When a phone or camera runs the first few layers of a neural network and sends the intermediate features to a server that runs the rest, a practice called split computing, the designer faces a trade-off. Splitting early keeps the on-device model small, but the features are bulky. Splitting late gives compact features, but the device must run a large model. The benchmark that formalized this trade-off, SC2 Benchmark: Supervised Compression for Split Computing, treats the wireless link as an error-free pipe with a fixed bit budget. Real links fade: their capacity changes from moment to moment. Multi-antenna research has shown the opposite of what that simplification ignores. When the receiving base station has many antennas, the random fluctuations average out and the link gain becomes nearly constant. This effect is called channel hardening.
We propose to use channel hardening as the bridge between these two literatures. Our claim is that a hardened uplink turns the fading channel into a predictable per-frame bit budget. That budget is a deterministic function of the number of antennas and of how correlated they are, and it can be computed from slowly varying channel statistics. If this holds, the split point, the feature encoder, and the transmission rate can all be chosen offline, without tracking the channel instant by instant. A single measurable quantity, the leftover gain variance, would tell the designer how much safety margin to leave and when a noise-robust analog feature encoder is worth its complexity. The result would be a design rule stating how the best split point moves earlier as the array grows. We would also test whether that rule survives spatial correlation, imperfect channel estimates, and many users sharing the link.
Core Claim
Consider a single-antenna edge device sending to a base station with $N$ antennas. The station combines the antenna signals by weighting each with its channel estimate, which is called maximum-ratio combining. If the channel vector is Gaussian with spatial covariance $R$ (an $N \times N$ matrix describing how correlated the antennas are), a standard quadratic-form calculation gives the normalized gain variance
$$ \kappa(R) = \frac{\mathrm{Var}(\|\mathbf{h}\|^2)}{\left(\mathbb{E}\|\mathbf{h}\|^2\right)^2} = \frac{\mathrm{tr}(R^2)}{\mathrm{tr}(R)^2}. $$
This equals $1/N$ when the antennas are independent. It decays more slowly under line-of-sight or dense packing, which is the sub-linear regime analyzed in Channel Hardening of IRS-Aided Multi-Antenna Systems: How Should IRSs Scale?. A frame of $T$ channel uses then carries roughly $B \approx T \log_2(1 + \bar{\rho}\,\mathbb{E}\|\mathbf{h}\|^2)$ bits, where $\bar{\rho}$ is the per-antenna signal-to-noise ratio. The spread of this quantity is controlled by $\kappa$. Choosing a rate backed off by a few multiples of $\sqrt{\kappa}$ bounds the fraction of frames that fail.
The testable claim has three parts. First, a split-computing design chosen using only $\mathbb{E}\|\mathbf{h}\|^2$ and $\kappa$ matches a design that adapts to the instantaneous channel, once $\kappa$ falls below a threshold we will measure. Second, the best split depth under a fixed latency budget moves earlier as $N$ grows. Third, $\kappa$, rather than $N$ itself, predicts both the threshold and the accuracy loss under correlated channels.
Relation to Prior Work
Two bodies of work motivate this proposal. On the wireless side, Twenty-Five Years of Signal Processing Advances for Multiantenna Communications establishes that hardening lets systems rely on long-term statistics rather than instantaneous channel knowledge. Rethinking Federated Learning Over the Air: The Blessing of Scaling Up shows the analogous effect for model training: with 100 clients summing gradients over the air, "blind" transmission without power control reaches accuracy negligibly different from perfect-CSI channel inversion. Combating Interference for Over-the-Air Federated Learning: A Statistical Approach via RIS uses a 256-element reflecting surface to make aggregation errors vanish at rate $O(1/N)$ using only statistical CSI. Both papers concern aggregating gradients during training. We concern the inference-time transmission of features from a partitioned network, where each device's own features matter, not only their sum.
On the learning side, SC2 and its companion SC2: Supervised Compression for Split Computing supply our method and metric. We reuse their Entropic Student encoder, a small device-side network trained by distillation with a learned entropy model, and their encoder-size-times-rate versus distortion metric. We replace their idealized bit pipe with a physical uplink.
Two papers constrain the design. Deep learning for massive MIMO channel state acquisition and feedback introduces AnalogDeepCMC, which maps learned features directly to analog symbols and degrades gracefully instead of failing abruptly when a digital code's rate exceeds capacity. This is the strongest argument for analog transmission. Hardening weakens that argument, because a predictable channel removes most rate mismatch. We treat this as a hypothesis to test rather than assume. Over-the-Air Collaborative Inference with Feature Differential Privacy sends VGG11 features from 12 devices over an analog multi-access channel, but it compresses them with a linear matrix and never exploits hardening. Task-oriented CSI feedback work, namely Precoding-oriented Massive MIMO CSI Feedback Design: A Task-Oriented Approach and Deep Learning for Distributed Channel Feedback and Multiuser Precoding in FDD Massive MIMO, learns features of the channel itself. We learn features of images and treat the channel as the constraint.
Research Program
We will build a simulator coupling the sc2bench codebase (yoshitomo-matsubara/sc2-benchmark) to a multi-antenna uplink. The network is ResNet-50 on ImageNet, split at every stage the benchmark supports. Each image must arrive within a budget of $T$ channel uses. The uplink model sweeps $N$ from 8 to 256. It uses three channel families. The first is independent Rayleigh fading. The second is exponentially correlated arrays with a tunable correlation coefficient. The third is a 3GPP TR 38.901 urban macro-cell model, the same family used by TypeII-CsiNet: CSI Feedback with TypeII Codebook. Channel estimates come from finite pilots, and we add pilot contamination from $K$ co-scheduled users with $K$ up to 16, combined by maximum-ratio or zero-forcing receivers.
