Ad platforms run auctions on their own predictions: a bid is a value per click multiplied by the model's predicted click probability. These predictions are noisy, and the noise is not harmless to the seller. Auction revenue can fall when inputs change in ways that should help. Adding bidders, raising valuations, or refining a prediction model can all lower revenue. This note argues that the bias noise causes in revenue can be computed exactly, and that it is the same quantity a classical edge detector computes. The auction's welfare function has creases, located where the winning allocation switches. Feeding the auction noisy values is the same as blurring that function with a Gaussian. The resulting revenue bias equals an integrated Laplacian-of-Gaussian response, which is the Marr–Hildreth edge detector applied to the creases.
The practical consequence is that a platform can build an "edge map" of where in prediction space its revenue is fragile to noise. It can then denoise its predictions selectively: shrink estimates hard near allocation boundaries and leave them sharp elsewhere. This is the structure-aware denoising that edge detection learned to prefer over isotropic blurring. The open question this answers comes from Model Monotonicity in Autobidding Auctions. That paper shows that refining prediction models can cut revenue by 6.2% under VCG with target-cost-per-acquisition bidders, 16.8% under budgeted first-price auctions, and 66% in a MAX-CPA setting. It assumes perfectly calibrated predictions and explicitly leaves noisy, uncalibrated models unmodeled.
Core Claim
Take the Vickrey–Clarke–Groves (VCG) auction, in which each winner pays the welfare loss they impose on everyone else. Let $W(v)$ be the maximum total welfare given the value vector $v$, and let $W_{-i}(v)$ be the maximum welfare with bidder $i$ removed. Total VCG revenue is then $R(v) = \sum_i W_{-i}(v) - (n-1)W(v)$, where $n$ is the number of bidders. In unit-demand matching markets each $W$ is a maximum over allocations, so it is convex and piecewise linear. Its curvature is concentrated entirely on the boundaries where the optimal assignment changes.
Suppose the mechanism instead sees $v + \Sigma^{1/2} z$, where $z$ is standard Gaussian noise and $\Sigma$ is the covariance of the prediction errors. Expected revenue is then $R$ smoothed by a Gaussian kernel. Because Gaussian smoothing solves the heat equation, the bias has an exact form:
$$\mathbb{E}[R(v+\Sigma^{1/2}z)] - R(v) = \tfrac12\int_0^1 \mathrm{tr}\!\left(\Sigma\, \nabla^2 (G_{t\Sigma} * R)\right)(v)\, dt .$$
Here $G_{t\Sigma}$ is a Gaussian with covariance $t\Sigma$, and $\nabla^2$ is the matrix of second derivatives. The integrand is an anisotropic Laplacian-of-Gaussian response evaluated on the welfare creases.
Each term $W$ and $W_{-i}$ contributes a nonnegative Jensen gap, but the full-market welfare enters with weight $-(n-1)$. The sign of the bias therefore depends on which crease set dominates near the operating point. The full market has more competing allocations than any market with one bidder removed. We therefore conjecture that noise biases revenue downward in thin markets near marginal conflicts. That is exactly the regime of Hartline's examples in Revenue Non-monotonicity in Matching Markets, where a small change resolves a conflict and VCG revenue collapses from \$1 to \$0.
The conjecture has a concrete reading. Noise acts like a spurious edge: it randomly breaks ties between competing allocations, and every broken tie erases an externality. Model refinement feeds this effect, because finer user segments rest on less data and so have larger entries in $\Sigma$. Refinement therefore raises revenue through better targeting while also lowering it through a noise term we can now compute. The second term is the one the calibrated analysis cannot see.
Research Program
The first strand is theory. We will make the conjecture precise for unit-demand matching and for single-parameter settings. The goal is to characterize when the bias is negative and how it scales with the noise covariance and with distance to the nearest allocation boundary. For small noise the bias is approximately $\tfrac12\mathrm{tr}(\Sigma \nabla^2 R)$, which is what the heat-equation expansion gives at leading order. Because $\nabla^2 R$ is a measure supported on creases, the bias should decay quickly with distance from the nearest boundary, roughly like a Gaussian tail in that distance divided by the noise scale.
