A converter that turns the analog antenna signal into bits (an analog-to-digital converter, or ADC) uses power roughly proportional to $2^B$, where $B$ is its bit depth. Receivers that can switch resolution on the fly already exist in silicon. The catch is that the algorithms deciding how many bits each antenna should use all run at one central node that holds the full, instantaneous channel for every antenna. That arrangement makes sense in a single base station. It does not scale to cell-free massive MIMO, where dozens of small access points (APs) spread over a neighborhood all feed a central processing unit (CPU) over a limited "fronthaul" link. There, shipping instantaneous channels to the CPU is precisely the traffic that decentralized designs exist to remove.
This note proposes splitting the decision across two timescales. The CPU looks only at slowly changing, large-scale channel statistics and occasionally broadcasts one number per AP: an exchange rate between converter power and useful signal. Each AP then picks its own ADC resolution and fronthaul quantization depth in every coherence block, using only its own instantaneous channel estimate and that number. The claim is that this recovers most of the energy-efficiency gain of centralized bit allocation. The comparison counts the control bits that centralized allocation itself would require, which prior work never charges against it.
Core Claim
Resolution-adaptive ADC work divides into two camps that have not met. On the optimization side, joint combiner and bit allocation for a collocated mmWave array is solved with fractional programming and penalty dual decomposition in Energy Efficiency Optimization for Millimeter Wave System With Resolution-Adaptive ADCs, which finds that energy efficiency saturates around 3 average bits. Bits, reflecting-surface phases and decoders are optimized jointly by block coordinate descent in Uplink Achievable Rate Maximization for Reconfigurable Intelligent Surface Aided Millimeter Wave Systems with Resolution-Adaptive ADCs. Both papers assume one node with perfect global channel knowledge. They motivate the proposal by showing the gain exists, and they constrain it by showing that bits and combiner weights are tightly coupled. On the hardware side, A Resolution-Adaptive 8 mm2 9.98 Gb/s 39.7 pJ/b 32-Antenna All-Digital Spatial Equalizer for mmWave Massive MU-MIMO in 65nm CMOS switches 64 converters between 3 and 6 bits and reports roughly 3x better energy per bit in the low mode. We reuse its 3/6-bit mode set and its power scaling as our hardware model.
Cell-free work, meanwhile, shows how to avoid centralizing fast channel information. In large-scale fading decoding (LSFD), each AP combines its own signals locally and the CPU weights the resulting per-user scalar estimates using only long-term statistics. Uplink Precoding Design for Cell-Free Massive MIMO with Iteratively Weighted MMSE and Over-the-Air Computation in Cell-Free Massive MIMO Systems show that LSFD retains much of the centralized performance. Both assume ideal, infinite-precision converters.
The closest bridge is Max-Min Rate of Cell-Free Massive MIMO Uplink With Optimal Uniform Quantization. It quantizes the fronthaul scalars at a fixed, uniform resolution, derives the optimal step size through the Bussgang decomposition, and solves filters and power centrally. It does not adapt the physical ADC, and it does not adapt resolution over time.
The new element is structural. LSFD already tells us which information the CPU legitimately has, namely large-scale statistics. We show that this information is enough to set a per-AP price that makes purely local, per-block resolution choices nearly as good as centralized ones.
Concretely, the Bussgang model writes the output of a $B$-bit quantizer as $(1-\rho_B)y + d$. Here $y$ is the input, $\rho_B$ is a known distortion factor that shrinks roughly as $2^{-2B}$, and $d$ is uncorrelated distortion noise. In block $t$, AP $m$ chooses ADC bits $b_m$ from the hardware mode set and fronthaul bits $q_m$ per forwarded scalar by solving
$$ \max_{b_m, q_m}\; \sum_{k} a_{mk}\, \mathrm{SINR}_{mk}\!\left(\hat{\mathbf h}_{mk}(t), \rho_{b_m}, \rho_{q_m}\right) \;-\; \lambda_m \left(c\, 2^{b_m} N_m + \kappa\, q_m K_m\right). $$
In this objective, $\mathrm{SINR}_{mk}$ is the local signal-to-interference-plus-noise ratio of user $k$'s scalar estimate at AP $m$, computed from the AP's own instantaneous channel estimate $\hat{\mathbf h}_{mk}(t)$. The weight $a_{mk}$ is the LSFD weight the CPU already computes from statistics. The term $c\,2^{b_m}N_m$ is the ADC power for $N_m$ antennas. The term $\kappa q_m K_m$ is the fronthaul cost of forwarding $K_m$ user scalars. The CPU updates the price $\lambda_m$ on the slow timescale, using dual ascent on a network energy-efficiency target that is evaluated on statistics alone. Each AP is therefore charged, in its local currency, for how much its contribution matters to each user after the CPU combines all APs.
Two costs that prior formulations set to zero are kept explicitly. The first is converter reconfiguration, handled with a switching penalty and hysteresis. The second is the fact that the ADC resolution and the fronthaul resolution interact. Digitizing at 6 bits is wasted if the forwarded scalar is then quantized to 2 bits.
Research Program
The first step is analysis. We will derive, in closed form under the Bussgang model, the uplink SINR after LSFD with cascaded ADC and fronthaul distortion. From that expression we will show when the centralized energy-efficiency problem decouples exactly into the per-AP problems above given the right prices, and when it decouples only approximately. We expect exact decoupling when inter-AP interference enters the LSFD output only through statistics. The expected gap then comes from how instantaneous local channels correlate with other APs' distortion, and we will bound that gap.
