---
title: 'Match-2 Elimination: Pairwise Removal Mechanisms'
url: https://www.emergentmind.com/topics/match-2-elimination
type: topic
---

# Match-2 Elimination: Pairwise Removal Mechanisms

Searching arXiv for the cited papers and closely related material on elimination and matching.
arXiv search query: 1107.2314 OR 2203.12011 OR 1208.4854 OR 1001.0529 OR 2210.08844 OR 2402.13366
“Match-2 Elimination” (Editor's term) is a convenient label for elimination problems in which pairing or removal in units of two is central. In the materials considered here, the term spans at least four technically distinct settings: destruction or survival of non-crossing perfect matchings in complete geometric graphs; tournament formats in which exactly two participants are eliminated per round; pair-removal dynamics in the solitaire game of Memory; and elimination procedures in anonymous games, voting games, and statistical curriculum learning that are explicitly related to pairing, matching, or repeated pruning. This suggests that the unifying theme is not a single canonical model, but a family of elimination mechanisms whose behavior depends on geometry, order of play, information structure, and stochastic uncertainty [1107.2314][2203.12011][1208.4854][1001.0529][2210.08844][2402.13366].

## 1. Geometric-graph formulation and the basic edge-removal threshold

In the geometric setting, a complete geometric graph on \(2n\) points is a set \(G\) of \(2n\) points in general position in the plane, with all possible straight-line edges drawn between them. A non-crossing perfect matching is a set of \(n\) non-crossing edges that covers all \(2n\) vertices, pairing every vertex with exactly one other. The associated edge-removal number \(h(G)\) is the maximum integer \(h\) so that removal of any set of \(h\) edges from \(G\) leaves a non-crossing perfect matching in the remaining graph [1107.2314].

For arbitrary configurations, the basic lower bound is tight: removing any set of \(n-1\) edges always leaves a perfect non-crossing matching, but there exist configurations where removing \(n\) edges can destroy all matchings. In the convex case, where all points lie on the convex hull, this tightness is realized, so the worst configuration satisfies \(h(G)=n-1\). The same paper also establishes that there exists a configuration where for any set of \(2n-2\) edges removed, a perfect matching remains; in that construction, \(2n-1\) points are placed in convex position plus one interior point appropriately, so \(h(G)=2n-2\) may be attainable for some graphs [1107.2314].

A concise summary of the geometric bounds given in the source is the following.

| Configuration | Max edges removable, always leaving a matching |
|---|---:|
| Convex position | \(n-1\) |
| General position | \([n-1,\,2n-2]\) |
| Special construction | \(2n-2\) |

These statements show that “match elimination” in the geometric sense is highly configuration-dependent. The same edge budget can be fatal in convex position and non-fatal in other embeddings, so the relevant combinatorial object is not merely the complete graph \(K_{2n}\), but the straight-line realization of that graph in the plane.

## 2. Convex-hull boundary size as a structural control parameter

The paper “Edge-Removal and Non-Crossing Perfect Matchings” isolates the boundary of the convex hull as a decisive parameter. If the convex hull of the \(2n\) points includes at most \(n+1\) points on its boundary, then any \(n\) edges can be removed and the resulting graph will always admit a non-crossing perfect matching. In the notation of the paper,
\[
\text{If } |\partial \operatorname{conv}(G)| \leq n+1 \implies h(G)\geq n.
\]
The result is sharp: examples are provided where \(n+1\) boundary points allow elimination of all matchings by removing \(n+1\) edges, but not with only \(n\) removable [1107.2314].

The same source emphasizes that any edge-removal set of size \(n\) that eliminates all matchings under these boundary-size conditions is highly constrained structurally. Most vertices have low degree in such a bad subgraph, and degrees \(\geq 2\) are restricted to boundary vertices. This is not just an extremal bound; it is a description of the geometry of obstruction.

A plausible implication is that convex-hull sparsity acts as a robustness certificate. When the boundary is “small” relative to \(n\), the configuration is more resistant to arbitrary edge deletions. By contrast, when many points lie on the hull, the threshold drops back toward the convex-position worst case.

