LNE: Diverse Concepts in Science

• LNE in metric geometry, defined as Lipschitz normally embedded, ensures that intrinsic path lengths are uniformly controlled by ambient distances.
• In machine learning, LNE encompasses Length Normalized Entropy, Longitudinal Neighborhood Embedding, and nearly Euclidean metrics to enhance model evaluation and training.
• LNE also appears in physics and game theory as the Layer Nernst Effect and local Nash equilibrium, and as the French metrology institute setting benchmarks in quantum and gravimetric standards.

LNE is a polysemous technical acronym. In recent arXiv literature it denotes several distinct concepts: Lipschitz normally embedded in metric and algebraic geometry, Length Normalized Entropy in large-language-model contamination detection, Longitudinal Neighborhood Embedding in self-supervised neuroimaging, Linearly Nearly Euclidean in geometric formulations of discretized neural-network training, Layer Nernst Effect in twisted moiré transport, Linearized Noise Equation in nonlinear optical communications, local Nash equilibrium in game theory and adversarial learning, and LNE as the French national metrology and testing laboratory in quantum benchmarking and gravimetry (Mendes et al., 2021, Hou et al., 18 Sep 2025, Ouyang et al., 2021, Chen et al., 2023, Hu et al., 2024, Zhang et al., 2012, Kosohorská et al., 8 Jun 2026, Barbaresco et al., 2024).

1. Acronymic scope across research domains

The acronym appears in sharply different technical traditions. The table summarizes the principal usages attested in the cited literature (Mendes et al., 2021, Hou et al., 18 Sep 2025, Ouyang et al., 2021, Chen et al., 2023, Hu et al., 2024, Zhang et al., 2012, Kosohorská et al., 8 Jun 2026, Barbaresco et al., 2024, Merlet et al., 2010).

Expansion of LNE	Research area	Defining idea
Lipschitz normally embedded	Metric geometry, singularity theory	Inner and outer metrics are equivalent
Length Normalized Entropy	LLM evaluation	Average token-level entropy on greedy decoding
Longitudinal Neighborhood Embedding	Neuroimaging ML	Neighborhood alignment of latent trajectory vectors
Linearly Nearly Euclidean	Discretized neural networks	Metric written as Euclidean metric plus small perturbation
Layer Nernst Effect	Condensed matter physics	Layer-resolved transverse thermoelectric response
Linearized Noise Equation	Optical fiber communications	Linearized stochastic noise propagation around a noise-free signal
local Nash equilibrium	Game theory, GANs	Local best-response equilibrium notion
LNE (institution)	Metrology	Independent and trusted third party maintaining benchmark suites

A common source of confusion is that these usages are not variants of a single theory. In geometry, LNE is a metric-regularity property of sets and germs. In machine learning and physics, it is a method name, a transport effect, or an equation class. In metrology, it is an institutional acronym.

2. Lipschitz normally embedded sets: definition and basic geometry

In the metric-geometric literature, LNE means Lipschitz normally embedded. For a path-connected subanalytic subset $X \subset \mathbb{R}^n$, the outer metric is the Euclidean restriction

$d(x,y)=\|x-y\|,$

while the inner metric is

$$d_X(x,y)=\inf\{\operatorname{Length}(\gamma)\mid \gamma \text{ is a rectifiable path in }X\text{ connecting }x\text{ and }y\}.$$

The set $X$ is LNE if there exists $C\ge 1$ such that

$$d_X(x,y)\le C\,\|x-y\| \qquad \text{for all } x,y\in X.$$

Since $\|x-y\|\le d_X(x,y)$ always holds, this is equivalent to bilipschitz equivalence of inner and outer metrics (Mendes et al., 2021).

For complex analytic germs, the same idea is expressed with the ambient Hermitian metric. If $\phi:(X,0)\hookrightarrow (\mathbb{C}^n,0)$ is an embedding, then

$d_o(x,y):=\|\phi(x)-\phi(y)\|,$

and

$$d_i(x,y):=\inf\{\operatorname{length}(\phi\circ\gamma): \gamma \text{ is a rectifiable path in } X \text{ from } x \text{ to } y\}.$$

A germ is LNE when the identity map is a bilipschitz homeomorphism between $d(x,y)=\|x-y\|,$0 and $d(x,y)=\|x-y\|,$1, equivalently when

$d(x,y)=\|x-y\|,$2

near the origin for some $d(x,y)=\|x-y\|,$3 (Neumann et al., 2018).

