Existential Positive Sparsification Conjecture

• The Existential Positive Sparsification Conjecture is a framework connecting finite model theory and convex geometry to characterize when sparse graph classes transform into dense ones via existential positive transductions.
• It utilizes combinatorial techniques like flips and subflips to construct sparse subgraphs while preserving structural parameters such as bounded tree-width and clique-width.
• In the convex setting, EPSC predicts tight sparsification bounds for sums in convex cones, generalizing spectral sparsification results to broader classes including PSD and hyperbolicity cones.

The Existential Positive Sparsification Conjecture (EPSC) collects recent advances in the theory of logical and geometric sparsification. The EPSC connects structural graph theory, finite model theory, and convex geometry, predicting broad structural sparsifiable phenomena for both graph classes and sums in convex cones. In logic, it characterizes when dense graph classes can be obtained by existential positive first-order transductions from sparse classes; in convex geometry, it predicts that sums from cones with good barrier structure always admit small positive sparsifiers. The conjecture has received independent formulation and supporting theorems in the contexts of graph theory (Mählmann et al., 22 Jan 2026) and convex analysis (Saunderson, 26 Dec 2025).

1. Formal Definition and Logical Framework

The EPSC in the model-theoretic/graph-theoretic context asserts an equivalence between two structural properties of graph classes. Let $\mathcal C$ be a class of reflexive graphs:

1. $\mathcal C$ is semi-ladder-free (equivalently co-matching-free) and monadically dependent (equivalently monadically stable and half-graph-free).
2. $\mathcal C$ is the existential positive (EP) first-order transduction of some nowhere dense class $\mathcal D$ of reflexive graphs.

An EP first-order formula uses only existential quantifiers, conjunction, disjunction, and equality (but neither negation, nor universal quantifiers, nor non-equality). An EP-transduction $T$ consists of coloring operations and an interpreting pair $(\nu(x), \eta(x,y))$ of EP-formulas, defining a class as $T(G)=\{\mathsf I(H) : H \in \Gamma_\sigma(G)\}$, where $\Gamma_\sigma(G)$ is the set of all vertex colorings of $G$ with palette $\sigma$. The reflexivity condition ensures that EP-formulas cannot distinguish through the absence of self-loops.

The EPSC predicts that the semi-ladder-free monadically stable graph classes are exactly those obtainable from the sparsest classes—nowhere dense classes—using EP-transductions (Mählmann et al., 22 Jan 2026).

2. Key Notions: Structural and Logical Dependence

Several concepts are central to the EPSC:

• Monadic dependence/stability: A class $\mathcal C$ is monadically dependent if it does not EP-transduce the class of all graphs; monadically stable if further no first-order formula exhibits the order property (cannot interpret linear orderings).
• Co-matching-free / semi-ladder-free: A class is co-matching-free if it omits arbitrarily large bipartite complement matchings; equivalently, it avoids large semi-induced co-matchings and, combined with monadic dependence, characterizes the minimal structure required for the conjecture.

These properties lie at the intersection of stability theory and the finite model theoretic “tameness” hierarchy, referencing classes like nowhere dense graphs, which are characterized by the nonexistence of large cliques as shallow minors and minimal combinatorial complexity.

3. Existential Sparsification in Convex Cones

The geometric analog of the EPSC concerns sparsification of sums in convex cones. Given $x_1,\dots,x_m$ in a closed convex cone $K \subseteq \mathbb R^n$ with $e = \sum_i x_i \in \operatorname{relint} K$, an $\epsilon$-sparsifier is a reweighted sum $\sum_{i \in S} \lambda_i x_i$ that approximates $e$ in the $K$-partial order: $$(1-\epsilon)e \le_K \sum_{i \in S} \lambda_i x_i \le_K (1+\epsilon)e$$ The sparsification function $\varphi_K(\epsilon)$ records, in the worst case over all such configurations, the minimal support size $|S|$ of an $\epsilon$-sparsifier. The conjecture predicts that for all proper cones $K$ admitting a $\nu$-logarithmically homogeneous self-concordant barrier, $\varphi_K(\epsilon)$ is $O(\nu/\epsilon^2)$, and in cones with the additional “pairwise” self-concordant barrier property—such as the positive semidefinite (PSD) or hyperbolicity cones—$$\varphi_K(\epsilon) \le \lceil 4\nu/\epsilon^2\rceil$$ (Saunderson, 26 Dec 2025).

