Fractional Hypertree Width

• Fractional Hypertree Width is defined by assigning each tree decomposition bag its minimal fractional edge-cover cost and minimizing the maximum over all bags.
• It serves as a tractability criterion for CSPs by supporting dynamic programming strategies that exploit bounded fractional cover numbers.
• Research demonstrates that fractional covers can outperform integral covers, leading to more efficient recognition and approximation algorithms in complex hypergraphs.

Fractional hypertree width, usually denoted $fhw$, is a hypergraph width measure defined by assigning to each bag of a tree decomposition the minimum fractional edge-cover cost of that bag and then minimizing the maximum such cost over all tree decompositions. In the structural theory of conjunctive queries and constraint satisfaction problems, it occupies a central position between integral edge-cover-based notions such as generalized hypertree width and more general semantic parameters such as submodular width. It is simultaneously a decomposition parameter, a tractability criterion, and a point of comparison in the study of exact recognition, approximation, semantic reformulation, and the boundary between syntactic and semantic structure (Grohe et al., 2017, Chen et al., 2020).

1. Formal definition

Let $H=(V(H),E(H))$ be a hypergraph. A tree decomposition of $H$ is a pair $(T,(B_u)_{u\in T})$ such that every hyperedge is contained in some bag and, for every vertex $v$, the set of nodes whose bags contain $v$ is connected in $T$. More generally, if $f:2^{V(H)}\to \mathbb{R}^+$, the $f$-width of a tree decomposition is

$\sup\{f(B_u)\mid u\in T\},$

and the $H=(V(H),E(H))$0-width of the hypergraph is the minimum of this quantity over all tree decompositions.

For a set $H=(V(H),E(H))$1, a fractional edge cover is a weight function $H=(V(H),E(H))$2 such that for every $H=(V(H),E(H))$3,

$H=(V(H),E(H))$4

where $H=(V(H),E(H))$5 is the set of hyperedges incident to $H=(V(H),E(H))$6. Its size is $H=(V(H),E(H))$7. The minimum size of a fractional edge cover of $H=(V(H),E(H))$8 is denoted $H=(V(H),E(H))$9; the corresponding minimum integral cover size is $H$0. Fractional hypertree width is then defined by

$H$1

that is,

$H$2

Equivalently, a fractional hypertree decomposition is a triple $H$3 where $H$4 is a tree decomposition, each $H$5, and $H$6; the width is $H$7 (Chen et al., 2020, Grohe et al., 2017).

A useful technical point is that the fractional guard $H$8 is a weight function on all hyperedges of $H$9, not merely on hyperedges induced by the bag. In practice, the cleanest reading of the definition is therefore the bag-wise one: each bag is measured by $(T,(B_u)_{u\in T})$0, and the decomposition is judged by the largest such value (Grohe et al., 2017).

2. Position in the width hierarchy

Fractional hypertree width belongs to a family of width notions obtained by varying the bag-cost function.

Notion	Bag cost	Definition
Primal treewidth	$(T,(B_u)_{u\in T})$1	$(T,(B_u)_{u\in T})$2
Generalized hypertree width	$(T,(B_u)_{u\in T})$3	$(T,(B_u)_{u\in T})$4
Fractional hypertree width	$(T,(B_u)_{u\in T})$5	$(T,(B_u)_{u\in T})$6
Submodular width	$(T,(B_u)_{u\in T})$7-width	$(T,(B_u)_{u\in T})$8

The standard hierarchy is

$(T,(B_u)_{u\in T})$9

Here $v$0 denotes hypertree width. Unlike $v$1 and $v$2, $v$3 is not expressible in the $v$4-width framework, because it uses the same edge-cover function as $v$5 but imposes an additional special condition on decompositions. Nevertheless, boundedness on classes coincides: $v$6 is bounded if and only if $v$7 is bounded. By contrast, there exist classes with bounded $v$8 and unbounded $v$9, so the fractional relaxation can be strictly more permissive than hypertree width on classes (Chen et al., 2020).

The comparison with integral guarding is not merely formal. The paper introducing the CSP-side decomposition view gives examples showing that $v$0 and the global fractional edge-cover number $v$1 are incomparable, while $v$2 subsumes both as a decomposition parameter. In particular, there are hypergraphs $v$3 with $v$4 but $v$5, showing that fractional covers can be exponentially more economical than integral covers. At the same time, $v$6 if and only if $v$7, so the acyclic case is unchanged (Grohe et al., 2017).

