acceptodds
Under review as a conference paper at ICLR 2027

FedPHM: Communication-Efficient Personalization through Parameterized Hypercomplex Multiplication

Abstract

Personalized federated learning requires transferable shared information and client-specific adaptation under a limited communication budget. We propose FedPHM, a structured selective transmission framework based on Parameterized Hypercomplex Multiplication (PHM). Specifically, we decompose fully-connected layers into shared basis matrices and client-specific private interaction matrices via Kronecker factorization: shared bases and other shared network parameters are transmitted for aggregation, while private interactions persist locally. Per-step spectral projection bounds the interaction matrices and their influence on shared gradients. Theoretically, we establish a finite-time joint-stationarity bound for the PHM subsystem under partial client participation. With a horizon-dependent common step size, the residual is bounded by , where bounds the private-gradient noise variance. An exact quadratic-risk decomposition further separates the approximation cost of shared bases from the benefit of private interactions. Experiments on CIFAR-10, CIFAR-100, and FMNIST demonstrate that FedPHM achieves excellent personalized accuracy with less communication, and is robust under high heterogeneity and low participation scenarios.

Primary area:
optimization
Keywords:
Federated Learning, Personalized Federated Learning, Communication-Efficient Learning, Parameterized Hypercomplex Multiplication, Kronecker Factorization, Non-IID Data

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_yabAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FedPHM, a personalized federated learning method designed to reduce communication costs while adapting to differences in client data distributions. It uses the existing PHM parameterization of a fully connected weight matrix as a sum of Kronecker products, shares the basis matrices across clients, and keeps small interaction matrices private. Clients use an equivalent blockwise forward computation and jointly update the shared bases, private interactions, and other public network parameters using local SGD. After every local step, the private interaction matrices are projected onto a spectral-norm ball. Only the shared bases and other public parameters are uploaded and aggregated by client sample count; private interactions remain local across rounds. The paper analyzes which client-specific weight differences the parameterization can represent and discusses the potential benefits of private interactions under a simplified loss model. Its main experiments use a CNN on CIFAR-100, with additional Fashion-MNIST and CIFAR-10 results in the appendix, to evaluate accuracy, upload payload, parameter count, and the effects of individual design components.

Critical strengths
  1. Reduced communication through compact shared bases. FedPHM transmits compact PHM basis matrices together with the other public network parameters, reducing the payload associated with large fully connected weights. This directly addresses the upload bottleneck in federated training. In Table 1, FedPHM uploads 2.62 MiB per selected client per round, compared with 9.62 MiB for FedAvg, a reduction of approximately 72.8%. At Dirichlet alpha 0.1, the corresponding accuracies are 54.08% and 34.06%. The method therefore provides a useful way to reduce upload cost while retaining personalized predictive performance.

  2. Client personalization with few private parameters. FedPHM allows each client to combine the shared basis matrices using its own interaction matrices, producing weights adapted to its local data distribution. These private parameters are retained and updated across training rounds. With a PHM order of 8, each layer requires only 512 private interaction parameters, keeping local personalization storage costs low. In Table 2, under strong heterogeneity, this configuration achieves 54.03% accuracy, compared with 31.95% for the fully shared PHM-Full baseline, demonstrating the benefit of private interactions for personalized learning.

Critical weaknesses
  1. Limited methodological novelty. The parameter compression largely comes from the existing PHM parameterization, while separating globally shared and client-private parameters is an established personalization strategy. The principal change in Sections 3.1–3.4 is to share the PHM bases, keep private interaction matrices, and apply a standard spectral-norm projection. I view this as a specific combination and application of existing techniques, with a limited methodological increment. The practical results are useful, but the paper does not yet articulate a sufficiently strong new technical insight to support its methodological contribution in my assessment.
  2. The experimental evidence leaves important parts of the assessment unresolved. The main external baselines were mostly published in 2021 or earlier, while the original LoRA work cited for the manuscript's FedLoRA configuration appeared at ICLR 2022. Although Table 1 includes a valuable upload matched LoRA-private-base control, the absence of recent applicable personalized or communication-efficient FL comparators limits the assessment of current competitiveness for an ICLR 2027 submission. Separately, Table 3 varies the projection threshold but does not include a genuinely unprojected control. Consequently, the independent effect of spectral projection on accuracy or stability remains unclear.
Questions for authors
  1. Please clarify the main methodological advance of FedPHM over the original PHM parameterization and existing shared/private personalization frameworks. What new technical insight is provided by the specific choice to share the bases and keep the interactions private?

