Certified Parity Synthesis: Methods & Guarantees

• Certified parity synthesis is a framework for constructing artifacts that include machine-checkable parity guarantees across various domains such as MDPs, circuit design, and distributed networks.
• It employs strategy templates, invariant-preserving constructions, and quantitative certificates to ensure both deterministic and probabilistic correctness.
• Applications span reversible and quantum circuit synthesis, certified controller synthesis in stochastic games, and efficient proof logging in SAT solving, enabling robust system verification.

Certified parity synthesis concerns synthesis procedures in which parity-related properties are produced together with explicit guarantees, strategy witnesses, proof logs, or local certificates. In the game-theoretic literature, parity is an $\omega$-regular objective defined by the maximal priority seen infinitely often; in reversible and quantum circuit synthesis, parity is the XOR of bit values and becomes an invariant of the implemented transformation; in proof logging and distributed certification, parity is the target of machine-checkable derivations or local labels (Berthon et al., 2017, Paul et al., 2013, Gocht et al., 2022, Bousquet et al., 3 Jun 2026). Across these settings, the recurring requirement is that the synthesized artifact is not merely correct in extension, but accompanied by a certificate that can be checked under the semantics of the model.

1. Parity as an objective and as an invariant

In finite MDPs and parity games, a priority function is a map

$p : S \to \{1,2,\dots,d\}$

or, in vertex notation,

$\Omega : V \to \{0,\dots,d\}.$

The associated parity objective is the set of infinite plays for which the maximum priority occurring infinitely often is even: $$\text{Parity}(p) = \{ \pi \in \text{Plays}(G) \mid \max_{s \in \inf^\omega(\pi)} p(s) \text{ is even} \}.$$ This is the standard encoding of $\omega$-regular specifications, including LTL via automata (Berthon et al., 2017).

In reversible logic, parity is instead the bitwise XOR of an input or output vector: $$\mathsf{par}(x) = x_1 \oplus x_2 \oplus \cdots \oplus x_n.$$ A reversible Boolean function $f:\{0,1\}^n\to\{0,1\}^n$ is parity-preserving iff

$$\forall x\in\{0,1\}^n:\quad \mathsf{par}(x) = \mathsf{par}(f(x)).$$

A parity-preserving reversible specification is therefore a bijective truth table in which input and output parity match on every row (Paul et al., 2013).

These two uses of parity differ formally, but both are certification-friendly. In one case, parity is an acceptance condition over infinite executions; in the other, it is a row-wise algebraic invariant. This suggests a common pattern: certification attaches a checkable witness to a parity property, whether the witness is a winning strategy, a template, an auxiliary truth-table construction, or a proof object.

2. Beyond-worst-case parity synthesis in Markov decision processes

A central formulation of certified parity synthesis in stochastic systems is the beyond worst-case parity problem for finite MDPs

$\mathcal{M} = (G=(S,E), S_1, S_2, \delta),$

where Player 1 is the controller and Player 2 is either probabilistic or antagonistic, depending on the semantics under consideration (Berthon et al., 2017). The paper introduces three operators over events $\mathcal{A}\subseteq \text{Plays}(G)$: sure satisfaction $p : S \to \{1,2,\dots,d\}$0, threshold probability $p : S \to \{1,2,\dots,d\}$1 or $p : S \to \{1,2,\dots,d\}$2, and almost-sure satisfaction $p : S \to \{1,2,\dots,d\}$3. A certified controller is then a strategy $p : S \to \{1,2,\dots,d\}$4 such that it guarantees a primary parity property on all plays and certifies a quantitative level of satisfaction of a secondary parity objective under the stochastic environment.

The canonical problem is

$p : S \to \{1,2,\dots,d\}$5

and the threshold variant is

$p : S \to \{1,2,\dots,d\}$6

The same strategy must witness both conjuncts. This is the distinctive certified aspect: the hard guarantee against an antagonistic environment and the probabilistic guarantee against the stochastic model are attached to a single synthesized strategy (Berthon et al., 2017).

The structural core of the almost-sure case is the notion of ultra-good end-components (UGECs). An end-component $p : S \to \{1,2,\dots,d\}$7 is ultra-good if, inside the restricted MDP $p : S \to \{1,2,\dots,d\}$8, Player 1 can both ensure $p : S \to \{1,2,\dots,d\}$9 while almost surely reaching a maximal even $\Omega : V \to \{0,\dots,d\}.$0-region, and satisfy $\Omega : V \to \{0,\dots,d\}.$1. Writing $\Omega : V \to \{0,\dots,d\}.$2 for the union of UGECs, the paper proves the equivalence

$\Omega : V \to \{0,\dots,d\}.$3

For threshold problems, the weaker notion of very-good end-components (VGECs) is used; every UGEC is a VGEC, but not conversely (Berthon et al., 2017).

