Partial Gap: Selective Incompleteness

• Partial gap is a state of selective incompleteness where a subset of a system is undefined or suppressed while key residual components remain operative.
• In electronic systems, partial gaps manifest as the suppression of specific spectral weights, leading to coexistence of metallic transport channels alongside gapped states.
• Across disciplines, partial-gap regimes inform both experimental measurements and algorithmic methods, guiding optimal performance in settings like spectroscopy, multi-hop reasoning, and engineered devices.

“Partial gap” denotes a regime of selective absence, depletion, or under-specification in which only a proper subset of a relevant structure is removed, unresolved, or constrained, while a residual component remains operative. In condensed-matter spectroscopy, this means that only part of the low-energy electronic spectrum or Fermi surface is suppressed, leaving metallic transport channels or surviving pockets intact; in multi-hop reasoning, it denotes the missing relational bridge between provided evidence and a candidate answer; in alloy theory, it describes effective band-gap closure induced by a partial order–disorder transition; and in geometric or analytic settings it can refer to a spectral gap near zero or to a rule restricting which embeddings are permitted (Liu et al., 2023, Khot et al., 2019, Troppenz et al., 2020, Piovani, 9 Feb 2026). Taken together, these usages suggest a common pattern: the operative distinction is not between “gap” and “no gap,” but between full removal and structured residuality.

1. Core semantics of selective incompleteness

A recurring feature across disciplines is that a partial gap is defined against a stronger limiting case. In La$_3$Ni$_2$O$_7$, the transition at $T^\ast \simeq 115$ K does not open a full insulating gap, because one Drude channel survives; in OpenBookQA, a core science fact is provided but is usually insufficient on its own, so the model must identify the missing relation between a salient span and an answer choice; in multi-agent code generation, agents fail not because they cannot write code, but because the specification leaves shared internal representations implicit; and in partial-to-partial shape matching, both shapes are incomplete and the overlap region is itself unknown (Liu et al., 2023, Khot et al., 2019, Sartori, 25 Mar 2026, Ehm et al., 2024).

This terminological spread also marks an important misconception. A partial gap is not always an energy gap. In some literatures it is a deficit in observability, recoverability, or coordination rather than a depletion of states. In satellite image time series, the “gap” consists of masked observations propagated through partial convolutions; in probabilistic shaping and PAUC optimization, “gap” measures separation from an ideal objective such as capacity or exact constrained ranking rather than a missing spectral interval (Appel, 2022, Gültekin et al., 2019, Jiang et al., 1 Dec 2025). The term therefore functions as a structural descriptor of incompleteness with residual continuity, not as a single physical invariant.

2. Partial gaps in electronic spectra and correlated transport

In correlated-electron materials, the phrase has its most literal spectroscopic meaning. Optical measurements on La$_3$Ni$_2$O$_7$ show that below $T^\ast$ the Drude peak is suppressed and the spectral weight is transferred to high frequency, with reflectivity suppression between roughly $400$ and $1200$ cm$_2$0, and the zero-crossing of $_2$1 yielding $_2$2 meV. The low-energy conductivity is described by two Drude components: D1 remains essentially metallic with $_2$3, whereas D2 is abruptly suppressed below $_2$4, vanishes on further cooling, and shows $_2$5. The measured coherent kinetic energy is also strongly renormalized, with $_2$6, placing the system in the proximity of the Mott phase (Liu et al., 2023).

Ca$_2$7Ru$_2$8O$_2$9 exhibits a different but related partial-gapping phenomenology. Magnetotransport and thermoelectric measurements on an untwined crystal indicate a correlated semimetal ground state with small surviving pockets rather than a fully gapped low-temperature phase. Shubnikov–de Haas oscillations with $_7$0 indicate pockets of order $_7$1 of the Brillouin zone, while opposite thermopower signs, $_7$2 and $_7$3, provide direct evidence that electron and hole pockets coexist. Hall and Nernst coefficients further suggest a temperature- and momentum-dependent partial gap below the structural transition $_7$4, with a possible Lifshitz transition at $_7$5 (Xing et al., 2018).

Heavy-fermion CeNiGe$_7$6 supplies an SDW-linked variant. Below $_7$7 K and $_7$8 K, Arrott plots, nonlinear dc susceptibility, resistivity, and heat capacity collectively indicate a partial gap opening associated with magnetic instability and possible spin density wave formation. The heat-capacity ratio $_7$9 is close to the BCS-like SDW value $T^\ast \simeq 115$0. Y substitution on the Ce site suppresses the gap opening and shifts it toward lower temperature, while no zero-field quantum critical point is established up to 40% dilution (Singh et al., 2018).

Across these cases, a partial gap does not imply a simple metal–insulator transition. The decisive signature is coexistence: depleted spectral weight, reconstructed pockets, or a removed conduction channel, alongside a residual metallic sector. This is why the relevant contrast is with full gapping, not merely with disorder-broadened metallic narrowing.

