IE 598 GTO · Spring 2026 · Nick Trapp

When Approximate Solvers Break Incentive Compatibility

A profitable bid manipulation against a time-limited VCG mechanism — and the branch-and-bound anchor effect that drives it.
VCG is strategyproof only when the auctioneer solves the winner-determination problem exactly. Real auctions don't — the WDP is NP-hard, so solvers run under a time limit. This study shows that a bidder who simply overbids by 5×, without observing any competitor, turns a profit 72–75% of the time when the solver runs at 5–10% of the time it needs to be exact — and that the exploit vanishes once the solver is given a quarter of that time.
72–75%
profitable at 5–10% solve time
+11.8
mean utility gain (×5 overbid)
~25%
solve time where it collapses

The finding: a manipulation window that slams shut

Each bidder commits to a fixed overbid multiplier before the auction, knowing the distribution of competitors but never their realized bids. We sweep the solver's time limit as a fraction of the mean exact solve time and measure, across 100 seeds, how often overbidding beats truthful bidding.

Overbid ×5 Overbid ×2 Overbid ×1.25
Left: fraction of 100 instances where overbidding beat truthful bidding. Right: mean utility gain over truthful. The manipulation lives entirely in the shaded window where the solver runs at under ~20% of the time it needs to be exact; past 25% it is a net loss. Common-value instances; x-axis is time limit ÷ mean exact solve time.

Under exact solving the picture inverts, exactly as theory predicts: overbidding is unprofitable (0% at unlimited time), confirming VCG is incentive-compatible when the solver reaches proven optimality. The exploit is a pure artifact of stopping early.

Why it happens: the anchor effect

The VCG payment charges a winner the welfare the rest of the market loses by their presence: p = W(B₋ᵢ) − (W(B) − vᵢ). Overbidding pays off only if the solver underestimates W(B₋ᵢ) — the market's welfare without the bidder — more than it underestimates W(B). Under time pressure, it does, and there's a structural reason.

With focal bidder — W(B)

~4% below exact at 5% time limit

The 5× overbid is a high-value bid on a small bundle. Branch-and-bound grabs it into an early incumbent and spends the rest of its budget on the smaller leftover problem. The overbid anchors the search.

Without focal bidder — W(B₋ᵢ)

~6% below exact at 5% time limit

No anchor. The solver must find the best allocation among all ~1,600 competing bids from scratch, and with ~0.15 s it explores far fewer branches — returning a more suboptimal, lower welfare.

That ~2-point asymmetry — roughly 8 welfare units — flows straight through the payment formula. The bidder is charged for a market that the solver has quietly undervalued, so they pay far less than the incentive-compatible price.

The price of the exploit

For a 5× overbid, exact VCG would charge a payment of about 34 — more than the bidder's mean true value of ~29, which is precisely why honest mechanisms make overbidding unprofitable. Under a 5% time limit the charged payment collapses to about 14:

Payment @ 5% limit
≈ 14
Mean true value
≈ 29
IC payment (exact)
≈ 34

Winning a ~29-value bundle for a ~14 payment nets roughly +15 in utility. By 25% of exact solve time the payment recovers to ~33 and the gain is gone.

How it was measured

Instances60 items · 40 competing bidders · 40 XOR bids each (1,601 binary vars)
Value modelShared item values U[1,10]; 70% complementary / 30% substitutable synergy; common-value
SolverGurobi 13 MIP, fixed seed; exact solve ≈ 3.0 s mean
Time limitsλ · mean-exact-time, λ ∈ {5, 10, 25, 50, 75, 100}%
StrategiesFixed overbid ×{1.25, 2.0, 5.0}, committed before the auction
Seeds100 independent · 95% confidence intervals throughout
On novelty. That approximate winner determination can break VCG strategyproofness is known in general (Nisan & Ronen 2000; Lehmann, O'Callaghan & Shoham 2002). Prior work models adversarial or worst-case approximation; this study characterizes a specific, standard algorithmic implementation — Gurobi branch-and-bound under a wall-clock limit — and identifies the anchor mechanism behind it. To my knowledge that particular mechanism has not been characterized before. The anchor effect is a diagnostic account consistent with the data, not a formal proof.