The algorithmic picture is tight. Deciding reachability-under-parity, almost-sure parity under parity, and threshold parity under parity all lie in $\Omega : V \to \{0,\dots,d\}.$4, the same class as classical parity games. For $\Omega : V \to \{0,\dots,d\}.$5 and $\Omega : V \to \{0,\dots,d\}.$6, witness strategies exist whenever the answer is positive, but infinite memory is in general necessary. The certificate described in the synthesis viewpoint consists of the states from which $\Omega : V \to \{0,\dots,d\}.$7 holds, the relevant UGECs or VGECs, and the strategy structure specifying which local strategy is used in each region (Berthon et al., 2017).

3. Certified artifacts in parity games: templates, contracts, and partial solvers

In parity games, certification often takes the form of symbolic strategy descriptions rather than single positional policies. One approach computes permissive winning strategy templates

$\Omega : V \to \{0,\dots,d\}.$8

where $\Omega : V \to \{0,\dots,d\}.$9 is a safety template forbidding unsafe edges, $$\text{Parity}(p) = \{ \pi \in \text{Plays}(G) \mid \max_{s \in \inf^\omega(\pi)} p(s) \text{ is even} \}.$$0 is a live-group template of the form

$$\text{Parity}(p) = \{ \pi \in \text{Plays}(G) \mid \max_{s \in \inf^\omega(\pi)} p(s) \text{ is even} \}.$$1

and $$\text{Parity}(p) = \{ \pi \in \text{Plays}(G) \mid \max_{s \in \inf^\omega(\pi)} p(s) \text{ is even} \}.$$2 is a co-live template

$$\text{Parity}(p) = \{ \pi \in \text{Plays}(G) \mid \max_{s \in \inf^\omega(\pi)} p(s) \text{ is even} \}.$$3

A strategy follows the template if all its plays satisfy $$\text{Parity}(p) = \{ \pi \in \text{Plays}(G) \mid \max_{s \in \inf^\omega(\pi)} p(s) \text{ is even} \}.$$4. For safety games the resulting template is maximally permissive, and for parity games the recursive algorithm ParityTemplate computes a winning strategy template in time $$\text{Parity}(p) = \{ \pi \in \text{Plays}(G) \mid \max_{s \in \inf^\omega(\pi)} p(s) \text{ is even} \}.$$5, matching the asymptotic complexity of Zielonka’s classical algorithm (Anand et al., 2023).

These templates serve as certificates because every strategy that respects them is winning from the computed region. They also support incremental synthesis and fault-tolerant control. For generalized parity obtained by conjunction of multiple parity objectives, the composition algorithm repairs conflicts by modifying priorities on conflict vertices; empirically, the prototype implementation returned the full winning region in all 1400 benchmark instances reported in the study (Anand et al., 2023).

A stronger assume-guarantee version appears in two-objective parity games. There, each player receives a contracted strategy-mask (CSM)

$$\text{Parity}(p) = \{ \pi \in \text{Plays}(G) \mid \max_{s \in \inf^\omega(\pi)} p(s) \text{ is even} \}.$$6

where $$\text{Parity}(p) = \{ \pi \in \text{Plays}(G) \mid \max_{s \in \inf^\omega(\pi)} p(s) \text{ is even} \}.$$7 is an assumption template over the other player’s edges and $$\text{Parity}(p) = \{ \pi \in \text{Plays}(G) \mid \max_{s \in \inf^\omega(\pi)} p(s) \text{ is even} \}.$$8 is a strategy template over the player’s own edges. Compatibility of two CSMs means that each player can realize the conjunction of its own strategy template and the other player’s assumption template. From compatible, adequately permissive CSMs, one derives contracted local specifications

$$\text{Parity}(p) = \{ \pi \in \text{Plays}(G) \mid \max_{s \in \inf^\omega(\pi)} p(s) \text{ is even} \}.$$9

and these are shown to be independently realizable and maximally cooperative (iRmaC): their conjunction preserves exactly the cooperative winning plays of the original two-objective game, and their winning regions intersect exactly on the cooperative winning region. The local routine ParityTemp computes adequately permissive CSMs in time $\omega$0, and the global negotiation procedure terminates in time $\omega$1 (Anand et al., 2023).

A different certification layer is provided by partial solvers for parity and generalized parity games. A partial solver returns sets $\omega$2 and $\omega$3 together with a residual subgame. BÜchiSolver, GoodEpSolver, LaySolver, and their generalized counterparts are polynomial-time and sound by construction: every classified vertex is genuinely winning for the declared player. When combined with Zielonka’s recursive algorithm under the paper’s escape condition, the resulting solver remains complete. This provides a certificate structure based on attractors, traps, layered or good-episode fixpoints, and, in the generalized case, antichain representations of closed sets (Bruyère et al., 2019).

