Enhanced IMF-SIC for MIMO Detection

• The paper introduces IMF-SIC, which recursively applies shadow area constraints and dynamic feedback to significantly mitigate error propagation in MIMO detection.
• It employs multiple feedback loops and branch searches over constellation neighbors to achieve near-optimal detection with manageable computational complexity.
• Simulation results show gains up to 2 dB over conventional SIC, demonstrating effective performance improvements in varying MIMO configurations.

Improved Multiple Feedback Successive Interference Cancellation (IMF-SIC) is an advanced detection algorithm designed for symbol vector detection in spatially multiplexed multiple-input multiple-output (MIMO) systems. It addresses the performance limitations of classical Successive Interference Cancellation (SIC) by introducing recursively applied multiple feedback and dynamic reliability checks at each detection layer, thereby strongly mitigating error propagation and enabling near-optimal detection performance with tractable computational complexity (Li et al., 2013, Mandloi et al., 2015).

1. System Model and Detection Formulation

Consider a MIMO communication system with $N_t$ transmit and $N_r$ receive antennas. The received signal vector is modeled by

$\mathbf{y} = \mathbf{H} \mathbf{s} + \mathbf{n}$

where $\mathbf{y} \in \mathbb{C}^{N_r}$ is the received vector, $\mathbf{H} \in \mathbb{C}^{N_r \times N_t}$ is the channel matrix (typically i.i.d. $\mathcal{CN}(0,1)$ entries), $\mathbf{s} \in \mathbb{A}^{N_t}$ is the transmit symbol vector drawn from a constellation $\mathbb{A}$ (e.g., QAM), and $$\mathbf{n}\sim \mathcal{CN}(0, \sigma^2 \mathbf{I})$$ is additive white Gaussian noise. The maximum-likelihood (ML) detection problem is

$$\hat {\mathbf{s}}_{ML} = \arg\min_{\mathbf{s} \in \mathbb{A}^{N_t}} \|\mathbf{y} - \mathbf{H}\mathbf{s}\|^2$$

which is infeasible for large $N_r$0 or high-order constellations due to exponential complexity (Mandloi et al., 2015).

2. Conventional SIC and Multiple Feedback (MF-SIC) Framework

SIC detects symbols sequentially, applying a (typically MMSE) filter to obtain soft estimates, making hard decisions, and cancelling detected symbols from $N_r$1. The $N_r$2th layer applies the filter

$N_r$3

to the residual vector

$N_r$4

producing soft output $N_r$5 and hard output $N_r$6. Successive incorrect decisions drive error propagation and severe BER floors (Li et al., 2013).

The MF-SIC technique introduces the Shadow Area Constraint (SAC): for each layer, the Euclidean distance $N_r$7 is computed. If $N_r$8 (the “shadow region”), a set of $N_r$9 nearest constellation neighbors are used to branch the search for possible symbol candidates. The best candidate is selected by the minimum Euclidean cost (Mandloi et al., 2015).

3. Recursive SAC and the IMF-SIC Algorithm

MF-SIC’s limitation is that, within each branched candidate, subsequent layers default back to conventional SIC without SAC evaluation; this can induce error propagation if lower layers also fall in the shadow region. IMF-SIC addresses this by recursively applying SAC at each layer within every feedback branch, invoking a recursive routine if unreliability is detected. Specifically, for layer $\mathbf{y} = \mathbf{H} \mathbf{s} + \mathbf{n}$0:

• Compute $\mathbf{y} = \mathbf{H} \mathbf{s} + \mathbf{n}$1 and $\mathbf{y} = \mathbf{H} \mathbf{s} + \mathbf{n}$2.
• If $\mathbf{y} = \mathbf{H} \mathbf{s} + \mathbf{n}$3, set $\mathbf{y} = \mathbf{H} \mathbf{s} + \mathbf{n}$4.
• If $\mathbf{y} = \mathbf{H} \mathbf{s} + \mathbf{n}$5, branch over $\mathbf{y} = \mathbf{H} \mathbf{s} + \mathbf{n}$6 nearest neighbors and apply the IMF-SIC subroutine recursively to each, up to maximum recursion depth $\mathbf{y} = \mathbf{H} \mathbf{s} + \mathbf{n}$7 (Mandloi et al., 2015).

The recursive strategy enforces reliability screening of every symbol decision, substantially mitigating error propagation compared to MF-SIC. The candidate with the overall minimum ML cost is ultimately chosen.

4. Ordered IMF-SIC (OIMF-SIC) and Dynamic Detection Ordering

OIMF-SIC refines the detection process by dynamically reordering the decision sequence at each stage based on a log-likelihood ratio (LLR) metric, prioritizing the most reliable symbols first. For each undecoded index $\mathbf{y} = \mathbf{H} \mathbf{s} + \mathbf{n}$8: $\mathbf{y} = \mathbf{H} \mathbf{s} + \mathbf{n}$9 where $\mathbf{y} \in \mathbb{C}^{N_r}$0 and $\mathbf{y} \in \mathbb{C}^{N_r}$1. After each new decision, the order is re-updated, ensuring that reliable estimates are detected and cancelled first. This further reduces the likelihood of early errors, especially in ill-conditioned channels or at low SNR (Mandloi et al., 2015).

