Choosing among many options is much harder, for people, swarms and algorithms alike, than choosing between two. Models that are tidy for two options become messy or intractable with three or more. In honeybee swarms, the signaling strength needed to break a tie among $N$ equally good nest sites grows like $N^2$, while the signaling that best picks out one superior site grows only like $N$ (Model of the best-of-N nest-site selection process in honeybees). Closed-form drift-diffusion approximations stop at single-boundary go/no-go decisions, because evidence for competing options is correlated (Approximating Drift-Diffusion Models for User Decisions under Nudging and External Information). This proposal borrows one move from physics. The renormalization group (a procedure that repeatedly averages away fine-scale detail and tracks how the parameters of the remaining coarse description must change to stay accurate) can be applied to the choice set itself.
The idea in two sentences: merge similar options into a single "effective option" and replace the merged group with one accumulator, correcting its drift, noise and inhibition parameters so that choice probabilities and decision times are preserved. Repeating the merge gives flow equations that describe how a decision problem with $N$ options maps onto one with $N/2$, $N/4$, and so on. Those equations should predict how accuracy, decision time and deadlock thresholds scale with $N$ from the two-option dynamics alone.
Core Claim
The claim has two parts. The first is that leaky competing accumulator models are approximately closed under a block-merging operation, provided the parameters are renormalized rather than simply averaged. These are models where each option has a noisy evidence tally that leaks, inhibits the other tallies, and triggers a choice when it crosses a threshold. The same family includes the gaze-dependent variant of Gaze-dependent evidence accumulation predicts multi-alternative risky choice behaviour and the mean-field swarm equations of Reina et al.
The second part is that the resulting flow reproduces known multi-option scaling laws as exponents. In particular it should recover the $O(N^2)$ deadlock-breaking and $O(N)$ value-sensitivity thresholds of the swarm model. The flow should also predict a schedule that removes the tension between them: decide coarsely at low signaling, then refine.
If the merged models do not reproduce full-$N$ behavior, or if a simple refit of a nested grouping does equally well, the claim fails. The experiments are built to show either outcome.
Mechanism
Each option $i$ has accumulator $x_i$ obeying
$$dx_i = \big(v_i - \lambda x_i - w\sum_{j\ne i} x_j\big)\,dt + \sigma\, dW_i,$$
The symbols are:
- $v_i$ is the option's drift, meaning its value or evidence strength.
- $\lambda$ is the leak.
- $w$ is the cross-inhibition between options.
- $\sigma$ is the noise level and $dW_i$ is independent Brownian noise.
A choice occurs when some $x_i$ first reaches threshold $a$.
Merging a block $B$ of $k$ similar options splits the block's dynamics into two parts. One is the block-mean mode, which is kept as the effective accumulator. The other is the set of within-block relative modes, which are integrated out. The block wins when its leading member crosses threshold, not its average member. Integrating out the relative modes therefore shifts the effective threshold downward, by roughly the expected maximum of $k$ correlated relative fluctuations. It also renormalizes the leak and the self-inhibition, because members inhibit each other. For a scale variable $\ell = \log(N_0/N)$, where $N_0$ is the original number of options and $N$ the current number, this yields flow equations
$$\frac{d\theta}{d\ell} = \beta(\theta),$$
Here $\theta = (\Delta v/\sigma,\; w/\lambda,\; a/\sigma)$ collects the dimensionless drift spread, inhibition-to-leak ratio and threshold-to-noise ratio.
The fixed points of this flow correspond to the regimes Reina et al. found by bifurcation analysis: deadlock, multistability and symmetry breaking. The swarm model's $N^2$ law then has a concrete reading. Under a halving of $N$, the critical inhibition ratio must flow by a factor of about four.
Physics supplies the method only at the level of the template. Match-and-run procedures in particle physics integrate out heavy fields at a threshold and resume running with matched couplings (Next-to-minimal SOFTSUSY). Wavelet and neural coarse-grainings show how to decimate degrees of freedom while tracking the retained couplings (Wavelet Conditional Renormalization Group, Machine Learning Renormalization Group for Statistical Physics). None of those works touches choice. The nearest decision-side physics analogy, a double-well landscape for two-alternative choice (A Discrete Formulation of Two Alternative Forced Choice Decision Dynamics Derived from a Double-Well Quantum Landscape), stays binary and has no notion of scale. Here the scale is the resolution at which alternatives are distinguished, not energy or space.
Research Program
The work has three stages, ordered from most to least controlled.
Stage 1: derive and check the decimation rule in simulation. We first derive the block-merging rule analytically for the accumulator above and for the Reina swarm equations, treating the leading-member threshold correction with extreme-value approximations. We then verify the rule numerically. Full-$N$ systems with $N$ from 2 to 64 are simulated as ground truth. Each is compared against its successively merged surrogates on three measures:
- choice probabilities;
- the full decision-time distribution, measured by KL divergence;
- deadlock probability.
The option sets vary drift spread and similarity structure: equal options, one best among many, and clustered options.
Stage 2: test the scaling predictions. We integrate the flow and read off how deadlock-breaking and value-sensitivity thresholds should scale with $N$. We then check whether the exponents 2 and 1 reported by Reina et al. are recovered. We also test a coarse-to-fine policy that the flow suggests: first choose among blocks with low signaling, then run signaling up within the winning block. Reina et al. conjectured a time-varying signaling ramp but gave no law for it, and this policy is meant to supply one.
