Localizability: Networks, Robotics & Complexity

• Localizability is the measure of how uniquely and reliably positions or improvements can be inferred from local data, applicable in sensor networks, robotics, and quantum field theory.
• It encompasses graph rigidity, Fisher information, and local search methodologies, influencing network realizability, pose estimation, and algorithmic optimization.
• It bridges practical deployments and theoretical frameworks in areas from cooperative robotics to multimodal learning by clarifying when local observations reveal global structure.

Localizability denotes the extent to which location can be uniquely determined, reliably estimated, or meaningfully represented from available observations, but the term is not uniform across disciplines. In network localization it concerns unique realizability of node positions under anchor and distance constraints; in cooperative robotics it is a geometric-statistical property of the Fisher information or of scan-registration structure; in optimization it measures how effectively global improvement is exposed by local exchanges; in complexity theory it describes whether approximation-method lower bounds persist under small local oracle augmentations; in relativistic quantum theory it concerns the existence of a position representation on a one-particle Hilbert space; and in contemporary multimodal learning it refers to whether model features or sensory cues reveal where a target action, object, or geographic origin can be localized (Ping et al., 2021, Ny et al., 2018, Fujii, 2020, Pich, 2022, Yokomizo et al., 2010, Bao et al., 2024, Zhang et al., 6 Jan 2026).

1. Network, graph, and anchor-based localizability

In wireless and sensor-network localization, localizability is classically tied to uniqueness of geometric realization. A network is modeled as a graph with anchors of known position and free nodes of unknown position; network localizability means that the entire network has a unique realization consistent with the measured distances and anchor positions, while node localizability means that a specific node’s position is uniquely determined even if the rest of the network is not. In the nonlinear localization schema, the 2D network-level condition is explicit: a network is localizable if and only if it contains at least 3 anchors and it is globally rigid, equivalently 3-connected and redundantly rigid (Ping et al., 2021).

Distributed WSN work sharpens this graph-theoretic viewpoint. A distance graph is localizable if it is globally rigid in $\mathbb{R}^2$, and centralized testing can therefore rely on 3-connectivity and redundant rigidity. Because centralized rigidity algorithms are costly in resource-limited WSNs, distributed methods instead identify sufficient localizable subclasses. The “triangle bar” family strictly contains trilateration graphs and wheel extensions; it is closed under overlaps of at least three common vertices and under adding a new node adjacent to at least three vertices of an existing triangle bar, and every triangle bar is generically globally rigid (Sau et al., 2013). “Triangle Extension” further turns this idea into a linear-time distributed localizability detector: nodes grow minimally rigid branches from pairs of beacons, then detect when a third non-collinear beacon or a dual-V configuration closes the structure into a globally rigid subgraph, allowing node-level localizability to be certified with only one-hop exchanges (Wu et al., 2018).

Barycentric coordinate-based linear localization alters the meaning of localizability because the relevant structure is no longer only the measurement graph $\mathcal{G}$ but the generated barycentric graph $\mathcal{G}^A$. In this schema, a free node must have at least $d+1$ mutually connected neighbors in $\mathbb{R}^d$ for barycentric coordinates to exist, and in 2D all free nodes are localizable if every free node can find at least three node-disjoint paths to anchors using only edges in $\mathcal{G}^A$. At the node level the paper gives a necessary clique condition and the sufficient “Recursive-3DP” condition, then detects all nodes satisfying that sufficient condition with an iterative max-flow procedure (Ping et al., 2021).

Collaboration changes the anchor sufficiency condition probabilistically. In stochastic-geometry models of cellular positioning, a device is non-collaboratively uniquely localizable if it hears at least 3 base stations for range-based positioning or at least 4 for range-difference positioning. With two-device collaboration, unique localizability can also occur when a device is one anchor short but the collaborator contributes enough distinct hearable anchors so that the union satisfies the required count and non-collinearity conditions. The resulting improvement is analytically tractable in PPP cellular models and is especially strong under shadowing and random frequency reuse (Schloemann et al., 2015).

