pith. sign in

arxiv: 2606.06640 · v1 · pith:CBU2GR4Lnew · submitted 2026-06-04 · 📡 eess.SP

SEMIKHORN: Globally balanced affinities for mmWave Localization in MU mMIMO systems

Abhisha Garg , Raghav Shukla , Suraj Srivastava , Aditya K. Jagannatham This is my paper

Pith reviewed 2026-06-27 23:58 UTC · model grok-4.3

classification 📡 eess.SP
keywords SEMIKHORNchannel chartingmmWave localizationt-SNEkhornentropic optimal transportsemisupervised learningMU mMIMOCSI fusion
0
0 comments X

The pith

SEMIKHORN achieves 6.86 percent mean localization error over a 100 m radius using less than 15 percent labeled CSI samples by fusing local dissimilarity matrices through globally balanced affinities.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents SEMIKHORN as a semisupervised channel charting framework for mmWave localization in multi-user massive MIMO systems with distributed base stations. It replaces standard t-SNE with t-SNEkhorn, which applies entropic optimal transport to produce pairwise similarities that are normalized globally rather than independently per data point. Each base station builds a local dissimilarity matrix from its channel state information, these matrices are combined into one global dissimilarity matrix, and manifold learning then embeds the users into a low-dimensional geometric map. The approach is tested in a simulated outdoor setting, with Bayesian optimization used to tune hyperparameters for minimal mean localization error. A reader would care because the method reaches useful positioning accuracy while depending on only a small share of labeled CSI, which is expensive to collect in practice.

Core claim

The central claim is that t-SNEkhorn generates globally balanced similarities via entropic optimal transport, enabling the fusion of local CSI dissimilarity matrices from distributed base stations into a single global matrix that, after manifold learning, yields user embeddings with a mean localization error of 6.86 percent inside a 100 m radius while using fewer than 15 percent labeled CSI samples.

What carries the argument

t-SNEkhorn, a doubly stochastic variant of t-SNE that employs entropic optimal transport to enforce globally balanced pairwise similarities instead of independent per-point normalization.

If this is right

  • Localization accuracy improves when local CSI dissimilarity matrices from multiple base stations are fused rather than used in isolation.
  • Hyperparameter tuning via Bayesian optimization can be used to minimize mean localization error for a given deployment radius.
  • The framework operates with a small fraction of labeled CSI samples while still producing usable position embeddings.
  • Manifold learning applied to the fused global dissimilarity matrix produces a geometric map suitable for user positioning in distributed mMIMO networks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same global-balancing step could be applied to other manifold-learning tasks that combine measurements from spatially separated sensors.
  • Reducing the labeled-sample requirement may lower the cost of deploying channel-charting solutions in new frequency bands or cell sizes.
  • The method's reliance on simulated environments leaves open whether the same error levels hold when real hardware impairments and dynamic scatterers are present.
  • Extending the fusion step to include time-varying CSI could support tracking of mobile users without increasing the labeled-sample budget.

Load-bearing premise

The simulated outdoor environment and the fusion of local dissimilarity matrices from the distributed base stations accurately capture the geometric relationships that the manifold embedding needs to reflect true user positions.

What would settle it

A measurement campaign in a real outdoor mmWave deployment in which the observed mean localization error exceeds 6.86 percent or the fraction of labeled CSI samples required exceeds 15 percent would falsify the reported performance.

Figures

Figures reproduced from arXiv: 2606.06640 by Abhisha Garg, Aditya K. Jagannatham, Raghav Shukla, Suraj Srivastava.

