Overlap Gap Property in Random Optimization

• Overlap Gap Property (OGP) is a geometric phenomenon in random optimization problems, where near-optimal solutions cluster with either very high or very low mutual overlap, excluding intermediate similarities.
• OGP signals algorithmic difficulty by showing that many standard methods (e.g., low-degree polynomials, AMP, local algorithms) fail when the solution space splits into distinct clusters.
• Variants such as ensemble OGP and multi-OGP further refine the concept by examining correlated instances or multiple solutions, deepening insights into clustering and algorithmic thresholds.

The Overlap Gap Property (OGP) is a structural statement about the geometry of near-optimal solutions in random optimization problems: after fixing an approximation scale, the set of pairwise overlaps of near-optimal assignments is disconnected, so intermediate similarities are absent. In random structures ranging from number partitioning, random constraint satisfaction problems, and mean-field spin glasses to graph optimization and statistical models, this disconnectivity has emerged as a geometric signature of algorithmic difficulty, especially for algorithmic families whose outputs vary continuously under perturbations of the input (Gamarnik, 2021).

1. Formal definition and variants

For an optimization problem over discrete assignments $\sigma\in\Sigma_n$ with random instance $\xi$ drawn from a distribution $\Xi_n$, let $H_n(\sigma)$ be the objective to be maximized and let

$\operatorname{OPT}_n=\max_{\sigma} H_n(\sigma).$

For $\varepsilon\ge 0$, the near-optimal set is

$$S_n(\varepsilon)=\bigl\{\sigma:\ H_n(\sigma)\ge (1-\varepsilon)\cdot \operatorname{OPT}_n\bigr\}.$$

When $\sigma,\tau\in\{\pm1\}^n$, the overlap and normalized Hamming distance are

$$R(\sigma,\tau)=\frac{1}{n}\sum_{i=1}^n \sigma_i\tau_i,\qquad d(\sigma,\tau)=\frac{1-R(\sigma,\tau)}{2}.$$

For general discrete assignments, “overlap” is interpreted as the fraction of coordinates on which two assignments agree.

The OGP asserts that, for an appropriate scaling of $\varepsilon$ with $\xi$0, there exist $\xi$1 such that, with high probability over $\xi$2, every pair $\xi$3 satisfies

$\xi$4

or equivalently

$\xi$5

Thus overlaps in the interval $\xi$6 are forbidden. Geometrically, near-optimal solutions split into clusters that are internally tight and externally far.

Two extensions are standard. In ensemble OGP (e-OGP), one considers correlated instances $\xi$7 and asks that near-optimal solutions across different instances have either very small or very large mutual overlap, with no intermediate regime. In multi-OGP (m-OGP), one considers $\xi$8 instances and $\xi$9 near-optimal solutions, and requires that one cannot find an $\Xi_n$0-tuple whose pairwise overlaps all lie strictly inside the forbidden interval. The case $\Xi_n$1 coincides with e-OGP. OGP was discovered in the study of random CSP solution spaces and later formulated as a general algorithmic barrier by work associated with Achlioptas, Ricci-Tersenghi, Mézard, Mora, Zecchina, Gamarnik, Sudan, Chen, Panchenko, and Rahman (Gamarnik, 2021).

Recent formulations adapt the overlap notion to model-specific geometry. In the large average subtensor problem, overlap is tracked per mode through fractions $\Xi_n$2, and m-OGP is phrased as the emptiness of the set of $\Xi_n$3 near-optimal subtensors whose per-mode overlaps all lie in a prescribed band $\Xi_n$4 (R. et al., 20 Jun 2025). In fixed-$\Xi_n$5 MAX-$\Xi_n$6-XOR-SAT, the relevant metric is a $\Xi_n$7-minimum Hamming semimetric built from $\Xi_n$8 disjoint coordinate blocks, and the resulting $\Xi_n$9-OGP forbids a window of blockwise distances for near-optimal assignments across correlated instances (Anschuetz et al., 18 Sep 2025).

2. Geometric meaning and relation to clustering

The basic geometric picture is that near-optimal solutions form well-separated basins. Within a basin, overlaps are large, typically at least $H_n(\sigma)$0; between basins, overlaps are small, at most $H_n(\sigma)$1. The forbidden band $H_n(\sigma)$2 is therefore a topological disconnect in overlap space rather than merely a low-density region.

