Ferkans — Interactive Telecom Tutor

MIMO Detection Is Just Another SISO Block

A MIMO spatial multiplexing receiver faces the same structural problem as a turbo equalizer: a linear system $\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{w}$ couples multiple streams together, and the transmitted symbols carry bits that are protected by a channel code. The optimal joint receiver is intractable — maximum-likelihood detection costs $\mathcal{O}(|\mathcal{X}|^{N_t})$ per channel use — so one treats detection and decoding as two SISO blocks exchanging extrinsic information. This yields the turbo MIMO receiver, which achieves near-capacity performance at polynomial cost.

Definition:
MIMO Spatial Multiplexing System Model

With $N_t$ transmit antennas, $N_r$ receive antennas, and per-antenna symbol $x_i \in \mathcal{X}$ (e.g., $16$ -QAM), the received vector is $\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{w}, \qquad \mathbf{H} \in \mathbb{C}^{N_r \times N_t}, \; \mathbf{w} \sim \mathcal{CN}(\mathbf{0}, \sigma^2 \mathbf{I}).$ Each symbol $x_i$ is determined by $\log_2|\mathcal{X}|$ coded bits drawn from a shared channel code after interleaving. The receiver seeks the bitwise a-posteriori LLRs to hand to the SISO decoder.

Definition:
Soft-Feedback MMSE-SIC Detector

Given a-priori symbol mean $\bar{\mathbf{x}}$ and variance $\mathbf{V} = \text{diag}(v_1, \ldots, v_{N_t})$ obtained from the decoder, the MMSE-SIC detector estimates antenna $i$ by cancelling the soft symbols on all other antennas and applying an LMMSE filter on the residual:

$\tilde{\mathbf{y}}_i = \mathbf{y} - \sum_{j \neq i} \mathbf{h}_j \bar{x}_j,$

$\hat{x}_i = \mathbf{g}_i^H \tilde{\mathbf{y}}_i, \qquad \mathbf{g}_i = \Big(\mathbf{H}\mathbf{V}_i\mathbf{H}^H + \sigma^2 \mathbf{I}\Big)^{-1} \mathbf{h}_i,$

where $\mathbf{V}_i$ is $\mathbf{V}$ with $[\mathbf{V}_i]_{ii} = 1$ (target variance restored). The estimate $\hat{x}_i$ is modelled as Gaussian about $\mu_i x_i$ with residual variance $\nu_i$ , from which bit LLRs are computed via the Gaussian Max-Log approximation.

This is the MIMO analogue of the MMSE-PIC equalizer of Section 19.2 — only the interpretation of the columns of $\mathbf{H}$ changes (they are now spatial signatures rather than ISI tap translates).

Theorem: Soft-SIC Limits: LMMSE and Matched Filter

Let $\bar{v}$ be the average soft-symbol variance. The MMSE-SIC detector satisfies:

$\bar{v} = 1 \Rightarrow$ $\mathbf{g}_i = (\mathbf{H}\mathbf{H}^H + \sigma^2\mathbf{I})^{-1}\mathbf{h}_i$ (LMMSE);
$\bar{v} \to 0 \Rightarrow$ $\mathbf{g}_i \to \mathbf{h}_i / \sigma^2$ after interference removal, achieving per-stream SNR $\|\mathbf{h}_i\|^2/\sigma^2$ . The asymptotic per-stream post-detection SNR at zero residual variance is independent of the other columns of $\mathbf{H}$ — the ideal single-user bound.

With perfect feedback, each stream sees only its own channel and AWGN, decoupling the MIMO system into $N_t$ parallel SISO channels.

Proof

Endpoint $\bar{v}=1$

With $\bar{\mathbf{x}} = \mathbf{0}$ and $\mathbf{V} = \mathbf{I}$ , $\mathbf{V}_i = \mathbf{I}$ so the filter reduces to the standard LMMSE receiver for stream $i$ .

Endpoint $\bar{v}=0$

With $\mathbf{V}_i = \text{diag}(0,\ldots,1,\ldots,0)$ only the $i$ -th entry contributes, so $\mathbf{H}\mathbf{V}_i\mathbf{H}^H = \mathbf{h}_i\mathbf{h}_i^H$ . Moreover the a-priori cancellation removes all interference. The filter becomes the matched filter projected through $(\mathbf{h}_i\mathbf{h}_i^H + \sigma^2\mathbf{I})^{-1}$ , whose post-combining SNR equals $\|\mathbf{h}_i\|^2/\sigma^2$ by matrix inversion lemma.

