When the center stops paying: how competitors split into niches
A market can have one obvious center of gravity—and still end up with competitors scattered across several specialized niches. In a mathematical model built to describe everything from job seekers to species and political candidates, the switch depends on how broadly each competitor can reach. When influence is broad, players converge on the resource-weighted center. When influence narrows past a threshold, the center becomes unstable: clusters split, competitors chase one another in cycles, and smaller contests emerge inside the larger one.
The paper calls this framework the influencer’s game. Its central insight is not simply that competition produces specialization. It is that the same resource landscape can generate convergence or fragmentation depending on the interaction between resource shape, competitor count, and the reach of each player’s influence. In the paper’s motivating example, job postings occupy a two-dimensional space defined by technology and finance relevance. A job seeker’s “position” is a skill profile; the jobs they can plausibly pursue depend on how close that profile is to the traits employers demand.
That is a useful way to think about a familiar dilemma. A broad skill set may make someone eligible for many opportunities, but also put them in competition with many others. A narrow profile may offer less coverage but help someone stand out in a specialized corner. The model does not predict actual hiring decisions or tell an individual what to study. It offers a general theory of when pressure to compete can turn a shared center into a collection of niches.
![Figure 1: The influencer’s game: a unifying framework for niche competition.
a, Three agents (i, j, k) choose strategic positions in a resource trait space 𝔹 over which a resource distribution B(b) is defined (e.g. consumer demand, voter support, or ecological resources).
b, Proportional allocation: each agent s=i,j,k receives a share Gs⋅B of the resource B(b) proportional to its relative influence (∑sGs=1), yielding a constant-sum payoff across all players ∑s∈Nus(𝐱)=u.
c, Agents adapt strategic positions via a learning mechanism.
d, Motivating example: LinkedIn job postings (2023–2024) scored on [0,1] for technology and finance relevance. Job seekers compete for postings by choosing strategic positions in the resource trait space.
e, Each job seeker exerts a multivariate Gaussian influence fi(xi,b) determining their proportional share of job opportunities via Gi(𝐱,b); proximity to a resource trait yields greater influence and more opportunities.
f, Arrows show how job seekers adjust strategic positions to maximize expected opportunities via adaptive dynamics or reinforcement learning.](https://arxiv.org/html/2611.17148v1/fig1_scaled.png)
The Science
Mark Lovett and Feng Fu, mathematicians at Dartmouth College, define a game in which each of several players chooses a position in a “trait space.” That space could represent job skills, consumer preferences, political ideology, or an ecological trait such as food preference. A resource distribution describes where opportunities or resources lie within it: voter support across ideologies, for example, or job openings across skill categories.
Each player exerts influence that is strongest near their chosen position and weakens farther away. At any point in the trait space, players divide the available resource in proportion to their influence there. A player’s total payoff is the sum of their shares across the whole landscape. They then adjust their positions to improve their payoffs, either by following the local gradient—moving in the direction that increases payoff—or by learning through trial and error.
This allocation rule matters. Unlike a model in which every person chooses the nearest provider and abruptly switches allegiance at a boundary, influence here can fade smoothly with distance. A job seeker might be a plausible match for several kinds of postings, but a stronger match for some than others. That makes it possible to study broad and overlapping competition without requiring all resources to belong exclusively to one player.
The authors first establish that, under appropriate conditions, there is a unique symmetric Nash equilibrium: a position where all players choose the same strategy, and no individual can improve their payoff by moving alone. The relevant condition is strict log-concavity of the influence kernel—a mathematical way of saying that influence has a single, smoothly declining peak. Gaussian influence, the familiar bell curve, is one example.
But having a unique symmetric equilibrium does not mean players will actually settle there. The key question is stability: if players are nudged away from the shared position, do their incentives pull them back, or push them farther apart? The paper derives a threshold that balances two forces. Resource “scatter” measures how strongly the resource landscape pulls players toward different directions; kernel curvature measures how sharply each player’s influence falls away from their chosen position. When the resource-driven push to separate overcomes the stabilizing curvature, the shared equilibrium loses stability.
This analysis works in any number of trait dimensions. The authors show that symmetry lets them reduce the stability calculation to two matrix problems whose size depends on the number of traits, not the number of players. One mode moves all players together; the other moves them apart. The first does not destabilize the equilibrium in their setting. Fragmentation begins through the second.