4. Parity-preserving synthesis in reversible and quantum circuits

In reversible circuit synthesis, certified parity synthesis is formulated at the specification level. The paper on parity-preserving reversible circuits proves that the number of $\omega$4-variable parity-preserving reversible Boolean functions is

$\omega$5

since the odd-parity inputs must permute among odd-parity outputs and the even-parity inputs among even-parity outputs (Paul et al., 2013).

Two constructive results are central. First, any reversible $\omega$6-variable specification can be converted to a parity-preserving reversible specification with the introduction of at most one extra variable. The construction fixes a new input bit to $\omega$7 and appends an output bit

$\omega$8

so that the augmented mapping preserves total parity. Second, for an irreversible specification with repeated output rows, the minimum number of extra bits needed to obtain a parity-preserving reversible specification is

$\omega$9

where $$\mathsf{par}(x) = x_1 \oplus x_2 \oplus \cdots \oplus x_n.$$0 is the multiplicity of output pattern $$\mathsf{par}(x) = x_1 \oplus x_2 \oplus \cdots \oplus x_n.$$1 and $$\mathsf{par}(x) = x_1 \oplus x_2 \oplus \cdots \oplus x_n.$$2 counts the rows in that group where input and output parity already match. Algorithm 1 realizes this bound and runs in $$\mathsf{par}(x) = x_1 \oplus x_2 \oplus \cdots \oplus x_n.$$3 time for an $$\mathsf{par}(x) = x_1 \oplus x_2 \oplus \cdots \oplus x_n.$$4-input, $$\mathsf{par}(x) = x_1 \oplus x_2 \oplus \cdots \oplus x_n.$$5-output specification (Paul et al., 2013).

The certification viewpoint is explicit: the transformed truth table itself is the witness. Injectivity and row-wise parity preservation can be checked directly, and the counts used in the construction act as a minimality certificate for the number of added bits. This suggests a specification-first certification workflow: first synthesize a parity-preserving reversible specification, then feed it to a reversible circuit synthesis engine (Paul et al., 2013).

In quantum optimization, parity network synthesis plays an analogous role for QAOA diagonal blocks. Given a set of parities $$\mathsf{par}(x) = x_1 \oplus x_2 \oplus \cdots \oplus x_n.$$6, the task is to construct a CNOT-only circuit such that every parity in $$\mathsf{par}(x) = x_1 \oplus x_2 \oplus \cdots \oplus x_n.$$7 appears as the XOR of current wire values at some point, and the final linear transformation is a specified $$\mathsf{par}(x) = x_1 \oplus x_2 \oplus \cdots \oplus x_n.$$8. The paper proposes a greedy parity network synthesis algorithm and a greedy Gaussian elimination algorithm for the inverse linear map. The stated application is the diagonal component of the QAOA circuit, generally the most expensive in terms of two-qubit gates, and the reported empirical comparisons show lower CNOT counts than off-the-shelf compiler tools on random, full, and graph-based optimization problems (Campbell et al., 2024).

The paper does not formalize correctness in theorem form, but it explicitly notes that correctness follows from standard linear algebra over $$\mathsf{par}(x) = x_1 \oplus x_2 \oplus \cdots \oplus x_n.$$9. Because both algorithms update basis matrices by CNOT-induced row operations, the resulting circuit, basis sequence, and parity coverage times are naturally suited to matrix-level replay checking. A plausible implication is that parity network synthesis can support a certified compilation pipeline in which coverage of target parities and inversion of the final basis are independently checkable (Campbell et al., 2024).

5. Machine-checkable parity reasoning with pseudo-Boolean proofs

Certified parity synthesis also appears as proof-producing parity reasoning inside SAT solving. The pseudo-Boolean framework represents Boolean constraints as inequalities

$f:\{0,1\}^n\to\{0,1\}^n$0

supports cutting-planes rules, PB-RUP, and extension variables introduced by reification. The key reification

$f:\{0,1\}^n\to\{0,1\}^n$1

is encoded by two PB constraints, and their introduction is certified by a redundance-based strengthening rule checked by Algorithm 3.1. This turns VeriPB into a strict generalization of DRAT (Gocht et al., 2022).

For XOR constraints

$f:\{0,1\}^n\to\{0,1\}^n$2

the paper uses the PB equality

$f:\{0,1\}^n\to\{0,1\}^n$3

A chain of one-bit full adders derives this encoding from canonical CNF encodings of short XORs. Once such PB encodings are present, Gaussian elimination modulo $f:\{0,1\}^n\to\{0,1\}^n$4 becomes a sequence of linear combinations of PB equalities, and XOR-based conflicts or propagations are certified by derived clausal PB constraints. The method was integrated into a CDCL solver with lazy reason generation: an XOR engine performs Gaussian elimination, while proofs are emitted only when a reason or conflict must be justified (Gocht et al., 2022).