In spin-symmetric models this picture is often sharpened by global sign symmetry. If near-optimal states occur in pairs $H_n(\sigma)$3, then $H_n(\sigma)$4, and the physically relevant overlaps concentrate near a few extremal values rather than varying continuously. In terms of distance, the identity

$H_n(\sigma)$5

makes the missing middle regime explicit.

OGP is closely related to, but stronger than, clustering. Weak clustering can coexist with exceptional “tunnels” between distant regions of solution space, and such tunnels may still permit algorithmic success. Strong clustering excludes such exceptions and is tightly linked to OGP. This distinction is central in the symmetric binary perceptron: weak clustering with singleton clusters appears for all $H_n(\sigma)$6, but strong OGP emerges only at sufficiently high density, well below the satisfiability threshold, and this stronger geometry underlies algorithmic barriers (Gamarnik, 2021).

A further refinement is ultrametric organization. In the symmetric binary perceptron, recent work studies $H_n(\sigma)$7-level ultrametric OGPs with block-constant overlap values

$H_n(\sigma)$8

together with hierarchical cluster sizes $H_n(\sigma)$9. That framework yields upper bounds $\operatorname{OPT}_n=\max_{\sigma} H_n(\sigma).$0 and $\operatorname{OPT}_n=\max_{\sigma} H_n(\sigma).$1 at $\operatorname{OPT}_n=\max_{\sigma} H_n(\sigma).$2, and proposes a conjectural identification of the algorithmic threshold with $\operatorname{OPT}_n=\max_{\sigma} H_n(\sigma).$3 and with the limit of parametric RDT lifting thresholds (Stojnic, 21 Apr 2026).

3. Canonical models and threshold phenomena

The survey literature records a strikingly consistent pattern: OGP, e-OGP, or m-OGP appears in regimes already associated with apparent algorithmic hardness, while no-OGP phases often align with the reach of known algorithms (Gamarnik, 2021).

In the random Number Partitioning Problem with $\operatorname{OPT}_n=\max_{\sigma} H_n(\sigma).$4 and cost $\operatorname{OPT}_n=\max_{\sigma} H_n(\sigma).$5, the optimum satisfies $\operatorname{OPT}_n=\max_{\sigma} H_n(\sigma).$6 with high probability. For target values $\operatorname{OPT}_n=\max_{\sigma} H_n(\sigma).$7 with any $\operatorname{OPT}_n=\max_{\sigma} H_n(\sigma).$8, the near-optimal assignments exhibit an overlap gap: pairs either satisfy $\operatorname{OPT}_n=\max_{\sigma} H_n(\sigma).$9 or $\varepsilon\ge 0$0. More generally, m-OGP holds for every $\varepsilon\ge 0$1, extending down to subexponential target scales of order $\varepsilon\ge 0$2 when $\varepsilon\ge 0$3 grows (Gamarnik, 2021). In the planted random number partitioning problem, this framework extends to a planted model: m-OGP is established and then used to rule out stable algorithms seeking energy $\varepsilon\ge 0$4, which is the first such planted-model application in the supplied literature (Kızıldağ, 2023).

For Max-Clique in $\varepsilon\ge 0$5, the existential clique size is approximately $\varepsilon\ge 0$6, while known algorithms plateau near $\varepsilon\ge 0$7. For cliques of size $\varepsilon\ge 0$8 with $\varepsilon\ge 0$9, any two such cliques have intersection either at least $$S_n(\varepsilon)=\bigl\{\sigma:\ H_n(\sigma)\ge (1-\varepsilon)\cdot \operatorname{OPT}_n\bigr\}.$$0 or at most $$S_n(\varepsilon)=\bigl\{\sigma:\ H_n(\sigma)\ge (1-\varepsilon)\cdot \operatorname{OPT}_n\bigr\}.$$1, where $$S_n(\varepsilon)=\bigl\{\sigma:\ H_n(\sigma)\ge (1-\varepsilon)\cdot \operatorname{OPT}_n\bigr\}.$$2. Across an edge-resampling interpolation, the corresponding e-OGP undergoes a phase transition at $$S_n(\varepsilon)=\bigl\{\sigma:\ H_n(\sigma)\ge (1-\varepsilon)\cdot \operatorname{OPT}_n\bigr\}.$$3 (Gamarnik, 2021).

