Ferkans — Interactive Telecom Tutor

Beamspace: The Fourth Sparsifying Transform

OTFS sparsifies the channel in two dimensions (delay and Doppler) by mapping time-frequency to delay-Doppler via the symplectic Fourier transform. The MIMO dimension adds two more: angle of arrival and angle of departure. At large arrays, the channel is also sparse in angle — each path illuminates a narrow beam direction. The beamspace transform is the DFT-like mapping from antenna indices to beam indices that exposes this angular sparsity. Combined with the DD transform, it yields a 4D sparse representation where a channel with $P$ paths is literally a sum of $P$ point masses.

,

Definition:
Beamspace Transform

For an $N$ -element ULA, the beamspace matrix is the DFT matrix $\mathbf{F}_{BS} \in \mathbb{C}^{N \times N}$ with entries $[\mathbf{F}_{BS}]_{n, b} \;=\; \frac{1}{\sqrt{N}} e^{j 2\pi n b / N}, \qquad n, b \in \{0, \ldots, N-1\}.$ The beamspace channel is obtained by applying $\mathbf{F}_{BS}$ at both ends of the MIMO channel: $\mathcal{H}^{\mathrm{BS}}_{\mathrm{DD}}[\ell, k] \;=\; \mathbf{F}_{BS}^H\, \mathcal{H}_{\mathrm{DD}}[\ell, k]\, \mathbf{F}_{BS}.$

Physical interpretation: the $b$ -th beamspace index corresponds to the direction $\sin\theta_b = 2b/N - 1 \in [-1, 1]$ . The $(\ell, k, b_r, b_t)$ -th entry of the 4D beamspace-DD tensor is nonzero only when a physical path exists at delay $\ell/M$ , Doppler $k/N$ , arrival $b_r$ , and departure $b_t$ .

Theorem: Beamspace-DD 4D Sparsity

For a MIMO-OTFS channel with $P$ resolvable paths at angles $(\theta_i, \phi_i)$ aligned to the beamspace grid, the beamspace-DD tensor $\mathcal{H}^{\mathrm{BS}}_{\mathrm{DD}}$ has at most $P$ nonzero entries out of $MN \cdot N_r \cdot N_t$ total entries. The sparsity ratio is $\frac{P}{MN N_r N_t}.$ For $P = 10$ , $MN = 1024$ , $N_t = N_r = 64$ : sparsity ratio $\approx 2 \times 10^{-6}$ . The beamspace-DD tensor is ultra-sparse.

For off-grid angles, each path spreads over $\sim 3$ - $5$ adjacent beamspace bins (leakage), so the effective support grows to $\sim 5P$ — still extremely sparse.

The MIMO-OTFS channel, when viewed in the beamspace-DD domain, is a 4D tensor that is mostly zero. All the energy concentrates at a handful of grid points corresponding to the actual propagation paths. This is the cleanest mathematical statement of the "every path is a point" intuition, extended to include angle. Compressed sensing algorithms exploit this structure directly: the number of measurements needed scales with the number of paths, not with the ambient tensor size.

Proof

Per-path support

A path at angle $\theta_i$ aligned to beamspace yields $\mathbf{F}_{BS}^H \mathbf{a}_r(\theta_i) = \mathbf{e}_{b_r}$ (a standard basis vector). Similarly for departure. So the path's beamspace contribution is a point mass at $(\ell_i, k_i, b_{r,i}, b_{t,i})$ .

Aggregate support

Summing $P$ paths: at most $P$ nonzero entries.

Off-grid leakage

For arbitrary angle $\theta$ , $\mathbf{F}_{BS}^H \mathbf{a}_r(\theta)$ is approximately a Dirichlet kernel peaked at the nearest beamspace index. The 3-dB support is $\sim 3$ bins. $\blacksquare$

,

Key Takeaway

Beamspace-DD gives a 4D compressed-sensing-friendly channel. For a ULA at both ends with $N_t = N_r = N_a$ , the beamspace-DD tensor has $N_a^2 \cdot MN$ entries but only $P$ are nonzero. This is the foundation for compressed-sensing channel estimation (OMP, AMP, ANM) that needs $\mathcal{O}(P \log(MN N_a^2))$ pilot measurements.

Definition:
Virtual Channel Representation

The virtual representation is the beamspace-DD tensor $\mathcal{H}^{\mathrm{BS}}_{\mathrm{DD}}$ interpreted as a discrete approximation of the continuous spreading function. Each nonzero entry $\mathcal{H}^{\mathrm{BS}}_{\mathrm{DD}}[\ell, k, b_r, b_t]$ represents the channel gain for the virtual path: $(\tau, \nu, \theta, \phi) \;\approx\; \left(\frac{\ell}{W},\, \frac{k}{T},\, \sin^{-1}\!\frac{2 b_r}{N_r} - \frac{\pi}{2},\, \sin^{-1}\!\frac{2 b_t}{N_t} - \frac{\pi}{2}\right).$ This quantized grid defines the resolution of the virtual representation. Fine enough grids capture the channel accurately; coarse grids produce leakage (§5 discusses the tradeoff).

Example: Beamspace Sparsity for an Urban MIMO Channel

Consider a 28-GHz urban microcell: BS with $N_t = 64$ UPA, UE with $N_r = 4$ ULA. OTFS frame $M = 256$ , $N = 16$ . 3GPP Urban Micro channel: $P = 12$ paths spanning $360°$ azimuth, $60°$ elevation.

(a) Count beamspace-DD tensor entries. (b) Estimate nonzero support. (c) Compute sparsity ratio.