2. Statistical and stochastic-geometric formulations

A different but closely related meaning appears in cooperative robotics, where localizability is not only uniqueness but the quality of the geometry for estimation under noise. For a mobile robotic network with noisy range measurements, the unknown robot positions are stacked in $\tilde{\mathbf p}\in\mathbb R^{2n}$, the Fisher Information Matrix is

$$F(\mathbf p) = -\mathbb E \left[ \frac{\partial^2}{\partial \tilde{\mathbf p}^2} \ln h(\Theta \mid \mathbf p) \right],$$

and the Cramér–Rao bound gives $$\operatorname{Cov}(\hat{\mathbf p}) \succeq F(\mathbf p)^{-1}$$. Localizability is then encoded by scalar functions of $F$: $\mathcal{G}$0, $\mathcal{G}$1, $\mathcal{G}$2, and $\mathcal{G}$3. Lower values correspond to better localizability, and the extended FIM admits the rigidity-theoretic factorization $\mathcal{G}$4, making localizability and weighted rigidity two views of the same geometric constraint structure (Ny et al., 2018).

A stochastic-geometry treatment of vision-based localization with indistinguishable landmarks makes the same idea probabilistic. Landmarks are modeled as a homogeneous PPP $\mathcal{G}$5 on $\mathcal{G}$6; a location $\mathcal{G}$7 induces a local landmark pattern $\mathcal{G}$8, and a measurement map $\mathcal{G}$9 extracts counts, ranges, relative locations, or unordered range sets. With bounded measurement uncertainty represented by $\mathcal{G}^A$0, two locations are distinguishable if $\mathcal{G}^A$1. The paper defines

$\mathcal{G}^A$2

with complement $\mathcal{G}^A$3, and derives explicit formulas or bounds for several vision-inspired measurement models, including ordered range vectors and sets of relative landmark locations. A central asymptotic result is that localizability probability approaches one as the landmark intensity tends to infinity, so error-free localization is achievable in that limiting regime (Hu et al., 2024).

These two lines of work use different machinery—FIM/CRB in one case, PPP measurement distinguishability in the other—but both treat localizability as an intrinsic property of sensing geometry before any particular estimator is fixed. This suggests a broad division between existence-type definitions, which ask whether a unique solution exists, and informativeness-type definitions, which ask how strongly the available data exclude nearby alternatives.

3. Localizability-aware robotic perception and control

In LiDAR odometry and registration, localizability becomes explicitly direction-dependent. One formulation defines, for a scan $\mathcal{G}^A$4, a binary 6-DOF localizability vector

$\mathcal{G}^A$5

where $\mathcal{G}^A$6 means the corresponding translational or rotational direction is localizable and $\mathcal{G}^A$7 means it is non-localizable. Ground-truth labels are obtained by repeatedly sampling nearby child poses, running point-to-plane ICP, and averaging the resulting registration error over those trials; localizability is thus defined operationally as whether scan-to-scan registration will succeed for 3D pose estimation. A sparse 3D ResUNet then predicts this localizability vector directly from a single LiDAR scan, enabling early failure detection without evaluating the underlying registration optimization (Nubert et al., 2022).

X-ICP instead extracts localizability from the registration problem itself. In scan-to-map point-to-plane ICP, correspondences induce a Hessian or information matrix

$\mathcal{G}^A$8

whose eigenpairs identify principal directions of the optimization. Large eigenvalues correspond to well-constrained directions; small eigenvalues indicate weakly constrained or degenerate directions. X-ICP performs a fine-grained localizability analysis in this eigenspace and then projects the ICP update onto the well-constrained subspace, so that drift-free pose updates are generated by enforcing controlled updates or leaving the degenerate directions unchanged (Tuna et al., 2022).

Navigation work uses a more online, task-coupled operationalization. In localizability-enhanced navigation in dynamic human environments, localizability is “the capability of acquiring reliable localization” and is represented through AMCL’s covariance-derived variances $\mathcal{G}^A$9, the pose error, and a lost condition. The policy state augments the goal representation with these uncertainty terms and a pedestrian map, while the reward scales goal-approach progress by $d+1$0, rewards reductions in pose error, and penalizes lost events. The learned planner thereby prefers geometrically informative regions such as walls and corners over open or occlusion-prone regions (Chen et al., 2023).

Taken together, these robotics formulations separate localizability from the final state estimate. It is a property of the scene, the motion increment, or the current correspondence structure that determines whether a localization backend should trust LiDAR geometry, preserve only certain eigendirections, or actively seek more informative viewpoints.

