Capture-Quiet Decomposition (CQD)

• Capture-Quiet Decomposition (CQD) is a method that divides chess endgame positions into terminal, capture, and quiet classes, enabling precise WDL labeling.
• It verifies correctness by using retrograde constraints with capture positions anchored by verified sub-models and quiet positions checked internally.
• The approach reduces computational cost by isolating material-changing moves, yielding efficiency gains that scale with increasing piece counts.

Searching arXiv for the cited CQD papers to ground the article in the relevant preprints. Capture-Quiet Decomposition (CQD) is a structural theorem and verification method for Win-Draw-Loss (WDL) endgame tablebases. It partitions positions into terminal, capture, and quiet classes, and verifies correctness by combining terminal checks, sub-model lookups on material-changing moves, and retrograde consistency on non-material-changing moves. The central claim is necessary and sufficient: a WDL labeling is correct if and only if terminal positions are labeled correctly, capture positions are consistent with verified sub-models of smaller piece count, and quiet positions satisfy retrograde consistency within the same endgame. In this formulation, capture positions provide external anchors that break the “all-draw” fixpoint trap and reduce the cost of verification (Pavlov, 9 Apr 2026). In unrelated transport-theory literature, the acronym “CQD” denotes “Consistent Quasidiffusion”; that usage is distinct from Capture-Quiet Decomposition (Wang, 2024).

1. Formal setting and partition

CQD is defined on a fixed $N$-piece endgame $E$, with $P = P(E)$ the finite set of all legal positions in that endgame, including side-to-move and all rule-relevant state. A position $p \in P$ is legal if both kings are on the board and not in illegal check, all moves obey chess rules including en passant legality and promotions, and terminal outcomes recognized are checkmate and stalemate. The legal move set is $M(p)$, and the transition function is $\phi: P \times M \to P$, where $\phi(p,m)$ is the unique successor after move $m$ (Pavlov, 9 Apr 2026).

Terminal positions are those with no legal moves. If $p \in T$ and the side to move is in check, the position is checkmate and therefore a game-theoretic loss for the side to move; if $p \in T$ and the side to move is not in check, the position is stalemate and therefore a draw. A WDL labeling is a function $E$0, interpreted from the perspective of the side to move, and correctness means coincidence with the minimax value under perfect play on the game graph induced by chess rules and the endgame definition (Pavlov, 9 Apr 2026).

The decomposition itself is a partition of $E$1 into terminal, capture, and quiet positions:

$E$2

with

$E$3

$E$4

and

$E$5

Here, $E$6 is the multiset of piece types and colors present in $E$7. Capture positions therefore include not only ordinary captures but also promotions, since promotions change material. Quiet positions are precisely those non-terminal positions with no material-changing moves. The partition is disjoint and exhaustive (Pavlov, 9 Apr 2026).

A notable feature of the definition is that classification depends solely on the availability of material-changing moves, not on whether the side to move is in check. A position in check can still be quiet if all legal evasions are non-capturing and non-promoting. This makes CQD a structural decomposition of the move graph rather than a tactical or semantic categorization (Pavlov, 9 Apr 2026).

2. Verification theorem and retrograde constraints

For non-terminal positions, CQD uses the standard WDL retrograde constraints. Writing $E$8, the local optimality conditions are:

$E$9

$P = P(E)$0

and

$P = P(E)$1

These equalities are used as verification constraints rather than as a generation procedure (Pavlov, 9 Apr 2026).

The CQD theorem states that, assuming every strictly smaller-piece endgame $P = P(E)$2 has a verified-correct labeling $P = P(E)$3, a labeling $P = P(E)$4 on $P = P(E)$5 is correct if and only if three conditions hold. First, terminal correctness: every checkmate is labeled $P = P(E)$6 and every stalemate is labeled $P = P(E)$7. Second, capture consistency: for every $P = P(E)$8, the retrograde constraints hold when capture successors are evaluated through the already verified sub-model labelings $P = P(E)$9, while non-capture successors are evaluated by $p \in P$0 within the current endgame. Third, quiet consistency: for every $p \in P$1, the retrograde constraints hold using only successors within the same endgame (Pavlov, 9 Apr 2026).

The capture clause is the distinctive feature. A capture move from $p \in P$2 reaches a successor $p \in P$3 whose material multiset differs from that of $p \in P$4. If a piece is removed, then $p \in P$5 lies in a strictly smaller-piece endgame $p \in P$6 with $p \in P$7, or fewer if promotions or other material changes alter piece count differently. In CQD, those successors are not evaluated self-referentially within the current graph; they are evaluated against a verified external model $p \in P$8. This establishes a boundary condition across material strata (Pavlov, 9 Apr 2026).

