D-regular Max-k-XORSAT

• D-regular Max-k-XORSAT is an optimization problem that maximizes satisfied parity constraints in k-uniform instances where each variable appears in exactly D constraints.
• It integrates methodologies from bounded-degree approximation, LDPC decoding, and quantum optimization, with rigorous analyses via QAOA on large-girth regular hypergraphs.
• The framework reveals structural insights like core clustering, overlap-gap barriers, and algorithmic limitations, guiding both classical and quantum decoding approaches.

D-regular Max-k-XORSAT is the optimization problem of maximizing the number of satisfied parity constraints in a $k$-uniform instance where each variable appears in exactly $D$ constraints. In the $\{\pm1\}$ hypergraph formulation often used in the regular setting, with signs $J_{i_1\ldots i_k}\in\{\pm1\}$, the cost function is

$$C^{\text{XOR}}_J(z)=\sum_{(i_1,\dots,i_k)\in E} \frac12\bigl(1 + J_{i_1\ldots i_k}\, z_{i_1}\cdots z_{i_k}\bigr),$$

and MaxCut is recovered at $k=2$ with all $J_{ij}=-1$ (Basso et al., 2021). The subject sits at the intersection of bounded-degree approximation, sparse linear algebra over $\mathbb F_2$, random regular hypergraph geometry, LDPC decoding, and quantum optimization. In the bounded-degree worst case, the optimal additive scale above random assignment is $1/\sqrt D$ (Barak et al., 2015, Kramer et al., 11 Jun 2026); in average-case regular ensembles, the problem admits detailed analyses via QAOA and via reductions to decoding of regular LDPC codes (Basso et al., 2021, Shutty et al., 27 Apr 2026).

1. Formal model and regular-ensemble conventions

In the standard Boolean optimization form, Max-k-XORSAT is a system $Bx=v \pmod 2$ with $D$0, each row having Hamming weight $D$1, and objective

$D$2

equivalently maximizing the number of satisfied constraints $D$3. In a $D$4-regular instance, every row has exactly $D$5 ones and every column has exactly $D$6 ones, so $D$7; the associated Tanner graph is $D$8-biregular. In the LDPC language used for Gallager ensembles, the primal code is $D$9 and the dual code is $\{\pm1\}$0 (Shutty et al., 27 Apr 2026).

Two regular conventions coexist in the literature. In bounded-degree approximation results, $\{\pm1\}$1 is the variable-degree bound itself, so a D-regular instance is the special case $\{\pm1\}$2 for all variables (Barak et al., 2015). In the large-girth QAOA analysis, the hypergraph is $\{\pm1\}$3-regular, with $\{\pm1\}$4, and the cost Hamiltonian is normalized by $\{\pm1\}$5; the resulting formulas therefore use $\{\pm1\}$6 as the asymptotic degree parameter even though the literal vertex degree is $\{\pm1\}$7 (Basso et al., 2021). The distinction is purely notational but matters when comparing formulas across papers.

On random regular instances, the dominant ensemble in the decoding literature is Gallager’s $\{\pm1\}$8 ensemble, which produces a random $\{\pm1\}$9-regular bipartite graph and thereby a regular LDPC code (Shutty et al., 27 Apr 2026). In the QAOA literature, the dominant ensemble is instead a large-girth regular hypergraph, where girth $J_{i_1\ldots i_k}\in\{\pm1\}$0 ensures that every depth-$J_{i_1\ldots i_k}\in\{\pm1\}$1 neighborhood is a $J_{i_1\ldots i_k}\in\{\pm1\}$2-ary $J_{i_1\ldots i_k}\in\{\pm1\}$3-uniform hypertree (Basso et al., 2021).