3. Disorder-driven closure and finite-gap degeneration

In Ba$T^\ast \simeq 115$1Al$T^\ast \simeq 115$2Si$T^\ast \simeq 115$3, partiality enters through configurational thermodynamics. The well-ordered ground state is semiconducting, with an indirect DFT band gap of about $T^\ast \simeq 115$4 eV, valence-band maximum near $T^\ast \simeq 115$5, conduction-band minimum at $T^\ast \simeq 115$6, and no Al–Al bonds. Yet for 44 calculated configurations, the band gap decreases almost linearly with increasing total energy above the ground state and becomes zero already at around $T^\ast \simeq 115$7 meV/atom above it. As temperature rises, a partial order–disorder transition transfers approximately $T^\ast \simeq 115$8 Al atoms from the $T^\ast \simeq 115$9 to the $_3$0 sublattice, increases Al–Al bonds, and drives the finite-temperature effective spectral function toward nonzero weight at $_3$1. The macroscopic transition temperature extracted by finite-size scaling is $_3$2, with the spectral function at the Fermi level becoming nonzero above about $_3$3 K and the density-of-states gap effectively closed around $_3$4 K (Troppenz et al., 2020).

The crucial point is that the gap is not said to vanish for a unique static crystal structure. The effective spectral function is canonically averaged,

$_3$5

so gap closure is an ensemble phenomenon produced by thermal occupation of more disordered, metallic or nearly gapless configurations (Troppenz et al., 2020). A plausible implication is that “partial gap” here names a selective persistence of ordered sublattice structure rather than a partial depletion of an already metallic Fermi surface.

A closely related residual-structure motif appears in finite-gap KdV theory. Partial degeneration of a genus $_3$6 finite-gap solution collapses only $_3$7 spectral bands, leaving a residual genus-one elliptic curve. The resulting solution decomposes into a shifted elliptic or cnoidal background plus a Kay–Moses-type determinant carrying the solitonic content, with explicit Jacobi-theta-function formulas for group velocities and pairwise phase shifts. Here, too, the defining feature is preservation of a nondegenerate background while another part of the spectral data collapses into localized excitations (Bertola et al., 2022).

4. Partial knowledge, partial observation, and algorithmic gap filling

In multi-hop question answering, the paper “What’s Missing: A Knowledge Gap Guided Approach for Multi-hop Question Answering” formalizes a partial-gap setting in which the system receives only a relevant core science fact, not all facts needed to answer directly. The missing piece is the knowledge gap between a key span in the fact and an answer option, captured by relations such as $_3$8 and their inverses. GapQA first predicts the key span with a BiDAF-style span predictor, then retrieves ConceptNet tuples and ARC corpus sentences conditioned on the span $_3$9 and answer choice $_2$0, and finally scores each choice by

$_2$1

The model is trained with answer prediction plus auxiliary relation prediction. On OpenBookQA, GapQA $_2$2 reaches $_2$3 on OBQA-Short versus $_2$4 for KER $_2$5, and $_2$6 on OBQA-Full versus $_2$7, showing that explicit identification of what is missing outperforms baselines that rely on fact relevance alone (Khot et al., 2019).

Satellite time series gap filling uses a different formalization of partiality. Spatiotemporal partial convolutions take both an input block $_2$8 and a binary mask $_2$9, renormalizing the convolution by the number of valid entries and updating the mask so that an output becomes valid when at least one valid observation contributes. Implemented in a 3D U-Net-like model for Sentinel-5P CO data, this approach yields competitive gap-filling errors, with STpconv at MAE $_7$0 and RMSE $_7$1, and one-step-ahead prediction at MAE $_7$2 and RMSE $_7$3; prediction time per block is $_7$4 s, versus $_7$5 s for stmra and $_7$6 s for gapfill (Appel, 2022).

Partial-to-partial shape matching generalizes the same theme to geometry. Because the overlap region between incomplete shapes is unknown, the method is formulated on a triangle product space with binary variables selecting matched triangle or degenerate-triangle pairs subject to orientation-consistent neighborhood constraints $_7$7 and one-use inequalities $_7$8, $_7$9. The resulting integer non-linear program minimizes normalized mean cost and outperforms Sm-comb and DPFM, reaching mean IoU $T^\ast$0 on CP2P TEST and $T^\ast$1 on PARTIALSMAL (Ehm et al., 2024).

In multi-agent code generation, the “specification gap” is a coordination failure under partial knowledge. Across 51 class-generation tasks, single-agent performance drops from $T^\ast$2 at L0 to $T^\ast$3 at L3, whereas two-agent Split performance drops from $T^\ast$4 to $T^\ast$5, leaving a persistent $T^\ast$6–$T^\ast$7 pp coordination gap. An AST-based conflict detector reaches $T^\ast$8 precision at L3, but the factorial recovery experiment shows that restoring the full specification alone recovers the single-agent ceiling, with Spec-Only at $T^\ast$9 and Guided equal to Blind at $400$0. The paper therefore treats specifications as the primary coordination mechanism rather than conflict reports as a sufficient repair instrument (Sartori, 25 Mar 2026).