4. Localizability in optimization and complexity theory

Outside physical localization, the term is used for the relation between local and global structure. For monotone set-function maximization, a non-negative monotone set function $d+1$1 is $d+1$2-localizable with size $d+1$3 if for every pair $d+1$4 with $d+1$5 and every bijection $d+1$6,

$d+1$7

A more general $d+1$8-localizability notion allows multi-element exchanges of size at most $d+1$9 and asymmetric coverage multiplicities. This property is precisely what allows local search guarantees under matroid, $\mathbb{R}^d$0-matroid intersection, and $\mathbb{R}^d$1-exchange constraints. Linear functions are $\mathbb{R}^d$2-localizable, monotone submodular functions are $\mathbb{R}^d$3-localizable, and in sparse optimization the induced set function $\mathbb{R}^d$4 becomes localizable when $\mathbb{R}^d$5 satisfies restricted strong concavity and restricted smoothness (Fujii, 2020).

In complexity theory, “localizability” appears in an almost opposite sense: not as the power of local search, but as a limitation of lower-bound methods. For Razborov’s approximation method, a lower bound is localizable if it extends to circuits augmented with arbitrarily powerful oracles of small fan-in. The paper proves that, for many legitimate approximation models, lower bounds obtained by the approximation method persist against such oracle-augmented circuits, both for general circuits and for constant-depth circuits with quantitative trade-offs. This yields a “locality barrier” for hardness magnification: one cannot use the approximation method to derive an extremely strong constant-depth lower bound and then magnify it to an $\mathbb{R}^d$6 lower bound for an explicit function when the magnification theorem relies on small local oracles (Pich, 2022).

These two uses are formally unrelated, but they share a common template: localizability measures whether global behavior is already visible from bounded local moves or bounded local augmentations. In optimization that visibility is an algorithmic resource; in complexity it is a barrier.

5. Particle localizability in quantum field theory

In relativistic and curved-spacetime quantum theory, localizability concerns whether one can consistently assign position probabilities to one-particle states. For a massive scalar field in $\mathbb{R}^d$7, the one-particle Hilbert space $\mathbb{R}^d$8 is selected by the Hadamard/Bunch–Davies positive-frequency decomposition, and localizability is formulated through a family of unitary maps

$\mathbb{R}^d$9

so that $\mathcal{G}^A$0 is interpreted as a Newton–Wigner-type position wavefunction on the spatial circle. Rotational covariance, parity, time reversal, and a large-mass limit analogous to the flat-space Newton–Wigner construction uniquely fix the representation, yielding a de Sitter analogue of the position operator. The resulting scheme is not fully de Sitter covariant and depends on the chosen vacuum, but it establishes that a viable notion of one-particle localizability exists for massive scalar fields in this setting (Yokomizo et al., 2010).

A broader QFT analysis asks whether a relativistic field theory can be consistently described as a theory of localizable particles at all. With a Klein–Gordon field, a UV cutoff at the Compton scale is used to define a controlled non-relativistic regime. The analysis compares local field operators with Newton–Wigner-type annihilation operators, constructs POVM elements $\mathcal{G}^A$1 for bandlimited $\mathcal{G}^A$2-particle sectors, and shows that one recovers Schrödinger dynamics and a Born-rule interpretation in this approximate regime. At the same time, structural obstructions persist: the local and global Fock decompositions remain inequivalent, compact spatial support is impossible under bandlimitation, and ground-state entanglement between local degrees of freedom survives at order $\mathcal{G}^A$3. Localizability therefore remains approximate and scheme-dependent even after the non-relativistic reduction (Papageorgiou et al., 2019).

This literature sharply distinguishes localizability from simple possession of a position operator. The existence of a probabilistic spatial representation can depend on vacuum choice, spacetime symmetry, and the localization scheme itself, while entanglement and Reeh–Schlieder-type effects obstruct a fully classical particle ontology.