Figure 1
Figure 1. Figure 1: Semi-supervised multi-point CC framework received at the b-th BS, the corresponding CSI Rb i is set to be zero. In a static radio environment with omnidirectional antennas and fixed transmit power, the channel covariance solely depends on the UEs spatial location. Thus, CC assumes the existence of a continuous mapping from the spatial location to the covariance-based CSI [3]. Therefore, the CSI samples {{R… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of low-dimensional CC for different state-of-the-art algorithms (a) considered scenario (b) SEMIKHORN (c) t-SNE (d) Laplacian Eigenmaps Algorithm 3: SEMIKHORN framework Input: Dissimilarity matrix D, ground-truth coordinates {y1, · · · , yF }, labeled set L = {1, · · · , F} Output: Low-dimensional matrix Z (L) 1 Network Parameters: Perplexity ϕ, max iterations I, learning rate η 2 Initialization… view at source ↗
Figure 3
Figure 3. Figure 3: Correlation between high dimensional space and true distance (a) Euclidean metric (b) Manhattan metric MLE function q(ck). Let Jk = {ck, q(ck)} K k=1 represent the complete dataset of evaluated hyperparameters. The posterior distribution for a new sample is then given by p(qnew|cnew, JK) = N (qnew|µ(cnew), σ˜ 2 (cnew)), (17) where µ(·) and σ˜ 2 (·) denote the mean and variance of the pre￾dicted MLE, respec… view at source ↗
Figure 4
Figure 4. Figure 4: (a) Comparison of continuity for various CC algorithms (b) Comparison of continuity for various CC algorithms (c) CDF for localizaion errors (d) Effect of perplexity in SEMIKHORN Manhattan distance metric: We consider the Manhattan dis￾tance metric [11] as our dissimilarity measure, which is defined by df (A, B) = PM i=1 PN j=1 |Ai,j − Bi,j | [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
read the original abstract

This work conceives SEMIKHORN, a semisupervised channel charting (CC) framework for mmWave localization, which leverages t-SNEkhorn, a doubly stochastic variant of t-distributed Stochastic Neighbor Embedding (t-SNE) that utilizes entropic optimal transport to construct pairwise similarities. Unlike standard t-SNE, which normalizes affinities independently for each data point, t-SNEkhorn generates globally balanced similarities ensuring consistent neighborhood representation. We consider wireless networks with distributed base stations (BSs) equipped with multiple antennas, where each BS constructs a local dissimilarity matrix from the channel state information (CSI). These local dissimilarity matrices are then fused to obtain a single global dissimilarity matrix, which is processed through manifold learning to embed users onto a geometric map. The performance is evaluated in a simulated outdoor environment, and Bayesian optimization is employed on the framework hyperparameters to minimize the mean localization error (MLE). Experimental results demonstrate that the proposed framework achieves an MLE of 6.86% in a circular vicinity of radius 100m, requiring less than 15% of labeled CSI samples.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes SEMIKHORN, a semisupervised channel charting framework for mmWave localization in multi-user massive MIMO systems with distributed base stations. It introduces t-SNEkhorn, a doubly stochastic variant of t-SNE that employs entropic optimal transport to generate globally balanced pairwise similarities from local CSI-derived dissimilarity matrices at each BS. These are fused into a single global dissimilarity matrix that is then processed via manifold learning to produce a geometric embedding of user positions. Hyperparameters are tuned via Bayesian optimization, and the method is evaluated in one simulated outdoor environment, where it reports an MLE of 6.86% within a 100 m radius using fewer than 15% labeled CSI samples.

Significance. If the fusion step reliably encodes Euclidean geometry, the globally balanced affinities could reduce labeled-sample requirements in distributed mmWave channel charting while improving neighborhood consistency across BSs. The adaptation of entropic OT to t-SNE is a concrete technical contribution that builds directly on established manifold-learning tools; the reported numerical result, if reproducible with full parameter disclosure, would supply a useful benchmark for semisupervised localization methods.

major comments (2)
  1. [Experimental evaluation] The central performance claim (MLE of 6.86% with <15% labels) rests on the assumption that the fused global dissimilarity matrix preserves the geometric relationships present in the true user positions. No independent diagnostic—such as a reported correlation between the fused affinities and ground-truth Euclidean distances—is described in the experimental evaluation, leaving open the possibility that the embedding succeeds numerically without reflecting physical locations (see skeptic note on fusion of local CSI matrices).
  2. [Abstract and § on experiments] The abstract and evaluation section state a specific MLE figure but supply no details on simulation parameters (number of users, antenna counts, path-loss model, array response), baseline methods, number of Monte-Carlo trials, error bars, or data-exclusion rules. These omissions are load-bearing for assessing whether the reported result generalizes beyond the single simulated outdoor scenario.
minor comments (2)
  1. The acronym MLE is introduced for mean localization error; a brief parenthetical clarification would avoid potential confusion with maximum-likelihood estimation.
  2. [Method section] Notation for the local-to-global fusion operation could be made more explicit (e.g., an equation showing how the per-BS matrices are combined before t-SNEkhorn).

