← Blog/blog/thompson-tempering-dimension-tax

Cool the posterior, widen the search—and pay a dimension tax

Thompson sampling chooses an action by drawing one plausible model from the posterior and acting as if that draw were true. It is an elegant exploration rule—until the posterior becomes so concentrated that an optimistic draw is vanishingly unlikely in many dimensions.

Jaiswal, Pati, Bhattacharya, and Mallick explain a familiar theoretical fix as coherent Bayesian machinery. Raise every likelihood contribution to a power α below one. The posterior then learns more cautiously and keeps sampling unexplored possibilities. The paper's sharpest lesson is also its least plug-and-play one: α must shrink with dimension, and the proof only starts after a dimension-scale warm-up.

01

One exponent turns data into softer evidence

A fractional posterior replaces the usual likelihood L(θ) with L(θ)α. For Gaussian linear regression, this multiplies every observation's precision contribution by α. The same sufficient statistics are used; only their strength changes.

def fractional_gaussian_posterior(prior_mean, prior_precision, noise_variance, sum_x2, sum_xy, alpha):
    scaled_precision = alpha / noise_variance
    precision = prior_precision + scaled_precision * sum_x2
    mean = (prior_precision * prior_mean + scaled_precision * sum_xy) / precision
    variance = 1.0 / precision
    return mean, variance

In a weak-prior, large-sample regime, covariance is multiplied by 1/α and standard deviation by 1/√α. With d = 16, the recommended α is 0.0625: asymptotically 16× the covariance and 4× the posterior standard deviation.

ordinary posterior (α=1)tempered posterior (α=1/16)
Core-computed, seed-free Gaussian illustration with unit prior precision and unit noise. At d=16, tempering treats 80 physical observations as only 5 likelihood-equivalents. This is not a paper experiment.
02

Wider uncertainty means more optimistic draws

In a linear bandit, each action is a feature vector and expected reward is its dot product with an unknown coefficient vector. Regret is the reward lost by choosing something other than the oracle's best action. Thompson sampling explores only when its random coefficient draw makes an uncertain action look best.

P(draw > 0.3), α=1P(draw > 0.3), α=1/16
Core-computed synthetic mechanism check. Repeated unit-feature rewards average 0.2, while a competing action is fixed at 0.3. The tempered posterior keeps a larger probability of sampling above the competitor. No random draw or paper-reported number is used.

Tempering does not make the observed mean better. It makes the learner less certain that the mean is settled. That distinction matters: variance inflation buys exploration, not evidence.

03

Dimension sets the temperature—and the tax

Paper statementFormInterpretation
Tempered likelihoodL(θ)^α, 0 < α < 1Likelihood information is deliberately discounted
Posterior covarianceinflated by α⁻¹Posterior draws explore a wider parameter region
Dimension-aware choiceα = d⁻¹Keeps the optimism term from decaying exponentially in d
Upper boundO(d³ᐟ² √T log T)Matches the best known LinTS rate under stated conditions
Constructed lower boundΩ(d³ᐟ² √T)The extra √d is unavoidable for this algorithm class
Paper-reported theoretical results. Big-O and big-Ω suppress constants; they are not empirical regret measurements.
Core-computed asymptotic standard-deviation inflation under the paper's α=1/d choice. Exploration grows wider as the feature dimension grows; these bars visualize the formula, not measured performance.

Why not temper even more? The contraction term scales roughly as d/α, so choosing α = d−(1+ε) widens the upper bound to O(d3/2+ε/2 √T log T). Why not temper less? Terms involving αd reintroduce exponential dimension dependence. The proof's balance point is therefore α = 1/d, not a free tuning suggestion.

Proof conditionWhat it narrows
Warm-upt₁ = Ω(d), sufficiently large for the approximation error
Action spaceCompact subset of ℝᵈ; the analysis targets linear bandit geometry
PriorPositive, continuous, bounded density; sub-Gaussian tails elsewhere
Exponential familyVariance bounded above and away from zero on the parameter domain
Sub-Gaussian familyBounded noise plus additional smoothness of the error density
Not coveredMisspecification, non-stationarity, and infinite-dimensional models
Conditions and open directions stated in the paper. The exponential-family result excludes unbounded-variance regions such as unrestricted Poisson means.
04

What a practical reproduction should probe

  • Compare α = 1, 1/√d, and 1/d on the same seeded bandit instances, reporting early and late regret separately.
  • Vary the number and design of warm-up actions instead of treating t₁ = Ω(d) as a sufficient recipe.
  • Stress misspecified rewards and non-stationary coefficients, which the paper lists as open extensions.
  • Measure posterior-sampling cost for nonconjugate generalized linear models; the theorem does not make those samples computationally cheap.
  • Report simple regret and estimation quality as well as cumulative regret, because extra exploration can help one objective and hurt another.

The linear regression page gives the Gaussian update behind the exact core above, while logistic regression is the most familiar generalized-linear case. The durable insight is that posterior variance is part of the decision policy: cool the likelihood and the learner searches longer—but the right amount of cooling depends on geometry, not taste.

References

  1. Prateek Jaiswal, Debdeep Pati, Anirban Bhattacharya, Bani K. Mallick (2026). Posterior Tempering Explains Variance Inflation in Linear and Generalized Linear Thompson Sampling. arXiv preprint
  2. Shipra Agrawal, Navin Goyal (2013). Thompson Sampling for Contextual Bandits with Linear Payoffs. Proceedings of ICML, PMLR 28