Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Tail-Latency-Aware Federated Learning with Pinching Antenna: Latency, Participation, and Placement

Yushen Lin, Zhiguo Ding

TL;DR

The paper tackles tail latency in synchronous wireless FL under non-IID data by introducing PASS, which reshapes uplink latencies via PA placement along a waveguide. It jointly optimizes PA location and client participation to minimize the expected time-to-accuracy, deriving a tail-latency premium in the KKT conditions and revealing a two-class phase transition that governs slow-class participation. The outer PA placement is characterized by a piecewise envelope derivative, and a breakpoint-and-root global search algorithm yields the optimal placement. Simulations confirm theoretical predictions, showing that PASS enables more eligible participation and materially improves wall-clock accuracy, especially under stringent deadlines.

Abstract

Straggler synchronization is a dominant wall-clock bottleneck in synchronous wireless federated learning (FL). Under non-IID data, however, aggressively sampling only fast clients may significantly slow convergence due to statistical heterogeneity. This paper studies PASS-enabled FL, where a radiating pinching antenna (PA) can be activated at an arbitrary position along a dielectric waveguide to reshape uplink latencies. We consider a joint optimization of PA placement and client participation to minimize the expected time-to-accuracy, coupling the exact expected maximum round latency via order statistics with a heterogeneity-aware convergence factor. We derive first-order optimality conditions that reveal an explicit tail-latency premium in the KKT recursion, quantifying how latency gaps are amplified by maximum-order-statistic synchronization. Under a latency-class structure, we obtain a within-class square-root sampling law and establish a two-class phase transition where slow-class participation collapses under an explicit heterogeneity-threshold condition as the per-round sample size grows. For PA placement, we prove a piecewise envelope-derivative characterization and provide an exact breakpoint-and-root candidate-enumeration procedure. Simulation results verify the theoretical findings and show that PASS enables more eligible participation, yielding higher wall-clock accuracy.

Tail-Latency-Aware Federated Learning with Pinching Antenna: Latency, Participation, and Placement

TL;DR

The paper tackles tail latency in synchronous wireless FL under non-IID data by introducing PASS, which reshapes uplink latencies via PA placement along a waveguide. It jointly optimizes PA location and client participation to minimize the expected time-to-accuracy, deriving a tail-latency premium in the KKT conditions and revealing a two-class phase transition that governs slow-class participation. The outer PA placement is characterized by a piecewise envelope derivative, and a breakpoint-and-root global search algorithm yields the optimal placement. Simulations confirm theoretical predictions, showing that PASS enables more eligible participation and materially improves wall-clock accuracy, especially under stringent deadlines.

Abstract

Straggler synchronization is a dominant wall-clock bottleneck in synchronous wireless federated learning (FL). Under non-IID data, however, aggressively sampling only fast clients may significantly slow convergence due to statistical heterogeneity. This paper studies PASS-enabled FL, where a radiating pinching antenna (PA) can be activated at an arbitrary position along a dielectric waveguide to reshape uplink latencies. We consider a joint optimization of PA placement and client participation to minimize the expected time-to-accuracy, coupling the exact expected maximum round latency via order statistics with a heterogeneity-aware convergence factor. We derive first-order optimality conditions that reveal an explicit tail-latency premium in the KKT recursion, quantifying how latency gaps are amplified by maximum-order-statistic synchronization. Under a latency-class structure, we obtain a within-class square-root sampling law and establish a two-class phase transition where slow-class participation collapses under an explicit heterogeneity-threshold condition as the per-round sample size grows. For PA placement, we prove a piecewise envelope-derivative characterization and provide an exact breakpoint-and-root candidate-enumeration procedure. Simulation results verify the theoretical findings and show that PASS enables more eligible participation, yielding higher wall-clock accuracy.
Paper Structure (21 sections, 10 theorems, 44 equations, 7 figures, 1 algorithm)

This paper contains 21 sections, 10 theorems, 44 equations, 7 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. PASS
  3. Straggler and dropout mitigation in federated learning
  4. PASS-empowered edge learning and PASS-FL
  5. Organization.
  6. System Model
  7. Straggler Order Statistics
  8. Tail-Latency Coupling
  9. Coupling with Non-IID Convergence
  10. Optimal Participation for PA Location
  11. Class-wise Reduction
  12. Pinching-Antenna Placement: Exact Characterization and Global Search
  13. Envelope Derivative
  14. Global Placement Algorithm
  15. Simulation
  16. ...and 6 more sections

Key Result

Theorem 1

Under i.i.d. $K$-client sampling with replacement using probabilities $q_i$, let $R_\epsilon$ denote the number of communication rounds required to reach a target accuracy $\epsilon$. The following upper bound can be obtained: where $\omega \triangleq \frac{E}{K}\frac{L_{\mathrm{sm}}}{m_{\mathrm{cv}}^2}$, $c_i \triangleq p_i^2 G_i^2$, $V_{\mathrm{var}}\triangleq \sum_{i=1}^N p_i^2 \chi_i^2$, $V_{

Figures (7)

  • Figure 1: Tail-latency premium induced by synchronous straggling. (a) Optimal slow-class mass $\delta^\star$ versus latency gap $\Delta$. (b) Tail-latency premium $K\Delta(1-\delta^\star)^{K-1}$ at the optimizer. (c) Probability of sampling at least one slow client, $1-(1-\delta^\star)^K$. Larger $K$ amplifies the straggler penalty, driving $\delta^\star$ toward the $\Theta(1/K)$ scale.
  • Figure 2: Complementary CDF of synchronous round time $T_{\mathrm{round}}$ under conventional fixed placement and PASS-optimized placement, for $N\in \{10, 20, 30\}$ clients. PASS shifts the tail left, reducing extreme straggler rounds that dominate wall-clock training.
  • Figure 3: Phase transition of slow-class participation. (a) Optimal $\delta^\star$ versus $K$ for different $C_s$ values. (b) Normalized product $K\delta^\star$: bounded behavior indicates $\delta^\star=\Theta(1/K)$; linear growth indicates $\delta^\star=\Theta(1)$. (c) Ratio $\delta^\star/\delta^\star_{\mathrm{stat}}$, quantifying deviation from the statistics-only optimum.
  • Figure 4: Piecewise-smooth envelope objective versus PA position. Top panel: per-client latency profiles $t_i(x)$; crossings induce ordering changes. Bottom panel: envelope objective $J^\star(x)$; breakpoints occur where $t_i(x)=t_j(x)$. At the optimum $x^\star$ (red line), straggling concentrates on different clients.
  • Figure 5: Latency--convergence trade-off space. Each marker represents a design point in the $(f,g)$ plane, where $f(q,x)$ is the expected straggler latency and $g(q)$ is the convergence factor. Marker shape indicates the method (Conventional, PASS-Random, PASS-Joint) and marker size indicates sample size $K\in\{10,20,30\}$.
  • ...and 2 more figures

Theorems & Definitions (25)

  • Theorem 1
  • proof
  • Lemma 1
  • proof
  • Lemma 2
  • proof
  • Theorem 2
  • Remark 1
  • Remark 2
  • Lemma 3
  • ...and 15 more