The analysis has three steps. First, we derive the $\kappa$-based rate backoff and the resulting frame-failure probability for each channel family, and check the prediction against simulated failures. Second, for each $(N, \kappa, T)$ we pick the split point and entropy-model rate that maximize accuracy under the predicted budget. This gives a curve of best split depth against $\kappa$. Third, we train a noise-robust analog variant of the Entropic Student. Its quantizer is replaced by power-normalized analog symbols, and it is trained with gain perturbations drawn at the measured $\kappa$. This lets us locate the region of $\kappa$ where analog transmission beats digital transmission and the region where it does not.
Baselines
The naive reference is SC2's own practice: the Entropic Student trained for an ideal bit pipe, deployed at a fixed rate chosen for average capacity with no margin, at its default split. It shows what ignoring the channel costs. It is run on the identical simulator and images.
The strongest competitor is an oracle-adaptive digital system. The same Entropic Student family is used, but the rate and split point are re-chosen every frame from the instantaneous post-combining SNR, as if perfect channel feedback were free. It is the strongest because any statistics-only design can at best match it. Our claim is precisely that hardening closes the gap. Both systems share checkpoints, the channel realizations, and the per-configuration tuning budget. Beating it is not required. Matching it within 0.5 top-1 points at $\kappa$ below the predicted threshold, with zero instantaneous feedback, counts as success. There is no reported result on Lacuna for Entropic Student top-1 accuracy under a fading uplink. The only reported figure the file gives is that its encoder is about 40 times smaller than a mean-scale hyperprior codec (SC2 Benchmark: Supervised Compression for Split Computing).
Input compression sends JPEG, WebP, or BPG images at the same statistical rate backoff, with full inference on the server. SC2 reports that it loses to supervised compression on an ideal pipe, and we test whether that ordering holds under fading.
A deep analog baseline, in the spirit of AnalogDeepCMC and trained on unhardened Rayleigh fading with no knowledge of $\kappa$, isolates whether calibrating training noise to $\kappa$ matters.
The key ablation replaces $\kappa$ with $1/N$ in the rate backoff. If designs chosen with $1/N$ fail under correlated or urban channels while $\kappa$-calibrated designs do not, the leftover variance is doing the work, not the antenna count.
Why not just…?
Why not just use instantaneous rate adaptation, which already works? It does work, and it is our upper baseline. Its cost is per-frame channel estimation and feedback plus rate re-selection on the device. Rethinking Federated Learning Over the Air: The Blessing of Scaling Up found such machinery unnecessary at scale for aggregation. If our oracle comparison shows a persistent gap even at small $\kappa$, the proposal's main claim fails, and we will report that.
Why not just send a well-coded digital bitstream, since hardening removes outages anyway? This is the most serious threat, and it may be right. If the channel is predictable, a separate source code plus channel code loses little, and the analog encoder's graceful degradation may be worth nothing. We therefore do not assume analog wins. The third step of the program measures the $\kappa$ region where it does. We expect that region to be large $\kappa$, correlated channels, and short frames, and possibly empty. A finding that digital suffices once $\kappa$ is small would still be useful, because it tells designers to skip analog co-design in hardened regimes.
Why not just train the feature encoder with generic fading noise and skip the theory? Generic noise training fixes one noise level. The deep analog baseline trained on Rayleigh tests exactly this. The question is whether a single robust model transfers across $N$ and correlation, or whether $\kappa$ must be known. If the generic model matches the $\kappa$-calibrated one everywhere, the calibration ingredient is unnecessary, and we will drop it.
Why not just compress the raw image with JPEG? SC2 already reports that supervised feature compression beats input codecs on ImageNet, COCO, and PASCAL VOC on an ideal link. We include input codecs to check that fading does not reverse this. We expect it will not, since both suffer the same rate backoff.
Decisive Experiments
The first check is predictive. Across all three channel families, the frame-failure rate predicted from $\kappa$ must agree with simulation within a factor of two at a target of 1%. If it does not, the budget is not predictable and the program stops there.
The second check is the central one. We plot top-1 accuracy of the statistics-only design against the oracle-adaptive design as a function of $\kappa$. The main claim holds if a gap under 0.5 points appears below a threshold that is consistent across independent, correlated, and urban channels when indexed by $\kappa$, but not when indexed by $N$.
The third check tests the split-point shift. With latency and device parameters fixed, the chosen split should move toward earlier stages as $\kappa$ shrinks. The reduction in on-device encoder parameters will be reported in SC2's metric, and we will confirm the trend with a second backbone from sc2bench.
The final check places the analog-versus-digital boundary on the same $\kappa$ axis, including a multi-user setting with pilot contamination, where estimation error, rather than fading, may dominate what is left of the uncertainty.
References
- SC2 Benchmark: Supervised Compression for Split Computing
- Channel Hardening of IRS-Aided Multi-Antenna Systems: How Should IRSs Scale?
- Twenty-Five Years of Signal Processing Advances for Multiantenna Communications: From theory to mainstream technology
- Rethinking Federated Learning Over the Air: The Blessing of Scaling Up
- Combating Interference for Over-the-Air Federated Learning: A Statistical Approach via RIS
- Deep learning for massive MIMO channel state acquisition and feedback
- Over-the-Air Collaborative Inference with Feature Differential Privacy
- Precoding-oriented Massive MIMO CSI Feedback Design
- Deep Learning for Distributed Channel Feedback and Multiuser Precoding in FDD Massive MIMO
- TypeII-CsiNet: CSI Feedback with TypeII Codebook