The theorem is both necessary and sufficient. The “only-if” direction follows because the correct minimax labeling necessarily satisfies terminal definitions and one-step retrograde optimality. The “if” direction proceeds by strong induction on piece count, with the base case $p \in P$9 identified as $M(p)$0, where all legal positions are draws or terminal draws. For the inductive step, capture positions are forced by already verified smaller endgames, and quiet positions are then constrained by retrograde consistency on the quiet subgraph with those capture nodes acting as boundary conditions (Pavlov, 9 Apr 2026).

3. Fixed-point interpretation and anchoring

CQD is formulated explicitly against the fixed-point structure of retrograde verification. Let the poset of labelings on $M(p)$1 be ordered pointwise by $M(p)$2. The standard retrograde operator $M(p)$3 acts on labelings $M(p)$4 by assigning terminal values on $M(p)$5, assigning $M(p)$6 if some successor is labeled $M(p)$7, assigning $M(p)$8 if all successors are labeled $M(p)$9, and assigning $\phi: P \times M \to P$0 otherwise. This operator is monotone on the finite lattice of labelings and therefore has least and greatest fixpoints by Knaster–Tarski (Pavlov, 9 Apr 2026).

Within that framework, self-consistency alone is insufficient. The all-draw labeling $\phi: P \times M \to P$1, which assigns $\phi: P \times M \to P$2 to every position, can satisfy $\phi: P \times M \to P$3 when considered in isolation. CQD addresses this by replacing self-closure on the whole graph with boundary conditions on capture edges. Capture successors are evaluated in smaller, already verified models, so a non-draw value in a sub-model forces a corresponding local constraint at the predecessor. This is the mechanism by which capture positions “anchor” the verification and destroy degenerate fixpoints such as the all-draw solution (Pavlov, 9 Apr 2026).

The associated anchoring lemma is stated informally: if at least one capture successor $\phi: P \times M \to P$4 has $\phi: P \times M \to P$5, then an all-draw labeling on the current endgame violates the retrograde constraints at the predecessor capture node $\phi: P \times M \to P$6. If $\phi: P \times M \to P$7, then $\phi: P \times M \to P$8 must be $\phi: P \times M \to P$9; if $\phi(p,m)$0, then $\phi(p,m)$1 cannot remain $\phi(p,m)$2 when all relevant successors force otherwise. The significance of this observation is structural rather than heuristic: it identifies the exact source of non-circularity in the verification theorem (Pavlov, 9 Apr 2026).

This fixed-point viewpoint also clarifies the role of the quiet subgraph. Quiet nodes are not externally anchored directly; instead, they are retrograded within the same endgame subject to the already anchored values on capture-adjacent structure and terminal seeds. The theorem states that, under these conditions, the induced fixpoint on the quiet component is the unique correct labeling. This suggests that CQD’s primary conceptual contribution is not a new retrograde rule, but a new decomposition of where retrograde must be internal and where it can be externally grounded (Pavlov, 9 Apr 2026).

4. Verifier construction and computational properties

The verifier takes as input a candidate labeling $\phi(p,m)$3 for an endgame $\phi(p,m)$4 and a chain of verified sub-model labelings for all strictly smaller endgames. Its high-level workflow is to compute $\phi(p,m)$5, $\phi(p,m)$6, and $\phi(p,m)$7; check terminal correctness on $\phi(p,m)$8; check capture consistency on $\phi(p,m)$9 using sub-model values on material-changing successors; and check quiet consistency on $m$0 using internal successors within $m$1. It reports violations, meaning positions at which one of the required conditions fails (Pavlov, 9 Apr 2026).

The implementation-ready classification rule is simple. For each position $m$2, legal successors are generated. If there are no legal moves, the position is terminal. Otherwise, if at least one successor changes $m$3, the position is in $m$4; if all successors preserve $m$5, the position is in $m$6. Capture successors are evaluated via the appropriate verified sub-model, determined by the material configuration of the successor; non-capture successors are evaluated by the candidate labeling $m$7 within the current endgame (Pavlov, 9 Apr 2026).

The local verification rule is then applied according to the label assigned to $m$8. If $m$9, some evaluated successor must be $p \in T$0. If $p \in T$1, all evaluated successors must be $p \in T$2. If $p \in T$3, no evaluated successor may be $p \in T$4, and at least one evaluated successor must be $p \in T$5. Violations are categorized as terminal mismatches, capture inconsistencies, or quiet inconsistencies (Pavlov, 9 Apr 2026).

The stated complexity is $p \in T$6 to enumerate positions and successors and check constraints once, with move generation dominating $p \in T$7. Memory usage can be streaming, since the entire graph need not be stored. External sub-model access is $p \in T$8 per capture successor via table lookups, with Syzygy-style probing given as an example. These implementation notes are integral to CQD’s practical framing: the method is not only a theorem, but also an implementation-ready verifier (Pavlov, 9 Apr 2026).

