Partial Constraint Satisfaction Problems (PCSPs)

• Partial Constraint Satisfaction Problems (PCSPs) are promise-based problems that distinguish inputs with homomorphism mappings between two relational structures.
• They employ algebraic frameworks—using polymorphisms, minions, and minor conditions—to analyze problem tractability and guide reduction techniques.
• PCSPs are applied in areas like approximate graph coloring, satisfiability variants, and compiler optimization, revealing diverse algorithmic paradigms and hardness mechanisms.

The acronym PCSP has two distinct uses. In the recent algebraic-complexity literature, it denotes a Promise Constraint Satisfaction Problem: for similar relational structures $A,B$ with a homomorphism $A \to B$, the task is to distinguish inputs $X$ with $X \to A$ from inputs with $X \nrightarrow B$, under the promise that one of these alternatives holds (Barto et al., 2018). In a separate optimization literature, especially on soft constraints and compiler analysis, PCSP refers to a partial or weighted constraint problem in which violations are permitted at a cost (Cai et al., 3 Feb 2026). The modern theoretical theory is centered on the promise interpretation, where polymorphisms, minions, minor conditions, and homomorphism “sandwiches” organize both tractability and hardness (Barto, 2019).

1. Formal model and basic examples

For a fixed relational structure $A$ over a finite signature, $\mathrm{CSP}(A)$ asks, given a structure $X$ of the same signature, whether there exists a homomorphism $X \to A$ (Deng et al., 2020). A promise template is a pair $(A,B)$ of similar structures together with a homomorphism $A \to B$0, and the associated problem is

$A \to B$1

under the promise that exactly one of these cases holds (Deng et al., 2020). The special case $A \to B$2 recovers ordinary CSP.

A canonical example is approximate graph coloring: one asks to distinguish graphs admitting a homomorphism to the clique $A \to B$3 from graphs not even admitting a homomorphism to $A \to B$4, with $A \to B$5 (Deng et al., 2020). Other standard examples include promise versions of satisfiability, hypergraph coloring, and Boolean templates such as $A \to B$6 (Barto, 2019).

The recent literature consistently emphasizes that PCSP in this sense is a promise problem rather than a partial-satisfaction problem (Barto et al., 2018). This distinction matters because the main algebraic invariants are not optimization weights or penalties but homomorphisms between two templates and the closure properties of operations from $A \to B$7 to $A \to B$8.

2. Algebraic framework: polymorphisms, minions, and minor conditions

The algebraic theory replaces direct analysis of instances by analysis of polymorphisms. For a template pair $A \to B$9, an $X$0-ary polymorphism is a homomorphism $X$1; equivalently, $X$2 maps tuples from each relation of $X$3 coordinatewise into the corresponding relation of $X$4 (Barto et al., 2018). The family $X$5 is closed under taking minors, so it forms a minion rather than, in general, a clone (Barto, 2019). This is a key difference from ordinary CSP: composition is typically ill-defined for $X$6 operations, whereas taking minors remains meaningful.

Minion homomorphisms are the basic reduction mechanism. If there is a minion homomorphism

$X$7

then $X$8 log-space reduces to $X$9 (Barto et al., 2018). This extends the classical algebraic approach from clones and pp-interpretations to the promise setting, and it is one of the central structural results of the subject (Barto, 2019).

The same theory admits a reformulation in terms of bipartite minor conditions and an algebraic version of Label Cover. Every fixed-template PCSP is log-space equivalent to a promise minor-condition satisfiability problem for its polymorphism minion (Barto et al., 2018). From a categorical viewpoint, the polymorphism object itself can be described as a right Kan extension, and standard gadget constructions become adjunctions between presheaf categories (Hadek et al., 13 Mar 2025). This categorical reframing does not change the computational content, but it clarifies why “complexity depends only on polymorphisms” and why minion homomorphisms subsume gadget reductions (Hadek et al., 13 Mar 2025).

3. Sandwiches, finite templates, and the necessity of infinite intermediates

A major organizing principle is the sandwich paradigm. If there exists a structure $X \to A$0 with homomorphisms

$X \to A$1

then $X \to A$2 reduces, by the identity map on instances, to $X \to A$3 (Deng et al., 2020). Many tractable PCSPs known at present are solved in exactly this way (Kazda et al., 2021).

Affine sandwiches are particularly important. If $X \to A$4 is affine—equivalently, its relations are cosets of subgroups and $X \to A$5 is a polymorphism—then $X \to A$6 reduces to solving linear equations and is polynomial-time tractable (Deng et al., 2020). This yields explicit linear-algebraic algorithms and also forces rich symmetry in $X \to A$7: affine sandwiches imply block-symmetric polymorphisms of arbitrarily large width, and finite affine sandwiches imply symmetric polymorphisms of arbitrarily large arity (Deng et al., 2020).