In random $$S_n(\varepsilon)=\bigl\{\sigma:\ H_n(\sigma)\ge (1-\varepsilon)\cdot \operatorname{OPT}_n\bigr\}.$$4-SAT and NAE-SAT, for large $$S_n(\varepsilon)=\bigl\{\sigma:\ H_n(\sigma)\ge (1-\varepsilon)\cdot \operatorname{OPT}_n\bigr\}.$$5 the satisfiability threshold is approximately $$S_n(\varepsilon)=\bigl\{\sigma:\ H_n(\sigma)\ge (1-\varepsilon)\cdot \operatorname{OPT}_n\bigr\}.$$6, while known algorithms stall near $$S_n(\varepsilon)=\bigl\{\sigma:\ H_n(\sigma)\ge (1-\varepsilon)\cdot \operatorname{OPT}_n\bigr\}.$$7. Weak clustering appears near $$S_n(\varepsilon)=\bigl\{\sigma:\ H_n(\sigma)\ge (1-\varepsilon)\cdot \operatorname{OPT}_n\bigr\}.$$8, whereas strong clustering and OGP emerge around $$S_n(\varepsilon)=\bigl\{\sigma:\ H_n(\sigma)\ge (1-\varepsilon)\cdot \operatorname{OPT}_n\bigr\}.$$9 and are used to rule out low-degree polynomials and certain local algorithms beyond the algorithmic threshold (Gamarnik, 2021). A recent fixed-$\sigma,\tau\in\{\pm1\}^n$0 development proves the first explicit fixed-$\sigma,\tau\in\{\pm1\}^n$1 chaos and $\sigma,\tau\in\{\pm1\}^n$2-OGP thresholds for MAX-$\sigma,\tau\in\{\pm1\}^n$3-XOR-SAT on transposed Gallager LDPC-type ensembles, with

$\sigma,\tau\in\{\pm1\}^n$4

for the optimum value, and an explicit forbidden-distance threshold

$\sigma,\tau\in\{\pm1\}^n$5

for $\sigma,\tau\in\{\pm1\}^n$6-OGP (Anschuetz et al., 18 Sep 2025).

In mean-field spin glasses, especially pure even $\sigma,\tau\in\{\pm1\}^n$7-spin models, the ground-state landscape exhibits an algorithmic boundary aligned with OGP. Subag and Montanari gave polynomial-time algorithms that reach near-ground states in no-OGP regimes, whereas Alaoui, Montanari, and Sellke constructed algorithms that attain the best energy accessible before the onset of OGP. Disorder chaos furnishes the corresponding e-OGP mechanism, and low-degree polynomials, AMP, and Langevin dynamics are ruled out in the OGP regime (Gamarnik, 2021). For AMP specifically, a broad class of AMP-type iterations is shown to fail under OGP assumptions in $\sigma,\tau\in\{\pm1\}^n$8-spin models with even $\sigma,\tau\in\{\pm1\}^n$9 (Gamarnik et al., 2019).

The symmetric binary perceptron provides a particularly explicit m-OGP threshold. With

$$R(\sigma,\tau)=\frac{1}{n}\sum_{i=1}^n \sigma_i\tau_i,\qquad d(\sigma,\tau)=\frac{1-R(\sigma,\tau)}{2}.$$0

the paper “Algorithms and Barriers in the Symmetric Binary Perceptron Model” shows that for sufficiently small $$R(\sigma,\tau)=\frac{1}{n}\sum_{i=1}^n \sigma_i\tau_i,\qquad d(\sigma,\tau)=\frac{1-R(\sigma,\tau)}{2}.$$1 the model exhibits m-OGP above

$$R(\sigma,\tau)=\frac{1}{n}\sum_{i=1}^n \sigma_i\tau_i,\qquad d(\sigma,\tau)=\frac{1-R(\sigma,\tau)}{2}.$$2

and that this matches the best known algorithmic threshold up to logarithmic factors as $$R(\sigma,\tau)=\frac{1}{n}\sum_{i=1}^n \sigma_i\tau_i,\qquad d(\sigma,\tau)=\frac{1-R(\sigma,\tau)}{2}.$$3 (Gamarnik et al., 2022).