  2. Most external baselines evaluated in the paper date from 2021 or earlier. How was this baseline set selected to reflect the relevant literature available at submission time? Could the authors add a recent, applicable personalized or communication-efficient FL baseline under the same backbone and data splits, and compare accuracy at matched cumulative upload budgets?

  3. Could the authors include an ablation that removes spectral projection entirely while keeping the other hyperparameters and random seeds fixed? Please report accuracy across seeds, training-loss curves, and spectral-norm trajectories of the private interaction matrices. These results would help quantify the contribution of projection to accuracy and training stability.

  4. Appendix B.3 lists PHM-Full accuracies of 9.99%, 62.48%, and 23.88% at 10% participation. These values imply a sample standard deviation of approximately 27.20 percentage points, whereas Table 6 reports 32.12 ± 2.72. Could the authors verify the underlying per-seed results and reconcile the inconsistent statistics?

1 comment

1. Methodological advance and the role of private interactions

FedPHM assigns distinct learning roles to its two PHM factors: Wk=∑i=1nAk,i⊗SiW_k = \sum_{i=1}^{n} A_{k,i} \otimes S_i.

Shared bases SiS_i supply reusable transformations; private interaction matrices Ak,iA_{k,i} govern how each client combines them across blocks of its effective weight.

The (a,b)(a,b) block is [W_k]ab = Σ{i=1}^n [A_{k,i}]_ab S_i.

Clients share transformations while keeping block-mixing coefficients local. For PHM order nn, each layer needs only n3n^3 private parameters, independent of the shared-basis dimensions. These compact private states can produce full-rank weight differences across clients.

If only one interaction term differs between clients kk and ℓ\ell,

then Wk−Wℓ=(Ak,1−Aℓ,1)⊗S1W_k-W_\ell = (A_{k,1}-A_{\ell,1}) \otimes S_1, and rank⁡(Wk−Wℓ)=rank⁡(Ak,1−Aℓ,1)rank⁡(S1)\operatorname{rank}(W_k-W_\ell) = \operatorname{rank}(A_{k,1}-A_{\ell,1})\operatorname{rank}(S_1).

Full-rank factors can therefore yield full-rank weight differences, enabling layer-wide personalization through small local states.

Our theory uses a quadratic-risk decomposition to separate approximation costs of the shared PHM structure from heterogeneity captured by private interactions. Shared-gradient disagreement is decomposed into data-driven effects and private-state dispersion; spectral projection constrains gradient amplification from private factors.

The PHM-Full ablation, with fixed architecture and parameter count, validates this sharing strategy:

MethodCIFAR-100, α=0.1\alpha=0.1CIFAR-100, α=0.5\alpha=0.5
PHM-Full (n=8n=8)31.47±0.4431.47 \pm 0.4435.81±0.5135.81 \pm 0.51
FedPHM (n=8n=8)54.07±0.8754.07 \pm 0.8741.41±0.4241.41 \pm 0.42

Results are client-matched accuracy (%), reported as mean ±\pm standard deviation over seeds {42,123,456,2026,2027}\{42,123,456,2026,2027\}. Private interactions improve accuracy by 22.60 percentage points at α=0.1\alpha=0.1 and 5.60 points at α=0.5\alpha=0.5.

2. Baseline selection and an additional recent comparator

Our original baselines cover fully shared optimization, personalized representations, regularized local training, and parameter-efficient adaptation. They probe FedPHM's key dimensions: global sharing, client-specific parameters, and communication reduction through structured factorization.

We evaluate DA-PFL (2024) under the same CIFAR-100 partitions, evaluation metric, five seeds, 200 rounds, and 20 local steps per selected client:

MethodCIFAR-100, α=0.1\alpha=0.1CIFAR-100, α=0.5\alpha=0.5Uplink (MiB/client/round)
DA-PFL (2024)51.73±0.8451.73 \pm 0.8431.34±0.2131.34 \pm 0.219.616
FedPHM54.07±0.8754.07 \pm 0.8741.41±0.4241.41 \pm 0.422.616

At 200 rounds, FedPHM improves accuracy by 2.34 percentage points at α=0.1\alpha=0.1 and 10.07 points at α=0.5\alpha=0.5, while reducing per-round uplink by approximately 72.8%.

3. Ablation without spectral projection

We conduct an ablation that disables spectral projection while keeping the seeds, data split, network, and hyperparameters unchanged. Models are trained on CIFAR-100 with α=0.1\alpha=0.1 for 200 rounds.