2. Solution-space geometry, cores, and clustering

The sharpest structural results currently available concern the standard random $J_{i_1\ldots i_k}\in\{\pm1\}$4-XORSAT ensemble with fixed clause size $J_{i_1\ldots i_k}\in\{\pm1\}$5 and asymptotically Poisson variable degrees, not D-regular Max-k-XORSAT directly. In that model, the exact solution set $J_{i_1\ldots i_k}\in\{\pm1\}$6 undergoes a sharp clustering transition at $J_{i_1\ldots i_k}\in\{\pm1\}$7, and this threshold coincides with the 2-core threshold. Below $J_{i_1\ldots i_k}\in\{\pm1\}$8, the entire solution set has large conductance under polylogarithmic Hamming moves; above $J_{i_1\ldots i_k}\in\{\pm1\}$9, the solution space decomposes into exponentially many clusters separated by linear Hamming distance, while each individual cluster still has large internal conductance because it admits a polylogarithmically sparse basis (Ibrahimi et al., 2011).

That picture is not a theorem for D-regular Max-k-XORSAT, but it is the natural structural template. In a D-regular family, the density parameter is $$C^{\text{XOR}}_J(z)=\sum_{(i_1,\dots,i_k)\in E} \frac12\bigl(1 + J_{i_1\ldots i_k}\, z_{i_1}\cdots z_{i_k}\bigr),$$0, the incidence graph is locally tree-like with fixed offspring rather than Poisson offspring, and the same notions of 2-core, backbone, and periphery remain meaningful. This suggests that a corresponding regular-core threshold should control the onset of shattering, with exact solutions or near-optimal assignments organized around core configurations and extensive periphery degrees of freedom (Ibrahimi et al., 2011).

A closely related algorithmic phenomenon is known for sequential local algorithms in the Poisson-degree random $$C^{\text{XOR}}_J(z)=\sum_{(i_1,\dots,i_k)\in E} \frac12\bigl(1 + J_{i_1\ldots i_k}\, z_{i_1}\cdots z_{i_k}\bigr),$$1-XORSAT model. There, for any $$C^{\text{XOR}}_J(z)=\sum_{(i_1,\dots,i_k)\in E} \frac12\bigl(1 + J_{i_1\ldots i_k}\, z_{i_1}\cdots z_{i_k}\bigr),$$2, sequential local algorithms with certain local rules fail with high probability; this includes Unit Clause Propagation for $$C^{\text{XOR}}_J(z)=\sum_{(i_1,\dots,i_k)\in E} \frac12\bigl(1 + J_{i_1\ldots i_k}\, z_{i_1}\cdots z_{i_k}\bigr),$$3 and local rules that compute exact marginals on trees, including Belief Propagation and Survey Propagation, for $$C^{\text{XOR}}_J(z)=\sum_{(i_1,\dots,i_k)\in E} \frac12\bigl(1 + J_{i_1\ldots i_k}\, z_{i_1}\cdots z_{i_k}\bigr),$$4 (Yung, 2024). A plausible implication is that once a regular 2-core with analogous overlap-gap structure emerges, local decimation-based methods should encounter the same topological barrier in D-regular ensembles.

3. Worst-case approximability at bounded degree

For bounded-degree Max-k-XOR, the central worst-case fact is that degree control changes the approximation landscape. For any odd $$C^{\text{XOR}}_J(z)=\sum_{(i_1,\dots,i_k)\in E} \frac12\bigl(1 + J_{i_1\ldots i_k}\, z_{i_1}\cdots z_{i_k}\bigr),$$5, there is a randomized polynomial-time algorithm that, on any instance of Max-kXOR with degree at most $$C^{\text{XOR}}_J(z)=\sum_{(i_1,\dots,i_k)\in E} \frac12\bigl(1 + J_{i_1\ldots i_k}\, z_{i_1}\cdots z_{i_k}\bigr),$$6, finds an assignment satisfying at least

$$C^{\text{XOR}}_J(z)=\sum_{(i_1,\dots,i_k)\in E} \frac12\bigl(1 + J_{i_1\ldots i_k}\, z_{i_1}\cdots z_{i_k}\bigr),$$7