## 3. Random point sets and asymptotic robustness of non-crossing matchings

The random version considers \(2n\) points drawn uniformly at random from a convex body \(\Gamma\). In this setting, the main theorem states that, with probability tending to \(1\) as \(n\to\infty\), at least
\[
h(G)\geq n+\frac{n}{3\log(2n)}
\]
edges can be removed and yet a non-crossing perfect matching will remain. Equivalently, for \(n\to\infty\), the probability that
\[
h(G)\leq n+\frac{n}{3\log(2n)}
\]
tends to \(0\); almost surely,
\[
h(G)>n+\frac{n}{3\log(2n)}.
\]
The paper describes this as showing that random configurations are much more robust than the extremal convex examples [1107.2314].

The proof is summarized in terms of geometric-probabilistic mechanisms: with high probability, no large subset of points lies on the convex hull, attempts to remove edges can only locally block options, and recursive Ham Sandwich divisions preserve the existence of non-crossing matchings unless many more edges are removed. The paper also notes an upper-bound connection through the expected number of convex hull points \(C(\Gamma,2n)\): by removing edges incident to boundary points, all non-crossing perfect matchings can often be blocked, and the expected boundary size governs this regime [1107.2314].

This random-case analysis corrects a common overgeneralization from convex-position examples. The worst-case lower bound \(n-1\) is sharp, but it is not typical. In random geometric graphs, destroying all non-crossing perfect matchings generally requires more than \(n+\Theta(n/\log n)\) removals.

## 4. Fixed-rate elimination in tournaments and sequential voting

A second meaning of match-oriented elimination appears in tournament design. “Merging Knockout and Round-Robin Tournaments: A Flexible Linear Elimination Tournament Design” proposes a linear elimination tournament that combines knockout tournaments and round-robin tournaments. Its central design principle is subtractive elimination: players are matched in pairs each round, all players play every round, eliminations occur among the lowest-ranked losers, and the number of eliminations per round is distributed as evenly as possible. The paper explicitly treats Match-2 as the special case in which two participants are eliminated per round, so
\[
N_{t+1}=N_t-2,
\]
and for \(N_0=2m\) the tournament takes \(m-1\) rounds. In this case, the number of matches in round \(t\) is \(N_t/2\), and total matches are
\[
\sum_{t=0}^{T-1}\frac{N_t}{2}.
\]
The same framework uses snake matching, \(r_i+r_j=N+1\), and a reranking objective
\[
\max_{\boldsymbol{r}} \sum_{n=1}^{N}|r_n-n|,
\]
while the elimination distribution is governed by
\[
\min_{\boldsymbol{N}} \sum_t \left| (N_t-N_{t+1})-(N_{t-1}-N_t)\right|.
\]
The paper also states that eliminated players can continue to play in a separate league to determine the full ranking rather than just the champion [2203.12011].

Sequential elimination voting supplies a game-theoretic analogue. In that model, voters act in sequence and eliminate one remaining candidate per round until only one remains. Given a fixed elimination sequence, the outcome is the same in all subgame-perfect Nash equilibria, and the SPNE outcome is the sincere-play outcome for the reversed sequence. The outcome is polynomial-time computable. For palindromic sequences, strategic and sincere outcomes coincide, and the paper gives tight worst-case bounds for the Price of Anarchy and the Sincerity Ratio in terms of Borda score [2210.08844].

The voting paper does not study Match-2 variants in full, but it states that there are analogs and close relatives to variations where multiple candidates are eliminated per round, and that the framework and results strongly indicate that similar backward-induction and SPNE computation methods apply. This suggests that fixed-rate elimination by twos inherits many of the same concerns as one-by-one elimination: sequence design, susceptibility to manipulation, and the tradeoff between low communication and welfare guarantees.

## 5. Matrix elimination, anonymous games, and matched paths

A third technical strand arises from iterated elimination of weakly dominated actions in anonymous games. For two-action anonymous games, the payoff structure can be encoded as an \((n+1)\times n\) matrix \(X=(x_{i,j})\). Matrix elimination proceeds by repeatedly deleting an active column together with one boundary row: if the column is increasing on the current interval of rows, the top row is removed; if decreasing, the bottom row is removed. The remaining submatrix is always a contiguous interval. The matrix elimination problem asks whether one can eliminate all columns and all rows except the final one; the related column eliminability problem asks whether a particular column is eliminable via some elimination sequence [1001.0529].

