Precoding and Detection for MIMO-OTFS

Precoder and Detector, DD-Style

Classical MIMO-OFDM decomposes the frequency-selective channel into parallel flat subcarriers; each subcarrier sees a flat MIMO channel and is precoded/detected independently. MIMO-OTFS does not have this decoupling β€” the DD-domain channel is a 2D convolution in delay and Doppler, so precoding must account for the convolution structure. The result is a joint precoder-detector pair that operates on the full DD tensor. This section develops the natural MMSE/ZF precoders, shows how MP detection extends to MIMO, and discusses hybrid structures.

,

Definition:

Spatial Precoder and Combiner

Consider a MIMO-OTFS transmitter encoding NsN_s spatial streams into NtN_t antennas. The spatial precoder is a matrix V∈CNtΓ—Ns\mathbf{V} \in \mathbb{C}^{N_t \times N_s} applied per DD cell: x[β„“,k]β€…β€Š=β€…β€ŠV s[β„“,k],s[β„“,k]∈CNs.\mathbf{x}[\ell, k] \;=\; \mathbf{V}\, \mathbf{s}[\ell, k], \qquad \mathbf{s}[\ell, k] \in \mathbb{C}^{N_s}. At the receiver, the spatial combiner U∈CNrΓ—Ns\mathbf{U} \in \mathbb{C}^{N_r \times N_s} (typically ZF or MMSE) extracts per-stream estimates: s^[β„“,k]β€…β€Š=β€…β€ŠUH y[β„“,k].\hat{\mathbf{s}}[\ell, k] \;=\; \mathbf{U}^H\, \mathbf{y}[\ell, k]. Typical choices:

  • Full-rank ZF: V\mathbf{V} spans the NsN_s strongest right singular vectors of a channel averaged over DD.
  • Block-diagonal MMSE: V\mathbf{V} chosen per DD cell via MMSE criterion. Higher complexity, better performance in selective fading.
  • Constant-envelope: V\mathbf{V} is a DFT-like matrix for hybrid analog-digital beamforming (mmWave).

Theorem: SVD Precoder for MIMO-OTFS

For a MIMO-OTFS channel tensor HDD\mathcal{H}_{\mathrm{DD}}, define the effective MIMO channel averaged over DD: HΛ‰β€…β€Š=β€…β€Š1MNβˆ‘β„“,kHDD[β„“,k]β€…β€Š=β€…β€Šβˆ‘i=1Paiar(ΞΈi)at(Ο•i)Hβ‹…1[pathΒ iΒ accessible].\bar{\mathbf{H}} \;=\; \frac{1}{MN} \sum_{\ell, k} \mathcal{H}_{\mathrm{DD}}[\ell, k] \;=\; \sum_{i=1}^{P} a_i \mathbf{a}_r(\theta_i) \mathbf{a}_t(\phi_i)^H \cdot \mathbb{1}[\text{path } i \text{ accessible}]. Let HΛ‰=UHΞ£HVHH\bar{\mathbf{H}} = \mathbf{U}_H \boldsymbol{\Sigma}_H \mathbf{V}_H^H be its SVD. The SVD precoder is Vβ€…β€Š=β€…β€ŠVH[β‹…,1:Ns],Uβ€…β€Š=β€…β€ŠUH[β‹…,1:Ns].\mathbf{V} \;=\; \mathbf{V}_H[\cdot, 1:N_s], \qquad \mathbf{U} \;=\; \mathbf{U}_H[\cdot, 1:N_s]. This achieves capacity of the averaged channel, within a per-stream loss bounded by the Doppler spread.

The averaged channel Hˉ\bar{\mathbf{H}} is what the transceiver sees after pulse-shaping and block processing. Its SVD gives parallel sub-channels. The precoder aligns streams to these sub-channels, and the combiner separates them. For low-to- moderate Doppler, this is nearly capacity-optimal. For high Doppler (LEO, V2V), per-DD-cell precoding does better, at the cost of compute.