What They Found
For Gaussian influence, the threshold has a particularly clear interpretation. If a player’s competitive reach is broad enough relative to the spread of resources, the symmetric equilibrium is stable. If reach falls below a critical value, the equilibrium destabilizes. In one dimension, the threshold is the resource distribution’s standard deviation multiplied by a factor determined by the number of players. More players make stability harder to maintain; with just two players, the destabilizing factor is zero, so the symmetric equilibrium remains locally stable under the paper’s assumptions.
That two-player result is a striking limit. Even when the resource landscape has multiple peaks, two symmetric competitors using log-concave influence do not spontaneously break away from their common equilibrium through the modeled gradient dynamics. The instability requires at least three players—or asymmetries outside the model’s symmetric setting.
The paper then follows the dynamics as reach narrows. In a one-dimensional example with three players and two resource peaks, broad influence leads all three to the same central position. As influence becomes more localized, they first spread into distinct positions. At still narrower reach, they can enter a persistent cycle: each competitor’s move changes the incentives facing the others, creating an endless chase rather than a settled arrangement. Narrowing reach further produces asymmetric clusters, such as two players sharing one region while the third occupies another.
The sequence is not just a set of possible endpoints. It is a cascade of bifurcations: qualitative changes in a system’s behavior as a parameter crosses a threshold. The initial split from the symmetric state is a pitchfork bifurcation, where a stable central state gives way to separated alternatives. Later, a Hopf bifurcation creates a limit cycle, a repeating orbit in which players continue to move rather than settle. In the paper’s three-player example, the second transition occurs at competitive reach \(\sigma_1=0.158\); the authors report that 97% of 10,000 simulated starting conditions cycle at reach 0.125. At reach 0.05, every run in that experiment reaches one of the two asymmetric cluster arrangements.
Three-player outcomes as competitive reach narrows
Reported simulation outcomes at the stated reach values for the three-player bimodal-resource example. Values are percentages of 10,000 random initial conditions.
| Label | Value |
|---|---|
| Reach 0.25: symmetric equilibrium | 100 percent |
| Reach 0.18: dispersed equilibrium | 100 percent |
| Reach 0.125: limit cycles | 97 percent |
| Reach 0.05: asymmetric cluster equilibria | 100 percent |
The number of players changes the richness of the picture. In a six-player game, clusters can split into arrangements such as three-and-three, four-and-two, or five-and-one. Different initial positions can lead to different outcomes even when the resource distribution and influence reach are identical. The paper reports that, for one six-player setting, a three-and-three arrangement occurred in 60% of 20,000 initial conditions at reach 0.18; at reach 0.10 it occurred in 40%, with four-and-two arrangements accounting for the rest. These percentages describe simulated starting conditions in that specific model, not the odds of real-world job seekers ending up in those groups.
A particularly suggestive pattern is hierarchical splitting. A cluster of competitors can behave like a smaller game within the larger one. As influence becomes more localized, a large group may divide into smaller groups, which can then split again. In this sense, the system’s organization repeats at different scales: the same kinds of competitive dynamics recur inside a niche that first formed as part of a broader market.
The authors demonstrate the framework with LinkedIn job postings from 2023–2024. A language model scored postings for technology and finance relevance, placing each in a two-dimensional trait space. The density of postings across that space becomes the resource landscape. In simulations with 1,000 players, narrowing Gaussian influence produces increasingly distinct clusters around areas with many postings. The figure illustrates the model’s application; it is not evidence that actual job seekers behave this way or that the measured skill dimensions capture the full labor market.
Six-player outcomes at two competitive-reach values
Frequency of reported equilibrium configurations among 20,000 random initial conditions at alpha 0.5.
| Label | Value |
|---|---|
| Reach 0.18: (3,3) configuration | 60 percent |
| Reach 0.10: (3,3) configuration | 40 percent |
The results are not confined to smooth, continuous movement. The authors also test independent Q-learning, a reinforcement-learning method in which agents update choices based on rewards from experience, while observing only partial information and choosing from discretized strategies. The broad sequence of stable shared positions, clustering, and cycles persists. They also extend the analysis to other influence kernels and to strategy spaces shaped like a simplex—a constrained space where trait shares sum to one. In the simplex example, six players move from one shared cluster to four-and-two and eventually three two-player clusters as reach narrows.
