Basis Exchange Walk in Matroids

Updated 12 July 2026

Basis Exchange Walk is a local reconfiguration process on matroid bases using one-element symmetric exchanges that preserve the structure of a matroid.
It connects deterministic and stochastic frameworks by underpinning polynomial-time oracle algorithms and Markov-chain sampling through precise exchange mechanisms.
The method informs critical conjectures such as those of White and Gabow and offers practical insights into fair division, efficient sampling, and representability in diverse matroid classes.

Basis exchange walk is the local reconfiguration process on matroid bases generated by one-element exchanges that preserve basishood. In the deterministic reconfiguration setting, the states are either individual bases or ordered pairs of bases, and adjacency is defined by symmetric exchanges. In the stochastic setting, the same local mechanism underlies the down–up Markov chain on the set of bases. The topic connects matroid structure, exchange distance, oracle algorithms, sampling, and discrete curvature, and it is closely tied to long-standing conjectures of White and Gabow on the global geometry of basis families (Bérczi et al., 2023, Detherage, 23 Sep 2025).

1. Local exchange mechanisms

Let $M=(E,\mathcal B)$ be a matroid. A basis is a maximal independent set, and all bases have the same cardinality $r(M)$ . If $I$ is independent and $x\notin I$ but $I\cup\{x\}$ is dependent, then there is a unique circuit contained in $I\cup\{x\}$ , denoted $C(x,I)$ . In particular, if $B$ is a basis and $e\notin B$ , then $C(e,B)$ is the unique circuit in $r(M)$ 0. For two bases $r(M)$ 1, a symmetric exchange is a pair $r(M)$ 2 with $r(M)$ 3 and $r(M)$ 4 such that both $r(M)$ 5 and $r(M)$ 6 are bases. The admissibility condition can be expressed circuit-theoretically: if $r(M)$ 7, then $r(M)$ 8 is allowed exactly when $r(M)$ 9. The basis-exchange graph, denoted $I$ 0 in one common notation, has vertex set $I$ 1 and an edge between two bases when they differ by one exchange (Kotlar, 2011).

A parallel notation writes this graph as $I$ 2: one node for each basis $I$ 3, and an edge $I$ 4 whenever $I$ 5 for some $I$ 6 and $I$ 7. For single bases, the graph distance satisfies

$I$ 8

This identity makes the single-basis walk comparatively rigid. By contrast, once the state space is enlarged from bases to basis pairs, the corresponding exchange-distance problem becomes genuinely nontrivial and is the source of several open and recently resolved reconfiguration questions (Bérczi et al., 2022).

2. Basis pairs, compatibility, and classical conjectures

A basis pair is an ordered pair $I$ 9 of bases, not necessarily disjoint. A symmetric exchange in $x\notin I$ 0 chooses elements $x\notin I$ 1 and $x\notin I$ 2 so that both updated sets remain bases; for pairs, this produces a new state $x\notin I$ 3. The basis-pair exchange graph $x\notin I$ 4 has one node for each ordered pair of bases and an edge for each such symmetric exchange. Two basis pairs $x\notin I$ 5 are compatible if every element of $x\notin I$ 6 has the same total multiplicity in $x\notin I$ 7 and $x\notin I$ 8. White’s conjecture for $x\notin I$ 9 asks whether compatibility is also sufficient: any two compatible basis pairs should lie in the same connected component of $I\cup\{x\}$ 0 (Bérczi et al., 2023).

For disjoint pairs, this suggests that compatibility reduces to preservation of the same $I\cup\{x\}$ 1-element union, which is the form used in weighted exchange-distance problems. Several general results precede the recent regular-matroid breakthrough. Kotlar proved that for any two bases of a matroid there exist three consecutive symmetric exchanges. The same paper shows that every pair of bases in a rank- $I\cup\{x\}$ 2 matroid admits a full serial symmetric exchange of length at most $I\cup\{x\}$ 3. It also gives a characterization of binary matroids via the circuit identity

$I\cup\{x\}$ 4

required to hold for every basis $I\cup\{x\}$ 5, every distinct $I\cup\{x\}$ 6, and every $I\cup\{x\}$ 7 (Kotlar, 2011).

3. Regular matroids and polynomial exchange walks

For regular matroids, Bérczi, Mátrovölgyi, and Schwarcz verified White’s conjecture for basis sequences of length two. If $I\cup\{x\}$ 8 is a regular matroid of rank $I\cup\{x\}$ 9 and $I\cup\{x\}$ 0, $I\cup\{x\}$ 1 are compatible basis pairs, then $I\cup\{x\}$ 2 and $I\cup\{x\}$ 3 are connected in $I\cup\{x\}$ 4 by a walk of length at most $I\cup\{x\}$ 5; no element is used in more than $I\cup\{x\}$ 6 of the exchanges; and such a sequence can be found in polynomial time under an independence-oracle model. As a byproduct, the same framework proves Gabow’s conjecture in the regular case: for every pair of disjoint bases in a regular matroid of rank $I\cup\{x\}$ 7, there is a symmetric-exchange sequence of exactly $I\cup\{x\}$ 8 steps transforming $I\cup\{x\}$ 9 into $C(x,I)$ 0, again in polynomial time (Bérczi et al., 2023).