Proof
Effective channel

After DD-domain transmission and reception, the effective channel matrix is Hˉ\bar{\mathbf{H}} to first order in the Doppler spread.

Capacity-achieving precoder

Standard MIMO result: SVD precoding + MMSE combining achieves βˆ‘ilog⁑(1+Οƒi2P/Οƒw2)\sum_i \log(1 + \sigma_i^2 P/\sigma_w^2) per-stream capacity.

Doppler penalty

The averaging ignores DD-cell-specific channel variations. Per- cell variation ∼νmax⁑T\sim \nu_{\max} T. Loss: O(Ξ½max⁑T)2\mathcal{O}(\nu_{\max} T)^2 in capacity. For mobile scenarios: ∼0.1\sim 0.1-11 dB. β– \blacksquare

,

Key Takeaway

SVD precoding works for moderate Doppler. When Ξ½max⁑T≀0.1\nu_{\max} T \leq 0.1 (up to ~100 km/h at 5G NR mmWave), averaging the channel over DD and applying SVD precoding loses ∼1\sim 1 dB vs capacity. For higher Doppler, use per-DD-cell MMSE or the MP precoder of the next algorithm.

Definition:

MIMO-MP Detector

The MIMO message-passing detector runs on the DD-spatial factor graph: each variable node is a transmitted symbol sj[β„“,k]s_j[\ell, k] for stream jj, DD cell (β„“,k)(\ell, k). Each observation node is a received vector y[β„“,k]∈CNr\mathbf{y}[\ell, k] \in \mathbb{C}^{N_r}.

Messages: Gaussian approximation of the a-posteriori marginal for each symbol. Updates:

  • Observation β†’ symbol: interference cancellation using current estimates of other symbols, then soft demodulation per stream.
  • Symbol β†’ observation: updated mean and variance.

Converges in 5-10 iterations. Complexity per iteration: O(MNβ‹…Pβ‹…Nsβ‹…Nr)\mathcal{O}(MN \cdot P \cdot N_s \cdot N_r) β€” linear in DD frame size, linear in paths, linear in streams.

,

MIMO-OTFS Message-Passing Detector

Input: DD observations y[β„“,k] ∈ β„‚^{N_r}, channel tensor Δ€_DD,
constellation Ξ©, N_iter
Output: Detected symbols ŝ[β„“,k] ∈ Ξ©^{N_s}
1. INITIALIZE per-symbol means and variances:
ΞΌ_j[β„“,k] = 0, σ²_j[β„“,k] = E[|s|Β²]
2. FOR iter = 1, …, N_iter:
a. FOR each observation (β„“, k):
(i) Compute expected interference from other symbols:
I[ℓ,k] = Σ_{(ℓ',k') ≠ (ℓ,k)} H̃[ℓ-ℓ',k-k'] · μ[ℓ',k']
(ii) Residual:
r[β„“,k] = y[β„“,k] - I[β„“,k]
(iii) Soft demodulate each stream j in β„‚^{N_s}:
ΞΌ_j[β„“,k] = argmax_Ο‰ Pr(r[β„“,k] | s_j = Ο‰, Δ€_DD)
σ²_j[β„“,k] = variance from a-posteriori distribution
b. Damping: mix new and old estimates to stabilize.
3. Return hard decisions: ŝ_j[β„“,k] = argmax_Ο‰ ΞΌ_j[β„“,k] (hard)
Complexity per iteration: O(MN Β· P Β· N_s Β· N_r). Iterations: 5-10.
Memory: O(MN Β· N_s Β· N_r) for message state.
Convergence: Gaussian damping ensures stability for realistic P.

Theorem: MIMO-MP Convergence

The MIMO-MP detector converges to the minimum-MSE estimate E[s∣y,HDD]\mathbb{E}[\mathbf{s} | \mathbf{y}, \mathcal{H}_{\mathrm{DD}}] within Ο΅\epsilon tolerance in Niterβ€…β€Šβ‰€β€…β€Šclog⁑ ⁣(1Ο΅)N_{\text{iter}} \;\leq\; c \log\!\left(\frac{1}{\epsilon}\right) iterations, under mild regularity conditions (Gaussian noise, sparse channel, small inter-symbol correlation).