Solution

Total entries

$MN \cdot N_r \cdot N_t = 256 \cdot 16 \cdot 4 \cdot 64 = 1.05 \times 10^6$ .

Nonzero support

$P = 12$ aligned paths give 12 nonzero entries. Off-grid leakage spreads each to $\sim 3$ - $5$ bins. Total support: $\sim 60$ nonzero entries.

Sparsity ratio

$60 / 1.05 \times 10^6 \approx 6 \times 10^{-5}$ . The beamspace-DD tensor is 99.994% zero.

Implication for estimation

OMP-based channel estimation recovers this tensor from $\sim K \log(N) = 60 \log(10^6) \approx 1200$ pilot measurements. Classical estimation would need $\sim MN N_t = 10^6$ . A 1000× reduction.

Beamspace-DD 4D Channel Visualization

Visualize the beamspace-DD tensor as a 2D slice: Doppler-angle (fixed delay) or delay-angle (fixed Doppler). Sliders: number of paths, channel power profile, beamspace grid size.

Parameters

P

8

N_t

32

2D slice

Theorem: Compressed-Sensing Channel Estimation

Let $\mathcal{H}^{\mathrm{BS}}_{\mathrm{DD}}$ be $P$ -sparse in the beamspace-DD tensor. Then compressed-sensing recovery (e.g., OMP, ANM, AMP) with $m$ random pilot measurements achieves exact recovery with high probability provided $m \;\geq\; c P \log\!\left(\frac{MN \cdot N_r \cdot N_t}{P}\right)$ for a universal constant $c$ . For practical OTFS deployments ( $P = 10$ , $MN N_r N_t = 10^6$ ): $m \approx 10 \cdot 14 = 140$ pilots suffice for exact recovery — a 4 orders of magnitude reduction from dense estimation.

Compressed sensing formalizes the intuition that "sparse signals need few measurements". The MIMO-OTFS channel has at most $P$ degrees of freedom; random sampling at rate $\mathcal{O}(P \log N)$ is sufficient to recover all of them. The practical impact is enormous: a 5G mmWave UE with dense-estimation pilot overhead of $10\%$ drops to $< 0.1\%$ with compressed-sensing estimation.

Proof

Restricted Isometry Property

Random pilot matrix $\boldsymbol{\Phi}$ satisfies RIP of order $2P$ with high probability when $m \geq c_1 P \log(N/P)$ .

Exact recovery

Under RIP, $\ell_1$ -minimization recovers the true sparse channel exactly. OMP achieves similar guarantees in practice.

Pilot design

Random (Bernoulli or Gaussian) pilots satisfy RIP with high probability. Structured pilots (Zadoff-Chu sequences) offer similar performance at the cost of tighter constraints.

Computational complexity

OMP runs in $\mathcal{O}(P m N \log N)$ . For realistic numbers, $\sim 10$ ms on commercial SoC. $\blacksquare$

,

OMP Beamspace-DD Channel Estimation

Input: Pilot measurement matrix Φ ∈ ℂ^{m × N_total}

Pilot observations y ∈ ℂ^m

Sparsity P, tolerance ε

Output: Sparse channel estimate ĥ ∈ ℂ^{N_total}

1. INITIALIZE:

r = y (residual)

S = ∅ (support set)

2. FOR k = 1, 2, …, P:

a. Correlate: c = Φ^H r

b. Select: i* = argmax_i |c_i|

c. Update support: S = S ∪ {i*}

d. Least-squares: ĥ|_S = (Φ_S)^+ y

e. Update residual: r = y - Φ_S ĥ|_S

f. If ||r|| < ε: BREAK

3. Return ĥ (zero outside S).

Complexity: O(P · m · N_total) per iteration, P iterations total.

For m = 200, N_total = 10⁶, P = 10: 2 × 10⁹ ops, ~10 ms on GPU.

🔧Engineering Note

Compressed Sensing in 5G/6G

Compressed-sensing channel estimation is a research mainstay but has slow standardization:

5G NR (Rel. 15-17): Uses structured pilots (DMRS, SRS) with classical LS estimation. Compressed sensing is vendor-proprietary.
5G NR (Rel. 18-19): Introduces hints of sparse estimation via "beam management optimization" — not standardized, but vendor implementations use OMP/AMP under the hood.
6G (expected 2028+): Explicit support for compressed-sensing channel estimation. Standardized pilot designs for OTFS beamspace-DD recovery.

Practical barrier: compressed-sensing requires more compute per pilot estimation cycle ( $\sim 10\times$ classical LS). Modern mmWave SoCs handle this, but legacy sub-6 GHz chips do not. Migration path: low-band remains classical; mmWave and future 6G adopts compressed sensing.

Practical Constraints

•
5G NR: classical LS estimation (vendor-optional CS)
•
6G: standardized compressed-sensing for OTFS
•
Compute ~10× classical, acceptable on modern mmWave SoC

Common Mistake: Beamspace Grid Mismatch

Mistake:

Assuming physical angles align to the beamspace DFT grid. In practice, scatterers are at arbitrary angles; a path at $\theta = 32.7°$ does not land at any integer beamspace index.

Correction:

Use atomic-norm minimization (ANM) or grid-free recovery algorithms that estimate off-grid angles via Newton refinement. Alternatively, over-sample the beamspace grid by $2\times$ - $4\times$ (refined DFT) to reduce leakage. Trade-off: finer grid → lower leakage bias but higher compute cost. Typical: $2\times$ oversampling + OMP-Newton hybrid.

Beamspace-DD Representation