Ferkans — Interactive Telecom Tutor

Three Dimensions, One Tensor

A SISO OTFS channel is two-dimensional: delay $\tau$ and Doppler $\nu$ . With antenna arrays at the two ends, there is a third dimension — angle — at both transmitter and receiver. The channel is therefore a tensor. But unlike the nominally enormous time- varying MIMO matrix, this tensor is structurally sparse: each physical path contributes exactly one point in the $(\tau, \nu, \theta, \phi)$ cube. The point of this section is to make the tensor explicit and to harvest the sparsity for later algorithms.

,

Definition:
MIMO-OTFS Channel Tensor

For a MIMO system with $N_t$ transmit and $N_r$ receive antennas, served by a DD grid of $M$ delay bins and $N$ Doppler bins, the MIMO-OTFS channel tensor is $\mathcal{H}_{\mathrm{DD}}[\ell, k] \;=\; \sum_{i=1}^{P} a_i\, \mathbf{a}_r(\theta_i)\, \mathbf{a}_t(\phi_i)^H \, \delta[\ell - \ell_i, k - k_i]\, e^{-j 2\pi \nu_i \ell_i/(MN)},$ taking values in $\mathbb{C}^{N_r \times N_t}$ at each DD cell $(\ell, k)$ . Here $\mathbf{a}_t(\phi), \mathbf{a}_r(\theta)$ are transmit and receive array steering vectors, $P$ is the number of resolvable paths, and each path $i$ has complex gain $a_i$ , integer delay $\ell_i$ , integer Doppler $k_i$ , departure angle $\phi_i$ , and arrival angle $\theta_i$ .

The vectorized input-output relation stacks $\mathbf{y}[\ell, k]$ into $\mathbf{y} \in \mathbb{C}^{MN N_r}$ and $\mathbf{x}[\ell, k]$ into $\mathbf{x} \in \mathbb{C}^{MN N_t}$ : $\mathbf{y} \;=\; \mathbf{H}_{\mathrm{DD}}\, \mathbf{x} + \mathbf{w}, \qquad \mathbf{H}_{\mathrm{DD}} \in \mathbb{C}^{MN N_r \times MN N_t}.$ The matrix $\mathbf{H}_{\mathrm{DD}}$ is block-sparse: only $P$ of its $MN N_r N_t$ entries (per DD row) are nonzero.

,

Theorem: Tensor Rank of the MIMO-OTFS Channel

The MIMO-OTFS channel tensor $\mathcal{H}_{\mathrm{DD}}$ has multilinear rank at most $P$ in the spatial-DD decomposition. Specifically, the unfolding along any single mode has rank bounded by $P$ : $\mathrm{rank}\!\left(\text{unfold}_{\mathrm{delay}}(\mathcal{H}_{\mathrm{DD}})\right) \;\leq\; P, \qquad \text{and similarly along Doppler, Tx-angle, Rx-angle.}$ Consequence. The tensor admits a sparse CP (CANDECOMP/PARAFAC) decomposition with $P$ terms, each corresponding to one propagation path. For $P \ll \min(MN, N_t, N_r)$ , this provides dramatic dimensionality reduction over the dense MIMO channel.

The point is that the channel tensor is not merely a large array of numbers — it carries geometric structure from the $P$ scattering paths. Each path is a rank-one tensor: a Kronecker product of array steering vectors in space, delta functions in delay/Doppler, and a scalar gain. The full channel is a sum of $P$ such rank-one terms. Any algorithm that acknowledges this geometry (compressed sensing, tensor methods, path-based estimation) runs in $\mathcal{O}(P)$ complexity rather than $\mathcal{O}(MN N_t N_r)$ .

Proof

CP decomposition

Each term in the path sum is a rank-one tensor: $a_i \mathbf{a}_r(\theta_i) \otimes \mathbf{a}_t(\phi_i)^* \otimes \mathbf{e}_{\ell_i, k_i}$ , where $\mathbf{e}_{\ell, k}$ is the indicator on the $(\ell, k)$ DD cell (modulated by the Doppler phase).

Sum of rank-one

Summing $P$ such terms yields a tensor of CP-rank $\leq P$ . Uniqueness of the decomposition holds under the Kruskal condition, generically satisfied for $P \leq N_t + N_r + MN - 2$ .

Mode unfolding

Unfolding along any mode produces a matrix whose rank is bounded by the CP-rank. Hence each unfolding has rank $\leq P$ . $\blacksquare$

Key Takeaway

The MIMO-OTFS channel is a low-rank tensor. The rank is bounded by the number of resolvable paths $P$ , not by the ambient dimensions $MN N_t N_r$ . This is the single most important fact for the chapter: every efficient MIMO-OTFS algorithm — pilot design, detection, precoding — exploits this low-rank geometry.

Definition:
Array Steering Matrices

For a uniform linear array (ULA) with $N$ elements spaced $d = \lambda/2$ apart, the steering vector at angle $\theta$ is $\mathbf{a}(\theta) \;=\; \begin{pmatrix} 1 \\ e^{j\pi \sin\theta} \\ e^{j 2\pi \sin\theta} \\ \vdots \\ e^{j(N-1)\pi \sin\theta} \end{pmatrix} \in \mathbb{C}^{N}.$ The steering matrix stacks $P$ such vectors: $\mathbf{A}_r \;=\; [\mathbf{a}_r(\theta_1), \ldots, \mathbf{a}_r(\theta_P)] \in \mathbb{C}^{N_r \times P}, \quad \mathbf{A}_t \;=\; [\mathbf{a}_t(\phi_1), \ldots, \mathbf{a}_t(\phi_P)] \in \mathbb{C}^{N_t \times P}.$ For a uniform planar array (UPA), the steering vector is the Kronecker product of two ULA steering vectors (azimuth $\times$ elevation).

