisms that generate non-Gaussian fitness. Inferential tools that distinguish "heterogeneous selection strength" from "genuinely non-Gaussian selection" from "maladaptation" would let empirical workers finally adjudicate among these competing explanations with data. And the new family of stochastic differential equation models means that the OU process is no longer the only game in town — future phylogenetic comparative analyses can now fit models where the shape of the landscape, not just the position of its peak, evolves.

There is a deeper, almost philosophical point hiding in the title's whimsy. The Little Prince's puzzle was about perception: the adult saw a hat, the child saw the elephant. For fifty years, evolutionary biology has looked at fitness data and mostly seen the hat — the tidy Gaussian, the single peak, the well-behaved OU spring. This paper argues that what we've been looking at is really a boa constrictor containing an elephant, and that the shape we refused to see — bulky, irregular, alive — is not a pathology but the natural signature of a world in flux. The boat has been steady for half a century; the elephant, the authors suggest, was there all along.

That reframing is the paper's deepest contribution. Not a new equation for its own sake, but a new lens: the fitness landscape is not a static sculpture to be discovered, but a living surface whose peaks wander, flatten, and split as selection itself breathes in and out across time and space. Understanding that breathing is now not just possible, but newly tractable. And for a planet pushing species hard against rapidly shifting selection gradients, that understanding could not arrive a moment too soon.