The main structural fact is that stepwise iterated dominance in two-action anonymous games is logically equivalent, under the reductions stated in the paper, to matrix elimination with \(m=n+1\). Yet the complexity of the unrestricted two-action case remains open. The paper explicitly states that this complexity is “surprisingly difficult to characterize.” By contrast, the three-action case is NP-complete [1001.0529].

The same work establishes a connection to a matched-path problem on labeled layered grid graphs. Vertices correspond to intervals of matrix rows; edges correspond to elimination of the top or bottom row; edge labels record which columns are active at that transition. Finding an elimination sequence is equivalent to finding a path together with a matching between path edges and labels so that each column is used at most once. Matched path is NP-complete in general, even on grid graphs with restricted label sizes, but matrix elimination induces a highly structured subclass with directed convex labeling. Under forward-closure or backward-closure conditions, the problem decomposes into simpler matching problems on convex bipartite graphs and becomes tractable, with solutions in NC\(^2\) or nondeterministic logspace [1001.0529].

This line of work shows that elimination processes that look operationally simple can conceal difficult combinatorial structure. In particular, a two-action or pair-based elimination rule is not automatically algorithmically benign.

## 6. Stochastic pair removal in Memory and elimination learning in statistics

The solitaire game of Memory gives a stochastic pair-removal model. A shuffled deck contains \(n\) pairs of identical cards laid face down; each move flips two cards, and matching cards are removed. Under optimal play with perfect memory, the optimal strategy is unique in the sense described in the source: if the locations of both members of a remaining pair are known, remove them; otherwise flip a previously unseen card, and if its mate has already been seen, remove the pair; if not, flip another unseen card. The asymptotic results are precise:
\[
E[\text{moves}]=(3-2\ln 2)n+\frac{7}{8}-2\ln 2+o(1)\approx 1.6137n,
\]
\[
E[\text{unwitting matches}]=\ln 2 \approx 0.6931,
\]
and the expected number of flips until two matching cards have been seen is asymptotically \(\sqrt{\pi n}\) [1208.4854].

Velleman and Warrington’s analysis shows that pair elimination can be decomposed into informed removals and lucky removals. The expected number of unwitting matches is independent of \(n\) asymptotically, while the waiting time to the first seen pair grows like \(\sqrt{\pi n}\). This gives a probabilistic benchmark for elimination systems in which information acquisition and pair removal interact.

A statistical analogue appears in curriculum learning. In the single-source case, the learner estimates a target parameter vector and may collect data either from the target model or from a similar but less noisy source model. The paper proposes an elimination learning method: use \(N/2\) samples from each model, compute empirical means, estimate \(Q_1^2=\|\theta_1-\theta_0\|^2\), and eliminate the source if
\[
\|\tilde{\theta}_0-\tilde{\theta}_1\|^2 \geq 10g(\delta)\frac{d\sigma_0^2}{N/2}.
\]
With probability at least \(1-\delta\),
\[
\|\hat{\theta}-\theta_0\|^2 \leq \frac{8g(\delta/2)}{\nu}\cdot \min\left\{Q_1^2+\frac{d\sigma_1^2}{N},\;\frac{d\sigma_0^2}{N}\right\},
\]
and the expected risk nearly matches the strong-oracle risk up to logarithmic factors. In the multiple-source case, the paper develops an adaptive multiple elimination-rounds CL algorithm, introduces an elimination curve \(\beta_\delta(\tau)\), and characterizes instance-dependent conditions under which the risk matches that of the weak-oracle learner [2402.13366].

The curriculum-learning paper explicitly describes its source elimination lemma and elimination rounds as a comparison approach akin to match-2 elimination, while also stressing that the procedure is not just pairwise but collective over the current set of models. This suggests a broader interpretation of Match-2 Elimination as an elimination paradigm in which survival depends on repeated comparison, thresholding, and pruning, even when the objects being removed are edges, players, strategies, cards, or source models.

Source: https://www.emergentmind.com/topics/match-2-elimination