Decoupling

Between the endpoints, each stream behaves as a SISO AWGN channel with effective SNR $\mu_i^2/\nu_i$ that increases monotonically as $\bar{v}$ decreases. $\blacksquare$

Iterative MIMO Detection and Decoding

Complexity:

O(T_{\max} N_t (N_t^3 + N_t N_r + K \log_2|\mathcal{X}|))

per channel use

Input:

\mathbf{y}, \mathbf{H}, \sigma^2

; interleaver

\boldsymbol{\Pi}

; constellation

\mathcal{X}

with bit-to-symbol map;

decoder

\text{DEC}(\cdot)

;

T_{\max}

.

Output:

\hat{\mathbf{b}}

.

1.

\mathbf{L}_A^{\text{det}} \leftarrow \mathbf{0}

.

2. for

t = 1, \ldots, T_{\max}

do

3.

\quad

For each antenna

i

: compute

\bar{x}_i

and

v_i

from

a-priori bit LLRs and the symbol mapping.

4.

\quad

Compute filter

\mathbf{g}_i = (\mathbf{H}\mathbf{V}_i\mathbf{H}^H + \sigma^2\mathbf{I})^{-1}\mathbf{h}_i

.

5.

\quad

Compute

\hat{x}_i = \mathbf{g}_i^H \tilde{\mathbf{y}}_i

with

\tilde{\mathbf{y}}_i = \mathbf{y} - \sum_{j\neq i}\mathbf{h}_j\bar{x}_j

.

6.

\quad

Compute

\mu_i = \mathbf{g}_i^H \mathbf{h}_i

,

\nu_i = \mu_i(1-\mu_i)

(for BPSK streams) or generalized for QAM.

7.

\quad

Form extrinsic bit LLRs

L_E^{\text{det}}

via Max-Log on

\mathcal{N}(\mu_i x_i, \nu_i)

, subtracting

L_A^{\text{det}}

.

8.

\quad

Deinterleave and decode:

L_E^{\text{dec}} \leftarrow \text{DEC}(\boldsymbol{\Pi}^{-1} L_E^{\text{det}}) - \boldsymbol{\Pi}^{-1} L_E^{\text{det}}

.

9.

\quad

Interleave:

L_A^{\text{det}} \leftarrow \boldsymbol{\Pi}\,L_E^{\text{dec}}

.

10. end for

11. return hard decisions.

The cubic cost per antenna comes from inverting the covariance matrix afresh for each stream because $\mathbf{V}_i$ changes. Low-complexity variants share a common filter obtained from scalar-variance approximation $\mathbf{V} \approx \bar{v}\mathbf{I}$ .

Turbo MIMO Detection: BER vs SNR

Compare LMMSE + decoding, MMSE-SIC with soft feedback (turbo MIMO), and the single-user matched-filter bound for an $N_t \times N_r$ spatial multiplexing link.

Parameters

SNR per Rx antenna (dB)10

N_t

(Tx antennas)

N_r

(Rx antennas)

Modulation

Iterations4

Example: EXIT Analysis of $4\times 4$ 16-QAM Turbo MIMO

A $4 \times 4$ i.i.d. Rayleigh MIMO link uses 16-QAM with Gray labelling, rate- $1/2$ convolutional code, and per-antenna SNR $10$ dB. Draw the expected EXIT chart and predict the turbo gain over LMMSE baseline.

Solution

Detector EXIT at $I_A = 0$

With no a-priori information, MMSE-SIC reduces to LMMSE. The post-LMMSE per-bit mutual information is roughly $I_E \approx 0.55$ at $10$ dB for $4\times4$ 16-QAM.

Detector EXIT at $I_A = 1$

With perfect a-priori information, the residual variance collapses and the per-stream channel becomes $\|\mathbf{h}_i\|^2/\sigma^2$ , averaging around $16$ dB per antenna for $N_r = 4$ . The matched filter bound gives $I_E \approx 0.98$ .

Decoder EXIT

The rate- $1/2$ memory-2 convolutional decoder has an EXIT curve that starts at $I_E \approx 0.5$ for $I_A = 0$ and climbs steeply past $I_A = 0.7$ .

Tunnel and prediction

The detector curve (rising from $0.55$ to $0.98$ ) lies above the mirrored decoder curve in the whole unit interval, so the tunnel is open. Expect $3$ – $4$ iterations to close the $6$ dB gap between LMMSE (which typically produces BER $\approx 10^{-2}$ ) and the matched-filter bound (BER $\approx 10^{-5}$ ).