The geometric content is that intrinsic path lengths inside the set are uniformly controlled by ambient Euclidean separation. The supplied literature formulates this as the absence of “hidden long corridors” or “metric folding” near the singular point (Neumann et al., 2018). The property is local: $d(x,y)=\|x-y\|,$4 is LNE at $d(x,y)=\|x-y\|,$5 if some neighborhood $d(x,y)=\|x-y\|,$6 satisfies that $d(x,y)=\|x-y\|,$7 is LNE (Mendes et al., 2021). A further subtlety is that topological conical structure does not automatically yield metric conical structure; this is why link-based characterizations are nontrivial (Mendes et al., 2021).

3. Link criteria, singularities, curves, and finite mappings

A major development is the introduction of link Lipschitz normal embeddedness (LLNE). For $d(x,y)=\|x-y\|,$8 and small $d(x,y)=\|x-y\|,$9, let

$$d_X(x,y)=\inf\{\operatorname{Length}(\gamma)\mid \gamma \text{ is a rectifiable path in }X\text{ connecting }x\text{ and }y\}.$$0

A set is LLNE at $$d_X(x,y)=\inf\{\operatorname{Length}(\gamma)\mid \gamma \text{ is a rectifiable path in }X\text{ connecting }x\text{ and }y\}.$$1 if there exists $$d_X(x,y)=\inf\{\operatorname{Length}(\gamma)\mid \gamma \text{ is a rectifiable path in }X\text{ connecting }x\text{ and }y\}.$$2 such that all sufficiently small slices $$d_X(x,y)=\inf\{\operatorname{Length}(\gamma)\mid \gamma \text{ is a rectifiable path in }X\text{ connecting }x\text{ and }y\}.$$3 are $$d_X(x,y)=\inf\{\operatorname{Length}(\gamma)\mid \gamma \text{ is a rectifiable path in }X\text{ connecting }x\text{ and }y\}.$$4-LNE. For a closed subanalytic germ with connected punctured germ, the central theorem is

$$d_X(x,y)=\inf\{\operatorname{Length}(\gamma)\mid \gamma \text{ is a rectifiable path in }X\text{ connecting }x\text{ and }y\}.$$5

For a nonconnected punctured germ, the corresponding criterion adds the separation condition $$d_X(x,y)=\inf\{\operatorname{Length}(\gamma)\mid \gamma \text{ is a rectifiable path in }X\text{ connecting }x\text{ and }y\}.$$6 for all small $$d_X(x,y)=\inf\{\operatorname{Length}(\gamma)\mid \gamma \text{ is a rectifiable path in }X\text{ connecting }x\text{ and }y\}.$$7. The same work proves that every $$d_X(x,y)=\inf\{\operatorname{Length}(\gamma)\mid \gamma \text{ is a rectifiable path in }X\text{ connecting }x\text{ and }y\}.$$8-regular cell is LLNE at $$d_X(x,y)=\inf\{\operatorname{Length}(\gamma)\mid \gamma \text{ is a rectifiable path in }X\text{ connecting }x\text{ and }y\}.$$9 provided its link at $X$0 is connected, and that bounded subanalytic sets admit finite decompositions into pieces whose closures are LLNE at $X$1 (Mendes et al., 2021).

For normal surface singularities, LNE is characterized via generic projections, polar curves, and test curves. If $X$2 is a generic linear projection and $X$3 is a nodal test curve, then $X$4 is LNE if and only if two conditions hold for principal components $X$5 of $X$6: $X$7 and for distinct principal components $X$8,

$X$9

Here $C\ge 1$0 and $C\ge 1$1 are inner and outer contact exponents read from the asymptotics of distances on small spheres (Neumann et al., 2018).

Within rational surface singularities, the metric condition is exceptionally rigid: a rational surface singularity is LNE if and only if it is minimal (Neumann et al., 2015). For superisolated hypersurface singularities of the form $C\ge 1$2, the criterion becomes projective: the singularity is LNE if and only if its projectivized tangent cone $C\ge 1$3 has only ordinary singularities (Misev et al., 2018).

For complex curves, the local criterion is simpler. A complex analytic curve germ is LNE if and only if it is a finite union of smooth complex curve germs meeting pairwise transversally. Globally, a connected complex affine algebraic curve $C\ge 1$4 is LNE if and only if each singular germ has that transverse smooth-branch form and the projective closure $C\ge 1$5 intersects the hyperplane at infinity in exactly $C\ge 1$6 points. The same work shows that for irreducible affine LNE curves, Lipschitz equivalence is topological and is completely encoded by the invariant

$C\ge 1$7

up to the stated equivalence relation (Costa et al., 2023).