SettingClient-matched accuracy (%)Final maximum private spectral norm across seeds
Without projection54.05±0.9054.05 \pm 0.900.6984--0.7697
Projection, τ=0.4\tau=0.454.07±0.8754.07 \pm 0.87approximately 0.400001

Accuracy is reported as mean ±\pm standard deviation over five seeds. Projection enforces the theoretical norm constraint with negligible accuracy cost. Unprojected models reach larger spectral norms but achieve similar accuracy; projection primarily provides analysis-driven norm control rather than an accuracy gain.

4. Audit and five-seed extension of Table 6

We revisited the Table 6 experiments and added seeds 2026 and 2027, following the original protocol. This gives five seeds for each method and participation rate. The updated results use CIFAR-10, 10 clients, α=0.1\alpha=0.1, and five local epochs per selected client.

Method10%30%50%100%
FedAvg23.86±5.6523.86 \pm 5.6565.80±4.6865.80 \pm 4.6872.22±3.2172.22 \pm 3.2174.38±2.2174.38 \pm 2.21
PHM-Full (n=4n=4)28.83±22.0628.83 \pm 22.0665.95±2.3365.95 \pm 2.3370.50±3.1770.50 \pm 3.1772.77±1.8072.77 \pm 1.80
FedPHM (n=4n=4)90.04±1.7790.04 \pm 1.7791.30±1.7791.30 \pm 1.7791.41±1.3091.41 \pm 1.3091.63±1.1791.63 \pm 1.17

Entries are client-matched accuracy (%), reported as mean ±\pm sample standard deviation over seeds {42,123,456,2026,2027}\{42,123,456,2026,2027\}. Across all participation rates, the five-seed results show higher accuracy for FedPHM than for FedAvg and PHM-Full.

Reviewer_CSAtOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes FedPHM to reduce communication in personalized federated learning. Using the existing PHM formulation, it represents a fully connected layer through shared matrices and small private matrices that control how each client combines them. Clients exchange only the shared parameters, keeping their private matrices locally for personalization. The private matrices are constrained during training to limit their effect on shared updates.

The theoretical analysis studies convergence when only some clients train in each round and uses a simplified model to explain when personalization can outweigh the loss of flexibility from compression. Experiments on CIFAR-100, CIFAR-10, and Fashion-MNIST compare accuracy and communication costs, with the largest reported benefits occurring when clients have substantially different class distributions.

Critical strengths

The main strength is that FedPHM provides personalization through a small set of private parameters while sharing a compressed representation across clients. This is useful when clients have limited communication and storage resources. The comparisons with the same model trained entirely through shared parameters provide evidence that retaining these small private components improves accuracy when clients have different class distributions.

Critical weaknesses
  1. The novelty is limited. The work combines the existing PHM formulation with shared–private parameter decomposition in personalized FL. Closely related methods, particularly pFedPara (ICLR 2022), already separate shared and private factors within layers but are neither discussed nor compared. Although the paper provides theoretical and numerical support, it does not convincingly establish a substantial advance over these existing ideas.

  2. The experiments support the value of private interaction matrices, but the source of the reported accuracy–communication gains could be clearer. The communication reduction in Eqs. (7)–(8) mainly follows from PHM compression, whereas most personalized baselines in Table 1 use uncompressed layers. The PHM-Full comparison in Table 2 shows that keeping interaction matrices private improves personalization within the same compressed architecture. However, the advantage over the communication-matched LoRA control in Table 1 varies with the level of heterogeneity. Comparisons with existing personalization methods using similarly compressed layers would help establish how much of the claimed benefit comes specifically from FedPHM’s sharing strategy.

Questions for authors
  1. What is the main technical distinction between FedPHM and existing factorized personalized FL methods such as pFedPara [1] and Factorized-FL [2], beyond the choice of PHM? Which result best demonstrates the benefit of this distinction?

  2. The PHM-Full comparisons support keeping interaction matrices private, while Eqs. (7)–(8) attribute most communication savings to PHM compression. What existing evidence distinguishes the benefit of FedPHM’s sharing strategy from combining similar compression with an established personalization method?

  3. The reported comparisons do not include local-only training. Could you include a baseline where each client trains independently on its own data, using matched architectures, total local training steps, and the same evaluation metric? Comparing this baseline with FedAvg and FedPHM would clarify whether sharing information across clients is beneficial and whether FedPHM improves over both independent local training and a shared global model.

