---
title: Activity-Aware SIC for mMTC Uplinks
url: https://www.emergentmind.com/topics/activity-aware-sic-aa-mf-sic
type: topic
---

# Activity-Aware SIC for mMTC Uplinks

Activity-aware Multiple Feedback Successive Interference Cancellation (AA-MF-SIC) is a low-complexity detection strategy for massive machine-type communications (mMTC), specifically targeting grant-free code division multiple access (CDMA) uplinks in scenarios where user activity is sparse and random. AA-MF-SIC incorporates a priori activity probabilities into both its filter design and its decision logic, thereby achieving robustness and performance improvements over conventional SIC-based approaches while maintaining computational efficiency [1903.10643].

## 1. System and Signal Model

AA-MF-SIC is designed for under-determined grant-free CDMA uplinks with $N$ machine-type devices (MTCDs) and a spreading factor $M$ at the receiver. Each active device transmits one symbol per signaling interval, drawn from a finite constellation $\mathcal{A}$, while inactive devices transmit zero. The augmented transmit vector is
\[
\mathbf{x} = [x_1, \dots, x_N]^T, \qquad x_n \in \mathcal{A}_0, \;\; \mathcal{A}_0 = \mathcal{A} \cup \{0\},
\]
and the received signal is modeled as
\[
\mathbf{y} = \mathbf{H}\,\mathbf{x} + \mathbf{n},
\]
where $\mathbf{H} \in \mathbb{C}^{M \times N}$ combines spreading sequences with flat Rayleigh-fading coefficients, and $\mathbf{n} \sim \mathcal{CN}(\mathbf{0},\,\sigma_n^2\mathbf{I})$ is circular complex Gaussian noise.

The activity pattern is sparse: each $x_n$ is nonzero (active) with probability $p_n \ll 1$, otherwise $x_n = 0$, with user activity assumed independent across devices.

## 2. Detection Problem Formulation

The optimal maximum a posteriori (MAP) rule selects
\[
\hat{\mathbf{x}} = \arg\min_{\mathbf{x}\in\mathcal{A}_0^N} \left\| \mathbf{y} - \mathbf{H} \mathbf{x} \right\|_2^2 + \sum_{n=1}^N \lambda_n\,\mathbf{1}_{\{x_n\neq 0\}},
\]
where
\[
\lambda_n = \ln \left( \frac{1 - p_n}{p_n / |\mathcal{A}|} \right),
\]
and $\mathbf{1}_{\{x_n\neq 0\}}$ is the indicator function. This cost augments least-squares detection with an activity-related penalty term, enforcing sparsity in the estimated signal.

Direct minimization is intractable for large $N$, motivating the use of low-complexity SIC, where users are detected sequentially. At each step, interference from previously detected users is cancelled, and the activity prior ($\lambda_n$) shapes both the detection and reliability assessment.

## 3. Algorithmic Structure and Innovations

### 3.1 MMSE Filter with Activity-aware $\ell_1$ Regularization

At stage $n$ of SIC, the residual is
\[
\mathbf{y}_n = \mathbf{y} - \sum_{j=1}^{n-1} \mathbf{h}_j'\, \hat x_j,
\]
with $\mathbf{H}'$ a possibly permuted version of $\mathbf{H}$. The filter $\mathbf{w}_n \in \mathbb{C}^M$ is chosen to minimize
\[
J(\mathbf{w}_n) = \mathbb{E}\left[ |x_n - \mathbf{w}_n^H \mathbf{y}_n|^2 \right] + 2\lambda_n \|\mathbf{w}_n\|_1,
\]
where $\ell_1$-regularization is used to account for sparsity and is approximated by a weighted quadratic
\[
\|\mathbf{w}_n\|_1 \approx \mathbf{w}_n^H \boldsymbol\Lambda\, \mathbf{w}_n, \quad \boldsymbol\Lambda = \mathrm{diag}\left( \frac{1}{|w_{n,i}| + \epsilon} \right).
\]
Setting the derivative to zero yields the activity-aware MMSE filter
\[
\mathbf{w}_n = \left( \mathbf{R}_n + \frac{2\lambda_n}{\sigma_x^2} \boldsymbol\Lambda \right)^{-1} \mathbf{h}_n',
\]
with
\[
\mathbf{R}_n = \mathbf{H}_n' \mathbf{H}_n'^H + \frac{\sigma_n^2}{\sigma_x^2} \mathbf{I},
\]
where $\mathbf{H}_n'$ consists of columns $n$ to $N$. The soft symbol estimate is $z_n = \mathbf{w}_n^H \mathbf{y}_n$, followed by hard quantization $\hat x_n^{\text{(hard)}} = \mathcal{Q}[z_n]$ over $\mathcal{A}_0$.

### 3.2 Activity-aware Shadow Area Constraint (SAC)

Decision reliability at each user is dynamically set according to activity priors. The minimum distance to the nearest constellation point is
\[
d_n = \min_{a \in \mathcal{A}_0} |z_n - a|.
\]
The reliability threshold is
\[
d_n^{\text{th}} = 
\begin{cases}
1/\lambda_n, & \text{if the closest point is } 0, \\
1 - (1/\lambda_n), & \text{otherwise}.
\end{cases}
\]
If $d_n \le d_n^{\text{th}}$, the hard decision is accepted; otherwise, AA-MF-SIC enters a multi-feedback (MF) stage.

### 3.3 Multi-feedback (MF) Stage

A candidate set $\{c_1, ..., c_F\} \subset \mathcal{A}_0$ of the $F$ nearest symbols (including $0$) is established around $z_n$. For each candidate $c_f$, a trial-path is constructed: $\tilde x_n = c_f$ is cancelled from the residual, and standard SIC is executed for subsequent users, producing a trial estimate $\mathbf{b}_f$. The final decision for $x_n$ is selected as
\[
f^* = \arg\min_{f=1,\dots,F} \|\mathbf{y} - \mathbf{H}' \mathbf{b}_f\|_2^2,
\]
with $\hat x_n = b_{n,f^*}$.