OGP has also propagated into newer high-dimensional models. In the large average subtensor problem, for sufficiently large tensor order $$R(\sigma,\tau)=\frac{1}{n}\sum_{i=1}^n \sigma_i\tau_i,\qquad d(\sigma,\tau)=\frac{1-R(\sigma,\tau)}{2}.$$4 and both $$R(\sigma,\tau)=\frac{1}{n}\sum_{i=1}^n \sigma_i\tau_i,\qquad d(\sigma,\tau)=\frac{1-R(\sigma,\tau)}{2}.$$5 and $$R(\sigma,\tau)=\frac{1}{n}\sum_{i=1}^n \sigma_i\tau_i,\qquad d(\sigma,\tau)=\frac{1-R(\sigma,\tau)}{2}.$$6, the maximum average is concentrated around

$$R(\sigma,\tau)=\frac{1}{n}\sum_{i=1}^n \sigma_i\tau_i,\qquad d(\sigma,\tau)=\frac{1-R(\sigma,\tau)}{2}.$$7

and for any $$R(\sigma,\tau)=\frac{1}{n}\sum_{i=1}^n \sigma_i\tau_i,\qquad d(\sigma,\tau)=\frac{1-R(\sigma,\tau)}{2}.$$8 and $$R(\sigma,\tau)=\frac{1}{n}\sum_{i=1}^n \sigma_i\tau_i,\qquad d(\sigma,\tau)=\frac{1-R(\sigma,\tau)}{2}.$$9 there exist $\varepsilon$0 such that the corresponding m-OGP holds above $\varepsilon$1 with probability at least $\varepsilon$2 (R. et al., 20 Jun 2025).

4. Algorithmic barriers

The algorithmic force of OGP comes from stability. Along a correlated interpolation $\varepsilon$3, an algorithm $\varepsilon$4 is $\varepsilon$5-stable if consecutive outputs remain close, for example

$\varepsilon$6

or, in continuous Gaussian interpolation, if a Lipschitz-type bound holds with high probability for small parameter increments. When the ensemble exhibits e-OGP with gap parameters $\varepsilon$7, the interpolation endpoints are independent, and $\varepsilon$8, a standard “cannot jump the gap” argument shows that no such stable algorithm can remain $\varepsilon$9-optimal all along the interpolation (Gamarnik, 2021).

This barrier rules out several major algorithmic classes in OGP regimes. Low-degree polynomial methods are stable for many random optimization problems up to degrees as large as $\xi$00, which excludes them in OGP phases of $\xi$01-spin glasses, independent set, and random $\xi$02-SAT. AMP-like iterative algorithms obey analogous continuity properties and are ruled out in $\xi$03-spin models. Local algorithms of bounded depth are ruled out for independent set and clique-type problems on sparse and dense random graphs via m-OGP. Langevin dynamics at linear time scales is also blocked by e-OGP in spin-glass regimes, and extending these barriers to longer time scales remains open (Gamarnik, 2021).

Quantum variants inherit the same logic when they admit an appropriate stability notion. For QAOA, OGP implies that optimizing via tree parameters cannot recover the Parisi optimum in even-$\xi$04 Max-$\xi$05-XORSAT under limit swapping, and the obstruction already applies at logarithmic depth $\xi$06 (Goh, 2024). In MAX-$\xi$07-XOR-SAT on Gallager-transposed ensembles, approximately Lipschitz quantum algorithms in the quantum Wasserstein metric are ruled out beyond the OGP or chaos threshold. Under the stated decoding-threshold assumption, Decoded Quantum Interferometry is stable and therefore topologically obstructed; the same paper compares this barrier with AMP and depth-$\xi$08 QAOA, which attain higher large-$\xi$09 average-case approximation ratios in that setting (Anschuetz et al., 18 Sep 2025).

OGP-based barriers are not limited to offline or fully global algorithms. In discrepancy minimization and the symmetric binary perceptron, geometric barriers with m-tuples across correlated instances also rule out online algorithms, in some cases even when the allowed success probability is exponentially small. In the Gaussian discrepancy setting with $\xi$10, ensemble m-OGP furthermore rules out stable algorithms for constant discrepancy solutions (Gamarnik et al., 2023).

5. Proof mechanisms and structural connections

A standard route to OGP begins with counting near-optimal pairs or $\xi$11-tuples at a prescribed overlap. In number partitioning, if $\xi$12 have overlap $\xi$13, then

$\xi$14

so for $\xi$15 the exponent becomes negative near $\xi$16, creating a forbidden region (Gamarnik, 2021). In Max-Clique, the analogous exponent under correlated resampling is

$\xi$17

whose roots determine the two allowed intersection regimes and the phase transition at $\xi$18 (Gamarnik, 2021).

Interpolation is the second recurring ingredient. Graph problems use edge-resampling paths; spin glasses use Gaussian interpolation; tensor and discrepancy problems use correlated Gaussian ensembles. Once an overlap gap is identified, continuity of stable algorithms along the interpolation converts the geometric statement into an impossibility theorem. In models with many candidate overlaps, Ramsey-type arguments and extraction of almost-equilateral subfamilies are often used to pass from pairwise to m-ary contradictions (Gamarnik, 2021).