The second step is a feedback-loop study that neither prior RADC paper addresses. Channel estimates are themselves acquired through quantized front ends, so a rule that lowers resolution when the channel looks weak can corrupt the very estimate that told it to do so. We will compare two regimes. In one, pilots are always received at the highest mode (6 bits), which the Castañeda chip's per-symbol switching makes feasible. In the other, pilots follow the data resolution. We will measure whether the second regime collapses under channel aging.
The third step is signaling accounting. For every method, we count control bits per coherence block, including bit-allocation commands, channel-state uploads and price broadcasts, and subtract them from fronthaul capacity before computing rate.
Baselines
The naive reference is fixed-resolution ADCs at every AP combined with LSFD and uniform fronthaul quantization at the Bashar step size. We sweep the fixed resolution over the available modes and keep the best single value per scenario. This is the baseline to fear. Sheng et al. report saturation near 3 bits, so a well-chosen static 3-bit design may already sit close to the optimum. Beating it means higher energy efficiency, with ADC and fronthaul power both counted, at equal or better 5th-percentile user rate.
The strongest existing method is a centralized bit allocator in the style of JBQA. It gets global instantaneous channels at the CPU, runs combiner and bit optimization per block, and sends bit commands back to each AP. It gets the same power model and the same iteration budget as our method, and its channel upload is charged at the same fronthaul quantization. We report it twice, once with signaling ignored as an upper bound and once with signaling charged. No reported energy-efficiency number was found on Lacuna for JBQA or for any cell-free RADC scheme. The file gives only the qualitative 3-bit saturation, so we will reimplement it. The winning condition is to reach at least 90% of its signaling-free energy efficiency while using well under 1% of its control traffic, and to exceed it outright once signaling is charged in short-coherence settings. As a sanity anchor we will reproduce the Bashar setup, which reports a per-user rate of 2.442 bits/s/Hz under limited fronthaul, before adding adaptive converters.
The key ablation removes the price. Each AP instead uses a local rule that maximizes its own SINR per watt, with a fixed global price. This isolates whether the CPU's statistics-derived $\lambda_m$ adds anything beyond local greed. A second ablation fixes the fronthaul bits and adapts only the ADC, which tests whether the cascade coupling matters.
Why not just…?
Why not just fix every ADC at 3 bits? This is a serious threat. Sheng et al. place the saturation point there, and Castañeda et al. show 3-bit modes already capture most of the efficiency gain. The case for adaptation rests on heterogeneity. In cell-free deployments most APs are far from most users, while a few AP–user pairs have high SNR and need more bits to avoid the rate saturation that On Throughput of Millimeter Wave MIMO Systems with Low Resolution ADCs documents for coarse converters. If the best static mode comes within a few percent of our method across load and SNR sweeps, we will report that adaptation is not worth its control complexity in cell-free uplinks.
Why not just let each AP pick bits from its own SNR, with no CPU involvement? The price-free ablation tests exactly this. Our expectation is that local greed over-provisions APs whose contribution the LSFD weights will largely discard, and under-provisions weak-but-unique links for cell-edge users. The effect should show up in max-min rate more than in sum rate. If the ablation matches the full method, the contribution shrinks to the cascade analysis alone, and we will say so.
Why not just centralize with fiber fronthaul? With signaling ignored, the centralized allocator is our upper bound, so we do not claim to beat it on raw efficiency. The question is whether its channel upload and roughly 10-iteration solve, the convergence behavior reported by Xiu et al., survive short coherence times. Mobile Cell-Free Massive MIMO: Challenges, Solutions, and Future Directions documents severe losses from stale coordination under mobility. Our Doppler sweep measures where the centralized allocator's aged decisions fall below local decisions made on fresh channel estimates.
Proposed Experiments
The main testbed follows the Weichselberger cell-free setup of Wang et al.: a 1×1 km² area with wrap-around, 20 APs and 10 users. It is extended with per-AP converters that switch between 3 and 6 bits, plus a finer 1–6 bit variant, with power scaled from the 39.7 pJ/b figure and the ~3x low-mode gain. For robustness we add the 3GPP urban-microcell setup of Chen et al. and QuaDRiGA mmMAGIC UMi channels at 60 GHz. Channel aging follows a first-order Doppler model, from pedestrian speeds up to 350 km/h. Fronthaul capacity is swept from generous to binding.
We report energy efficiency in bits per joule including converter, fronthaul and control power; sum and 5th-percentile spectral efficiency; control bits per coherence block; and resolution-switching frequency. The decisive checks are three. First, whether the price-guided method stays within 10% of centralized signaling-free efficiency while beating the best static mode by a margin larger than seed-to-seed variation. Second, whether it overtakes the signaling-charged centralized scheme as coherence time shrinks. Third, whether always acquiring pilots at 6 bits is necessary to avoid collapse from the estimation feedback loop.
References
- Energy Efficiency Optimization for Millimeter Wave System With Resolution-Adaptive ADCs
- Uplink Achievable Rate Maximization for Reconfigurable Intelligent Surface Aided Millimeter Wave Systems With Resolution-Adaptive ADCs
- A Resolution-Adaptive 8 mm2 9.98 Gb/s 39.7 pJ/b 32-Antenna All-Digital Spatial Equalizer for mmWave Massive MU-MIMO in 65nm CMOS
- Uplink Precoding Design for Cell-Free Massive MIMO With Iteratively Weighted MMSE
- Over-the-Air Computation in Cell-Free Massive MIMO Systems
- Max-Min Rate of Cell-Free Massive MIMO Uplink With Optimal Uniform Quantization
- On Throughput of Millimeter Wave MIMO Systems with Low Resolution ADCs
- Mobile Cell-Free Massive MIMO: Challenges, Solutions, and Future Directions