6. Semantic localizability in multimodal learning

In open-vocabulary action detection, localizability becomes a latent property of pretrained vision-LLMs. The central claim is that modern video VLMs contain not only strong semantic alignment between video and text embeddings but also underexploited spatial and temporal localizability in their internal visual representations. OpenMixer makes this explicit by computing text-conditioned patch similarity

$\mathcal{G}^A$4

reversing it as $\mathcal{G}^A$5, and using the resulting heatmap as a prior over actor locations for DETR-style queries. Temporal decoding is then conditioned on the video-level feature $\mathcal{G}^A$6, and final open-vocabulary classification aligns dynamically fused query features with text embeddings. In this setting, localizability is the ability of VLM features to indicate where in space and time the evidence for a textual action concept lies (Bao et al., 2024).

An audio-language analogue defines localizability as the geographic informativeness of an audio clip. Each recording $\mathcal{G}^A$7 receives a scalar score

$\mathcal{G}^A$8

where $\mathcal{G}^A$9 is the fraction of the clip occupied by sound category $\tilde{\mathbf p}\in\mathbb R^{2n}$0, $\tilde{\mathbf p}\in\mathbb R^{2n}$1 and $\tilde{\mathbf p}\in\mathbb R^{2n}$2 are positive and negative AudioSet categories learned from ALM success and reasoning traces, and $\tilde{\mathbf p}\in\mathbb R^{2n}$3 are their contribution strengths. A clip is “highly localizable” if $\tilde{\mathbf p}\in\mathbb R^{2n}$4 with $\tilde{\mathbf p}\in\mathbb R^{2n}$5. Speech emerges as the strongest positive contributor, while generic noise sources such as engine or train horn and globally pervasive natural sounds such as rain and wood act as strong negative contributors. The metric is used to filter crowd-sourced recordings and curate a benchmark of 1,444 audio clips spanning 72 countries and territories (Zhang et al., 6 Jan 2026).

Both examples detach localizability from final prediction accuracy. A VLM or ALM may have strong category semantics, yet poor localizability if its internal features do not identify where the evidence lies or if the sensory signal lacks geographically distinctive content. Conversely, localizability metrics can be used to curate datasets, initialize detectors, or interpret failure modes.

7. Cross-disciplinary themes and distinctions

Across these literatures, localizability is not a single invariant property. In graph-based localization it is typically a uniqueness property tied to rigidity, anchor sufficiency, and graph connectivity. In cooperative estimation it is a geometric-statistical property of the FIM, CRB, or measurement distinguishability set. In robotic perception it is often estimator-specific and direction-specific, expressing whether scan geometry constrains particular pose components. In optimization it quantifies whether global improvement is accessible through local exchanges. In complexity theory it characterizes whether a proof method survives local oracle augmentation. In QFT it asks whether position probabilities can be consistently defined on a one-particle space. In multimodal learning it measures whether sensory cues or latent features reveal spatial, temporal, or geographic evidence (Ping et al., 2021, Ny et al., 2018, Nubert et al., 2022, Fujii, 2020, Pich, 2022, Yokomizo et al., 2010, Bao et al., 2024, Zhang et al., 6 Jan 2026).

Several recurrent distinctions nevertheless appear. First, localizability is frequently schema-dependent: a graph can be localizable under nonlinear distance constraints yet fail barycentric linear localizability, and a field theory can admit one localization scheme while obstructing another (Ping et al., 2021, Papageorgiou et al., 2019). Second, localizability is often not synonymous with accuracy. CRB-based formulations treat it as the quality of achievable estimation, LiDAR formulations treat it as a property of the geometry before optimization succeeds or fails, and audio or VLM formulations treat it as informativeness of the signal or internal representation rather than the final prediction itself (Ny et al., 2018, Nubert et al., 2022, Zhang et al., 6 Jan 2026). Third, many works separate necessary from sufficient conditions: clique conditions versus Recursive-3DP in BLL, triangle-bar or triangle-extension constructions versus full global rigidity in WSNs, and good-versus-degenerate eigendirections in X-ICP versus a full-rank unconstrained ICP update (Ping et al., 2021, Sau et al., 2013, Wu et al., 2018, Tuna et al., 2022).

This suggests that any use of the term “localizability” should be interpreted relative to a specific observation model, structural prior, and inferential objective. A plausible implication is that cross-disciplinary comparisons are most meaningful not at the level of the word itself, but at the level of the underlying question: whether the available local information is sufficient to rule out ambiguity.