This mechanism injects alternative hypotheses only for uncertain decisions, thereby suppressing error propagation at a modest complexity.

### 3.4 Algorithm Pseudocode

The high-level workflow is
```text
Input: y,H', {p_n}, λ_n, F, ε
Initialize y^MF = y
For n = 1 to N:
  Compute R_n = H'_n H'_n^H + (σ_n^2/σ_x^2) I
  Compute Λ = diag{1/(|w_{n,i}|+ε)}
  Compute w_n = (R_n + 2λ_n/σ_x^2 Λ)^(-1) h'_n
  z_n = w_n^H y^MF
  d_n = min_{a in A_0}|z_n - a|
  Determine d^th_n by activity (1/λ_n or 1−1/λ_n)
  If d_n ≤ d^th_n:
    x̂_n = Q(z_n)
  Else:
    Build candidate set {c_1…c_F} around z_n
    For f=1…F:
      Set trial b_f[1:n−1]=previous estimates, b_f[n]=c_f
      y_f = y
      For i=n to N:
        y_f ← y_f − h'_i b_f[i]
        Compute z_i = w_i^H y_f, b_f[i]=Q(z_i)
    f* = argmin_f ||y – H' b_f||^2
    x̂_n = b_f*[n]
  y^MF ← y^MF − h'_n x̂_n
End
Output: x̂
```
This structure maintains per-stage low complexity, with additional MF branches invoked only as required.

## 4. Computational Complexity

The computational costs of AA-MF-SIC and comparative detectors are summarized below:

| Detector         | Complexity Order                                     | Notable Features                          |
|------------------|-----------------------------------------------------|-------------------------------------------|
| MMSE             | $O(N^2)$                                            | -                                         |
| SA-SIC           | $\approx \frac{1}{6}(3N^3 + 11N^2 + 21N)$           | No ordering                               |
| K-Best           | $K|\mathcal A_0| (N^3/3+2N^2+O(N)+\log^2(K|\mathcal A_0|))$ | Number of survivors $K$, alphabet size    |
| Ordered SA-SIC   | $O(N^3)$                                            | A-SQRD, etc.                              |
| AA-MF-SIC        | $O(N^3)$, adds $\sim 10N^2$ at low SNR              | Modest per-symbol MF branch overhead      |

At high SNR, the multi-feedback branch is rarely needed, so complexity remains limited by the cubic scaling of conventional SIC. At low SNR, the number of MF paths increases, adding a quadratic term. These characteristics ensure scalability for large-scale mMTC scenarios [1903.10643].

## 5. Performance Evaluation

AA-MF-SIC was evaluated under uncoded block-fading channels using QPSK-augmented $\mathcal{A}_0$, with $N=128$ devices and spreading length $M=64$. The device activation probabilities $p_n$ were drawn uniformly from $[0.1, 0.3]$. Results over $10\,000$ Monte Carlo trials with SNR in the range $0$–$20$ dB are characteristic.

Key performance metric: Net Symbol Error Rate (NSER) for active devices.

Key results:
- AA-MF-SIC consistently outperforms MMSE, SA-SIC, SA-SIC+ordering (A-SQRD), Iterative Reweighted, and K-Best detectors across the entire SNR range, achieving up to $\sim 3$ dB gain at NSER $= 10^{-2}$.
- At increased activity rates, AA-MF-SIC remains superior and eventually outpaces even an “Oracle MMSE” detector with perfect active set knowledge.
- Under imperfect CSI (with $\hat H = H + E$, $E_{ij} \sim \mathcal{CN}(0, \hat\sigma^2)$), the AA-MF-SIC preserves its performance advantage, with only a slight margin lost compared to perfect CSI.

## 6. Key Insights and Applications

AA-MF-SIC leverages activity-awareness in two critical algorithmic junctures:
- The MMSE filter regularization is modulated by each user’s activation probability via $\lambda_n$, directly informing the aggressiveness of interference cancellation and thresholding.
- Decision confidence is governed by the activity-tuned Shadow Area Constraint (SAC), ensuring that only unreliable soft estimates trigger multi-path exploration.

This architecture suppresses error propagation, a principal source of performance degradation in standard SIC, while maintaining low additional cost per user—particularly at moderate and high SNR.

AA-MF-SIC enables efficient, near-optimal detection in mMTC settings where sparse user activity and grant-free operation preclude traditional scheduling and coordinated access. The design is naturally extensible to adaptive candidate set selection, coded transmission schemes (e.g., integration with LDPC codes), higher-order constellations, MIMO uplinks, and joint channel estimation and detection for grant-free systems [1903.10643].

## 7. Trade-offs and Prospects

The balance between complexity and performance is governed by the parameter $F$ (number of multi-feedback branches). Larger $F$ mitigates unreliability at low SNR but increases computational burden. In practice, $F$ can be set adaptively according to SNR or empirical reliability statistics.

Potential lines of further research include:
- Optimizing MF candidate selection beyond nearest points.
- Integration with advanced channel-coding frameworks.
- Generalization to higher spectral efficiency and multi-antenna contexts.
- Joint detection and channel estimation, particularly for fully grant-free access architectures.

AA-MF-SIC thus provides a systematic, activity-informed framework for powerful, scalable detection in the emerging landscape of massive, uncoordinated, and sporadic machine-type communications [1903.10643].

Source: https://www.emergentmind.com/topics/activity-aware-sic-aa-mf-sic