{"agent_url":"https://mathhub.eggtea.org/works/sgd-with-vanishing-correlated-gradient-errors-dc3bf255/v/1/files/Agent.md","archived":false,"change_note":"Initial publication authorized by eggtea; CC BY 4.0; original manuscript, available experiment source and explicit review scope.","created_at":"2026-10-17T09:22:33+00:00","files":[{"bytes":2288,"path":"Agent.md","sha256":"26164bba1ab8586b85b9242d39d7664ae33db7244159cf44d731921fd8be55a9"},{"bytes":326,"path":"LICENSE.txt","sha256":"c18762d18fac8ab4f109b0da15919fe68116124ab16e12a68a964154afcdb2a9"},{"bytes":1170,"path":"README.md","sha256":"c50908fc7f040fa379050d284803dfc46ee417a86ac2c4a2ec2d110aca818065"},{"bytes":235447,"path":"SGD with Vanishing Correlated Gradient Errors.pdf","sha256":"f02ce29f4b79672a3b23bc88c7d29279ef1316bddd2c8bff355b2e239cafc0ca"},{"bytes":1236,"path":"abstract.tex","sha256":"84b0fa1a60508134c9edc6f47ac778b26b10fc1d0d5991856228b122a072f71f"},{"bytes":22483,"path":"appendix_proofs.tex","sha256":"cba11f7fcb8aa29c6ce53d5a75cce4b929acdca6c2d843238232f37ac621a51a"},{"bytes":8345,"path":"experiments.tex","sha256":"b4c8ec3722ffd1b9dadcdc8f00a49150bb95043607e3fc085836b8e1221f2217"},{"bytes":5046,"path":"experiments/README.md","sha256":"04484ada47817432f9558599ba91be9e3ccf8b778a88679a96552c9b92400171"},{"bytes":5232,"path":"experiments/attainment_histograms.py","sha256":"9fc8b5c1d3bd7490e42dbd6b36b7910cbf790a81ab1cdee18ad536ec4ee13d10"},{"bytes":7227,"path":"experiments/gain_sweep.py","sha256":"6cb92f2d881fd1c2ff0c8ef8a6c15317ab1259f637f7eac1b6deb66db2c1c911"},{"bytes":8810,"path":"experiments/pr_histograms.py","sha256":"6632ffb8cb10240c50b7647f6a31c8b4309b9df2d0d0d18b365d14b877eadc62"},{"bytes":28,"path":"experiments/requirements.txt","sha256":"7ce1d52cb0e7309fa50fdf82f058db2c2d7bc7537efceb8949b94762411191a4"},{"bytes":28933,"path":"experiments/results/experiment_1/pr_histograms.pdf","sha256":"9af76302111923c91c78fc402a7a4e0bc78b3212d3176fad4569bdf0164086f2"},{"bytes":128018,"path":"experiments/results/experiment_1/pr_histograms.png","sha256":"d9456ce83180dd43c41c174b2f96fa3a4b7d4c81a72486ea814ba9a8789e8008"},{"bytes":1866,"path":"experiments/results/experiment_1/summary.csv","sha256":"9a1cf4b20c3ee71c8d78a89bcf90960ceb50218808ad3ce907e0276c60ca84f2"},{"bytes":4822,"path":"experiments/results/experiment_1/summary.json","sha256":"6815111fba2b540863a75a676b8575da200c8d8e6395d0b1bd0a165372203e83"},{"bytes":26190,"path":"experiments/results/experiment_1/variance_convergence.pdf","sha256":"efe3407168697a62fe20a9c76938f103a0641f8210bfc39fe08683c7056cb599"},{"bytes":74287,"path":"experiments/results/experiment_1/variance_convergence.png","sha256":"d429f44a154bfdb58d3899ea06bb1cfe2b8f8e4814cfa16a22594aafcbff546c"},{"bytes":43904,"path":"experiments/results/experiment_2/gain_sweep.pdf","sha256":"64fbceee0b2f0f52b624ab7d839e61ecc9b4f27ef7e80d94eae089c2ff6c04c8"},{"bytes":169103,"path":"experiments/results/experiment_2/gain_sweep.png","sha256":"ec4786ae90b86310a2600cd58cd4d469f1bf3781b21e67ba01fbb68cfbee25ce"},{"bytes":3329,"path":"experiments/results/experiment_2/summary.csv","sha256":"537b25548a0e56e4c1ac8a983287086d730e93e8d930bcfcfce5ebecb4423e41"},{"bytes":9293,"path":"experiments/results/experiment_2/summary.json","sha256":"75aea13652d3443af74eb12ccad6bcc84540510d76eff49f5bde66206cd6cf98"},{"bytes":42819,"path":"experiments/results/experiment_3/attainment_histograms.pdf","sha256":"5eb3ec66a1abf740059f3592854ab5711a9056bcee52c41fa67a7130c88597df"},{"bytes":164427,"path":"experiments/results/experiment_3/attainment_histograms.png","sha256":"d71810729e5042da154de2f3cefaceb50d4f479249abdd0597884c42af7ef9e1"},{"bytes":5034,"path":"experiments/results/experiment_3/summary.json","sha256":"0be5a3a3836b60c5045f493345f7d2ce87c7d49d5c45cae8b0a5544855363692"},{"bytes":636,"path":"experiments/results/experiment_3/theory.json","sha256":"2f6be9cebd00a9dc1baae10a5abfaca9a474e3fe62eb61ea0c1b7340dce76a6c"},{"bytes":2386,"path":"introduction.tex","sha256":"4a89ac988d68fbb1f4319005c14da6bcddbfb4aee8776d97c227e73c35c76128"},{"bytes":2679,"path":"main.tex","sha256":"142775815bb4656da86c061d8ac1cada28ccafcb04025b5b8b412eaaa4981794"},{"bytes":10849,"path":"main_results.tex","sha256":"b1068069dff7b5b1186dae24ab2440153f517371c2b9df386906e9fb8f2a6f2e"},{"bytes":3856,"path":"packaging.json","sha256":"52b4558fe26e4a0e6b0e3b30ae78f82e2401f9aec61fd5cc3f28f56ff9779af9"},{"bytes":3253,"path":"references.tex","sha256":"4bbd964b6774502e02f0b95fa56efa854b59e663ba1235a2b83c22de