5. Gap conditions, spectral gaps, and partial-order formalisms

In analysis and geometry, a gap can be a literal spectral exclusion near zero. On a complete Kähler manifold, the $400$1-$400$2-Lemma proved in “An $400$3-$400$4-Lemma on a class of complete Kähler manifolds” assumes

$400$5

equivalently $400$6. Under this hypothesis, smooth square-integrable forms satisfy the full chain of equivalences between $400$7-exactness, $400$8-exactness, $400$9-exactness, membership in $1200$0, and $1200$1-exactness. The spectral gap is also equivalent to closed image properties for the relevant differentials, so it replaces the compactness-based discreteness that underlies the classical compact Kähler lemma (Piovani, 9 Feb 2026).

In order theory, “gap” often names an unresolved complexity interval rather than a missing spectrum. The paper “The Complexity of the Partial Order Dimension Problem - Closing the Gap” resolves the last open case left by Yannakakis by proving that 3DH2—deciding whether a height-2 partial order has dimension at most 3—is NP-complete, via an equivalence with bipartite triangle containment representations and a reduction from $1200$2-3-CON-3-SAT(4) (Felsner et al., 2015). In a different vein, “Maximal order types for sequences with gap condition” studies weak, strong, and Gordeev’s symmetric gap conditions on finite sequences over a well order $1200$3, proves that Gordeev’s symmetric and weak gap orders coincide, and shows over $1200$4 that the corresponding well-partial-order statements are equivalent to $1200$5, while also computing the maximal order types of the associated sequence and tree orders (Uftring, 29 Jul 2025).

Partial-order methods also bridge classical convergence gaps. “Partial Order Infinitary Term Rewriting” replaces the metric model on total infinitary terms by a partial order on partial terms with $1200$6, proves that metric and partial-order convergence coincide on total terms, and then shows that for orthogonal, left-finite systems the gap between strong metric convergence and strong partial-order convergence is bridged by Böhm extensions. In this setting, orthogonal systems become infinitarily confluent and infinitarily normalising, and the unique infinitary normal forms are Böhm trees (Bahr, 2014).

A more elementary numerical notion appears for multiplicatively closed subsets of $1200$7. For sets generated by at most two elements, arbitrarily large gap intervals can be constructed explicitly, with endpoints in the set and known prime factorization at both ends; this extends, in partial form, to more general finitely generated multiplicatively closed sets when they lie inside a doubly generated one (Kumar, 2016).

6. Approximation gaps and partial-gap regimes in engineered systems

In communication theory, a gap may quantify deviation from an exact but intractable objective. Under partial CSIT, the exact beamforming criterion is the Expected Weighted Sum Rate (EWSR), whereas the massive-MIMO surrogate is ESEI-WSR. The relevant termwise gap

$1200$8

is nonnegative by Jensen’s inequality and monotonically increasing in $1200$9, so the worst case is the infinite-SNR limit. In the i.i.d. MISO case the asymptotic bound behaves as $_2$00, explaining why ESEI-WSR becomes accurate as the number of transmit antennas grows (Gopala et al., 2017).

Probabilistic shaping makes the same logic operational at the modulation level. “Partial Enumerative Sphere Shaping” shows that shaping only a subset of amplitude bit-levels already closes most of the shaping gap to capacity. For 16-ASK at $_2$01 bit/symbol, 2-bit P-ESS is about $_2$02 dB better than uniform signaling at FER $_2$03, only about $_2$04 dB away from full 3-bit ESS, while reducing storage and computational complexity by factors of 6 and 3, respectively. The abstract summarizes the same result as shaping 2 amplitude bits of 16-ASK having almost the same performance as shaping 3 bits and being $_2$05 dB more power-efficient than uniform signaling (Gültekin et al., 2019).

In constrained ranking, “Closing the Approximation Gap of Partial AUC Optimization” treats the gap between exact PAUC objectives and their optimized surrogates. The paper gives two instance-wise minimax reformulations: one with asymptotically vanishing gap, $_2$06, based on softplus smoothing, and one exact unbiased reformulation using $_2$07. The resulting algorithms have linear per-iteration computational complexity with respect to sample size and a convergence rate of $_2$08 for typical one-way and two-way PAUCs, with a generalization bound of sharp order $_2$09 (Jiang et al., 1 Dec 2025).

Planet–disk interaction studies use “partial gap” in a still different but literal hydrodynamic sense. For planets embedded in 3D disks, the gap depth is measured by $_2$10, with shallow gaps corresponding to type-I-like behavior and deeper partial gaps weakening eccentricity and inclination damping approximately linearly. The simulations identify a break around $_2$11, damping suppression by up to a factor of $_2$12 near $_2$13, and vortex appearance roughly when $_2$14. The practical conclusion is that standard no-gap prescriptions overestimate damping efficiency for partial-gap-opening planets (Pichierri et al., 2024).

Across these engineering settings, a gap usually measures distance from an idealized limit—exact expectation, Gaussian-like shaping, exact constrained ranking, or classical type-I damping. The common structural lesson is that partial-gap regimes are intermediate regimes: too structured to be treated as fully ungapped, but not far enough from the reference limit to justify fully separate asymptotic theories.