  4. Could you provide the missing steps connecting the round-descent inequality in Appendix A.3 to Eq. (38), particularly the coefficient 12 on the local-update error term? Please also clarify what the convergence analysis adds beyond the existing FedSim analysis [3].

References:

[1] Hyeon-Woo et al. “FedPara: Low-Rank Hadamard Product for Communication-Efficient Federated Learning.” ICLR 2022. Includes pFedPara.

[2] Jeong and Hwang. “Factorized-FL: Personalized Federated Learning with Parameter Factorization & Similarity Matching.” NeurIPS 2022.

[3] Pillutla et al. “Federated Learning with Partial Model Personalization.” ICML 2022. Includes FedSim.

1 comment

We thank the reviewer for the constructive comments and for the opportunity to revise our manuscript.

1. Technical differences from pFedPara and Factorized-FL

FedPHM writes client kk's weight matrix as

Wk=∑i=1nAk,i⊗Si,W_k = \sum_{i=1}^{n} A_{k,i} \otimes S_i,

where SiS_i are shared bases aggregated on the server and Ak,iA_{k,i} are client-local interaction matrices with ∥Ak,i∥2≤τ\|A_{k,i}\|_2 \leq \tau.

pFedPara adopts a Hadamard low-rank factorization:

Wk=Wglobal⊙(1+Wlocal,k),W_k = W_{\mathrm{global}} \odot (\mathbf{1} + W_{\mathrm{local},k}),

where WglobalW_{\mathrm{global}} is shared and Wlocal,kW_{\mathrm{local},k} stays local. Its private term acts as an elementwise multiplicative gate, unlike FedPHM’s coefficients mixing Kronecker products of shared bases.

Factorized-FL uses Wk=UkVk+MkW_k = U_k V_k + M_k, personalizing via similarity-weighted aggregation over peer client models. FedPHM keeps a fixed global basis and persistent local interaction matrices, which are never uploaded or reweighted by other clients.

Our baseline PHM-Full fixes architecture and total parameters, only varying whether interaction matrices are private. The pFedPara comparison is communication-matched with different parameter counts.

MethodCIFAR-100, α=0.1\alpha=0.1CIFAR-100, α=0.5\alpha=0.5Uplink (MiB/client/round)Parameters
PHM-Full (n=8n=8)31.47±0.4431.47 \pm 0.4435.81±0.5135.81 \pm 0.512.618686,180
FedPHM (n=8n=8)54.07±0.8754.07 \pm 0.8741.41±0.4241.41 \pm 0.422.616686,180
pFedPara55.64±0.6455.64 \pm 0.6437.61±0.4737.61 \pm 0.472.600939,620

Results are mean ±\pm std over seeds {42,123,456,2026,2027}\{42,123,456,2026,2027\}. FedPHM surpasses PHM-Full by 22.60 pp (α=0.1\alpha=0.1) and 5.60 pp (α=0.5\alpha=0.5). It is 1.57 pp lower than pFedPara at α=0.1\alpha=0.1 and 3.80 pp higher at α=0.5\alpha=0.5.

2. Separating PHM compression from FedPHM's sharing strategy

We disentangle communication reduction and personalization gains. PHM-Full fixes architecture and total parameters, only toggling the locality of interaction matrices; the accuracy gap isolates the effect of this sharing design. pFedPara serves as a communication-matched reference for another factorized personalized FL method.

MethodCIFAR-100, α=0.1\alpha=0.1CIFAR-100, α=0.5\alpha=0.5Uplink (MiB/client/round)
Local-only51.89±0.9651.89 \pm 0.9630.68±0.1930.68 \pm 0.190
FedAvg33.99±0.3433.99 \pm 0.3439.58±0.6239.58 \pm 0.629.616
FedPHM54.07±0.8754.07 \pm 0.8741.41±0.4241.41 \pm 0.422.616

Results are mean ±\pm std over seeds {42,123,456,2026,2027}\{42,123,456,2026,2027\}. FedPHM outperforms local-only training by 2.18 pp (α=0.1\alpha=0.1) and 10.73 pp (α=0.5\alpha=0.5). It exceeds FedAvg by 20.08 pp and 1.83 pp, cutting uplink from 9.616 MiB to 2.616 MiB per client per round.