Finite sandwiches, however, need not be small. A Boolean PCSP was exhibited that admits a tractable affine sandwich on a 3-element domain but no tractable Boolean sandwich, assuming $X \to A$8 (Deng et al., 2020). More strongly, for every integer $X \to A$9 and every prime $X \nrightarrow B$0, there are finite templates $X \nrightarrow B$1 of size $X \nrightarrow B$2 for which $X \nrightarrow B$3 reduces to a tractable $X \nrightarrow B$4 with $X \nrightarrow B$5, but to no tractable $X \nrightarrow B$6 with $X \nrightarrow B$7 (Kazda et al., 2021). Thus the minimal tractable intermediate domain can be arbitrarily larger than the original ones.

The phenomenon can be even more extreme. For the template $X \nrightarrow B$8, polynomial-time solvability is available via a natural reduction to an infinite-domain CSP over linear equations, but any finite-domain CSP obtained by the same natural pp-construction route is NP-complete unless $X \nrightarrow B$9 (Barto, 2019). This result initiated the systematic study of finite intractability. Subsequent work characterized finite tractability for several symmetric Boolean families and showed that, in many tractable PCSPs, infinity is not an artifact of current proofs but an intrinsic feature of the sandwich method (Asimi et al., 2020).

There are also structural exceptions. For undirected graphs $A$0, if $A$1 reduces to a tractable finite digraph CSP $A$2, then already $A$3 or $A$4 is tractable (Kazda et al., 2021). This rules out the “small template requiring only a larger tractable intermediate graph” phenomenon in the graph setting.

4. Algorithmic paradigms and tractable polymorphism signatures

Several major algorithmic frameworks for PCSPs are known, each characterized by a specific symmetry pattern in $A$5.

Before the summary table, two facts are fundamental. First, BLP solves a PCSP exactly when the template has symmetric polymorphisms of all arities (Barto et al., 2018). Second, AIP solves a PCSP exactly when the template has alternating polymorphisms of all odd arities (Barto et al., 2018). These are among the cleanest tractability theorems in the area.

Paradigm	Algebraic characterization in the literature	Consequence
BLP	Symmetric polymorphisms of all arities	Polynomial-time solvability (Barto et al., 2018)
AIP	Alternating polymorphisms of all odd arities	Polynomial-time solvability (Barto et al., 2018)
BLP+AIP	2-block-symmetric polymorphisms of all odd arities	Polynomial-time solvability (Ciardo et al., 2021)
CLAP	Minion homomorphism from the CLAP minion; sufficient condition via $A$6-symmetric operations of arbitrarily large arity	Extends BLP+AIP and finite-domain reductions (Ciardo et al., 2021)
Bounded width	Width-1 corresponds to totally symmetric operations of all arities; bounded width implies WNUs of all large arities	Local-consistency algorithms (Barto et al., 2018, Atserias et al., 2021)
Robust SDP	Majority or alternating-threshold polymorphisms	Robust satisfaction algorithms (Brakensiek et al., 2022)

These paradigms do not collapse to a single notion in the promise setting. In CSP, bounded width and several LP-based criteria are closely aligned; in PCSP, they separate. Every bounded-width PCSP has weak near-unanimity polymorphisms of all large arities, but the converse fails, and there are PCSPs solvable already at level $A$7 of Sherali–Adams that do not have bounded or even sublinear width (Atserias et al., 2021). This is one of the sharpest contrasts with ordinary CSP.

A second recent development is CLAP, a combination of constraint-level LP propagation with affine IP. Its power is characterized by a minion of skeletal stochastic-affine matrices, and a sufficient condition is the existence of $A$8-symmetric polymorphisms of arbitrarily large arity (Ciardo et al., 2021). CLAP strictly extends both BLP+AIP and reduction to tractable finite-domain CSPs (Ciardo et al., 2021).

On the optimization side, the SDP framework has been extended from CSPs to PCSPs. For Boolean folded PCSPs, majority polymorphisms yield a randomized robust algorithm achieving weak satisfaction at least

$A$9

while alternating-threshold polymorphisms yield

$\mathrm{CSP}(A)$0

on instances that are $\mathrm{CSP}(A)$1-strongly satisfiable (Brakensiek et al., 2022). The same work conjectures that robust algorithms exist exactly for those PCSPs decided by the canonical basic SDP (Brakensiek et al., 2022).

5. Hardness mechanisms and dichotomy results

The hardness theory of PCSPs is comparably rich. One central theme is the search for a dichotomy analogous to Feder–Vardi and Schaefer, but now in the promise setting.

