Quantizer-Aware Successive Cancellation for Low-Resolution NOMA Receivers

Non-orthogonal multiple access (NOMA) lets several users transmit on the same frequencies at the same time. The receiver separates them by successive interference cancellation (SIC): it decodes one user, rebuilds that user's signal, subtracts it, and moves on to the next user. Power-hungry wideband receivers increasingly digitize with coarse analog-to-digital converters (ADCs) whose bit depth changes at run time. Coarse quantization adds distortion whose power scales with the total received signal, so most of it is caused by the strongest user. Standard SIC subtracts only a clean, linear copy of that user's signal. Its quantization distortion, which we will call its "shadow," stays in the residual and lands on the weaker users.

This note makes two claims. First, much of that shadow is predictable once the strong user has been decoded, because the ADC's transfer function is known or can be calibrated. A receiver that subtracts the expected quantized contribution of the decoded user, instead of its linear reconstruction, should remove most of the shadow. Second, whether this works decides the best decoding order and how few ADC bits the receiver can get away with. With naive subtraction, lowering the bit depth should push the optimal order toward "weak user first." With quantizer-aware subtraction it should swing back to "strong user first," and the receiver can drop to fewer bits for the same error rate. Neither the ordering literature nor the resolution-adaptive hardware literature has looked at this coupling.

Core Claim

Consider a receiver with one or a few antennas in the overloaded regime, where there are more users than antennas and NOMA actually matters. Each ADC output can be written with the Bussgang decomposition, which splits a nonlinear quantizer into a linear gain plus distortion that is uncorrelated with the input:

$$r = G y + d, \qquad y = \sum_k h_k x_k + n.$$

Here $r$ is the quantized output, $y$ the analog input, $G$ the Bussgang gain, $d$ the distortion, $h_k$ and $x_k$ user $k$'s channel and symbol, and $n$ thermal noise.

Suppose $d$ is treated as Gaussian noise whose power is proportional to total received power, as in the aggregate impairment factor $\kappa$ of MIMO Transmission with Residual Transmit-RF Impairments. With perfect subtraction, the chain rule of mutual information then makes the two-user single-antenna sum rate identical for either decoding order: both sum to $\log(1+(P_1+P_2)/N)$, where $P_k$ is user $k$'s received power and $N$ is thermal plus distortion power. Under that surrogate model, the decoding order can matter only through imperfect cancellation. That is exactly how Achievable Rate Analysis of the STAR-RIS-Aided NOMA Uplink in the Face of Imperfect CSI and Hardware Impairments arrives at weak-user-first. That paper adds a residual-cancellation fraction and $\kappa$-distortion to a two-user uplink and shows that decoding the weaker user first then maximizes sum rate. This contradicts the ideal-hardware strong-user-first rule of Joint Power Control and Beamforming for Uplink Non-Orthogonal Multiple Access in 5G Millimeter-Wave Communications.

Our point is that $d$ is uncorrelated with $y$ but not independent of the decoded symbol. When user 1 dominates, the quantizer output is mostly a function of $h_1x_1$. The conditional mean $\mathbb{E}[d \mid x_1]$ therefore carries most of the distortion energy. The receiver can compute it per ADC by integrating the quantizer's cells against the conditional distribution of the undecoded users plus noise. Quantizer-aware SIC subtracts

$$\hat r_1 = \mathbb{E}[\,r \mid x_1\,] = G h_1 x_1 + \mathbb{E}[\,d \mid x_1\,]$$

instead of $G h_1 x_1$ alone.

The effective residual-cancellation fraction is therefore not a fixed hardware constant. It depends on the bit depth $B$, the users' power ratio, the ADC's capacitor mismatch, and which subtraction rule is used. The rate-optimal order flips at a threshold resolution $B^\star$, and that threshold moves when quantizer-aware subtraction is switched on. Decoding strong-first lets the receiver predict the dominant distortion. Decoding weak-first gives almost no leverage on it. We therefore expect quantizer-aware SIC to favor strong-first and to reach a target error rate at lower $B$.

Research Program

The first component is analysis. For two users we derive per-stage signal-to-noise-and-distortion ratios (SNDRs) for three subtraction rules: linear, conditional-mean with the ideal quantizer, and conditional-mean with the calibrated mismatched quantizer. From these we locate $B^\star$ as a function of power ratio and mismatch level. For more than two users, closed forms for ordered channel statistics become intractable, a limitation also noted for Closed-Form BER Analysis for Uplink NOMA with Dynamic SIC Decoding. There we use the per-stage SNDR as a cheap scoring rule and compare it against exhaustive search over orders.

The second component builds the receiver. It has three parts. The first is a per-ADC lookup table that maps a decoded symbol's contribution and the residual variance to $\mathbb{E}[r \mid x_k]$. The second is an ordering rule that picks the next user by predicted post-subtraction SNDR rather than by channel gain. The third is a controller that jointly picks, per frame, the bit depth, the order, and the subtraction rule, choosing the lowest-energy configuration that meets a bit-error-rate (BER) target.

The third component handles fabrication mismatch in successive-approximation (SAR) ADCs. The Impact of SAR-ADC Mismatch on Quantized Massive MU-MIMO Systems shows that errors in the largest (most significant bit) capacitor shift conversion thresholds and create an uncoded BER floor that transmit power cannot remove. It models this with Bussgang decomposition under linear equalization only. Mitigating SAR-ADC Non-Idealities in Massive MU-MIMO Systems via Affine Models corrects gain and offset through an affine decomposition, again for linear detection only. We supply the method both papers leave unaddressed. We compute the conditional mean with each ADC's calibrated thresholds, estimated from pilots, and ask whether full-threshold calibration is needed or whether the affine model suffices inside SIC.