Concentration and Gaussian comparison tools supply the sharp probability estimates. In the large average subtensor problem, the analysis combines first and second moment methods, the Paley–Zygmund inequality, Slepian’s lemma, multivariate Gaussian tail bounds, and Borell–TIS concentration. The covariance between two candidate subtensors with per-mode intersections $\xi$19 is

$\xi$20

and the large-$\xi$21 regime makes this correlation structure tractable enough to prove both ground-state concentration and m-OGP (R. et al., 20 Jun 2025). In fixed-$\xi$22 MAX-$\xi$23-XOR-SAT, the first fixed-$\xi$24 OGP thresholds are proved with second-moment estimates adapted to Gallager structure, Kravchuk polynomial moment inequalities, Paley–Zygmund interpolation, and a Ramsey-type monochromatic-clique argument (Anschuetz et al., 18 Sep 2025).

Spin-glass theory provides the conceptual backbone. Replica symmetry breaking, the Parisi measure, and disorder chaos explain why OGP arises exactly where algorithmic performance degrades. In sparse PCA, free energy wells associated with overlap-restricted Gibbs measures yield hitting-time lower bounds for MCMC and coexist with an explicit OGP regime, thereby connecting Franz–Parisi landscape methods with overlap-gap barriers (Arous et al., 2020). A closely related synthesis appears in the analysis of shortest path landscapes, where the Franz–Parisi potential and OGP agree that local search should fail for paths but not for shortest path trees (Koehler et al., 24 Nov 2025).

6. Scope, limitations, and contemporary reinterpretations

OGP is a strong obstruction, but it is not an omnibus hardness theorem. Its presence rigorously rules out broad classes of stable, local, or approximately Lipschitz algorithms; its absence does not by itself imply polynomial-time solvability. The survey literature explicitly notes that hard problems may fail to exhibit OGP at currently accessible scales, and that weak clustering can coexist with algorithmic success because algorithms are not uniform samplers of the solution space (Gamarnik, 2021).

Recent work has sharpened this caution. The shortest $\xi$25-$\xi$26 path problem in sparse $\xi$27 and in complete graphs with i.i.d. exponential edge weights exhibits OGP, yet it remains polynomial-time solvable and even admits a degree-$\xi$28 polynomial estimator in sparse Erdős–Rényi graphs; a uniform approximate shortest path can also be sampled in polynomial time. This is identified as the first example in which OGP is not predictive of algorithmic intractability for a non-algebraic average-case optimization problem (Li et al., 2024). Complementarily, shortest path trees do not satisfy ensemble OGP: their overlap can be tuned continuously by the correlation parameter, the Franz–Parisi potential is quasiconvex, and local search succeeds in the tree landscape (Koehler et al., 24 Nov 2025).

The planted-model frontier has also expanded. In the stochastic block model, modularity exhibits OGP for sparse $\xi$29 communities, producing a far basin of high-scoring decoy partitions and implying slow mixing for a natural single-vertex-flip Gibbs chain. At the same time, restricting to balanced partitions removes the OGP, which explains why balanced modularity heuristics can behave differently (Bhamidi et al., 11 May 2026). In planted random number partitioning, m-OGP is established and then used to rule out stable algorithms seeking low-energy solutions, showing that OGP-based barriers extend beyond null models when the algorithmic target is near-optimality rather than exact recovery (Kızıldağ, 2023).

Two themes remain active. One is sharpness: in the large average subtensor problem, the onset of m-OGP and the algorithmic threshold become asymptotically identical as $\xi$30, and a forthcoming branching-OGP result is stated to locate the exact threshold at $\xi$31 for the corresponding polynomial-time algorithmic guarantee (R. et al., 20 Jun 2025). The other is hierarchy: ultrametric OGP constructions for the symmetric binary perceptron suggest a multi-level geometric description whose limiting threshold may coincide with independently derived algorithmic thresholds from parametric RDT (Stojnic, 21 Apr 2026).

Taken together, these developments place OGP in a more precise role. It is a rigorous and highly informative descriptor of solution-space geometry, often aligned with computational thresholds and capable of excluding major algorithmic paradigms, but it must be interpreted relative to the optimization landscape, the representation of solutions, and the stability class under consideration.