,

Example: Parameter Count for a 5G Base Station

A 5G mmWave BS has $N_t = 64$ antennas in an $8 \times 8$ UPA. UEs have $N_r = 4$ antennas. An OTFS frame uses $M = 512$ , $N = 16$ , and typical channels have $P = 6$ dominant paths.

(a) Count the degrees of freedom in the dense time-varying MIMO channel. (b) Count the DD-angle parameters. (c) State the sample-complexity implication for pilot-aided estimation.

Solution

Dense MIMO channel

Full time-varying MIMO matrix: $MN \cdot N_r \cdot N_t \cdot 2 = 512 \cdot 16 \cdot 4 \cdot 64 \cdot 2 = 4.2 \times 10^6$ real degrees of freedom.

DD-angle parameters

$P$ paths $\times (\tau, \nu, \phi, \theta, |a|, \arg a) = 6P = 36$ real parameters.

Sample complexity

Pilot-aided estimation requires at least as many equations as unknowns. Dense channel: $\mathcal{O}(MN N_t) \approx 5 \times 10^5$ pilot symbols — infeasible. DD-angle sparse estimation: $\mathcal{O}(P) = \mathcal{O}(10)$ pilot symbols — entirely feasible. The $10^5$ reduction is the operational value of the sparsity.

Tensor Sparsity vs Ambient Dimensions

Plot the ratio $\frac{7P}{2 M N N_r N_t}$ as the scene parameters vary. Shows the compression factor provided by the DD-angle representation.

Parameters

P

6

N_t = N_r

16

\log_2(MN)

12

Definition:
Doubly-Selective MIMO Channel

A doubly-selective MIMO channel is time-varying (Doppler- dispersive) and frequency-selective (delay-dispersive) and spatially structured (angle-dispersive). Its time-domain impulse response is $\mathbf{H}(t, \tau) \;=\; \sum_{i=1}^{P} a_i \, \mathbf{a}_r(\theta_i) \mathbf{a}_t(\phi_i)^H \, \delta(\tau - \tau_i) \, e^{j 2\pi \nu_i t}.$ Its spreading function in the DD domain is the MIMO tensor $\mathcal{H}_{\mathrm{DD}}(\tau, \nu)$ , related to $\mathbf{H}(t, \tau)$ by the symplectic Fourier transform (Chapter 3).

Bello's four functions (Telecom Ch. 6) extend directly to the MIMO case: the time-variant transfer function, the input delay-spread function, the output Doppler-spread function, and the delay-Doppler spread function — now each matrix-valued.

From Bello to DD

Bello's 1963 characterization of time-varying linear channels (Telecom Ch. 6) gave us the four system functions: time-variant impulse response, time-variant transfer function, Doppler-variant impulse response, and delay-Doppler spreading function. The last of these is the object we work with throughout OTFS. The MIMO extension adds a Kronecker outer product with the spatial channel — the same algebra, one extra dimension.

🔧Engineering Note

Common mmWave Array Geometries (2024)

Practical MIMO-OTFS deployments use a handful of canonical array configurations:

ULA at BS, single antenna at UE: legacy 5G Rel. 15/16. Good for early MIMO-OTFS, limited to azimuth beams.
UPA at BS, ULA at UE: mature 5G mmWave (Rel. 17+). $8\times 8$ or $16\times 16$ BS, $1 \times 2$ or $1 \times 4$ UE.
UPA at BS, UPA at UE: 6G target. $32 \times 32$ BS, $2 \times 2$ UE. Full 3D beamforming on both sides.
Distributed arrays (cell-free): $K$ APs, each with few antennas, jointly serving UEs. Chapter 17 develops this.

Across all geometries, the MIMO-OTFS channel tensor structure holds — the number of parameters is still $\mathcal{O}(P)$ regardless of the number of antennas. The antenna count affects the angular resolution, not the channel complexity.

Practical Constraints

•
ULA: azimuth only; suitable for early deployments
•
UPA-ULA: current 5G mmWave
•
UPA-UPA: 6G target (2028+)
•
Distributed: cell-free (Ch 17)

Common Mistake: Don't Assume Sparsity Without Checking

Mistake:

Assuming the MIMO-OTFS channel is sparse in all scenarios. In dense urban microcell with heavy diffuse scattering (late-arriving echoes, non-specular reflections), the effective path count can reach $P \sim 100$ . At that scale, "sparse" algorithms lose their complexity advantage.

Correction:

Distinguish geometric paths (specular reflections from buildings, vehicles) from diffuse scattering. The geometric paths are $P \sim 5$ - $20$ for most scenarios; diffuse scattering appears as a residual noise floor. MIMO-OTFS algorithms typically estimate the $K$ strongest paths ( $K = 10$ - $20$ ) and absorb residual diffuse energy into the noise term. This gives tractable complexity without sacrificing the dominant channel structure. Check 3GPP TR 38.901 channel models for realistic $P$ values in your deployment scenario.

Why This Matters: From Tensor to Network

This section set up the MIMO-OTFS tensor as a local property of a single transceiver pair. Chapter 17 extends it to a network of transceivers — cell-free massive MIMO with distributed APs, each seeing a different angular slice of the same scatterers. The combined tensor has even richer structure: the angular diversity across APs reveals paths that would be invisible from any single AP. This is the spatial analog of macro-diversity, implemented at the DD-angle level.

The 3D MIMO-OTFS Channel Tensor