3. Tractability and algorithmic role

A basic algorithmic fact is that bounded fractional hypertree width yields tractability for CSP evaluation. One formulation states: if $v$8 is a class of CSP instances of bounded $v$9, then $T$0 is tractable. In decomposition terms, each bag has small fractional cover number, which supports dynamic-programming-style evaluation over a fractional hypertree decomposition (Chen et al., 2020).

The mechanism underlying this tractability is a sharp solution bound from fractional edge covers. If every constraint relation of a CSP instance $T$1 has at most $T$2 tuples, then

$T$3

This bound, proved via Shearer’s Lemma, implies that when a bag $T$4 satisfies $T$5, the bag-local solution set is polynomially bounded for fixed $T$6. The bounded-$T$7 algorithm therefore enumerates local solutions for every bag-induced instance and combines them by tree-decomposition dynamic programming. When a bounded-width fractional hypertree decomposition is given as input, this directly yields polynomial-time solvability; without such a decomposition, the 2017 paper invokes Marx’s approximation algorithm, obtaining a polynomial-time procedure for every fixed width bound $T$8, with runtime of the form $T$9 (Grohe et al., 2017).

Exact computation is harder. There is an exact exponential algorithm that computes fractional hypertree width in time $f:2^{V(H)}\to \mathbb{R}^+$0 on an $f:2^{V(H)}\to \mathbb{R}^+$1-vertex, $f:2^{V(H)}\to \mathbb{R}^+$2-edge hypergraph. Its method reduces hypergraph width computation to tree-decomposition optimization over the Gaifman graph, using minimal separators and potential maximal cliques, while evaluating each candidate bag by the linear program for $f:2^{V(H)}\to \mathbb{R}^+$3 (Moll et al., 2011).

A common misconception is that good algorithmic behavior of bounded $f:2^{V(H)}\to \mathbb{R}^+$4 should imply easy recognition for fixed $f:2^{V(H)}\to \mathbb{R}^+$5. In fact, exact recognition is hard in general: checking whether $f:2^{V(H)}\to \mathbb{R}^+$6 is NP-complete, and the same reduction also proves NP-completeness of checking $f:2^{V(H)}\to \mathbb{R}^+$7. Thus the tractability of bounded-width classes and the complexity of recognizing bounded width are sharply separated (Fischl et al., 2016).

4. Semantic width, cores, and the role of equivalence

Fractional hypertree width also plays a central role in the semantic theory of width. For a width notion $f:2^{V(H)}\to \mathbb{R}^+$8, semantic width is defined by

$f:2^{V(H)}\to \mathbb{R}^+$9

where semantic equivalence for CSP templates is homomorphic equivalence. For $f$0, the decisive result is core minimality: $f$1 Equivalently, on conjunctive queries,

$f$2

Thus semantic optimization for fractional hypertree width reduces exactly to passing to the core, rather than searching over infinitely many equivalent reformulations (Chen et al., 2020, Gottlob et al., 2018).

The proof uses a homomorphism from a structure onto its core together with a bag-projection argument: if $f$3 is a decomposition of the original hypergraph, one intersects each bag with the core vertex set, and a fractional cover of the original bag can be pushed forward through the homomorphism without increasing total weight. This establishes that the core never has larger $f$4 than the original formulation (Chen et al., 2020).

This semantic behavior matters because plain hypergraph width can substantially overestimate the real computational complexity of a problem class. The semantic-width papers emphasize examples where syntactically complicated templates are homomorphically equivalent to very small cores. The same phenomenon conceptually applies to $f$5: a given instance may have large syntactic fractional hypertree width while its semantic $f$6 is small (Gottlob et al., 2018, Chen et al., 2020).

At the same time, semantic $f$7 is not the final parameterized characterization of tractability. Assuming ETH, the exact FPT characterization for recursively enumerable classes of CSPs is bounded semantic submodular width, not bounded semantic fractional hypertree width. Since $f$8, bounded semantic $f$9 remains a sufficient condition, but it is not exact. The same pattern reappears for unions of conjunctive queries: the exact theorem is again in terms of semantic submodular width (Chen et al., 2020).