Why This Changes Things
The influencer’s game connects several old ideas without claiming that they are identical. In classic spatial competition, broad competitive reach can lead rivals toward a center: the principle of minimum differentiation. In the opposite limit, where influence becomes extremely local, the paper relates a two-player Gaussian model with a symmetric unimodal resource distribution to winner-takes-all competition at the median, echoing the median voter theorem. The new framework places these patterns in one family and shows how multi-player systems can move between them.
That bridge matters because many real competitive settings are neither purely “everyone crowds the center” nor “everyone finds a separate niche.” The transition depends on how concentrated the opportunities are, how many competitors share them, and how portable each competitor’s influence is across the landscape. A political candidate with broad appeal may be pulled toward a central bloc; several candidates with narrower appeal may divide the electorate into more specialized constituencies. A species that can exploit many food types may overlap with others, while narrower foraging strategies may partition resources.
For labor markets, the framework offers a way to ask a sharper question than “Which skills are in demand?” It asks how a changing landscape of opportunities interacts with the breadth of competing skill profiles. If opportunities cluster into distinct regions and competitors’ effective reach becomes narrow, the model predicts stronger incentives for specialization. Yet the result is not a prescription to specialize: the model does not include training costs, wages, employer screening, changing skill acquisition, or the possibility that people create new demand. Its value is as a map of forces, not a career guide.
The paper also complicates a simple story about efficiency. Clustering around resource peaks could make competition more responsive to where opportunities actually exist. But cycles and multiple stable outcomes mean that players may not settle into one predictable allocation. Starting positions matter. In real systems, that could mean historical accidents, early platform choices, or initial distributions of skills have lasting effects. The model identifies this path dependence mathematically; it does not measure how large that effect is in any particular institution.
One of the deeper contributions is the way it separates shared movement from separation. Players can all be pulled toward the same resource center, or they can respond to the landscape by differentiating from one another. The stability threshold compares the strength of that separating incentive to the way influence changes around a player’s current position. This gives researchers a portable diagnostic: before simulating a huge population, they can ask whether the symmetric state should hold or whether it is vulnerable to fragmentation.
Job-market resource landscape: simulated reach settings
Competitive reach values used for the paper’s 1,000-player illustration based on LinkedIn postings scored for technology and finance relevance. These are parameter settings, not measured outcome values.
| Label | Value |
|---|---|
| Broadest illustrated reach | 0.45 |
| Reach setting | 0.24 |
| Reach setting | 0.2 |
| Reach setting | 0.19 |
| Narrowest illustrated reach | 0.1 |
What’s Next
The empirical illustration is a starting point, not a field validation. Job postings were scored by a language model on just two dimensions, and the resulting landscape depends on how those dimensions are defined and measured. A richer labor-market analysis would need to test alternative skill taxonomies, validate the text scores, account for geography and wages, and compare predicted clustering with observed career paths and hiring outcomes.
The idealized dynamics also leave out important sources of realism. Players are treated as having a common influence kernel in the core analysis, while real people and organizations differ in resources, information, constraints, and goals. Learning rules can be noisy; the paper’s reinforcement-learning experiments show robustness to some limits on observation and choice, but not every feature of human adaptation. Nor does a mathematically stable strategy necessarily correspond to a socially desirable one.
A next step is to use the model alongside real data in settings where positions, resources, and outcomes can be observed over time. In ecology, that could mean testing whether changes in resource concentration precede niche splitting. In online platforms, it could mean asking whether creators cluster around content topics as audience demand and recommendation systems shift. In politics, it could help distinguish convergence driven by broad constituencies from fragmentation driven by specialized blocs. Each application would require its own careful definitions and validation.
The larger promise is a shared language for a recurring problem: when does competition make everyone chase the same center, and when does it push them apart? Lovett and Fu’s answer is conditional, not universal. Broad reach and a stable resource landscape can hold competitors together; concentrated influence, many players, and a sufficiently uneven resource landscape can unravel that arrangement. The most useful next question is not whether specialization is inevitable, but which real-world forces move a system across that threshold—and whether people can shape those forces before a cascade begins.