Two existing results constrain this strand. Revenue Monotonicity Under Misspecified Bidders shows that matroid feasibility protects revenue against misspecified bidders and that non-matroid constraints permit unbounded losses. We will test whether the same boundary governs the sign and size of the noise bias. On Revenue Monotonicity in Combinatorial Auctions proves approximate monotonicity for XOS valuations, which suggests the bias should be bounded in that class. Neither paper models stochastic prediction error. Both supply the structural conditions we will test against.
The second strand turns the identity into a diagnostic. On a live or logged market, the platform can estimate the edge response by Monte Carlo. It perturbs predicted values with the estimated error covariance, re-clears the auction, and averages. This is cheap because the same assignment solver is reused. The output is a per-auction and per-segment "fragility score": the estimated revenue bias caused by prediction noise. This is the auction analogue of the significance maps in Multiscale Edge Detection and Fiber Enhancement Using Differences of Oriented Means. That work tests elongated filters at multiple scales and accepts fine structure only when coarser support confirms it.
The third strand is the intervention, which we call boundary-aware shrinkage. The platform has two estimates for each user segment: its fine-grained prediction and its coarser parent cluster's prediction. Instead of shrinking every fine estimate toward its parent by a fixed empirical-Bayes amount, it chooses the shrinkage scale per auction. Shrinkage is strong where the fragility score is large and negative, and weak where the market is far from any allocation boundary.
This intervention changes only the platform's own predicted click probabilities, never the values advertisers report. Its incentive effects therefore pass only through how autobidders respond to predictions. That is the same channel Model Monotonicity in Autobidding Auctions already studies, and we will re-solve autobidder equilibria rather than assume them fixed. The analogy to anisotropic diffusion is literal. Smoothing stops at genuine structure, which here means confident, well-separated predictions. It proceeds where apparent structure is likely noise, which here means fine splits whose differences are small relative to their standard errors near a crease.
A fourth, smaller strand asks whether randomized or soft mechanisms already perform this smoothing implicitly. In Learning Revenue-Maximizing Auctions With Differentiable Matching, the matching is computed by the Sinkhorn algorithm, whose entropy parameter acts as a smoothing scale. The stepwise mechanisms of Stepwise randomized combinatorial auctions achieve revenue monotonicity recover monotonicity through randomization. Our identity predicts that both should show a smaller noise bias, at a cost in sharpness. We will measure that trade-off directly.
Baselines
The naive reference point is to deploy the refined prediction model as is, with no denoising. This is the setting whose calibrated version produced the 6.2%, 16.8%, and 66% revenue drops in Model Monotonicity in Autobidding Auctions. A second naive point is to deploy the coarse model only, which forgoes refinement entirely. Any useful method must beat both on revenue at no loss in advertiser welfare.
The strongest competitor on this specific question is uniform hierarchical empirical-Bayes shrinkage. Each fine segment's estimate is pulled toward its parent cluster in proportion to its estimated variance, independent of the auction. This is the textbook answer to noisy fine-grained estimates, and it already reduces the noise covariance everywhere. It will receive the same data splits, the same variance estimates, and the same hyperparameter search budget as our method. Its global shrinkage strength will be tuned directly for revenue on a validation split, which gives it every advantage.
Beating it means two things. First, higher revenue at matched prediction log-loss on held-out clicks. Second, revenue that stays monotone, or degrades less, as the refinement depth increases.
A second strong competitor is RobustBid. It derives closed-form robust bids under mean-squared-error uncertainty sets and reports higher conversion volume and lower cost per click than non-robust bidders on Synthetic, iPinYou, and BAT data. Its specific table values are not available in the file. We will run it as the bidder policy so that we can ask whether bidder-side robustness alone removes the platform-side bias.
A third competitor is the learned artificial competition of Increasing Revenue in Efficient Combinatorial Auctions by Learning to Generate Artificial Competition, together with plain reserve prices, both tuned on the same validation data.
The ablation that isolates the key ingredient keeps our shrinkage machinery but sets the shrinkage scale from a random permutation of the fragility scores. A second ablation sets the scale from prediction variance alone. If either ablation matches the full method, then knowing where the allocation boundaries are does not matter, and the central claim fails.
For the edge-map diagnostic itself, the baseline is the gradient-magnitude analogue: flag auctions with small bid gaps between the top competitors. This mirrors the comparison in Multiscale Edge Detection and Fiber Enhancement Using Differences of Oriented Means, where multiscale gradient magnitude scored an F-measure of 0.58 against 0.61 for the structure-aware method on BSDS.