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and will revise the manuscript accordingly to improve clarity and reproducibility.

read point-by-point responses

  1. Referee: [Experimental evaluation] The central performance claim (MLE of 6.86% with <15% labels) rests on the assumption that the fused global dissimilarity matrix preserves the geometric relationships present in the true user positions. No independent diagnostic—such as a reported correlation between the fused affinities and ground-truth Euclidean distances—is described in the experimental evaluation, leaving open the possibility that the embedding succeeds numerically without reflecting physical locations (see skeptic note on fusion of local CSI matrices).

    Authors: We agree that an explicit diagnostic would strengthen the claim that the fusion step preserves geometry. Although the reported MLE is consistent with effective embedding, we will add a correlation analysis (or similar metric) between the fused global dissimilarity matrix and ground-truth Euclidean distances to the experimental evaluation in the revision. revision: yes

  2. Referee: [Abstract and § on experiments] The abstract and evaluation section state a specific MLE figure but supply no details on simulation parameters (number of users, antenna counts, path-loss model, array response), baseline methods, number of Monte-Carlo trials, error bars, or data-exclusion rules. These omissions are load-bearing for assessing whether the reported result generalizes beyond the single simulated outdoor scenario.

    Authors: We acknowledge that the current presentation lacks sufficient experimental detail for full reproducibility and generalization assessment. In the revised manuscript we will expand the evaluation section with complete simulation parameters, baseline descriptions, Monte-Carlo trial counts, error bars, and data-handling rules. The abstract will be updated if space permits to reference these additions. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation applies established t-SNE/OT methods to fused CSI dissimilarities

full rationale

The paper constructs local dissimilarity matrices from CSI at each BS, fuses them into a global matrix, and embeds via manifold learning based on t-SNEkhorn (entropic OT variant of t-SNE). Hyperparameters are tuned by Bayesian optimization to minimize MLE on simulated data; the reported 6.86% MLE is the optimized simulation outcome, not a quantity forced by construction from the inputs. No self-definitional equations, no fitted parameters renamed as independent predictions, and no load-bearing self-citations appear in the derivation chain. The method remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Abstract-only review provides limited visibility into parameters and assumptions; Bayesian optimization implies fitted hyperparameters, and the framework rests on the unverified assumption that local CSI matrices can be fused without loss of geometric fidelity.

free parameters (1)
  • t-SNEkhorn and manifold learning hyperparameters
    Bayesian optimization employed to minimize mean localization error, indicating values chosen to fit the simulated data.
axioms (1)
  • domain assumption Local dissimilarity matrices constructed from CSI at distributed BSs can be fused into a single global matrix that preserves the underlying user geometry for manifold embedding.
    Central to the framework description in the abstract.

pith-pipeline@v0.9.1-grok · 5735 in / 1415 out tokens · 26493 ms · 2026-06-27T23:58:21.183350+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 1 canonical work pages · 1 internal anchor

  1. [1]

    IEEE Wireless Communications , volume=

    G. IEEE Wireless Communications , volume=. 2024 , publisher=

    2024

  2. [2]

    Polyanskiy, Yury and Wu, Yihong , year=

  3. [3]

    2024 , publisher=

    Stephan, Phillip and Euchner, Florian and Ten Brink, Stephan , journal=. 2024 , publisher=

    2024

  4. [4]

    Advances in Neural Information Processing Systems , volume=

    Van Assel, Hugues and Vayer, Titouan and Flamary, R. Advances in Neural Information Processing Systems , volume=

  5. [5]

    2008 IEEE international symposium on information theory , pages=

    P. 2008 IEEE international symposium on information theory , pages=. 2008 , organization=

    2008

  6. [6]

    Ruder, Sebastian , journal=

  7. [7]

    2021 , organization=

    Deng, Junquan and Shi, Wei and Hu, Jian and Jiao, Xianlong , booktitle=. 2021 , organization=

    2021

  8. [8]

