Home / Curated / Algorithms / pFedMe Personalized Federated Learning with Moreau Envelopes

pFedMe Personalized Federated Learning with Moreau Envelopes

Algorithms AIMech:PFedMePersonalizedFederatedLearningWithMoreauEnvelopes DRAFT

Aliases: pFedMe, Personalized Federated Learning with Moreau Envelopes, Moreau Envelope Personalized Federated Learning

A bi-level personalized federated learning algorithm (pFedMe) that formulates global coordination as minimizing the average Moreau envelope F(w) = (1/N) sum_{i=1}^N F_i(w) where each client's outer objective is F_i(w) = min_{theta_i in R^d} { f_i(theta_i) + (lambda / 2) ||theta_i - w||^2 } and its optimal personalized model is the proximal operator theta_hat_i(w) = prox_{f_i / lambda}(w). Using the exact envelope gradient identity nabla F_i(w) = lambda * (w - theta_hat_i(w)), each client i alternates across R local steps between computing a delta-approximate personalized solution theta_tilde_i(w_{i,r}^t) via K inner first-order steps and updating its local reference model w_{i,r+1}^t = w_{i,r}^t - eta * lambda * (w_{i,r}^t - theta_tilde_i(w_{i,r}^t)), followed by server interpolation w_{t+1} = (1 - beta) w_t + (beta / S) sum_{i in S^t} w_{i,R}^t.

Identity Notes & Concept Scope Distinct from FedProx (AIMech:FedProxProximalFederatedOptimization, which solves the same proximal subproblem to update a single shared global model w without decoupling a persistent personalized model theta_i from the global reference parameter w), from Per-FedAvg (AIMech:PerFedAvgModelAgnosticMetaLearningForPersonalizedFederatedLearning, which minimizes one-step MAML adaptation loss f_i(w - alpha nabla f_i(w)) requiring Hessian-vector products rather than first-order Moreau envelopes), and from Ditto (AIMech:DittoFairAndRobustPersonalizedFederatedMultiTaskLearning, where global optimization of w^* is completely independent of the personalized models v_k, whereas in pFedMe the outer global update depends directly on lambda * (w_{i,r}^t - theta_tilde_i)).

Claims & Grounded Evidence

4 Claims

Mechanism #r395-moreau-envelope-smoothness-and-strong-convexity-transfer

Proposition 1 proves that if f_i is L-smooth (and optionally mu-strongly convex, or nonconvex with lambda > 2L), then the Moreau envelope F_i(w) is L_F-smooth with L_F = lambda (and L_F = lambda * L / (lambda - L) < lambda for lambda > 2L) and mu_F-strongly convex with mu_F = lambda * mu / (lambda + mu), with exact gradient nabla F_i(w) = lambda * (w - theta_hat_i(w)), while Lemma 1 shows that K = O(kappa_h log(d / delta)) inner gradient steps suffice to compute a delta-accurate proximal iterate without second-order derivatives.

supports Source: Personalized Federated Learning with Moreau Envelopes · Locator: Section 3.1 Problem Formulation & Section 3.2 Algorithm (Proposition 1, Lemma 1, Eqs. 3-5), pp. 3-5

Mechanism #r395-strongly-convex-and-nonconvex-convergence-rates

Using the Moreau-envelope bounded diversity bound (Lemma 2), Theorem 1 and Corollary 1 establish O(1 / (T R N)^2) quadratic convergence speedup in total rounds T, local steps R, and clients N for mu-strongly convex objectives, while Theorem 2 and Corollary 2 establish O(1 / (T R S)^{2/3}) stationarity convergence for L-smooth nonconvex objectives when lambda >= sqrt(8 L^2 + 1).

supports Source: Personalized Federated Learning with Moreau Envelopes · Locator: Section 4 Convergence Analysis (Lemma 2, Theorems 1-2, Corollaries 1-2, Table 1), pp. 5-7

Limitation #r395-inner-loop-computational-overhead-and-lambda-threshold-limitation

Each outer local step r = 1,...,R requires solving an inner K-step proximal minimization subproblem (multiplying local gradient evaluations by K), and theoretical nonconvex smoothness of the Moreau envelope requires choosing the regularization parameter lambda > 2L (or lambda >= sqrt(8 L^2 + 1) in Theorem 2).

supports Source: Personalized Federated Learning with Moreau Envelopes · Locator: Section 3.1 Proposition 1, Section 3.2 Lemma 1, and Section 4.2 Theorem 2, pp. 3-7

Application #r395-heterogeneous-mnist-and-synthetic-personalization-application

Applied to non-i.i.d. MNIST and synthetic federated tasks with both strongly convex L2-regularized multinomial logistic regression and nonconvex 2-layer DNN models, outperforming FedAvg and Per-FedAvg in personalized test accuracy.

supports Source: Personalized Federated Learning with Moreau Envelopes · Locator: Section 5 Experimental Results and Discussion (Figures 1-2, Table 2), pp. 7-9

⚠️ Curated Limitations & Negative Results

Multiplies local computational cost by K inner gradient steps per outer local step r in 1..R.
Nonconvex Moreau envelope smoothness guarantees require choosing the proximal penalty parameter lambda > 2L.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:FederatedLearning` Federated Learning	narrow_match	pFedMe is a Moreau-envelope personalized federated optimization algorithm within FederatedLearning.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r395-primary-pdf`	Personalized Federated Learning with Moreau Envelopes Canh T. Dinh, Nguyen H. Tran, Tuan Dung Nguyen	primary paper full_text	ID: https://doi.org/10.48550/arXiv.2006.08848 sha256:07a9e178a874…
`r395-arxiv-abs`	Personalized Federated Learning with Moreau Envelopes (arXiv abstract) Canh T. Dinh, Nguyen H. Tran, Tuan Dung Nguyen	primary paper abstract_only	ID: https://arxiv.org/abs/2006.08848 sha256:55c1eddcd1ba…
`r395-metadata-cr`	Personalized Federated Learning with Moreau Envelopes (Crossref metadata) Canh T. Dinh, Nguyen H. Tran, Tuan Dung Nguyen	primary paper abstract_only	ID: https://doi.org/10.48550/arXiv.2006.08848 sha256:c7dba23f6451…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T07:58:00Z

Updated: 2026-10-19T07:58:00Z

Source File: data/curated/PFedMePersonalizedFederatedLearningWithMoreauEnvelopes.yaml · Schema: curated_concept.yaml

Explore across X-Mech

Related computational & biological Mechs: Embedding Map Network Graph ProteinTraitsMech (ESM-2 pLM) DUFMech (Domain pLM) TraitMech CultureMech All 14 X-Mech Repositories