The 2025 study of images of finite map germs extends the rigidity theme. For a finite map germ

$C\ge 1$8

with $C\ge 1$9 smooth and $$d_X(x,y)\le C\,\|x-y\| \qquad \text{for all } x,y\in X.$$0, the image $$d_X(x,y)\le C\,\|x-y\| \qquad \text{for all } x,y\in X.$$1 is LNE if and only if it is smooth, equivalently if and only if $$d_X(x,y)\le C\,\|x-y\| \qquad \text{for all } x,y\in X.$$2. For finite corank $$d_X(x,y)\le C\,\|x-y\| \qquad \text{for all } x,y\in X.$$3 maps this hypothesis is automatic, and for an injective holomorphic map germ $$d_X(x,y)\le C\,\|x-y\| \qquad \text{for all } x,y\in X.$$4, the image is LNE at $$d_X(x,y)\le C\,\|x-y\| \qquad \text{for all } x,y\in X.$$5 if and only if $$d_X(x,y)\le C\,\|x-y\| \qquad \text{for all } x,y\in X.$$6 is an embedding (Ballesteros et al., 29 Jul 2025).

Taken together, these results show that in singularity theory LNE is not a mild regularity label. In many important classes it is effectively a smoothness surrogate, or a metric reformulation of highly constrained local geometry.

4. Machine learning uses: entropy, longitudinal trajectories, and nearly Euclidean metrics

In large-language-model evaluation, LNE stands for Length Normalized Entropy. Given a prompt $$d_X(x,y)\le C\,\|x-y\| \qquad \text{for all } x,y\in X.$$7 and a model $$d_X(x,y)\le C\,\|x-y\| \qquad \text{for all } x,y\in X.$$8, greedy decoding produces $$d_X(x,y)\le C\,\|x-y\| \qquad \text{for all } x,y\in X.$$9, and the score is

$\|x-y\|\le d_X(x,y)$0

The method is based on the observation that contaminated or memorized outputs yield more peaked next-token distributions, hence lower entropy. The paper normalizes the score as

$\|x-y\|\le d_X(x,y)$1

and uses it to set the number of decoding-time Blocking operations through

$\|x-y\|\le d_X(x,y)$2

This is the core control rule in the LNE-Blocking framework, which is reported on HumanEval, GSM8K, GSM-Plus, and ACLSum, with task-specific $\|x-y\|\le d_X(x,y)$3 values $\|x-y\|\le d_X(x,y)$4, $\|x-y\|\le d_X(x,y)$5, and $\|x-y\|\le d_X(x,y)$6 respectively (Hou et al., 18 Sep 2025).

In neuroimaging, Longitudinal Neighborhood Embedding is a self-supervised representation-learning method for longitudinal MRI. Each subject contributes latent codes $\|x-y\|\le d_X(x,y)$7 and $\|x-y\|\le d_X(x,y)$8, together with a normalized trajectory vector

$\|x-y\|\le d_X(x,y)$9

A dynamic graph is built in each training iteration from starting points $\phi:(X,0)\hookrightarrow (\mathbb{C}^n,0)$0, neighbor relations are computed from Euclidean distances, and neighborhood trajectories are pooled as

$\phi:(X,0)\hookrightarrow (\mathbb{C}^n,0)$1

The training objective combines reconstruction with directional alignment,

$\phi:(X,0)\hookrightarrow (\mathbb{C}^n,0)$2

The method was evaluated on a healthy aging dataset of 274 subjects and on ADNI with $\phi:(X,0)\hookrightarrow (\mathbb{C}^n,0)$3, where the paper reports superior downstream performance relative to AE, VAE, SimCLR, and LSSL on age regression and diagnostic classification tasks (Ouyang et al., 2021).

In discretized neural-network training, LNE means Linearly Nearly Euclidean. The paper introduces an LNE metric as a small perturbation of the Euclidean metric, first in the abstract form

$\phi:(X,0)\hookrightarrow (\mathbb{C}^n,0)$4

and then through an information-geometric construction with

$\phi:(X,0)\hookrightarrow (\mathbb{C}^n,0)$5

The corresponding steepest descent is

$\phi:(X,0)\hookrightarrow (\mathbb{C}^n,0)$6

which reinterprets gradient mismatch in quantized networks as a metric perturbation rather than only as an ad hoc estimator error. Stability is then analyzed under Ricci flow and Ricci-DeTurck flow, with the paper claiming exponential decay of perturbations,

$\phi:(X,0)\hookrightarrow (\mathbb{C}^n,0)$7

This geometric framework is used to motivate RF-DNN, which the paper reports as more accurate and stable than several representative training-based quantization methods (Chen et al., 2023).

