Example Strict Value Comparison Breaks the Competitive Guarantee
example proved

Strict Value Comparison Breaks the Competitive Guarantee

Counterexample. If the sample-then-threshold auction uses strict value comparison $p < v$ as the acceptance rule (ignoring identities entirely), it is not $1/4$-competitive — even for $n = 2$ bidders with injective identities.

Construction

Fix $M > 0$ and set $n = 2$ with $v_0 = v_1 = M$ (two bidders with identical value). The sample-then-threshold rule observes the first bidder and sets threshold value $p = M$.

Welfare is zero on every arrival order

  • Bidder $0$ arrives first. Observed and rejected (observe phase). Threshold becomes $p = M$. Bidder $1$ arrives with value $M$; strict test $M < M$ is false. Rejected. Welfare $= 0$.

  • Bidder $1$ arrives first. Observed and rejected. Threshold becomes $p = M$. Bidder $0$ arrives with value $M$; $M < M$ is false. Rejected. Welfare $= 0$.

The competitive bound is violated

$$ \frac{1}{2!}\sum_{\sigma} \mathrm{welfare}(\sigma) = 0 \;<\; \frac{1}{4} \cdot M = \frac{1}{4} \cdot \max_i v_i. $$

Why it fails

Strict comparison cannot distinguish the observed bidder from a later bidder with the same value. In the sample-then-threshold skeleton, the threshold is set to the maximum observed value. Any later bidder whose value equals this maximum should be accepted — they are "as good as the best seen so far" — but strict comparison $p < v$ rejects them at the tie.

How lexicographic tie-breaking fixes this

The lexicographic acceptance rule ([[mechanism_design.auction.online.single_item_auction]]) tests $p < v_i \lor (p = v_i \land \bar{b} \le b_i)$, where $\bar{b}$ is the identity of the lex-max observed bidder. At the value tie $p = v_i = M$, acceptance falls to the identity comparison $\bar{b} \le b_i$. If the arriving bidder has a higher identity than the threshold, they clear — breaking the $M < M$ deadlock.

Comparison with weak comparison

Replacing strict comparison with weak comparison $p \le v$ (accept all ties) avoids this failure: with $v_0 = v_1 = M$, both arrival orders give welfare $M$.

However, weak comparison has its own failure mode on the needle profile $v = (M, 0, \ldots, 0)$: the threshold drops to $0$ and the test $0 \le 0$ accepts the first phase-2 arrival regardless of value, giving expected welfare only $(1/n) \cdot M$ ([[mechanism_design.auction.online.secretary_weak_comparison_needle]]).

Summary

Acceptance rule Equal-value welfare Needle welfare Competitive?
$p < v$ (strict) $0$ always No (this)
$p \le v$ (weak) $M$ $M/n$ No
lex $M$ $\ge M/4$ Yes

The lexicographic rule is the only one among these three that achieves a constant competitive ratio ([[mechanism_design.auction.online.secretary_quarter_competitive]]).

Remarks

Lean formalization

Strict comparison is modelled by StrictComparison.auction, which sets the identity component of the threshold to $\top$ — since $\top \le b$ fails for all $b < \top$, the acceptance condition degenerates to $p < v$. The theorem welfare_eq_zero proves welfare $= 0$ for $n = 2$, constant values, and any identities below $\top$.

References

  • [Roughgarden 2016, Lecture 2, Problem 2.1(c)] Tim Roughgarden, Twenty Lectures on Algorithmic Game Theory, Cambridge University Press. Analysis of the online auction (stated for distinct values, which avoids this failure).

Also in