We deliberately do not propose re-quantizing the residual between SIC stages. In a digital SIC receiver the ADC converts the composite signal once. The shrinking dynamic range after subtraction, which a naive "stage-adaptive bits" design would exploit, would require analog-domain cancellation. We treat that as out of scope.

Baselines

The naive reference is fixed strong-user-first ordering by average channel gain with linear subtraction. We run it at 6 bits, the top resolution of the ASIC in 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, and at every lower bit depth.

The strongest SIC competitor is dynamic ordering by instantaneous channel gain from Closed-Form BER Analysis for Uplink NOMA with Dynamic SIC Decoding. That paper proves this rule removes the fixed-order error floor with ideal hardware, and it reports that BPSK needs only a 1.63 dB power gap between users while 64-QAM needs 19.05 dB. We also run weak-user-first ordering driven by an equivalent-$\kappa$ model, following Achievable Rate Analysis of the STAR-RIS-Aided NOMA Uplink in the Face of Imperfect CSI and Hardware Impairments.

The strongest hardware-floor competitor is an affine-compensated front end followed by dynamic SIC, taken from Mitigating SAR-ADC Non-Idealities in Massive MU-MIMO Systems via Affine Models. In linear massive MU-MIMO detection with 4-bit ADCs, mismatch $\sigma_m=\Delta/2$, and 10 dB signal-to-noise ratio (SNR), that paper reports a 90%-quantile BER of $1.4\times10^{-3}$, down from $3.5\times10^{-2}$ uncorrected. No reported result was found on Lacuna for any SIC or NOMA receiver under SAR-ADC mismatch.

Two further baselines bound the problem from above. Bussgang-aware linear minimum-mean-square-error (MMSE) detection without SIC is the architecture of the resolution-adaptive ASIC. Joint maximum-likelihood detection over the exact quantized likelihood is an oracle upper bound, feasible for two users with up to 16-QAM. A 10-bit, mismatch-free receiver serves as the high-resolution reference.

All methods share the same channel draws, the same mismatch instances, the same pilot budget for calibration, and the same per-user power control. Every ordering rule gets the same channel state information.

Two ablations isolate the key ingredient. The first applies quantizer-aware subtraction with the order held fixed. The second applies the new ordering rule with linear subtraction. A third check uses the ideal-quantizer lookup table on a mismatched ADC, to test whether calibration is what matters.

The proposal beats the field if it meets a given BER target at least one bit lower than affine-compensated dynamic SIC. That margin corresponds to roughly halving ADC power under a $2^B$ power model. It should also close most of the gap to the joint maximum-likelihood oracle.

Why not just…?

Why not just model quantization as an equivalent $\kappa$ and apply the weak-user-first rule? This is the most serious threat. If the Gaussian surrogate is adequate, a one-line rule suffices and our lookup tables are unnecessary. The chain-rule identity above means the surrogate predicts that order matters only through a residual-cancellation fraction. The surrogate cannot say how that fraction depends on $B$ or on which subtraction rule is used. The ablation with fixed order but quantizer-aware subtraction settles this directly. If it gains nothing over linear subtraction at 2–4 bits, the surrogate wins, and we would report that as a clean negative result.

Why not just use high-resolution ADCs? The adaptive ASIC measured roughly 3x better energy efficiency when dropping to 3-bit ADCs, which is the whole motivation for resolution adaptation. More bits also leave the $1/\kappa$ ceiling from other radio impairments untouched, as shown in A New Look at Dual-Hop Relaying: Performance Limits with Hardware Impairments. We include the 10-bit receiver so the energy cost of simply adding bits is measured rather than argued.

Why not just use linear MMSE detection across antennas and skip SIC? With many more antennas than users, this is probably good enough, and the ASIC's success suggests as much. We therefore restrict the claim to overloaded and few-antenna regimes. We sweep the antenna-to-user ratio to find where linear detection catches up. If it does so already at modest overload, the proposal's scope shrinks accordingly.

Why not just run joint maximum-likelihood detection? Its cost grows exponentially with the number of users and the constellation size. We use it as the oracle that tells us how much of the achievable gain SIC-complexity processing recovers.

Proposed Experiments

The simulations cover two to four uplink users. Channels are drawn from i.n.i.d. Rayleigh fading, reusing the Monte Carlo setup of the dynamic-SIC paper and validating our unquantized results against its closed forms. They are also drawn from QuaDRiGa mmMAGIC urban microcell channels at 60 GHz, the scenario used by both the ASIC and SAR-mismatch papers. The access point has 1 to 8 antennas. Bit depth runs from 1 to 6 bits. Capacitor mismatch follows the SAR-ADC model of Guichemerre and Studer with $\sigma_m$ up to $\Delta/2$, and we report 90%-quantile BER over ADC instances, their yield criterion. Metrics are per-user uncoded BER, sum achievable rate computed from the quantized likelihood, the empirical threshold $B^\star$, and energy per bit under a $2^B$ ADC power model anchored to the ASIC's reported 39.7 pJ/b at full precision.

The first decisive check is whether, with ideal 4-bit ADCs, the measured order dependence grows as $B$ falls under linear subtraction but largely disappears under quantizer-aware subtraction. The surrogate model predicts no such asymmetry. The second is whether dynamic SIC, which removes the fading-inversion floor with ideal hardware, re-acquires a floor at $\sigma_m=\Delta/2$, and whether full-threshold conditional-mean subtraction removes that floor where affine correction does not. The third is whether the joint controller can drop to a lower bit depth for the same BER target in overloaded QuaDRiGa scenarios.