The significance is practical as well as formal. The reported experiments show substantial reductions in proof size and verification time compared with DRAT-based XOR proof logging, including cryptographic benchmarks on which DRAT runs reached a 100GB disk limit. Since PB constraints subsume clauses and the framework already supports 0–1 programming and constraint programming encodings, the paper positions parity certification as part of a unified proof-logging architecture for richer reasoning paradigms (Gocht et al., 2022).

6. Local certification of parity in distributed networks

In distributed computing, parity certification is studied through proof-labeling schemes. A prover assigns each node a certificate $f:\{0,1\}^n\to\{0,1\}^n$5, and a local verification algorithm inspects the radius-$f:\{0,1\}^n\to\{0,1\}^n$6 neighborhood, together with certificates and, in the ID model, identifiers. Completeness requires that graphs satisfying the property admit a certificate assignment accepted by all nodes; soundness requires that graphs not satisfying the property be rejected somewhere for every certificate assignment (Bousquet et al., 3 Jun 2026).

The property is

$f:\{0,1\}^n\to\{0,1\}^n$7

with parity as the case $f:\{0,1\}^n\to\{0,1\}^n$8. In general graphs equipped with identifiers and radius $f:\{0,1\}^n\to\{0,1\}^n$9, the local complexity of $$\forall x\in\{0,1\}^n:\quad \mathsf{par}(x) = \mathsf{par}(f(x)).$$0 is $$\forall x\in\{0,1\}^n:\quad \mathsf{par}(x) = \mathsf{par}(f(x)).$$1. The construction chooses a root, stores each node’s distance to the root modulo $$\forall x\in\{0,1\}^n:\quad \mathsf{par}(x) = \mathsf{par}(f(x)).$$2, and records the size of its rooted subtree modulo $$\forall x\in\{0,1\}^n:\quad \mathsf{par}(x) = \mathsf{par}(f(x)).$$3. The parent relation is defined implicitly from local ID comparisons, and each node checks a modular subtree-sum equation; root nodes additionally check that their subtree size is $$\forall x\in\{0,1\}^n:\quad \mathsf{par}(x) = \mathsf{par}(f(x)).$$4 (Bousquet et al., 3 Jun 2026).

In anonymous graphs with radius $$\forall x\in\{0,1\}^n:\quad \mathsf{par}(x) = \mathsf{par}(f(x)).$$5, parity becomes harder. The paper proves that certifying even network size requires $$\forall x\in\{0,1\}^n:\quad \mathsf{par}(x) = \mathsf{par}(f(x)).$$6 certificate size, and more explicitly at least $$\forall x\in\{0,1\}^n:\quad \mathsf{par}(x) = \mathsf{par}(f(x)).$$7 bits on $$\forall x\in\{0,1\}^n:\quad \mathsf{par}(x) = \mathsf{par}(f(x)).$$8-vertex graphs. The lower bound uses powerset graphs and a finite unions theorem argument producing disjoint sets $$\forall x\in\{0,1\}^n:\quad \mathsf{par}(x) = \mathsf{par}(f(x)).$$9 and $\mathcal{M} = (G=(S,E), S_1, S_2, \delta),$0 with the same certificate color, after which two certified even graphs are merged into an odd graph whose radius-1 anonymous views remain accepting (Bousquet et al., 3 Jun 2026).

The lower bound does not persist on bounded-expansion classes. For every fixed bounded-expansion class $\mathcal{M} = (G=(S,E), S_1, S_2, \delta),$1, the local complexity of $\mathcal{M} = (G=(S,E), S_1, S_2, \delta),$2 in the anonymous radius-1 model is

$\mathcal{M} = (G=(S,E), S_1, S_2, \delta),$3

where $\mathcal{M} = (G=(S,E), S_1, S_2, \delta),$4 depends only on the class. A bounded-degree special case yields $\mathcal{M} = (G=(S,E), S_1, S_2, \delta),$5. The enabling tool is the use of conflict-free colorings to encode a parent with constant-size local information, which allows the same modular subtree aggregation argument to be executed without identifiers (Bousquet et al., 3 Jun 2026).

Taken together, these results show that certified parity synthesis ranges from strategy synthesis under sure and probabilistic guarantees, through permissive templates and assume-guarantee contracts in parity games, to specification-level circuit constructions, PB proof logs for XOR reasoning, and locally checkable certificates for parity of network size. The certificate objects differ—UGECs and VGECs, winning regions and templates, added truth-table bits, matrix traces, PB derivations, or node labels—but in each case the parity property is accompanied by a formally checkable witness rather than an unaudited synthesized artifact.