where $$C^{\text{XOR}}_J(z)=\sum_{(i_1,\dots,i_k)\in E} \frac12\bigl(1 + J_{i_1\ldots i_k}\, z_{i_1}\cdots z_{i_k}\bigr),$$8; since a D-regular instance is a special case of degree at most $$C^{\text{XOR}}_J(z)=\sum_{(i_1,\dots,i_k)\in E} \frac12\bigl(1 + J_{i_1\ldots i_k}\, z_{i_1}\cdots z_{i_k}\bigr),$$9, this applies directly to D-regular Max-k-XORSAT (Barak et al., 2015). The same paper shows that the $k=2$0 dependence is optimal in general: already for Max-2XOR on a D-regular graph with random signs, with high probability every assignment has value at most $k=2$1 (Barak et al., 2015).

The bounded-degree hardness theory now matches this scaling. For any finite field $k=2$2, any $k=2$3, any $k=2$4, and sufficiently large $k=2$5, it is NP-hard under randomized reductions to approximate Max-Ek-LINSAT$k=2$6 of degree at most $k=2$7 within

$k=2$8

Specializing to Boolean parity constraints gives bounded-degree Max-Ek-XORSAT hardness beyond

$k=2$9

so the optimal worst-case dependence on $J_{ij}=-1$0 is $J_{ij}=-1$1 (Kramer et al., 11 Jun 2026).

The direct consequence for D-regular Max-k-XORSAT is that regularity does not improve the asymptotic approximation exponent in the worst case: any polynomial-time classical or quantum algorithm is confined to the $J_{ij}=-1$2 scale, and any genuine advantage can only appear in the constant prefactor. For Boolean instances this matches the algorithmic upper bound up to the $J_{ij}=-1$3-dependent constant; for general $J_{ij}=-1$4, the hardness extends but the corresponding $J_{ij}=-1$5 algorithmic guarantee is not yet known (Kramer et al., 11 Jun 2026).

4. QAOA on large-girth regular hypergraphs

The most explicit direct analysis of D-regular-type Max-k-XORSAT in the quantum optimization literature is the study of QAOA on large-girth regular hypergraphs. There the instance is Max-$J_{ij}=-1$6-XORSAT on a $J_{ij}=-1$7-regular $J_{ij}=-1$8-uniform hypergraph, with cost

$J_{ij}=-1$9

scaled Hamiltonian

$\mathbb F_2$0

and standard mixer $\mathbb F_2$1 (Basso et al., 2021).

Because girth $\mathbb F_2$2, every depth-$\mathbb F_2$3 neighborhood of a hyperedge is a $\mathbb F_2$4-ary $\mathbb F_2$5-uniform hypertree, and a sign-gauge argument shows that on such a hypertree the QAOA expectation is independent of the pattern of clause signs $\mathbb F_2$6. The expected satisfied-clause fraction therefore reduces to a single-hyperedge quantity: $\mathbb F_2$7 The paper derives an exact finite-$\mathbb F_2$8 hypertree recursion for $\mathbb F_2$9 with time complexity $1/\sqrt D$0 and memory $1/\sqrt D$1, and an infinite-$1/\sqrt D$2 matrix iteration with time $1/\sqrt D$3 and memory $1/\sqrt D$4 (Basso et al., 2021).

For Max-$1/\sqrt D$5-XORSAT numerics, the infinite-$1/\sqrt D$6 iteration was optimized up to $1/\sqrt D$7 and $1/\sqrt D$8. The resulting $1/\sqrt D$9 increases with $Bx=v \pmod 2$0. For even $Bx=v \pmod 2$1, however, locality and large girth impose a limitation: because $Bx=v \pmod 2$2, one remains in the regime $Bx=v \pmod 2$3, and overlap-gap results imply that $Bx=v \pmod 2$4 does not converge to 1 as $Bx=v \pmod 2$5 in this regular large-girth setting (Basso et al., 2021). In that sense, the paper identifies both an exact calculational framework and a structural obstruction.

