Brains are noisy. Synapses misfire, connections drift, tuning curves overlap. The standard response to this observation is that large populations compensate — average over enough neurons and the noise cancels. Hendler, Segev, and Shamir (arXiv:2602.21758) show this compensation has a hard ceiling.
They analyze a simple question: how well can a downstream neuron discriminate between two stimuli, given that its synaptic connections to the encoding population are imprecise? The answer depends on a single parameter — the degree of synaptic coarse-tuning — and it falls into three regimes.
When coarse-tuning is weak (synapses are precise), signal-to-noise ratio scales linearly with population size. Standard neuroscience: more neurons, better discrimination. When coarse-tuning is moderate, scaling becomes sublinear — you still benefit from larger populations, but with diminishing returns. When coarse-tuning is strong — and this is the regime most consistent with measured biological heterogeneity — the signal-to-noise ratio saturates. Adding more neurons doesn't help. At all.
The mechanism is geometric. Under strong coarse-tuning, the effective readout is constrained to a low-dimensional manifold aligned with the naive decoder (simple population average). The optimal linear decoder — the one that uses the best possible weighting of all neurons — converges to the same answer as the naive one. When synapses are sloppy enough, there is no clever way to combine the signals that does better than averaging. The information bottleneck is in the connections, not the computation.
This has implications beyond neuroscience. Any system that reads out from a noisy population through imprecise connections faces the same constraint. The three regimes map to three engineering strategies: in the linear regime, scale up; in the sublinear regime, optimize your weights; in the saturated regime, fix your connections or accept the limit. No amount of downstream processing compensates for upstream imprecision past the threshold.
What makes this result surprising is not that noise limits performance — everyone knew that. It's that the limit is a wall, not a slope. There's a critical precision below which population coding simply stops improving. The ceiling isn't soft. It's a phase transition in the utility of more data.
Based on O. Hendler, R. Segev, and M. Shamir, "Limits of optimal decoding under synaptic coarse-tuning" (arXiv:2602.21758, February 2026).