Reduction	Operation	Effect
Disjointness reduction	Contract $C(x,I)$ 1	Length and width unchanged when lifting
Tight-set reduction	Solve on $C(x,I)$ 2 and $C(x,I)$ 3 for tight $C(x,I)$ 4	Lengths add; width is the maximum
$C(x,I)$ 5-sum reduction	Align walks on $C(x,I)$ 6 and $C(x,I)$ 7	Combined length at most $C(x,I)$ 8 or $C(x,I)$ 9
Triad reduction	Preparatory exchanges, then delete/contract	Rank drops by one with small overhead

The proof proceeds by repeated simplification. Intersections of the two current bases are contracted. Nontrivial tight sets $B$ 0 with $B$ 1 are separated into the restriction $B$ 2 and the contraction $B$ 3. If $B$ 4, walks on the two sides are synchronized so that shared-element moves cancel. If a cocircuit $B$ 5 of size $B$ 6 is covered by $B$ 7, then after at most two preparatory exchanges one deletes and contracts two of the $B$ 8, reducing rank by one. These reductions eventually produce a $B$ 9-connected regular matroid with no tight sets and no cocircuits of size at most $e\notin B$ 0; by duality, also no circuits of size at most $e\notin B$ 1.

At that point, Aprile–Fiorini’s refinement of Seymour’s decomposition theorem yields a $e\notin B$ 2-sum decomposition $e\notin B$ 3, with $e\notin B$ 4 either graphic or cographic. A further argument based on McGuinness’s $e\notin B$ 5-operation allows one to assume $e\notin B$ 6 is graphic, and the no-tight/no-small-cocircuit constraints force it to arise from a simple $e\notin B$ 7-regular graph $e\notin B$ 8. The central graphic lemma states that in a bispanning graph whose edge set partitions into two spanning forests, any such partition can be transformed into any other by symmetric forest-exchanges in at most $e\notin B$ 9 steps while avoiding a prescribed set $C(e,B)$ 0 of at most $C(e,B)$ 1 edges. The two sides are then spliced: first a walk on the graphic side, then a walk on $C(e,B)$ 2 with every $C(e,B)$ 3-move replaced by a matching graphic move, then a final graphic walk.

The algorithmic framework mirrors this structure. It handles $C(e,B)$ 4 by case analysis; tests for $C(e,B)$ 5-, $C(e,B)$ 6-, and $C(e,B)$ 7-sum decompositions using a $C(e,B)$ 8-rank test via Truemper’s algorithm; finds tight sets by minimizing the submodular function $C(e,B)$ 9; searches for triads by querying all $r(M)$ 00-subsets in the dual; and otherwise applies the Aprile–Fiorini decomposition tree to extract the graphic component. At each stage, the rank decreases or the instance splits into smaller-rank minors, each subroutine runs in polynomial oracle time $r(M)$ 01, the recursion depth is $r(M)$ 02, and the total running time is polynomial in $r(M)$ 03 and $r(M)$ 04. The exceptional matroid $r(M)$ 05 is handled by finite verification, with any two basis pairs differing by at most $r(M)$ 06 exchanges; the same type of check yields a bound of at most $r(M)$ 07 steps for $r(M)$ 08.

4. Weighted exchange distance and special matroid classes

For a disjoint basis pair $r(M)$ 09, a symmetric exchange $r(M)$ 10 with $r(M)$ 11, $r(M)$ 12 updates the pair to $r(M)$ 13. Given a weight function $r(M)$ 14, the cost of this move is $r(M)$ 15, and the weighted exchange distance $r(M)$ 16 is the minimum total cost over all exchange sequences transforming $r(M)$ 17 into $r(M)$ 18. The weighted Hamidoune conjecture asserts that for compatible basis pairs,

$r(M)$ 19

When $r(M)$ 20, this becomes the unweighted bound $r(M)$ 21. The weighted conjecture has been proved for strongly base orderable matroids, split matroids, graphic matroids of wheels, and spikes; in each of these positive cases, each element is used in at most two swaps (Bérczi et al., 2022).

Class or setting	Guaranteed bound	Additional feature
Strongly base orderable	Cost $r(M)$ 22	Each element used at most twice
Split matroids	Length $r(M)$ 23, cost $r(M)$ 24	Each element used at most twice
Wheel graphic matroids $r(M)$ 25	Length $r(M)$ 26, cost $r(M)$ 27	Each edge used at most twice
Spikes	Length $r(M)$ 28, cost $r(M)$ 29	Each element used at most twice

The split-matroid case is especially detailed. Split matroids are characterized by connected components that are either uniform or elementary split, the latter described by a hypergraph representation with constraints $r(M)$ 30– $r(M)$ 31. For compatible basis pairs in a split matroid, there is a polynomial-time algorithm that computes a shortest exchange sequence. The method contracts any common intersection $r(M)$ 32, isolates the unique non-uniform connected component if one exists, performs a greedy strictly monotone phase, and when that stalls invokes a four-hyperedge obstruction $r(M)$ 33. Either the sequence can be extended by two more exchanges, or it is finished by a specially designed path of length $r(M)$ 34, where $r(M)$ 35. The resulting distance bounds are

$r(M)$ 36

and

$r(M)$ 37

and the algorithm uses $r(M)$ 38 oracle calls. These results verify White’s conjecture for length-two sequences and Gabow’s sequential symmetric exchange conjecture for split matroids, and therefore for paving matroids as well (Bérczi et al., 2022).