Stage 3: apply one merge step to human data. The three-gamble eye-tracking dataset of Molter et al. provides the empirical test. The attraction effect there is produced by a decoy that closely resembles a target, so it is exactly the kind of near-duplicate a merge operates on. We ask whether merging the decoy into the target with renormalized parameters, instead of fitting a separate distance-dependent inhibition term, reproduces the context effects with fewer free parameters.
With $N=3$, this dataset tests only one merge step. That limitation is real: the file lists no large-$N$ human choice dataset, so the large-$N$ claims rest on simulation.
Baselines
Full-$N$ simulation. The reference point is simulation of the unmerged model, with likelihoods obtained by simulation. It defines ground truth in Stages 1 and 2 and sets the compute cost the merged model must undercut. A merged model beats it only by matching its choice probabilities within sampling error, and its decision-time KL divergence within a pre-set tolerance, at substantially lower cost.
Naive truncation. The do-nothing reference drops the weakest options, or averages the drifts of merged options while leaving the other parameters unchanged. This is the obvious way to shrink a choice set.
Ablation of the renormalization. The key ablation is identical block-merging with the threshold, noise and inhibition corrections switched off. If this ablation performs as well as the full rule, the renormalization contributes nothing and the claim is false.
Nested logit. The strongest structural competitor is a nested-logit grouping, a standard discrete-choice model that groups similar options into nests. It is refitted freely at every scale with the same data and tuning budget as our model. It captures similarity grouping without any dynamics. Our model beats it only by predicting decision-time distributions and $N$-scaling that nested logit cannot produce, while matching it on choice shares.
GLA on the empirical task. The strongest empirical baseline is the gaze-dependent leaky accumulator. On the risky-choice data it predicted 74.3% of choices, against 57.9% for multialternative decision field theory, and it fit best for 90% of participants (Gaze-dependent evidence accumulation predicts multi-alternative risky choice behaviour). The merged model is given the same observed gaze input and is evaluated with the same cross-validation. It counts as beating GLA if it matches 74.3% with fewer free parameters, or exceeds it at equal parameter count. No reported result for any coarse-grained accumulator on this dataset was found on Lacuna.
Learned coarse-graining. An off-the-shelf learned coarse-graining, in the spirit of Mutual Information, Neural Networks and the Renormalization Group, is trained to compress option sets. It is given an equal compute budget, and we test whether it generalizes across $N$ as well as the analytic flow does.
Why not just…?
Why not just fit a leaky competing accumulator directly at every $N$? For fitting a single dataset of fixed size, this works, and it is what GLA does. It gives no account of how parameters must change as $N$ grows, so it cannot predict a 32-option regime from 4-option behavior. Reina et al. show that this change is where the qualitative trouble lies. The Stage 1 comparison against full-$N$ simulation measures exactly what the flow adds beyond a direct fit.
Why not just use nested logit or similarity grouping? This is the most serious threat, because nested logit already merges similar options. It has no decision times, deadlock or inhibition, so it cannot express the $N^2$ versus $N$ asymmetry. If it matches our choice-share predictions on the three-gamble data, we will report that the gain is confined to dynamics and scaling. That would be a narrower but still testable claim.
Why not just train a neural predictor? In the go/no-go setting, theory-grounded diffusion beat neural temporal point processes on decision-time distributions, improving KL divergence by 42–44% (Approximating Drift-Diffusion Models for User Decisions under Nudging and External Information). Flexible models also overfit sparse decision signals, as the KuaiRec results show (KuaiRec: A Fully-observed Dataset and Insights for Evaluating Recommender Systems). The learned-coarse-graining baseline tests directly whether the same holds across $N$.
Why not just use a hand-tuned signaling ramp? A tuned ramp may equal our schedule at any single $N$. It must, however, be re-tuned for every $N$. We will compare our coarse-to-fine schedule, with no per-$N$ tuning, against a ramp optimized separately at each $N$.
Decisive Checks
Three results settle the question.
Surrogate accuracy. After three merge steps from $N=64$, the renormalized surrogate must keep choice-probability error and decision-time KL divergence within tolerance where the non-renormalized ablation fails. If it does not, the core closure claim is false.
Recovered exponents. The integrated flow must recover the deadlock exponent near 2 and the value-sensitivity exponent near 1 without being fitted to them. If it misses either, the flow is not capturing the multi-option structure.
Human data. On the three-gamble data, merging the decoy must reproduce the attraction effect and approach GLA's 74.3% with fewer parameters. Failure on the first two checks refutes the proposal. Failure on only the third confines it to collective and artificial choice systems.
References
- Model of the best-of-<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi>N</mml:mi></mml:math>nest-site selection process in honeybees
- Approximating Drift-Diffusion Models for User Decisions under Nudging and External Information
- Gaze-dependent evidence accumulation predicts multi-alternative risky choice behaviour
- Next-to-minimal SOFTSUSY
- Wavelet Conditional Renormalization Group
- Machine learning renormalization group for statistical physics
- A Discrete Formulation of Two Alternative Forced Choice Decision Dynamics Derived from a Double-Well Quantum Landscape
- Mutual Information, Neural Networks and the Renormalization Group
- KuaiRec: A Fully-observed Dataset and Insights for Evaluating Recommender Systems