5. Structural frontier: exoticness, VC dimension, and sparse incidence structure

One of the most distinctive structural results about fractional hypertree width concerns its relation to hypertree width. A hypergraph class is called exotic if, for every integer $\sup\{f(B_u)\mid u\in T\},$0, it contains a hypergraph with an $\sup\{f(B_u)\mid u\in T\},$1-vertex subset $\sup\{f(B_u)\mid u\in T\},$2 such that the induced subhypergraph $\sup\{f(B_u)\mid u\in T\},$3 has at least $\sup\{f(B_u)\mid u\in T\},$4 distinct edges. The theorem is: if a class has unbounded hypertree width and bounded fractional hypertree width, then it is exotic. Equivalently, on non-exotic classes, bounded $\sup\{f(B_u)\mid u\in T\},$5 and bounded $\sup\{f(B_u)\mid u\in T\},$6 coincide. This formalizes the observation that known separations between fractional and ordinary hypertree width require a kind of exponential growth inside induced subhypergraphs (Chen et al., 2020).

The proof route passes through VC dimension. On classes of bounded VC dimension, the appendix establishes

$\sup\{f(B_u)\mid u\in T\},$7

Since non-exotic classes have bounded VC dimension, the distinction between fractional and integral edge-cover-based widths disappears there. This also leads to an algorithmic corollary: if a non-exotic class of CSPs has bounded semantic fractional hypertree width, then it is tractable (Chen et al., 2020).

A related comparison comes from sparse incidence structure. On hypergraphs whose incidence graphs lie in an apex-minor-free graph class, fractional hypertree width and incidence-graph treewidth are equivalent up to constant factors: $\sup\{f(B_u)\mid u\in T\},$8 For general minor-free incidence classes, direct equivalence with $\sup\{f(B_u)\mid u\in T\},$9 fails, but $H=(V(H),E(H))$00 still admits constant-factor approximation through the treewidth of suitable decomposition pieces. This places $H=(V(H),E(H))$01 in a second structural regime where it ceases to be dramatically more general than graph treewidth, namely on sufficiently sparse incidence classes (0809.3646).

Taken together, these results suggest a two-sided boundary. On the one hand, fractional hypertree width separates from hypertree width only on exotic classes. On the other hand, on apex-minor-free incidence classes it is essentially incidence-treewidth up to constants. The parameter is therefore most distinctive outside both of these collapse regimes.

6. Recognition complexity, restricted classes, and contemporary algorithms

After the NP-completeness of exact fixed-$H=(V(H),E(H))$02 recognition in general, a major research line has been to identify restricted hypergraph classes where $H=(V(H),E(H))$03 becomes algorithmically manageable. Exact recognition is polynomial-time decidable for fixed $H=(V(H),E(H))$04 on classes of bounded degree and on classes of bounded intersection; bounded rank is a corollary. On broader bounded multi-intersection classes, the 2020 support-compression result proves that every fractional edge cover of bounded weight can be replaced by one of the same weight bound and bounded support. This yields a polynomial-time algorithm for testing whether $H=(V(H),E(H))$05 on $H=(V(H),E(H))$06-hypergraphs for fixed $H=(V(H),E(H))$07 (Fischl et al., 2016, Gottlob et al., 2020).

Approximation and parameterized computation have advanced substantially. For bounded-rank hypergraphs, there is an FPT approximation algorithm parameterized by $H=(V(H),E(H))$08 and the rank $H=(V(H),E(H))$09 that either outputs a tree decomposition of fractional hypertree width at most $H=(V(H),E(H))$10 or correctly reports that $H=(V(H),E(H))$11. More recently, exact fixed-parameter tractability has been obtained on hypergraphs of bounded rank and bounded degree: there is an $H=(V(H),E(H))$12-time algorithm that computes an optimal-width decomposition whenever $H=(V(H),E(H))$13, giving the first exact FPT algorithm for fractional hypertree width under a nontrivial parameterization (Razgon, 2022, Lanzinger et al., 15 Jul 2025).

Approximation guarantees have also improved beyond the bounded-$H=(V(H),E(H))$14 regime. The 2024 paper gives a polynomial-time algorithm that, given a hypergraph $H=(V(H),E(H))$15 with $H=(V(H),E(H))$16, produces a tree decomposition of fractional hypertree width $H=(V(H),E(H))$17. Its second algorithm runs in time $H=(V(H),E(H))$18 and outputs a decomposition of width $H=(V(H),E(H))$19, improving both the approximation ratio and the running time over Marx’s earlier $H=(V(H),E(H))$20-time $H=(V(H),E(H))$21 approximation (Korchemna et al., 2024).

The broader complexity picture remains incomplete. Bounded $H=(V(H),E(H))$22 is a robust sufficient condition for polynomial-time tractability and a natural object of exact, approximate, and semantic analysis, but the full characterization of plain polynomial-time tractability for uniform CSP is still open. In the non-exotic regime, this motivates the conjecture that tractability coincides with bounded semantic hypertree width rather than with any purely fractional notion (Chen et al., 2020).