Why not just…?
Why not just apply uniform empirical-Bayes shrinkage to fine segments? This is the real threat. If noise bias were spread evenly across prediction space, uniform shrinkage would capture nearly all the gain. Our identity says the bias is concentrated on allocation boundaries, so uniform shrinkage must either over-smooth the many auctions far from any boundary, losing targeting value, or under-smooth the few near one. The matched-log-loss comparison settles this. If uniform shrinkage tuned for revenue comes within noise of boundary-aware shrinkage on iPinYou and on the counterexample instances, we will report that the edge view is explanatory but not actionable.
Why not just set reserve prices? Reserves raise the revenue floor, but they act on levels, not on the noise-induced erasure of externalities between two competitive bidders. Balcan et al. note that reserves can leave items unsold and sacrifice welfare. Their own fix requires relaxing individual rationality and assuming independent valuations. We will measure the residual noise bias under tuned reserves. If that bias shrinks to zero, reserves suffice and our intervention is unnecessary.
Why not just let bidders be robust, as RobustBid does? RobustBid protects each advertiser against its own uncertainty about conversions. The revenue bias we identify arises when the mechanism clears noisy predicted values, and it persists even when every bidder bids optimally. Running RobustBid bidders under the naive platform model directly tests whether the bias survives.
Why not just randomize the mechanism, as stepwise auctions do? Randomization is a form of smoothing, and our identity predicts it helps. The stepwise construction, however, requires knowing each single-minded bidder's desired bundle, and its guaranteed step size $\delta = 1/(n^2\gamma)$ shrinks with the number of bidders. Here $\gamma$ is the step parameter that sets how finely the winning probabilities are discretized. We will compare its revenue and welfare against ours on small combinatorial instances where it is computable.
Proposed Experiments
The first check verifies the identity. On synthetic unit-demand matching markets built around Hartline's instances, we will inject Gaussian noise of increasing scale into the values. We will compare the empirical revenue bias against the Monte Carlo edge-response prediction and against the small-noise approximation $\tfrac12\mathrm{tr}(\Sigma\nabla^2 R)$. We will also test whether the bias concentrates within a few noise-widths of allocation boundaries.
The second check evaluates the diagnostic as an edge detector. Ground-truth "fragile" auctions are those where re-running the auction with independently resampled prediction noise lowers revenue beyond a threshold. We will score predicted fragility maps with precision, recall, and F-measure, borrowing the BSDS protocol, against the bid-gap baseline.
The third check targets refinement. On the counterexample instances from Model Monotonicity in Autobidding Auctions, we will add sampling noise proportional to segment size. The question is whether the reported losses grow under noise and whether boundary-aware shrinkage recovers them.
The fourth check uses real data. We will use iPinYou, where logistic-regression click models reach an AUC as high as 0.988 for some campaigns, and BAT. On these logs we will build coarse-to-fine segment hierarchies, simulate VCG and budgeted first-price clearing with autobidders re-equilibrated, and report platform revenue, advertiser welfare, and held-out log-loss. There is no reported result on Lacuna for platform revenue under noisy predicted click rates on iPinYou or BAT, so the uniform-shrinkage baseline defines the bar.
The final check runs the Sinkhorn mechanism across entropy values on its 2-agent, 15-item benchmark, where it reports revenue 0.571 and regret 0.003. The test is whether lower noise bias trades off against revenue in the way our identity predicts.
The decisive result would be a revenue curve that rises monotonically as segments are refined under boundary-aware shrinkage, while staying flat or falling under uniform shrinkage at matched log-loss.
References
- Model Monotonicity in Autobidding Auctions: When Do Better Predictions Lead to Better Outcomes?
- Revenue Non-monotonicity in Matching Markets
- Revenue Monotonicity Under Misspecified Bidders
- On Revenue Monotonicity in Combinatorial Auctions
- Multiscale Edge Detection and Fiber Enhancement Using Differences of Oriented Means
- Learning Revenue-Maximizing Auctions With Differentiable Matching
- Stepwise randomized combinatorial auctions achieve revenue monotonicity
- Robust autobidding for noisy conversion prediction models
- Increasing Revenue in Efficient Combinatorial Auctions by Learning to Generate Artificial Competition
- Real-Time Bidding Benchmarking with iPinYou Dataset