Practical performance: at operational SNR (15-25 dB), 5-8 iterations suffice for 10βˆ’510^{-5} BER β€” well within the compute budget of 5G NR physical-layer chips.

Message passing on sparse graphs converges in a bounded number of iterations. The DD-domain sparsity (few paths) keeps the factor graph tree-like. Gaussian approximation of messages keeps each update tractable. The convergence guarantee is the same kind as for LDPC decoding, but on a different graph.

Proof
Gaussian BP

Under Gaussian noise + sparse channel, Gaussian BP converges to the correct marginals. Classical BP result.

Sparse graph

Factor graph has PP connections per observation (not MNMN). Sparsity ensures rapid convergence.

Damping

Damping factor α∈(0,1)\alpha \in (0, 1) mixes old and new estimates: ΞΌ(t+1)=Ξ±ΞΌ(t)+(1βˆ’Ξ±)ΞΌ~\mu^{(t+1)} = \alpha \mu^{(t)} + (1-\alpha) \tilde\mu. Prevents oscillation.

Iteration count

Log-convergence is standard for contractive iterations. β– \blacksquare

,

Example: 4Γ—4 MIMO-OTFS Detection Example

A MIMO-OTFS system with Nt=Nr=4N_t = N_r = 4, M=64M = 64, N=16N = 16, QPSK modulation, P=4P = 4 paths. Compare BER at SNR = 15 dB for: (a) Per-cell ZF. (b) SVD precoder + MMSE combiner. (c) MIMO-MP detector.

Solution
Per-cell ZF

Treat each DD cell as a flat MIMO channel; apply ZF. Ignores inter-cell interference (DD convolution). BER at 15 dB: ∼10βˆ’2\sim 10^{-2}. Poor β€” ZF amplifies noise when channel is ill-conditioned.

SVD + MMSE

SVD precoder on averaged channel + MMSE combiner. Handles low-Doppler well; averaging penalty for high Doppler. BER at 15 dB: ∼2Γ—10βˆ’3\sim 2 \times 10^{-3}.

MIMO-MP

Joint MP detection handles all inter-cell interference. Full DD + spatial diversity exploited. BER at 15 dB: ∼10βˆ’5\sim 10^{-5}. Best performance.

Tradeoff

MP is ~20x more compute than ZF per frame. For safety-critical applications (V2X, platooning): worth it. For infotainment: ZF or SVD precoder sufficient.

MIMO Detector BER Comparison

Plot BER vs SNR for per-cell ZF, SVD+MMSE, and MIMO-MP detection. Sliders: NtN_t, NsN_s, PP.

Parameters
4
2
4

Definition:

Hybrid Analog-Digital Beamforming

For large mmWave arrays (Ntβ‰₯32N_t \geq 32), fully-digital MIMO precoding is infeasible due to the cost of one RF chain per antenna. Hybrid beamforming decomposes the precoder: Vβ€…β€Š=β€…β€ŠVRF VBB,\mathbf{V} \;=\; \mathbf{V}_{RF}\, \mathbf{V}_{BB}, where VRF∈CNtΓ—NRF\mathbf{V}_{RF} \in \mathbb{C}^{N_t \times N_{RF}} is an analog (phase-shifter) precoder, VBB∈CNRFΓ—Ns\mathbf{V}_{BB} \in \mathbb{C}^{N_{RF} \times N_s} is a digital baseband precoder, and NRFβ‰₯NsN_{RF} \geq N_s is the number of RF chains (much smaller than NtN_t).

Constraint: ∣[VRF]n,r∣=1/Nt|[\mathbf{V}_{RF}]_{n, r}| = 1/\sqrt{N_t} (constant envelope for phase shifters).

Hybrid beamforming maintains most of the fully-digital capacity with ∼5Γ—\sim 5\times less hardware cost. Widely deployed in 5G mmWave.