For symmetric Boolean PCSPs with negations, Brakensiek and Guruswami established a dichotomy: tractability occurs when the polymorphism minion contains, for all odd arities, one of the parity, majority, or alternating-threshold families (or their negations); otherwise the problem is NP-hard (Brakensiek et al., 2017). Ficak, Kozik, Olšák, and Stankiewicz then removed the “allows negations” assumption and proved a dichotomy for all symmetric Boolean templates, with tractable side characterized by the presence of constants, $\mathrm{CSP}(A)$2, $\mathrm{CSP}(A)$3, alternating thresholds, XOR, or rational threshold families and their complements (Ficak et al., 2019).

Specialized Boolean regimes admit finer statements. For ordered Boolean PCSPs, where the template contains the predicate $\mathrm{CSP}(A)$4 and polymorphisms are monotone Boolean functions, a conditional dichotomy was proved under the Rich 2-to-1 Conjecture: the tractable side is characterized by the existence, for every $\mathrm{CSP}(A)$5, of monotone polymorphisms whose coordinates all have Shapley value at most $\mathrm{CSP}(A)$6, equivalently by threshold minors of arbitrarily large arity (Brakensiek et al., 2021).

The standard PCP-based hardness machinery has also been strengthened. A template was identified that is NP-hard but escapes the earlier layered-choice hardness condition; to handle it, an injective layered choice condition based on the smooth layered PCP theorem was introduced (Banakh et al., 2024). The same paper proves a P/NP-hard dichotomy for Boolean PCSPs whose polymorphisms lie in the minion of linear threshold functions, with the boundary given by a “low-weight” condition on threshold representations (Banakh et al., 2024).

Topological and combinatorial hardness methods have become prominent as well. A discrete-homotopy framework constructs minion homomorphisms from $\mathrm{CSP}(A)$7 to polymorphisms over discrete fundamental groups of simplicial complexes attached to the template, yielding NP-hardness whenever the resulting group is Abelian-bounded and a non-degeneracy condition holds (Beikmohammadi et al., 12 Oct 2025). This recovers known hardness results such as $\mathrm{CSP}(A)$8 and gives further one-dimensional homotopy-based hardness theorems (Beikmohammadi et al., 12 Oct 2025).

Approximate graph coloring remains a major benchmark. The algebraic framework was used to prove that for any $\mathrm{CSP}(A)$9, it is NP-hard to find a $X$0-coloring of a given $X$1-colorable graph (Barto et al., 2018). At the same time, width-based methods are sharply limited here: for all $X$2, $X$3 is not solvable in bounded or sublinear width (Atserias et al., 2021).

6. Terminological variants, optimization-oriented PCSPs, and open problems

A separate literature uses PCSP in a genuinely partial or soft-constraint sense. In possibilistic CSP, constraints carry necessity degrees, solutions are ranked by a possibility distribution on labelings, and optimal labelings maximize the induced possibility $X$4 or, equivalently, minimize the maximum violated necessity degree (Schiex, 2013). In compiler optimization, PCSP over a control-flow graph assigns one variable to each vertex, allows constraint violations at an edge-local cost $X$5, and seeks

$X$6

with an $X$7 dynamic program on Series-Parallel-Loop decompositions of structured control-flow graphs (Cai et al., 3 Feb 2026). These are substantive research directions, but they are conceptually different from the promise-template theory that dominates current algebraic work.

The promise theory also now interacts strongly with infinite-domain CSP. Recent results show that any non-trivially tractable template within the Bodirsky–Pinsker program can, after a Datalog-computable modification, serve as an infinite sandwich witnessing tractability of a finite-domain PCSP that is not finitely tractable (Pinsker et al., 10 Feb 2025). This suggests that the border between finite-domain promise problems and infinite-domain CSP techniques is structural rather than accidental.

Several open problems recur across the literature. One is the existence of a general Boolean or finite-domain PCSP dichotomy (Barto et al., 2018). Another is the decidability of affine sandwiches and the computability of the minimal finite sandwich size $X$8 (Deng et al., 2020). A third concerns the boundary between finite and infinite tractability, including whether stronger necessary and sufficient conditions can be stated directly at the level of minion identities (Asimi et al., 2020). On the algorithmic side, the conjecture that robust satisfiability is exactly characterized by the basic SDP remains open (Brakensiek et al., 2022). On the hardness side, the persistence of unresolved cases in approximate coloring and in non-symmetric Boolean families indicates that current PCP, algebraic, and topological methods are still incomplete.

Taken together, these developments show that PCSPs form a distinct layer above ordinary CSPs: they retain the homomorphism-based formalism of CSP, but their tractability may require larger or infinite sandwiches, their algebra is organized by minions rather than clones, and their algorithmic frontier spans local consistency, LP, affine and semidefinite relaxations, categorical constructions, and topological obstructions (Barto, 2019).