This recovers and generalizes the Batson–Spielman–Srivastava linear-sized spectral sparsification bound for PSD matrices to much broader classes of cones.

4. Main Theorems and Structural Results

For logical sparsification, all known semi-ladder-free, monadically stable graph classes with bounded structural width (tree-depth, clique-width, twin-width, merge-width) admit explicit EP-transductions to sparse subgraphs inside the original dense graphs. Theorems confirm that for such classes, EP-transductions $(\Sparsify, \Recover)$ can construct reflexive sparse subgraphs $G^*\subseteq G$ possessing stronger “width” properties, such as bounded tree-width or tree-depth. Quantitative bounds show that the sparsified preimages—often subgraphs—have structure parameters controlled by computable functions of the original class parameters.

In the convex setting, two main existential bounds hold for the sparsification function $\varphi_K(\epsilon)$:

• General proper cones (Theorem A): $$\varphi_K(\epsilon) \le \lceil (4\nu/\epsilon)^2\rceil$$
• Cones with pairwise self-concordant barriers (Theorem B): $$\varphi_K(\epsilon) \le \lceil 4\nu/\epsilon^2\rceil$$

Prominent examples include:

• $K = S_+^d$ (PSD matrices with $F(X) = -\log\det X$, $\nu=d$): matches span of Batson–Spielman–Srivastava bound.
• Hyperbolicity cones (Güler's theory): $\nu=d$, with pairwise property, yield the same $O(d/\epsilon^2)$ bound (Saunderson, 26 Dec 2025).

5. Combinatorial and Proof Techniques

Two combinatorial operations underpin the logical branch of EPSC: flips and the newly introduced subflip. A flip $G \mapsto G\oplus(P,F)$ complements edges between partition classes in $F$. The subflip $G\ominus P$ complements edges only between “fully adjacent” pairs, thus generating subgraphs. The property of subflip-flatness characterizes the co-matching-free classes: a class is co-matching-free plus monadically dependent if and only if it is $r$-subflip-flat for all $r$.

Proof strategies combine partition-refinement lemmas showing flips can be approximated by subflips (with budget increase controlled by the co-matching index), and construction of sparse preimages by iteratively removing layers of small neighborhood or separator covers via subflips.

In the geometric context, proofs fall into two categories:

• Frank–Wolfe/barrier-norm arguments: Reduction to convex combinations inside the cone and application of convergence properties of the Frank–Wolfe algorithm (curvature-based iteration bounds).
• BSS-style barrier-guided potentials: Barrier-based potential functions guide a greedy or inductive selection process, ensuring the reweighted sum stays within a prescribed band in the cone’s partial order.

Key lemmas establish stepwise control on potentials, Hessian inner products, and combinatorial existence required for the inductive construction of sparsifiers (Saunderson, 26 Dec 2025).

6. Strengthened Corollaries and Expressiveness

Further implications include strong structural decompositions for dense graphs with bounded width parameters, yielding canonical sparse subgraphs (e.g., bounded tree-width inside bounded clique-width). Resulting EP-formulas and transductions maintain bounded size and quantifier rank.

A general expressiveness result states that, on arbitrary relational structures, every existential positive (EP) monadic second order (MSO) formula is equivalent to an EP first-order (FO) formula. That is, in positive settings, MSO adds no expressive power beyond FO, since set quantifiers in positive contexts collapse by monotonicity (Mählmann et al., 22 Jan 2026).

7. Interconnections and Broader Impact

The EPSC sits at the interface of several major themes in contemporary discrete mathematics and logic: the tameness hierarchy (nowhere denseness, monadic stability), logical transduction theory, convex optimization, and spectral sparsification. Verification in broad cases—graph classes of bounded clique-width, twin-width, merge-width, and all classes characterized via self-concordant barriers in cones—demonstrates the universality of the sparsification phenomenon.

A plausible implication is that the EPSC provides a unifying lens linking “logical” and “geometric” sparsification, each motivated by minimizing complexity while preserving structural or combinatorial information under natural positive (i.e., monotonic, existential) operations. Its full scope continues to be investigated, with open questions remaining in both logic (the general case for all monadically stable classes) and convex geometry (tight algorithmic realizations and explicit constructions) (Mählmann et al., 22 Jan 2026, Saunderson, 26 Dec 2025).