Opening proposition
The idea in brief
Replicator models describe how the share of a type changes when types with higher payoffs become more common. If payoffs are uncertain, a better signal spreads the possible posterior payoffs farther apart while keeping their average fixed.
The replicator update is curved rather than linear. Below a population share of one half, that curvature rewards the extra spread created by better information. Above one half, the curvature bends the other way and the same increase in informativeness lowers the expected next share.
The surprising claim was not merely that information has costs. It was that information has no direction of its own: its expected evolutionary effect reverses at a precise geometric boundary.
Full exposition
Selection as a curved update
Replicator dynamics describe competition in terms of shares. A type that performs better than the population average gains representation; a type that performs worse loses it. The change is proportional both to performance and to the amount of population available to change. A type at zero cannot shrink further, and a type at one cannot grow further, so the update naturally bends near the boundaries.
That curvature matters whenever payoffs are uncertain. Suppose a signal does not change the average expected payoff but makes the possible posterior payoffs more dispersed: good news becomes more strongly good and bad news more strongly bad. If the population update were linear, those changes would cancel. In a curved update they do not. The average of the updated outcomes can differ from the update generated by the average payoff.
Why the sign turns at one half
Below a share of one half, the replicator response is locally convex in the payoff difference. Upside surprises produce a slightly larger gain than equally sized downside surprises produce a loss. More informative evidence therefore raises the expected next share. Above one half, the response is locally concave. The same spread of surprises now makes the downside dominate, lowering the expected next share.
The reversal is geometric rather than psychological. The signal has not become less accurate, and the type has not changed its preferences. What changes is the location from which uncertainty is converted into population movement. Rarity supplies more room for favourable variation to compound; dominance supplies more exposure for unfavourable variation to erode.
Information has a position
We often speak as though information carries an intrinsic quantity of value. A more revealing forecast is assumed to be better in the same direction for everyone. Information-Induced Reversal replaces that picture with a relational one. The effect of a signal belongs to the pair formed by the signal and the receiver’s current position in a nonlinear system.
The result is not that ignorance is preferable for dominant types. Better information can still improve long-run decisions, welfare or survival. The narrower claim is that its immediate expected effect on share can reverse. A population can become better informed while the presently common type loses ground, or remain poorly informed while incumbency is temporarily stabilised.
Where to look for the pattern
The cleanest demonstration would run the same stochastic payoff process twice, beginning once with the focal type at 25 percent and once at 75 percent. The more informative signal should push expected next-period share upward in the first experiment and downward in the second. Starting near one half should suppress the second-order effect.
Analogues may appear wherever uncertain advantages pass through a bounded, S-shaped update: market share, cultural adoption, organisational influence or ecological frequency. The practical lesson is to ask not only how much a signal reveals, but where the recipient sits on the response curve. Information policy that benefits an emerging minority may have a different immediate effect when that same group becomes the incumbent.
The minimal mathematical picture
Let a focal type occupy a share x of a population and let a payoff difference determine its next share. In many discrete replicator updates, the mapping from payoff difference to future share is bounded between zero and one and has an S-like shape. A more informative signal replaces one middling posterior estimate with a spread of more extreme posterior estimates while preserving their mean. The expected consequence then depends on the curvature of the update: averaging before applying the update is not equivalent to applying the update to each possible posterior and averaging afterward.
This is the role of Jensen’s inequality in plain language. Where the response curve bows upward, dispersion raises the average output; where it bows downward, dispersion lowers it. In the proposed update, the inflection occurs when the focal type occupies half the population. The same mean-preserving increase in information therefore has opposite second-order effects on either side of one half. The result does not require the signal to favour one type on average. It arises precisely when average expected advantage is held fixed.
What “better information” means here
Better information can mean many things: greater accuracy, lower noise, more detail, faster arrival or better calibration. The reversal concerns a narrower statistical relation. One signal is more informative when its posterior expectations are a mean-preserving spread of those generated by another signal. It sometimes produces more optimistic estimates and sometimes more pessimistic ones, but it does so without shifting their overall mean. The better signal is valuable because it distinguishes states that the coarser signal merges.
That definition is important because a signal whose average is more favourable would create an ordinary first-order advantage and could overwhelm the curvature effect. Likewise, a signal that changes behaviour strategically or costs resources introduces additional mechanisms. Information-Induced Reversal isolates the effect of resolution alone: two populations face the same underlying environments and the same mean payoff, but one population receives evidence that separates possible states more sharply.
Rarity and dominance are geometrically different
The threshold at one half is not a moral distinction between outsiders and incumbents. It marks a change in the geometry of a bounded share. A rare type has much more room to grow than to shrink in absolute population share, while a dominant type has much more room to shrink than to grow. The replicator rule converts payoff surprises through that asymmetric room. Extreme good news can accelerate the rare type into open space; extreme bad news can only push it toward a boundary it was already near. For the dominant type, the asymmetry is reversed.
This should not be confused with the familiar claim that small organisations are more agile or large organisations are more fragile. No difference in organisation, learning speed or risk preference is required. The types can be identical except for current frequency. Position itself changes the way volatility in assessed advantage becomes movement. That makes population share part of the informational environment rather than merely the outcome being predicted.
Short-run reversal is not long-run value
The result concerns an expected next-step change. It does not prove that dominant types should avoid information, nor that ignorance preserves dominance indefinitely. Better signals may improve survival, allocation and cumulative growth over longer horizons. They may allow a dominant type to correct weaknesses before a less informed rival could. Repeated updating can also move the population across the halfway point, causing the sign of the curvature effect to change during the process.
This separation prevents a seductive overreading. An intervention can reduce an incumbent’s expected immediate share while improving the accuracy of the whole system and the incumbent’s long-term prospects. Conversely, a temporary boost to a rare type need not imply lasting selection. The idea is most useful as a warning against assigning information a context-free directional effect. Time horizon, state variable and location on the response curve all belong in the claim.
Extensions and empirical signatures
A persuasive experiment would manipulate informativeness while matching posterior means, then repeat the manipulation on both sides of the frequency threshold. The distinctive signature is a three-part pattern: positive effect below one half, approximately zero second-order effect near one half, and negative effect above one half. Merely observing that information sometimes helps and sometimes hurts would be too weak, because costs, strategic adaptation and biased signals can produce that pattern for many reasons.
The same logic may extend beyond biological populations whenever a bounded share responds nonlinearly to uncertain relative performance. Candidate domains include market adoption, attention, political support and the allocation of organisational influence. The exact threshold need not always be one half; it will occur wherever the relevant response curve changes curvature. The general form is therefore portable even if the memorable fifty-percent boundary belongs to a particular update rule.
A thought experiment
Suppose two otherwise identical strategies receive a more revealing forecast about tomorrow’s environment. Starting at 25% of the population, the informed strategy is expected to gain more share. Starting at 75%, the same forecast can reduce its expected share.
Consequences
What the idea changes
We normally rank information as simply better or worse. This result suggests that the value of information can be a property of position: the identical signal is an advantage to an outsider and a disadvantage to an incumbent.
Ways to think with it
- It produced a clean sign reversal at a memorable threshold.
- The prediction could be tested by running the same process from 25% and 75% initial shares.
- It challenged the intuition that more informative evidence has a stable directional value.
Further reading
The intellectual neighbourhood
Notes on the idea’s provenance and editorial review are kept separately in the editorial appendix.