These three ML usages share the acronym only superficially. One is a decoding-time contamination signal, one is a self-supervised latent-trajectory method, and one is a Riemannian model for optimization on discretized networks.

5. Physics and engineering uses: layer thermoelectricity and optical-noise propagation

In condensed-matter transport, Layer Nernst Effect is a layer-resolved transverse thermoelectric response in twisted moiré multilayers. For a bilayer with a temperature gradient along $\phi:(X,0)\hookrightarrow (\mathbb{C}^n,0)$8, opposite transverse currents can flow in the two layers, giving a layer-contrasted current

$\phi:(X,0)\hookrightarrow (\mathbb{C}^n,0)$9

The theoretical mechanism identified in twisted moiré layers combines trigonal warping of the Fermi surface, a layer-contrasted pseudomagnetic field $d_o(x,y):=\|\phi(x)-\phi(y)\|,$0, and the resulting imbalance between left- and right-moving carriers. The paper defines the antisymmetric LNE coefficient as

$d_o(x,y):=\|\phi(x)-\phi(y)\|,$1

and expresses it through the layer velocity curvature

$d_o(x,y):=\|\phi(x)-\phi(y)\|,$2

In twisted bilayer graphene, the response vanishes when the relevant symmetry is restored, for example at $d_o(x,y):=\|\phi(x)-\phi(y)\|,$3 or in the chiral limit $d_o(x,y):=\|\phi(x)-\phi(y)\|,$4, and the predicted magnitude reaches $d_o(x,y):=\|\phi(x)-\phi(y)\|,$5 at $d_o(x,y):=\|\phi(x)-\phi(y)\|,$6 K, corresponding to roughly $d_o(x,y):=\|\phi(x)-\phi(y)\|,$7 for a bilayer thickness of about $d_o(x,y):=\|\phi(x)-\phi(y)\|,$8 nm (Hu et al., 2024).

In optical communications, Linearized Noise Equation describes the propagation of noise in nonlinear fiber systems when signal power is much larger than noise power. Writing

$d_o(x,y):=\|\phi(x)-\phi(y)\|,$9

and neglecting quadratic terms in $$d_i(x,y):=\inf\{\operatorname{length}(\phi\circ\gamma): \gamma \text{ is a rectifiable path in } X \text{ from } x \text{ to } y\}.$$0, one obtains a linear stochastic equation for the perturbation. In the EDFA-amplified setting, ASE noise is modeled as a forcing term and the resulting LNE preserves Gaussianity of the propagated noise. The 2012 paper contrasts earlier CW-approximation-based LNE methods with the more accurate LNE of Holzlöhner et al., computes the noise propagator directly from the accurate LNE via RK4IP, and then derives the moment-generating function of the filtered receiver current. The reported BER predictions agree well with experimental data for multi-span DPSK systems (Zhang et al., 2012).

The two physical uses are again independent. One concerns a transverse thermoelectric response in twistronics; the other is a linearized stochastic model for nonlinear optical fiber propagation.

6. Game-theoretic LNE: local Nash equilibrium, FONE, and adversarial learning

In game theory, LNE abbreviates local Nash equilibrium. For a continuous game

$$d_i(x,y):=\inf\{\operatorname{length}(\phi\circ\gamma): \gamma \text{ is a rectifiable path in } X \text{ from } x \text{ to } y\}.$$1

with compact convex strategy sets, a profile $$d_i(x,y):=\inf\{\operatorname{length}(\phi\circ\gamma): \gamma \text{ is a rectifiable path in } X \text{ from } x \text{ to } y\}.$$2 is an LNE if there exists $$d_i(x,y):=\inf\{\operatorname{length}(\phi\circ\gamma): \gamma \text{ is a rectifiable path in } X \text{ from } x \text{ to } y\}.$$3 such that

$$d_i(x,y):=\inf\{\operatorname{length}(\phi\circ\gamma): \gamma \text{ is a rectifiable path in } X \text{ from } x \text{ to } y\}.$$4