The constructive arguments differ significantly across classes. Strongly base orderable matroids use strong-order bijections and a weighted averaging argument over two exchange routes. Wheels use the cyclic organization of spokes and rims, with a distinction between same-orientation and opposite-orientation colorings. Spikes are treated by case analysis on the missing element and on whether a red basis is transversal or non-transversal. This diversity suggests that no single exchange mechanism is currently known to subsume all positive classes.

5. Markov-chain basis-exchange walks and curvature

In probabilistic usage, the basis-exchange walk is the down–up walk on the set of bases. Let $r(M)$ 39 be a rank- $r(M)$ 40 matroid. Starting from a basis $r(M)$ 41, one first performs a down step by deleting a uniformly chosen $r(M)$ 42, and then an up step by adding a uniformly chosen $r(M)$ 43, where

$r(M)$ 44

The transition kernel satisfies

$r(M)$ 45

and detailed balance implies that the stationary distribution is uniform on $r(M)$ 46 (Detherage, 23 Sep 2025).

Detherage studied the Ollivier–Ricci curvature of this chain. If $r(M)$ 47 is the shortest-path metric on the state graph, then for adjacent states it suffices to analyze

$r(M)$ 48

For a rank- $r(M)$ 49 matroid on $r(M)$ 50 elements, the curvature satisfies the lower bound

$r(M)$ 51

For adjacent bases $r(M)$ 52 and $r(M)$ 53, with

$r(M)$ 54

one also has the upper bound

$r(M)$ 55

In particular, if

$r(M)$ 56

then $r(M)$ 57.

The examples are mixed. Any rank- $r(M)$ 58 matroid has $r(M)$ 59, and if $r(M)$ 60 then $r(M)$ 61. Uniform matroids satisfy

$r(M)$ 62

The Vamos matroid has $r(M)$ 63, and graphic matroids whose underlying graph decomposes into edge-disjoint cycles also have non-negative curvature. On the other hand, there is a rank- $r(M)$ 64 linear matroid with $r(M)$ 65, and a graphic example on $r(M)$ 66 with $r(M)$ 67. These negative examples show that rapid mixing of basis-exchange walks cannot in general be established by Ollivier–Ricci methods alone, even though independent techniques based on spectral independence and strong log-concavity do guarantee rapid mixing uniformly for all matroids.

Exchange walks on basis pairs have consequences beyond reconfiguration. In the single-basis case, the basis-exchange random walk on $r(M)$ 68 has been shown to mix rapidly for every matroid after the Mihail–Vazirani conjecture on edge expansion. For matroid intersection, no general statement of this kind was known because the exchange graph need not be connected. In the special case $r(M)$ 69, common bases correspond to partitions of $r(M)$ 70 into two bases of $r(M)$ 71; White’s $r(M)$ 72 conjecture then becomes a connectivity statement for $r(M)$ 73, and the polynomial exchange-distance bound for regular matroids opens the door to mixing-time bounds for sampling such partitions in the regular case. Compatible exchange walks also imply fair-division statements, including the Equitability Conjecture for matroids and the existence of envy-free up to one good allocations under identical matroid constraints. For split matroids, the path from $r(M)$ 74 to $r(M)$ 75 likewise yields equitability (Bérczi et al., 2023, Bérczi et al., 2022).

At the structural level, the theory is intertwined with representability and circuit identities. Kotlar’s criterion characterizes binary matroids by the equality $r(M)$ 76, placing circuit behavior under exchange at the center of the binary/non-binary distinction (Kotlar, 2011). A plausible implication is that basis-exchange walks are not merely a reconfiguration gadget but also a probe of deep representability phenomena.

Several major questions remain open. The weighted Hamidoune conjecture is unresolved for arbitrary matroids, and the case of binary matroids is identified as a particularly attractive intermediate target (Bérczi et al., 2022). On the stochastic side, the curvature picture is incomplete: one open problem is to characterize exactly which matroids satisfy $r(M)$ 77, another is to narrow the gap between the general lower bound $r(M)$ 78 and the currently known negative examples, and a further direction is to extend the analysis to weighted down–up walks or up–down variants (Detherage, 23 Sep 2025). More broadly, the regular-matroid result for $r(M)$ 79 basis pairs highlights the main unresolved frontier: a general theory of short, constructive, and possibly weight-efficient exchange walks for arbitrary matroids.