When animals count brief flashes or clicks to decide which side had more, their errors grow in proportion to the count. The spread of their internal tally grows linearly with the number of pulses, not with its square root. This is Weber's law for number, called scalar variability. It contradicts the textbook model of decision-making, in which each pulse adds its own independent noise. It also sits awkwardly beside single-trial recordings in monkey parietal cortex, where the population decision signal spreads like ordinary diffusion. A separate literature on single neurons shows that a spike threshold turns steady, additive membrane fluctuations into firing variability that grows with the mean. We propose a specific way these two facts connect. The link does not run through per-pulse noise, which the threshold can make larger but can never make scale with the count. It runs through slow fluctuations in a sensory neuron's resting membrane potential. A threshold converts a slow fluctuation into a random gain that holds for the whole trial and multiplies every pulse alike. Summed over pulses, a shared gain yields a spread proportional to the count.
The idea matters because it makes a prediction that the competing accounts do not. The competitors are multiplicative top-down feedback, generic trial-to-trial variation in accumulation rate, and phenomenological scalar-noise models. Our prediction is that the Weber fraction of accumulated evidence should fall as the physical intensity of each pulse rises, following a form fixed by the threshold's power-law exponent. The same account says the scalar component should vanish when net evidence is zero. That offers a concrete reconciliation of the rat counting data with the monkey diffusion data.
Core Claim
Amplification of Trial-to-Trial Response Variability by Neurons in Visual Cortex modeled firing as a thresholded power of membrane potential, $R = k[V - V_{\text{thresh}}]_+^n$. Here $R$ is firing rate, $V$ is membrane potential, $V_{\text{thresh}}$ is the spike threshold, $k$ is a scale factor, $n$ is the exponent, and $[\cdot]_+$ keeps only the positive part. In their sample cell, firing-rate variance grew nearly 100-fold while membrane variance grew less than 4-fold. That analysis used 5–20 ms counting windows, and the paper lists behaviorally relevant timescales as unresolved.
Our observation is simple. If each pulse's noise is independent, a larger or mean-dependent per-pulse variance still gives total variance proportional to the number of pulses $N$, so the spread still grows as $\sqrt{N}$. Scalar variability needs a noise source that is shared across the pulses of a trial.
Suppose a sensory neuron's baseline potential carries a slow fluctuation $\xi$ with standard deviation $\sigma_\xi$ that is roughly constant within a trial. Each pulse of intensity $c$ adds a drive $\mu(c)$. Linearizing the power law around the operating point gives a trial gain whose coefficient of variation (standard deviation over mean) is approximately
$$\mathrm{CV}_g(c) \approx \frac{n\,\sigma_\xi}{\mu(c) - V_{\text{thresh}}}.$$
This follows from differentiating $(\mu+\xi-V_{\text{thresh}})^n$ with respect to $\xi$. If $m$ is the mean response per pulse and $\sigma_\epsilon^2$ is the independent per-pulse variance, the variance of the summed evidence is
$$\mathrm{Var}(S_N) \approx \mathrm{CV}_g(c)^2\, m^2 N^2 + N\sigma_\epsilon^2.$$
This one expression does three things. It gives sub-Weber behavior at small $N$, where the independent term dominates, and Weber behavior at large $N$. That is the transition reported in the towers task by Amplitude modulations of sensory responses, and deviations from Weber's Law in pulsatile evidence accumulation. Unlike any count-only account, it also makes the Weber fraction a decreasing function of pulse intensity, with a steepness set by $n$. At the highest intensities we expect a further drop, because the recurrent saturation described in Circuit-motivated generalized affine models characterize stimulus-dependent visual cortical shared variability pulls shared variance back toward an additive regime.
A second consequence is that a gain multiplies signal. If one gain acts on both evidence streams, the scalar term scales with the net evidence. It is therefore absent at zero net evidence and grows with signal strength squared. This suggests why Direct observation of the neural computations underlying a single decision saw variance proportional to time in parietal area LIP, and it makes that finding a testable constraint rather than a contradiction.
Relation to Prior Work
Three papers motivate the proposal by documenting departures from additive diffusion. Sources of noise during accumulation of evidence in unrestrained and voluntarily head-restrained rats fit more than 525,000 rat trials. They found spread linear in flash count, best captured by two accumulators with scalar noise. Their noise is phenomenological: they explicitly leave open whether it comes from encoding thresholds, feedback, or a variable internal clock. The towers paper found evidence- and choice-dependent gain in about half of cue-locked V1 neurons. It explained the sub-Weber-to-Weber transition with a feedback loop in which noise grows with the accumulated tally. That loop is our main mechanistic competitor. A divisive model of evidence accumulation explains uneven weighting of evidence over time showed that divisive inhibition with two time constants produces flat, primacy, recency, and bump weighting kernels, meaning curves of how much each moment of evidence sways the choice. It analyzed the mean kernel, not the noise, so its gain-control pool is a natural place where slow shared fluctuation could live.
Two papers constrain the design. Temporal integration is a robust feature of perceptual decisions showed across humans, monkeys, and rats that choices follow a true sum, and that large pulses do not count disproportionately. A gain model must therefore not smuggle in extrema-like weighting. The Steinemann LIP recordings constrain the regime in which the scalar term can be large.
The method comes from the single-neuron threshold analysis above and from the affine variance-versus-mean fitting of the V1–V2 work. We reuse both, but at trial timescales rather than millisecond windows.
Research Program
The work has three parts.
The first is modeling. We will build a two-stream accumulator whose encoder units pass a slow baseline fluctuation, fast per-pulse noise, and pulse-intensity drive through a power-law threshold. An optional saturating recurrent stage follows. We will derive the likelihood of choices given the full pulse sequence, including per-pulse intensity. We will fit variants in which the slow fluctuation is shared across the two streams or independent per stream. If it is shared, the scalar term scales with net evidence. If it is independent, it scales with each stream's own total.