,

Theorem: Hybrid Precoder for Sparse MIMO-OTFS

For a MIMO-OTFS channel with PP dominant paths, the optimal fully-digital precoder Vopt\mathbf{V}_{\text{opt}} lies in the span of the transmit array steering vectors {at(Ο•i)}i=1P\{\mathbf{a}_t(\phi_i)\}_{i=1}^P. Therefore, a hybrid decomposition with NRF=PN_{RF} = P RF chains achieves the fully-digital capacity: VRF=[at(Ο•1),…,at(Ο•P)]/Nt,VBB=adjustedΒ fromΒ SVD.\mathbf{V}_{RF} = [\mathbf{a}_t(\phi_1), \ldots, \mathbf{a}_t(\phi_P)] / \sqrt{N_t}, \qquad \mathbf{V}_{BB} = \text{adjusted from SVD}. Consequence. A 64-antenna mmWave BS serving Ns=4N_s = 4 streams over a P=6P = 6-path channel needs only NRF=6N_{RF} = 6 RF chains β€” not 64. Massive hardware savings.

The sparse-path channel lives in a low-dimensional angular subspace β€” the span of its steering vectors. The optimal precoder also lives in that subspace. The analog stage (phase shifters) implements the steering vectors for free; the digital stage handles the fine weighting. This sparsity-matched architecture is the hardware equivalent of the algorithmic sparsity exploitation.

Proof
Subspace argument

The channel Hˉ\bar{\mathbf{H}} has column space in span{at(ϕi)}\mathrm{span}\{\mathbf{a}_t(\phi_i)\}. Its right singular vectors also lie in this subspace.

Sufficient RF chains

Any vector in a PP-dimensional subspace is expressible as a PP-term combination. Hence NRF=PN_{RF} = P RF chains suffice.

Phase-shifter constraint

Each at(Ο•i)\mathbf{a}_t(\phi_i) has constant-modulus entries (ULA structure), so the phase-shifter constraint is automatically satisfied. β– \blacksquare

,
πŸŽ“CommIT Contribution(2018)

Iterative Interference Cancellation for OTFS Detection

P. Raviteja, K. T. Phan, Y. Hong, E. Viterbo, G. Caire β€” IEEE Trans. Wireless Communications

The Raviteja-Phan-Hong-Viterbo MP framework (with Caire's contributions on the sparse-channel analysis) provides the operational backbone of MIMO-OTFS detection. Three key results:

  1. Factor graph formulation for DD-channel + MIMO detection.
  2. Gaussian message-passing achieving near-MMSE performance with O(MNβ‹…Pβ‹…Nr)\mathcal{O}(MN \cdot P \cdot N_r) complexity per iteration.
  3. Convergence analysis under realistic Doppler and interference levels.

This is the standard reference for MIMO-OTFS detection. It extends the classical MP framework (turbo decoders, LDPC) to the DD- spatial factor graph structure, leveraging sparsity for tractable complexity. The algorithm is implemented in every serious MIMO-OTFS prototype.

commitmimo-otfsmp-detection

Common Mistake: MP May Diverge on Dense Channels

Mistake:

Running MP without damping on a dense channel. The Gaussian BP updates can oscillate or diverge when inter-symbol correlation is strong β€” typical of heavy multipath with many equally-strong paths.

Correction:

Always use damping: ΞΌ(t+1)=Ξ±ΞΌ(t)+(1βˆ’Ξ±)ΞΌ~(t+1)\mu^{(t+1)} = \alpha \mu^{(t)} + (1-\alpha) \tilde\mu^{(t+1)} with α∈[0.5,0.9]\alpha \in [0.5, 0.9]. Damping slows convergence but prevents oscillation. For channels with P>20P > 20 or high inter-path correlation, use Ξ±=0.9\alpha = 0.9. For sparse channels, Ξ±=0.5\alpha = 0.5 suffices. Check convergence: monitor message-difference norm across iterations; if it plateaus without dropping, increase Ξ±\alpha.

Beamspace-DD RepresentationSpatial Multiplexing and Diversity