5. Complexity Analysis and Parameterization

Let $\mathbf{y} \in \mathbb{C}^{N_r}$2. Conventional SIC requires $\mathbf{y} \in \mathbb{C}^{N_r}$3 complexity per symbol for filter operations. MF-SIC involves up to $\mathbf{y} \in \mathbb{C}^{N_r}$4 branches per shadow region (worst-case $\mathbf{y} \in \mathbb{C}^{N_r}$5), but the shadow event is rare with careful setting of $\mathbf{y} \in \mathbb{C}^{N_r}$6, so typical complexity remains close to SIC. IMF-SIC permits recursion to depth $\mathbf{y} \in \mathbb{C}^{N_r}$7, leading to worst-case branching $\mathbf{y} \in \mathbb{C}^{N_r}$8. OIMF-SIC adds LLR computations ($\mathbf{y} \in \mathbb{C}^{N_r}$9) and sorting ($\mathbf{H} \in \mathbb{C}^{N_r \times N_t}$0), which are negligible compared to branching complexity.

Practical settings:

• $\mathbf{H} \in \mathbb{C}^{N_r \times N_t}$1 for 4-QAM, higher (e.g., 0.5) for 16-QAM.
• $\mathbf{H} \in \mathbb{C}^{N_r \times N_t}$2 chosen as up to half the constellation size (e.g., $\mathbf{H} \in \mathbb{C}^{N_r \times N_t}$3 for 4-QAM, $\mathbf{H} \in \mathbb{C}^{N_r \times N_t}$4 for 16-QAM).
• Recursion depth $\mathbf{H} \in \mathbb{C}^{N_r \times N_t}$5–$\mathbf{H} \in \mathbb{C}^{N_r \times N_t}$6 typically suffices to approach ML performance (Mandloi et al., 2015).

6. Simulation Results and Comparative Performance

Simulation studies cover Rayleigh fading channels, perfect CSI, and 10,000 trial averages (Mandloi et al., 2015, Li et al., 2013):

• For $\mathbf{H} \in \mathbb{C}^{N_r \times N_t}$7 4-QAM ($\mathbf{H} \in \mathbb{C}^{N_r \times N_t}$8, $\mathbf{H} \in \mathbb{C}^{N_r \times N_t}$9, $\mathcal{CN}(0,1)$0), IMF-SIC and OIMF-SIC achieve $\mathcal{CN}(0,1)$11 dB gain over MF-SIC at BER $\mathcal{CN}(0,1)$2, closely trailing ML.
• For $\mathcal{CN}(0,1)$3 4-QAM, IMF-SIC/OIMF-SIC deliver $\mathcal{CN}(0,1)$42 dB gain over MF-SIC, with BER $\mathcal{CN}(0,1)$5 nearly matching ML detectors.
• For $\mathcal{CN}(0,1)$6 4-QAM ($\mathcal{CN}(0,1)$7, $\mathcal{CN}(0,1)$8, $\mathcal{CN}(0,1)$9), OIMF-SIC with deep recursion attains near-ML BER.
• With $\mathbf{s} \in \mathbb{A}^{N_t}$0-QAM, OIMF-SIC (with $\mathbf{s} \in \mathbb{A}^{N_t}$1–$\mathbf{s} \in \mathbb{A}^{N_t}$2, $\mathbf{s} \in \mathbb{A}^{N_t}$3) performs within $\mathbf{s} \in \mathbb{A}^{N_t}$4–$\mathbf{s} \in \mathbb{A}^{N_t}$5 dB of ML, depending on antenna configuration.

These results demonstrate that recursive SAC enforcement and dynamic ordering in IMF-SIC and OIMF-SIC effectively suppress error propagation and yield near-ML BER at polynomial complexity, even for large-scale MIMO.

7. Extensions and Integration with SISO Turbo Detection

IMF-SIC techniques can be embedded as front-end modules in soft-input, soft-output (SISO) turbo detection loops for coded MIMO systems. Here, initial iterations use IMF-SIC (or MB-MF-SIC) to provide extrinsic LLRs, while subsequent iterations typically revert to less complex soft-cancellation schemes (e.g., MMSE or PIC), or selectively re-apply MF-SIC for critical layers (Li et al., 2013). This integration allows coded systems to achieve BER performance within $\mathbf{s} \in \mathbb{A}^{N_t}$6–$\mathbf{s} \in \mathbb{A}^{N_t}$7 dB of the single-user, turbo-equalized bound using a manageable number of turbo iterations.

---

References:

• "Multi-Feedback Successive Interference Cancellation for Multiuser MIMO Systems" (Li et al., 2013)
• "Improved Multiple Feedback Successive Interference Cancellation Algorithm for Near-Optimal MIMO Detection" (Mandloi et al., 2015)