Review 1

Relevance and Technical Content and Quality of

Novelty and Originality
Timeliness Scientific Rigour Presentation

Marginal work and Some interesting ideas and

Substantial revision
Good. (4) simple contribution. results on a subject well
work is needed. (2)
Some flaws. (2) investigated. (3)

Strong Aspects (Comments to the author: What are the strong aspects of the paper?)
Exploring the impact of lossy compression in the trade-off between communication efficiency and
convergence speed for compression-aided Federated Learning schemes is a relevant and timely
problem.

Weak Aspects (Comments to the author: What are the weak aspects of the paper?)
-Some aspects of the proposed approach call for stronger justification or additional explanations.
-The evaluation results are less than comprehensive. -The presentation of the material needs
improvement. More details on these issues are provided in the next section of the review.

Recommended Changes (Recommended changes. Please indicate any changes that should be
made to the paper if accepted.)
Some parts of the proposed approach are obscure and require further justification and/or
elaboration. In particular: -The paper employs Lemma 1, through eq. (8), to determine the number
K^{SE} of participating devices that minimizes the number of communication rounds required for
convergence and states that this problem has complexity linear in the total number of devices K.
However, (in view of the factor \chi and the last sum in term B, within (6)) this is valid only for
symmetrical environments where all devices employ the same loss function and exhibit the same
distance between local and global loss. -The paper assumes that the communication between
participating devices and the BS employ a number S of subchannels and that each subchannel
may be used by only one of the participating devices. This arrangement imposes, as a
consequence of eqs. (11d-f), the strong limitation that the number of participating devices K^{SE}
cannot exceed the number of subchannels S. Why is such a restriction necessary? Can't the
participating devices share the available spectral resources by taking turns in the use of
subchannels when appropriate? Also, assuming that the policy of a single device per subchannel
is in force, why is the transmission time for a communication round expressed as a sum (in the rhs
of (10)) rather than a max operation? -The appropriateness of employing the coalition game
requires some further justification. Given that the participating devices have already been
determined, can't the search for the solution proceed in a more oriented way (e.g., by prioritizing
on the basis of channel gains or transmission power)? Also, the remark on the computational
complexity of the game is incomplete, as no bound is provided for the number of required
iterations. The discussion on the evaluation results omits several important system-related
characteristics. For example, no details are given about the Federated learning environment (e.g.,
characteristics of the local loss functions). The impact of compression is explored in connection
with the convergence rate, but not the communication efficiency (which provided the primary
motivation for introducing compression). Also, there is some inconsistency across results: While
the results about the convergence rate dictate that the optimal number K^{SE} of participating
devices should be about 40, this value is not feasible in the setup employed for studying the
 transmission time optimization (the number of available subchannels is kept below 30). The
presentation can be improved in several respects: -The manuscript needs a check for consistency
and the correct use of notation. For example, equations like (1), (2) and (10) should have their
sums span the range from 1 to K, not K^{SE}. -The meaning of symbols should be introduced
close to their first occurrence. (The meaning of quantity E is defined at the end of Section II-B,
almost half a page after its first encounter.) -While it is legitimate to omit the proof of Lemma 1
due to space restrictions, some insight should be provided about the nature of the result and how
it relates to earlier related results (e.g., those in [17] for an uncompressed setting). -The
manuscript should be checked for the correctness of factual claims therein. For example, it is
stated (more than once) that "each IoT device can only occupy one sub-channel", which is
different than the intended meaning (that each subchannel can be used by only one IoT device).
Review 2

Relevance and Technical Content and

Novelty and Originality Quality of Presentation
Timeliness Scientific Rigour

Some interesting ideas and Readable, but revision is

Valid work but limited
Good. (4) results on a subject well needed in some
contribution. (3)
investigated. (3) parts. (3)

Strong Aspects (Comments to the author: What are the strong aspects of the paper?)
This paper studied a resource allocation scheme for compression-aided federated learning. Both
energy efficiency and federated learning performance are jointly optimized.

Weak Aspects (Comments to the author: What are the weak aspects of the paper?)
The formulated original problem is decomposed into two sub problems, and the proposed
algorithm is designed. It needs to clarify how the proposed algorithm can find the most suitable
solution of the original problem. The simulation results can also show the optimality gap when
using the proposed algorithm.

Recommended Changes (Recommended changes. Please indicate any changes that should be
made to the paper if accepted.)
-It is mentioned that the proposed wireless network meets the 3GPP standard. However this
paper does not explain how/why this paper's network meets which clause(s) of TS38.300. -This
paper is mainly considering IoT networks. Please explain why this paper assumes to run the FL on
the IoT devices. IoT devices defined in the current wireless standards may not have enough
computing capability. -It is assumed that all data are IID. However the environment nearby could
be have some similarity. Moreover IoT devices are redundantly deployed and therefore their
sensory data can be correlated. How does the IID assumption always hold in the proposed system
model? -Please explain how the lossy model is implemented in the simulations. When the
proposed scheme is used, the simulation results can additionally show the system performance
with respect to the different degrees of error (between the original and decompressed models). -
It would be required to clarify why a certain number of devices are required/helpful to ensure the
FL convergence.
Review 3
 Relevance and Technical Content and Quality of
Novelty and Originality
Timeliness Scientific Rigour Presentation

Marginal work and Some interesting ideas and Readable, but revision
Good. (4) simple contribution. results on a subject well is needed in some
Some flaws. (2) investigated. (3) parts. (3)

Strong Aspects (Comments to the author: What are the strong aspects of the paper?)
- The paper provides the theoretical analysis of the proposed algorithm. - The topic of the paper
is timely.

Weak Aspects (Comments to the author: What are the weak aspects of the paper?)
- Not rigorous communication model and some strong assumptions - Questionable problem
formulation - It is not clearly shown that the proposed method thoroughly considers both
communication model and model compression in FL - Lack of performance evaluation and
comparisons

Recommended Changes (Recommended changes. Please indicate any changes that should be
made to the paper if accepted.)
1. The assumption of iid data is too strong. 2. The communication model is not rigorous. The
authors assumed that the proposed wireless network meets the 3GPP standard. However, it is not
clear that the communication model fits into the 3GPP standard because of the gap between
communication rounds in FL and timeslots in the wireless network. 3. There is no consideration of
local training time at devices and lagged data samples at unselected devices. 4. Why is minimizing
the transmission time so important? 5. It is questionable that the proposed algorithm is closely
related to the model compression since it does not jointly consider K^{SE} and r_t, \Theta_t. 6. The
performance evaluation is lacking. There is no comparison against state-of-the-art methods.