for every player $$d_i(x,y):=\inf\{\operatorname{length}(\phi\circ\gamma): \gamma \text{ is a rectifiable path in } X \text{ from } x \text{ to } y\}.$$5. In continuously differentiable games, any LNE satisfies the first-order inequalities

$$d_i(x,y):=\inf\{\operatorname{length}(\phi\circ\gamma): \gamma \text{ is a rectifiable path in } X \text{ from } x \text{ to } y\}.$$6

and the corresponding global variational inequality formulation

$$d_i(x,y):=\inf\{\operatorname{length}(\phi\circ\gamma): \gamma \text{ is a rectifiable path in } X \text{ from } x \text{ to } y\}.$$7

defines a first-order Nash equilibrium (FONE). The implication is one-way in general: LNE implies FONE, but FONE does not imply LNE. This distinction is central to the 2026 analysis of STON’R, which proves convergence to exact or approximate FONE in multiplayer general-sum games rather than to LNE, precisely because LNE may fail to exist even in zero-sum games (Kosohorská et al., 8 Jun 2026).

GAN theory uses the same acronym in a related but more algorithmic way. One line of work argues that current gradient-based GAN methods are at best guaranteed to converge to an LNE in parameter space, and that such LNEs can be arbitrarily far from a true Nash equilibrium. To remove the gap, adversarial networks are reformulated as finite zero-sum games in mixed strategies; in that setting every LNE is an NE, and the proposed Parallel Nash Memory method monotonically converges to a resource-bounded Nash equilibrium (RB-NE) (Oliehoek et al., 2018).

A separate 2022 study formulates GAN training as the problem of finding a stationary LNE, defined by the simultaneous vanishing conditions

$$d_i(x,y):=\inf\{\operatorname{length}(\phi\circ\gamma): \gamma \text{ is a rectifiable path in } X \text{ from } x \text{ to } y\}.$$8

It then applies nonlinear conjugate-gradient updates to generator and discriminator, proving convergence to the LNE problem under constant and diminishing learning-rate schedules and reporting improved FID relative to SGD and momentum SGD, with average performance better than Adam (Naganuma et al., 2022).

This body of work places LNE at the boundary between equilibrium theory and nonconvex-nonconcave optimization. The recurring theme is that local equilibrium notions are mathematically natural but may be weaker than the solution concepts ultimately desired.

7. Metrological and institutional usage: LNE in gravimetry and quantum benchmarking

In French metrology, LNE is the French national metrology and testing laboratory and the coordinator of MetriQs-France, the national program on Measurement, Evaluation, and Standards for Quantum Technologies. The BACQ project, funded within that framework, is devoted to application-oriented benchmarks for quantum computing. The consortium includes THALES, EVIDEN, CEA, CNRS, TERATEC, and LNE, and the benchmark suite is to be maintained by LNE as an independent and trusted third party. The methodology uses the MYRIAD-Q multi-criteria decision-analysis framework and a hierarchical aggregation model based on the 2-additive Choquet integral, with benchmark families in optimization, linear system solving, quantum physics simulation, and prime factorization (Barbaresco et al., 2024).

The same institutional acronym appears in precision gravimetry. The 2010 bilateral comparison between the LNE-SYRTE cold atoms gravimeter (CAG) and FG5#220 from Leibniz Universität Hannover was carried out in the LNE watt balance laboratory, specifically in its gravimetry room (GR). The two portable absolute gravimeters rely on different principles—atomic interferometry for the CAG and optical interferometry for the FG5—and their bilateral comparison showed an agreement of $$d_i(x,y):=\inf\{\operatorname{length}(\phi\circ\gamma): \gamma \text{ is a rectifiable path in } X \text{ from } x \text{ to } y\}.$$9, with the paper’s table also reporting

$d(x,y)=\|x-y\|,$00

The gravity unit is defined as

$d(x,y)=\|x-y\|,$01

The measurement points were separated by $d(x,y)=\|x-y\|,$02 m, their gravity difference transferred to the same height was measured with a Scintrex CG5 as $d(x,y)=\|x-y\|,$03, and the CAG final accuracy during the comparison was $d(x,y)=\|x-y\|,$04, with stated contributions from Coriolis shift, wavefront aberrations, and vertical alignment bias (Merlet et al., 2010).

In this institutional sense, LNE does not denote a theorem or an algorithm. It denotes the metrological organization that anchors comparison, stewardship, and standardization activities across domains ranging from gravimetry to quantum benchmarking.