    2018 52nd Asilomar Conference on Signals, Systems, and Computers , year=

    Deng, Junquan and Medjkouh, Sa. 2018 52nd Asilomar Conference on Signals, Systems, and Computers , year=

    2018

  9. [9]

    2011 , publisher=

    Jornet, Josep Miquel and Akyildiz, Ian F , journal=. 2011 , publisher=

    2011

  10. [10]

    Foundations and Trends

    Peyr. Foundations and Trends. 2019 , publisher=

    2019

  11. [11]

    2023 , publisher=

    Ferrand, Paul and Guillaud, Maxime and Studer, Christoph and Tirkkonen, Olav , journal=. 2023 , publisher=

    2023

  12. [12]

    IEEE Access , volume=

    Studer, Christoph and Medjkouh, Sa. IEEE Access , volume=. 2018 , publisher=

    2018

  13. [13]

    2019 IEEE 20th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC) , pages=

    Huang, Pengzhi and Casta. 2019 IEEE 20th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC) , pages=. 2019 , organization=

    2019

  14. [14]

    2022 , publisher=

    Kim, Jinhong and Ahn, Yongjun and Shim, Byonghyo , journal=. 2022 , publisher=

    2022

  15. [15]

    Advances in neural information processing systems , volume=

    Banach wasserstein gan , author=. Advances in neural information processing systems , volume=

  16. [16]

    Conditional Generative Adversarial Nets

    Conditional generative adversarial nets , author=. arXiv preprint arXiv:1411.1784 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    2020 , publisher=

    Zheng, Ming and Li, Tong and Zhu, Rui and Tang, Yahui and Tang, Mingjing and Lin, Leilei and Ma, Zifei , journal=. 2020 , publisher=

    2020

  18. [18]

    2021 , publisher=

    Li, Yantao and Luo, Jiaxing and Deng, Shaojiang and Zhou, Gang , journal=. 2021 , publisher=

    2021

  19. [19]

    2021 , publisher=

    Wang, Zhengwei and She, Qi and Ward, Tomas E , journal=. 2021 , publisher=

    2021

  20. [20]

    2021 , publisher=

    Balevi, Eren and Andrews, Jeffrey G , journal=. 2021 , publisher=

    2021

  21. [21]

    2025 , publisher=

    Zhang, Xinrong and Liu, Yu and Zhou, Shudong and Bian, Ji and Huang, Jie and Li, Xinru and Xin, Zhichao , journal=. 2025 , publisher=

    2025

  22. [22]

    Advances in neural information processing systems , volume=

    Generative adversarial nets , author=. Advances in neural information processing systems , volume=

  23. [23]

    2024 , publisher=

    Li, Dongbo and Liu, Xiangyu and Yin, Zhisheng and Cheng, Nan and Liu, Jie , journal=. 2024 , publisher=

    2024

  24. [24]

    2025 , organization=

    Wu, Yang and Pan, Yuhang and Ji, Jianghan and Wang, Cheng-Xiang and Li, Junling and Huang, Chen , booktitle=. 2025 , organization=

    2025

  25. [25]

    2016 , publisher=

    Heath, Robert W and Gonzalez-Prelcic, Nuria and Rangan, Sundeep and Roh, Wonil and Sayeed, Akbar M , journal=. 2016 , publisher=

    2016

  26. [26]

    2021 , organization=

    Zhang, Qianqian and Saad, Walid , booktitle=. 2021 , organization=

    2021

  27. [27]

    2021 , organization=

    Deng, Junquan and Tirkkonen, Olav and Zhang, Jianzhao and Jiao, Xianlong and Studer, Christoph , booktitle=. 2021 , organization=

    2021

  28. [28]

    2020 , publisher=

    Hu, Tianyu and Huang, Yang and Zhu, Qiuming and Wu, Qihui , journal=. 2020 , publisher=

    2020

  29. [29]

    2022 , publisher=

    Doshi, Akash S and Gupta, Manan and Andrews, Jeffrey G , journal=. 2022 , publisher=

    2022

  30. [30]

    2023 , publisher=

    Kazemi, Parham and Al-Tous, Hanan and Ponnada, Tushara and Studer, Christoph and Tirkkonen, Olav , journal=. 2023 , publisher=

    2023