3. Derivation of the local-update error coefficient 12 in Eq. (38)

Let q=M/Kq=M/K and x=LSEη∈[0,1]x=L_S E\eta \in [0,1]. The trajectory-error bound for the randomized residual is

= \frac{4}{q} + 2 \leq \frac{6}{q},$$ using $q \leq 1$. Converting to the exact projected-gradient residual: $$\mathcal{R}_t^{(E)} \leq 2\widetilde{\mathcal{R}}_t^{(E)} + 2\nu_A.$$ This doubles the bound: $$2\left(\frac{4}{q} + 2\right) = \frac{8}{q} + 4 \leq \frac{12}{q},$$ which gives coefficient 12 in Eq. (38): $$\frac{12L_{\mathrm{path}}^2 c_E \eta^2 G_{\mathrm{dir}}^2}{q}.$$ The same conversion contributes $2 \cdot 4 + 2 = 10$ to the private-gradient noise term. ### 4. What the convergence analysis adds beyond FedSim FedSim provides a framework for joint shared/personalized parameter optimization under partial client participation. We adapt it to our spectrally constrained PHM model with two extensions. First, we analyze constrained private updates via projected gradient mapping. On the spectral-norm ball boundary, nonvanishing unconstrained gradients can coexist with constrained stationarity. The projected residual encodes the first-order condition of FedPHM’s private update and accounts for stochastic gradient noise. Second, we decompose shared-gradient disagreement into two PHM-specific sources: data heterogeneity at a fixed private state and dispersion across clients’ local interaction matrices. PHM structure and spectral constraints govern how weight-gradient heterogeneity propagates to shared bases, linking the bound to PHM order, projection radius and private-state dispersion. We combine these into a finite-time bound supporting multiple local steps and partial sampling, separating sampling noise, local drift and private-gradient noise. Compared with FedSim, our analysis is constraint-aware and structure-dependent, connecting FedPHM’s design choices to its convergence guarantees.
Reviewer_6Yh8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FedPHM, a personalized federated-learning method that replaces a dense fully connected layer with Parameterized Hypercomplex Multiplication. Each effective client weight is expressed as a sum of Kronecker products between globally shared basis matrices and persistent client-private interaction matrices. This paper derives communication counts, shows that private PHM interactions can induce full-rank differences between client weights, gives a quadratic risk decomposition for shared versus private interactions, and provides a finite-time projected-stationarity analysis under partial participation. Experiments on CIFAR-100, CIFAR-10, and FMNIST compare compact and capacity-matched variants with fully shared and personalized baselines under Dirichlet label skew.

Critical strengths

Broad comparison within the chosen setting. FedAvg, FedProx, SCAFFOLD, FedRep, Ditto, pFedMe, FedLoRA, a budget-matched LoRA-private-base control, and capacity-matched FedPHM variants are included.

Statistical reporting is generally solid. Core experiments use five seeds and report means and standard deviations, with protocol and payload details supplied in Appendix B.

Critical weaknesses

FedPHM is not uniformly superior at matched budgets or capacity. In Table 1, the LoRA-private-base control is better at \alpha=0.1 (55.77% versus 54.08%), while FedPHM is better at \alpha=0.5. At capacity-matched full upload (Appendix Table 5), FedPHM remains slightly below FedLoRA on both heterogeneity settings. The result supports a trade-off, not general dominance; the abstract and discussion should state where each method is preferable.

Several experimental claims rely on legacy three-seed settings with high instability. For example, Appendix Table 6. This makes the dramatic partial-participation conclusion difficult to interpret without a five-seed replication.

Questions for authors

Can the author make it consistent for the seeds used across experimental on legacy to avoid the high instability?

2 comments

We thank the reviewer for raising the concern about seed consistency and instability in the legacy partial-participation experiments. We extended the experiment by adding two seeds, 2026 and 2027, to the original seeds 42, 123, and 456. All methods use the same CIFAR-10 protocol at each of the four participation rates. We recomputed the mean and sample standard deviation over the five runs and updated Table 6 accordingly. The resulting client-matched accuracies (%) are:

Participation rateFedAvgPHM-FullFedPHM
10%23.86 ± 5.6528.83 ± 22.1090.04 ± 1.77
30%65.80 ± 4.6865.95 ± 2.3391.30 ± 1.76
50%72.22 ± 3.2270.50 ± 3.1891.41 ± 1.30
100%74.37 ± 2.2172.77 ± 1.7991.63 ± 1.17

thanks for adding two more seeds. Adding more seeds turn to impact FedAvg and PHM-Full more than FedPHM, which is good. May I know which part in the paper has been updated? Table 6 still uses 3 seeds if I understand correctly.

Comments on the paper

1 comment

Dear reviewers,

The author has provided the corresponding response.

Please review the author's response and respond accordingly.

Thank you!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.