5. LDPC decoding, Regev reduction, and locally-quantum decoders

A different direct route to D-regular Max-k-XORSAT views the problem as nearest-codeword search in a regular LDPC code. For a $Bx=v \pmod 2$6-regular constraint matrix $Bx=v \pmod 2$7 and right-hand side $Bx=v \pmod 2$8, the objective is

$Bx=v \pmod 2$9

so maximizing satisfied constraints is equivalent to finding the codeword $D$00 closest to $D$01. Regev’s reduction converts this optimization task into decoding the dual code $D$02 under a coherent superposition of bit-flip errors, and thereby into a decoding problem on a regular LDPC code with parity-check matrix $D$03 (Shutty et al., 27 Apr 2026).

On Gallager’s $D$04 ensemble, the locally-quantum-decoding paper exploits a special block structure: there exists a subset of checks whose supports partition the $D$05 constraint bits into disjoint blocks of size $D$06. A blockwise “Fine-Grained Unambiguous Measurement” (FGUM) is then applied inside Regev’s reduction. The resulting effective channel is a block erasure channel, and after classical recovery on the erased blocks one obtains an expected satisfied fraction

$D$07

where $D$08 is the maximal block-erasure rate still decodable with high probability on the random Gallager code, and $D$09 is the average Hamming weight of a specially chosen coset-leader set $D$10 (Shutty et al., 27 Apr 2026).

This decoder strongly outperforms classical belief propagation on the same instances, and for some $D$11 the approximate optima achieved by the quantum decoder surpass both Prange’s algorithm and simulated annealing. However, the same work constructs an enhanced classical algorithm, “Turbo Prange”, which exactly matches the asymptotic performance of Regev+FGUM on Gallager-ensemble D-regular Max-k-XORSAT. In the large-$D$12 regime, both achieve

$D$13

so the locally-quantum decoder improves substantially over BP and standard Prange yet stops short of a proven quantum advantage (Shutty et al., 27 Apr 2026).

6. Annealing, regular embeddings, and broader algorithmic implications

Regular parity systems also expose a sharp distinction between algebraic tractability and local-search tractability. In a planar embedding of k-regular k-XORSAT built from reversible logic gates on a square grid, the resulting classical model has no finite-temperature phase transition, but thermal relaxation into the ground state remains glassy. Under quantum annealing, the same embedding supports a second-order quantum phase transition for $D$14 and a first-order transition for $D$15, corresponding respectively to polynomial and exponential scaling of the minimum gap and thus of time-to-solution (Patil et al., 2019).

That result is about k-regular k-XORSAT rather than D-regular Max-k-XORSAT, so any transfer is interpretive. The reliable lesson is that even when a parity CSP is solvable in polynomial time by Gaussian elimination, regular sparse embeddings can generate slow local dynamics, mixed-boundary-condition effects, and first-order quantum transitions under standard transverse-field interpolation. This suggests that D-regular Max-k-XORSAT should not be viewed as annealing-friendly merely because its underlying algebra is linear (Patil et al., 2019).

Taken together, the current picture is unusually coherent. Worst-case bounded-degree theory fixes the correct approximation scale at $D$16 (Barak et al., 2015, Kramer et al., 11 Jun 2026). Large-girth regular analyses show how finite-depth QAOA is controlled by a hypertree recursion and constrained by locality (Basso et al., 2021). Gallager-ensemble studies show that D-regular Max-k-XORSAT is naturally an LDPC decoding problem, enabling both strong classical heuristics and intrinsically quantum local measurements, while also revealing how hard it is to turn those quantum ingredients into a strict advantage (Shutty et al., 27 Apr 2026). The remaining open terrain is therefore not the exponent of $D$17, but the exact constants, the role of regular-core geometry in near-optimal states, and whether any nonclassical decoder or nonlocal driver can outperform the best structured classical algorithms on the same regular instances.