Several rule-sensitive details are explicit. Promotions count as material-changing moves and therefore place a position in $p \in T$9. En passant is also a capture and therefore belongs to $p \in T$0. Castling, if included by a modeling choice, does not by itself change material and would not move a position into $p \in T$1. Repetition and the 50-move rule are not modeled as terminal outcomes in the WDL notion used here; only checkmate and stalemate are terminal. Under that convention, CQD applies unchanged (Pavlov, 9 Apr 2026).

5. Empirical validation, examples, and edge cases

The empirical validation reported for CQD is exhaustive over all 35 three- and four-piece endgames, all 110 five-piece endgames, and all 372 six-piece endgames, for a total of 517 endgames. At the three- and four-piece level this corresponds to 42,036,972 positions. In every case, the decomposed verifier produced identical violation counts to a full retrograde baseline. A representative anomaly in $p \in T$2 yielded 121,188 violations in both methods, and those violations were attributed to the upstream decision-tree model used to propose labels rather than to CQD itself (Pavlov, 9 Apr 2026).

The reported efficiency observations distinguish small and larger material classes. At three to four pieces, runtime between full retrograde verification and CQD is comparable. At larger piece counts, the fraction of capture positions grows, reported as approximately $p \in T$3 at 4 pieces, approximately $p \in T$4 at 5 pieces, approximately $p \in T$5 at 6 pieces, and projected $p \in T$6 by 20 pieces. On that basis, CQD is expected to reduce the portion requiring retrograde reasoning and to outperform full retrograde’s multi-round propagation as piece count increases (Pavlov, 9 Apr 2026).

The paper gives three canonical examples. In $p \in T$7, a checkmated black king position with black to move, no legal moves, and black in check belongs to $p \in T$8 and must be labeled $p \in T$9. In $E$00, a position where White can play $E$01 belongs to $E$02; the successor lies in the smaller endgame $E$03, and CQD checks retrograde consistency at the predecessor using the verified $E$04 labeling. In $E$05 with no immediate captures, a position with only non-capturing legal moves belongs to $E$06, and CQD verifies its label solely through internal retrograde constraints (Pavlov, 9 Apr 2026).

The edge-case treatment is explicit and structurally important. A position in check may still be quiet if no capture or promotion is legal. Promotions are treated as capture-class moves because material changes. En passant is treated as a capture. The exclusion of repetition and 50-move-rule terminality means that the verified WDL value is the rule-agnostic game-theoretic value standard for tablebase WDL in this formulation. These choices determine the graph over which the theorem is stated and verified (Pavlov, 9 Apr 2026).

6. Relations, limitations, and extensions

CQD is presented in relation to full retrograde verification, tablebase generation practice, and possible extensions to richer labels. Full retrograde verification runs multi-round propagation over the entire graph until a fixpoint is reached. It is correct but expensive and, absent suitable seeding, admits trivial self-consistent fixpoints. Generation systems such as Syzygy often seed from checkmates and use captures as optimizations, but the decomposition itself is framed here as a verification theorem rather than merely an engineering heuristic (Pavlov, 9 Apr 2026).

The method’s principal limitations follow directly from its assumptions. CQD depends on a verified chain of smaller endgames; if a sub-model is incorrect, capture consistency can propagate error upward. Its correctness also relies on capture successors mapping into strictly smaller material models. The text notes that exotic rule encodings that alter material-equivalence, such as modeling 50-move counters as stateful dimensions, require a careful extension of what counts as a “smaller sub-model.” A further stated limitation is the possibility, in principle, of degenerate quiet-only components with no edges to capture positions and no terminal seeds, although the theorem’s conditions preclude such cases in the intended setting (Pavlov, 9 Apr 2026).

The proposed extensions are conservative rather than speculative. For depth-to-mate (DTM) and depth-to-zeroing-move (DTZ), the WDL-valued retrograde constraints are to be replaced with metric-labeled Bellman recurrences. The same decomposition remains applicable: capture positions obtain metric values from sub-models with a one-ply increment, while quiet positions satisfy within-endgame recurrences. For DTZ, the text notes that zeroing moves already align well with the capture/quiet dichotomy because captures and pawn moves are central to the metric (Pavlov, 9 Apr 2026).

More broadly, CQD is described as applicable beyond chess WDL to any game in which some moves cross “material boundaries,” reducing piece count or otherwise moving to a verified subspace. This suggests a general pattern: verification can be decomposed into externally anchored transitions across model strata and internally checked consistency within a fixed stratum. In the chess endgame setting, that pattern is formalized exactly by the partition $E$07, by the use of verified sub-models on capture successors, and by the theorem that these conditions are both necessary and sufficient for WDL correctness (Pavlov, 9 Apr 2026).