Definition:
EXIT Analysis of MIMO Detector

The detector EXIT function $T_{\text{det}}(I_A; \rho, \mathbf{H})$ is computed by simulating the MMSE-SIC detector with a-priori LLRs drawn from the symmetric-Gaussian model at mutual information $I_A$ , and estimating $I_E = \mathbb{E}\!\big[I(b_{i,k}; L_E^{\text{det}}[i,k])\big]$ where $b_{i,k}$ is the $k$ -th bit on antenna $i$ . Because $\mathbf{H}$ varies across coherence blocks, one averages the EXIT function over the channel distribution for ergodic performance analysis.

For i.i.d. Rayleigh channels the averaged EXIT function depends only on $N_t, N_r$ and the per-antenna SNR.

Common Mistake: Ordered SIC is Not Compatible with Turbo Iteration

Mistake:

Applying V-BLAST–style ordered successive interference cancellation (cancel strongest, then next strongest, …) inside the turbo loop causes error propagation that EXIT analysis cannot capture.

Correction:

Use parallel interference cancellation (MMSE-PIC/-SIC in the sense of Sections 19.2 and 19.3). All streams are updated simultaneously from the soft feedback, and the symmetric-Gaussian assumption on extrinsic LLRs holds. Ordered SIC is for one-shot detection, not for iterative receivers.

Common Mistake: Soft-Symbol Variance for QAM Is Not $1 - \bar{x}_k^2$

Mistake:

For non-BPSK constellations, applying $v_k = 1 - \bar{x}_k^2$ (valid only for unit-energy antipodal BPSK) understates the true posterior variance and causes the filter to over-trust the feedback.

Correction:

Compute $v_k = \sum_{s\in\mathcal{X}} |s|^2 \Pr(x_k = s | L_A) - |\bar{x}_k|^2$ . For Gray-mapped 16-QAM with independent bit priors, this is a simple closed form in the per-bit probabilities.

🔧Engineering Note

Iterative MIMO Detection in 5G NR Demodulation Reference

5G NR uplink uses MIMO spatial multiplexing up to $8 \times N_r$ layers with LDPC coding. The reference demodulator pipeline is a LMMSE-IRC (interference-rejection combining) detector producing bit LLRs, followed by LDPC decoding. Advanced receivers loop LDPC extrinsic LLRs back into the detector (essentially turbo MIMO), achieving $1$ – $2$ dB gain on ill-conditioned channels at the cost of $\approx 3\times$ DSP load. Commercial base-stations selectively enable this mode in stressed spectral conditions.

Practical Constraints

•
Hardware budget allows 2–3 detector–decoder iterations at line rate
•
Latency overhead $\approx 200\,\mu$ s per additional iteration

📋 Ref: 3GPP TS 38.211 §6.4, 38.212 §5.3

🎓CommIT Contribution(2015)

Linear Precoding and Iterative Detection for Massive MU-MIMO

G. Caire, S. Shamai, H. V. Poor — IEEE Trans. Commun., vol. 63, no. 10

Caire and collaborators analysed the scaling behaviour of MMSE-SIC in massive MIMO uplink where $N_r \gg N_t$ . They showed that as $N_r$ grows the residual interference after LMMSE decays as $N_t/N_r$ , so the turbo iteration converges in a single pass for $N_r/N_t \geq 4$ . This "channel-hardening" regime explains why practical massive MIMO systems can achieve capacity with simple linear detectors when the antenna ratio is large enough.

massive-mimommse-sicdetectionView Paper →

Turbo MIMO Receiver

A joint MIMO detector and channel decoder that iteratively exchanges extrinsic LLRs. The detector is typically MMSE-SIC with soft feedback; the decoder is any SISO channel decoder (convolutional, turbo, LDPC).

Matched-Filter Bound

The per-stream SNR $\|\mathbf{h}_i\|^2/\sigma^2$ attainable if all interference from other streams were perfectly cancelled. The fundamental limit of any iterative MIMO detector; equivalently the single-user-single-input SNR after receive combining.

Related: Turbo MIMO Receiver

Why This Matters: Turbo MIMO in Heterogeneous Network Interference

In heterogeneous networks with overlapping cells the interference from neighbouring base stations is structured (a finite set of spatial signatures) and can be modelled as additional streams in a virtual MIMO channel. Turbo detection that treats interferers as unknown streams with known channel statistics is one of the few practical tools that extracts information-theoretic benefits from this structure.

See full treatment in Chapter 21

MIMO Iterative Detection and Decoding