{"id":"arxiv/2606.11339","url":"https://arxiv.org/abs/2606.11339","tags":["math.OC","cs.AI","cs.LG","eess.SY","stat.ML"],"score":0,"title":"Quantized Stochastic Primal-Dual Methods for Distributed Optimization under Relaxed Global Geometry","domain":"math.OC","source":"arxiv","authors":["Susmit Sarkar","Abhinav Raghuvanshi","Kushal Chakrabarti","Mayank Baranwal"],"reviews":0,"subject":"math.OC","summary":"We study distributed optimization with stochastic gradients and finite-bit communication modeled by random (unbiased) quantization. We propose q-PDGD, a quantized stochastic primal-dual method, and analyze it under relaxed global geometry. Under restricted secant inequality (RSI), a constant step-size yields linear contraction to an explicit neighborhood determined by gradient noise, quantization distortion, and network connectivity, while a diminishing step-size achieves O(1/k) convergence without shared-minimizer assumptions. Under Polyak-Lojasiewicz (PL) inequality, we obtain…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-06-09","reactions":{}}