The second is behavioral. We will run this model on existing pulse datasets, then add one new manipulation: pulse intensity varied across trials at fixed counts. Flash luminance serves in the rat visual task and click level in the auditory task.
The third is neural. In the towers calcium imaging data, we will decompose each cue-locked response amplitude into two parts. One is a random effect for the whole trial, which is our shared gain. The other is a dependence on the evidence accumulated before that cue, which is the feedback account's signature. We will then ask which part predicts choice beyond cue counts. We will also test whether the trial-level part's variability shrinks when pre-cue baseline activity is higher, as the threshold formula requires.
Baselines
The naive reference is the additive accumulator in which every pulse contributes independent Gaussian noise. We will fit it as the Brunton-style model that A divisive model of evidence accumulation explains uneven weighting of evidence over time used as its comparison standard. It predicts spread growing as $\sqrt{N}$ and no dependence of the Weber fraction on intensity. Any candidate must beat it clearly, or the effect is not there.
The strongest phenomenological competitor is the dual-accumulator model with scalar variability from Sources of noise during accumulation of evidence in unrestrained and voluntarily head-restrained rats. It already fits count data well. That paper reports 70–91% correct but no likelihood values; no reported result found on Lacuna for its model-comparison scores.
The strongest mechanistic competitor is the multiplicative feedback model from Amplitude modulations of sensory responses, and deviations from Weber's Law in pulsatile evidence accumulation. No reported result found on Lacuna for its fit likelihoods either.
We will add a drift-variability model in which the accumulation rate varies randomly from trial to trial with a constant coefficient of variation. It is mathematically identical to our model at a single intensity. This serves as the ablation that removes the threshold. The only ingredient separating it from our model is the dependence of gain variability on the operating point.
All models will be fit to the same trials by maximum likelihood, with identical lapse and trial-history terms, since Scott et al. found history biases lasting three trials. Every model gets the same number of random restarts. Comparison will use cross-validated held-out log-likelihood per trial, alongside information criteria. Beating a baseline means a held-out likelihood gain that is consistent in sign across animals. It also means correctly predicting, out of sample, the Weber fraction at intensities withheld from fitting.
Why not just…?
Why not just add trial-to-trial drift variability to a standard diffusion model? This is the real threat. At fixed pulse intensity it is exactly our model, and it would explain the existing count data equally well. The difference lies entirely in the intensity manipulation. Constant-CV drift variability predicts a Weber fraction that does not change with intensity. The threshold account predicts one that falls roughly as $1/(\mu(c)-V_{\text{thresh}})$. If the new behavioral data show a flat Weber fraction across intensities, our specific mechanism is wrong, and the generic account wins. We will report that outcome as such.
Why not just accept the multiplicative feedback model? It reproduces the same sub-Weber-to-Weber curve, so count-only data cannot decide between the two. It differs in timing. Under feedback, a pulse's noise depends on how much evidence has accumulated when it arrives. Under a trial gain, early and late pulses are equally noisy and noise does not depend on history. Two tests separate them. The first is behavioral: sequences are matched on total counts but differ in when evidence arrived. The second is neural: the trial-level versus history-dependent decomposition of V1 cue responses in the towers data. The towers paper's own finding that V1 gain tracks current evidence is a point in feedback's favor. Both mechanisms could coexist, and the decomposition estimates each one's share rather than assuming either.
Why not just use divisive normalization, which already bends the temporal weighting kernel? Divisive normalization divides signal and independent noise alike. Its deterministic form therefore changes the mean kernel reported in A divisive model of evidence accumulation explains uneven weighting of evidence over time without producing spread linear in count. If its inhibitory pool fluctuates slowly, however, it becomes one implementation of a shared trial gain. We will treat it as a candidate location for $\xi$, not a rival. A shared gain predicts the scalar term scales with net evidence; per-stream gains predict it scales with each stream's total. That test says whether the gain acts on both streams jointly, as a pooled normalizer would, or on each separately.
Decisive Checks
The first check is the intensity sweep in rats. Counts are fixed at several levels, and flash luminance or click level spans roughly five steps from near-detection to strong. We estimate the scalar coefficient $\mathrm{CV}_g(c)$ separately at each intensity. The threshold account passes only if $\mathrm{CV}_g$ decreases with intensity with a shape fitted by a single exponent $n$ and offset, and flattens or drops further at the top. The per-sample intensity variation in the rat auditory task of Temporal integration is a robust feature of perceptual decisions gives a first look before new data are collected.
The second check is the zero-evidence prediction. In Scott-style data, spread at fixed total count should be smallest when left and right counts are equal if the gain is shared. It should be independent of the difference if gains are per stream. The same logic, applied to LIP population traces conditioned on motion strength, predicts a variance component growing with time squared only at nonzero coherence. That is a reanalysis we would request from the Steinemann recordings.
The third check is the towers decomposition. A significant trial-level gain whose variance shrinks with pre-cue baseline supports the threshold account. A purely history-dependent gain supports feedback.
If all three checks fail, scalar variability does not arise at the sensory encoding stage, which narrows the search to accumulator- or clock-level sources.
References
- Amplification of Trial-to-Trial Response Variability by Neurons in Visual Cortex
- Amplitude modulations of sensory responses, and deviations from Weber’s Law in pulsatile evidence accumulation
- Circuit-motivated generalized affine models characterize stimulus-dependent visual cortical shared variability
- Direct observation of the neural computations underlying a single decision
- Sources of noise during accumulation of evidence in unrestrained and voluntarily head-restrained rats
- A divisive model of evidence accumulation explains uneven weighting of evidence over time
- Temporal integration is a robust feature of perceptual decisions