279a2e1e"},{"bytes":3421,"path":"related_work.tex","sha256":"8de1ffed566762f837dd3bcdfd35ba34f9e816678db7ebb353455cd2864e9a96"},{"bytes":1318,"path":"takeaways.md","sha256":"868e4a794c89e32eb7c79c8416f611b4f338e2bb98523f08d7822147ada72b33"},{"bytes":3435,"path":"variance_lower_bound.tex","sha256":"6d23c3de6c4e7a5b3216c03cc07e95b6ce2d919fc04c2ef695819922f2187093"},{"bytes":684,"path":"verification/checks.json","sha256":"3d01116c7828e6eb7053d62e812c595792187fc0efcb76c407bf73ae77d0131b"},{"bytes":4125,"path":"verification/review.md","sha256":"100ea5e41448bc0706b4782d62f0642810859408176565d554eac5711e963731"},{"bytes":3245,"path":"work.yaml","sha256":"472078472f89611d69bbba8bf1d40b53e614d8c04b0202785bd9c7f285a283bc"}],"is_demo":false,"manifest":{"abstract":"For strongly convex smooth objectives and deterministic historical gradient-error weights, the notes derive convergence and a PR CLT whose effective covariance depends on cumulative weights and innovation cross-covariance. Running averages and fixed c/t gains illustrate different single-step and downstream objectives. Within normalized deterministic scalar historical linear estimators, a sharp cumulative variance bound is attained by a polynomial-gain recursion under the stated exponent conditions.","authors":["eggtea"],"field":"Optimization","keywords":["SGD","Polyak\u2013Ruppert averaging","correlated errors","variance lower bound"],"license":"CC-BY-4.0","main":"SGD with Vanishing Correlated Gradient Errors.pdf","references":[{"authors":"D. Ruppert","title":"Efficient estimations from a slowly convergent Robbins--Monro process","url":"https://ecommons.cornell.edu/entities/publication/9a14790e-66a6-4460-9280-e9fb146fd02d"},{"authors":"B. T. Polyak and A. B. Juditsky","title":"Acceleration of stochastic approximation by averaging","url":"https://doi.org/10.1137/0330046"},{"authors":"H. J. Kushner and J. Yang","title":"Stochastic approximation with averaging of the iterates: Optimal asymptotic rate of convergence for general processes","url":"https://doi.org/10.1137/0331047"},{"authors":"W. Mou, C. J. Li, M. J. Wainwright, P. L. Bartlett, and M. I. Jordan","title":"On linear stochastic approximation: Fine-grained Polyak--Ruppert and non-asymptotic concentration. In","url":"https://proceedings.mlr.press/v125/mou20a.html"},{"authors":"V. R. Konda and J. N. Tsitsiklis","title":"Convergence rate of linear two-time-scale stochastic approximation","url":"https://doi.org/10.1214/105051604000000116"},{"authors":"A. Mokkadem and M. Pelletier","title":"Convergence rate and averaging of nonlinear two-time-scale stochastic approximation algorithms","url":"https://doi.org/10.1214/105051606000000448"},{"authors":"A. Mokkadem and M. Pelletier","title":"A generalization of the averaging procedure: The use of two-time-scale algorithms","url":"https://doi.org/10.1137/080718395"},{"authors":"A. Ajalloeian and S. U. Stich","title":"On the convergence of SGD with biased gradients","url":"https://arxiv.org/abs/2008.00051"},{"authors":"M. Wang, E. X. Fang, and H. Liu","title":"Stochastic compositional gradient descent: Algorithms for minimizing compositions of expected-value functions","url":"https://doi.org/10.1007/s10107-016-1017-3"},{"authors":"F. Yu, R. Mehta, A. Luedtke, and Z. Harchaoui","title":"Stochastic gradients under nuisances","url":"https://arxiv.org/abs/2508.20326"},{"authors":"H. Robbins and D. Siegmund","title":"A convergence theorem for non negative almost supermartingales and some applications. In","url":"https://doi.org/10.1016/B978-0-12-604550-5.50015-8"}],"schema_version":1,"takeaways":"takeaways.md","title":"SGD with Vanishing Correlated Gradient Errors (technique notes)","type":"paper","verification":[{"date":"2026-10-17","model":"GPT-6 (Codex; exact runtime variant not exposed)","path":"verification/review.md","scope":"Main statements, convergence and PR proofs, variance bound and attainment; experiment implementation inspected; no full experiment rerun.","status":"materials-present-not-independently-verified","type":"review"}]},"parent":null,"pull_url":"https://mathhub.eggtea.org/api/works/sgd-with-vanishing-correlated-gradient-errors-dc3bf255/versions/1/bundle","resolves":[],"sha256":"9613a90e1448732498c5414f4d04361881eac769b32216b8a92a690ba021d247","submitted_by":"eggtea","url":"https://mathhub.eggtea.org/works/sgd-with-vanishing-correlated-gradient-errors-dc3bf255/v/1","version":1,"work_id":"sgd-with-vanishing-correlated-gradient-errors-dc3bf255"}