Angular-Domain Representation

From Space Domain to Angle Domain

The spatial covariance matrix Rt\mathbf{R}_t captures channel statistics, but working with it directly obscures physical intuition. A ULA's channel is fundamentally a superposition of plane waves arriving at discrete angles. The DFT matrix provides a natural orthonormal basis that transforms the spatial domain (antenna index) to the angular domain (spatial frequency / angle bin).

In the angular domain, the channel matrix H~\tilde{\mathbf{H}} has a sparse structure: most of its energy concentrates in a few angle bins corresponding to the physical scattering clusters. This sparsity is the key insight exploited by compressed sensing channel estimation, JSDM precoding, and hybrid beamforming design.

Definition:

ULA Steering Vector and Spatial Frequency

For a uniform linear array with NN elements and inter-element spacing dd, the steering vector at physical angle θ\theta (measured from broadside) is

a(θ)≜1N[1, ej2πdλsin⁡θ, …, ej(N−1)2πdλsin⁡θ]T∈CN.\mathbf{a}(\theta) \triangleq \frac{1}{\sqrt{N}} \left[1,\, e^{j\frac{2\pi d}{\lambda}\sin\theta},\, \ldots,\, e^{j(N-1)\frac{2\pi d}{\lambda}\sin\theta}\right]^T \in \mathbb{C}^N.

With half-wavelength spacing d=λ/2d = \lambda/2, the spatial frequency ψ≜dλsin⁡θ=12sin⁡θ∈[−1/2,1/2]\psi \triangleq \frac{d}{\lambda}\sin\theta = \frac{1}{2}\sin\theta \in [-1/2, 1/2], and a(ψ)=1N[1,ej2πψ,…,ej2π(N−1)ψ]T\mathbf{a}(\psi) = \frac{1}{\sqrt{N}}[1, e^{j2\pi\psi}, \ldots, e^{j2\pi(N-1)\psi}]^T.

The steering vector is the DFT basis vector at frequency ψ\psi. The mapping θ↦ψ=12sin⁡θ\theta \mapsto \psi = \frac{1}{2}\sin\theta is nonlinear: angle resolution is best at broadside (θ=0\theta = 0, ∂ψ/∂θ=1/2\partial\psi/\partial\theta = 1/2) and worst at endfire (θ=±90°\theta = \pm 90°, ∂ψ/∂θ→0\partial\psi/\partial\theta \to 0).

Definition:

Virtual Channel Representation (Angular Domain)

Let Ut∈CNt×Nt\mathbf{U}_t \in \mathbb{C}^{N_t \times N_t} and Ur∈CNr×Nr\mathbf{U}_r \in \mathbb{C}^{N_r \times N_r} be the DFT matrices:

[UN]mn=1Ne−j2πmn/N,m,n=0,…,N−1.[\mathbf{U}_N]_{mn} = \frac{1}{\sqrt{N}} e^{-j2\pi mn/N}, \quad m,n = 0, \ldots, N-1.

The virtual (angular-domain) channel matrix is

H~≜UrH H Ut∈CNr×Nt.\tilde{\mathbf{H}} \triangleq \mathbf{U}_r^H \, \mathbf{H} \, \mathbf{U}_t \in \mathbb{C}^{N_r \times N_t}.

The (p,q)(p,q) entry [H~]pq[\tilde{\mathbf{H}}]_{pq} represents the channel gain between transmit angle bin qq (spatial frequency ψqt=q/Nt\psi_q^t = q/N_t) and receive angle bin pp (spatial frequency ψpr=p/Nr\psi_p^r = p/N_r).

Since Ut\mathbf{U}_t and Ur\mathbf{U}_r are unitary, the transformation is information-preserving: ∥H~∥F=∥H∥F\|\tilde{\mathbf{H}}\|_F = \|\mathbf{H}\|_F. The capacity of the MIMO channel is unchanged.

Theorem: Sparsity of the Virtual Channel

Consider a physical channel with LL propagation paths: H=∑ℓ=1Lαℓ a^(ϕℓ)a(ψℓ)H,\mathbf{H} = \sum_{\ell=1}^L \alpha_\ell \, \hat{\mathbf{a}}(\phi_\ell) \mathbf{a}(\psi_\ell)^H, where αℓ∈C\alpha_\ell \in \mathbb{C} is the path gain, ϕℓ\phi_\ell is the angle of arrival, and ψℓ\psi_\ell is the angle of departure.

The virtual channel entry [H~]pq[\tilde{\mathbf{H}}]_{pq} is significant only when (p/Nr,q/Nt)≈(12sin⁡ϕℓ,12sin⁡ψℓ)(p/N_r, q/N_t) \approx (\frac{1}{2}\sin\phi_\ell, \frac{1}{2}\sin\psi_\ell) for some path ℓ\ell. For L≪min⁡(Nt,Nr)L \ll \min(N_t, N_r) paths, H~\tilde{\mathbf{H}} is approximately LL-sparse: at most O(L)O(L) entries have significant magnitude.

Each physical path concentrates energy in a specific (receive angle bin, transmit angle bin) pair in the virtual domain. With few paths (sparse scattering at mmWave), the virtual channel looks like a nearly-zero matrix with LL bright spots.

Show Hint

Substitute the path model into H~=UrHHUt\tilde{\mathbf{H}} = \mathbf{U}_r^H \mathbf{H} \mathbf{U}_t and recognize UrHa^(ϕ)\mathbf{U}_r^H \hat{\mathbf{a}}(\phi) as a DFT coefficient evaluation at frequency 12sin⁡ϕ\frac{1}{2}\sin\phi.

The result UNHa(ψ)=ek\mathbf{U}_N^H \mathbf{a}(\psi) = \mathbf{e}_k (a canonical basis vector) holds exactly when ψ=k/N\psi = k/N for integer kk — the "on-grid" case. Off-grid paths produce energy leakage to neighboring bins.

For large NN, the DFT concentrates the inner product ∣upHa^(ϕ)∣2|\mathbf{u}_p^H \hat{\mathbf{a}}(\phi)|^2 near p∗=⌊Nsin⁡ϕ/2⌉p^* = \lfloor N\sin\phi/2 \rceil with width ∼1/N\sim 1/N in spatial frequency.

Proof
Substitute path model

H~=UrH(∑ℓαℓa^(ϕℓ)a(ψℓ)H)Ut=∑ℓαℓ(UrHa^(ϕℓ))⏟fℓr(UtHa(ψℓ))H⏟(fℓt)H.\tilde{\mathbf{H}} = \mathbf{U}_r^H \left(\sum_\ell \alpha_\ell \hat{\mathbf{a}}(\phi_\ell) \mathbf{a}(\psi_\ell)^H\right) \mathbf{U}_t = \sum_\ell \alpha_\ell \underbrace{(\mathbf{U}_r^H \hat{\mathbf{a}}(\phi_\ell))}_{\mathbf{f}_\ell^r} \underbrace{(\mathbf{U}_t^H \mathbf{a}(\psi_\ell))^H}_{{(\mathbf{f}_\ell^t)}^H}.$

Evaluate DFT inner products

The pp-th entry of fℓr=UrHa^(ϕℓ)\mathbf{f}_\ell^r = \mathbf{U}_r^H \hat{\mathbf{a}}(\phi_\ell) is: [fℓr]p=1Nr∑n=0Nr−1ej2πn(p/Nr−sin⁡ϕℓ/2).[\mathbf{f}_\ell^r]_p = \frac{1}{N_r} \sum_{n=0}^{N_r-1} e^{j2\pi n (p/N_r - \sin\phi_\ell/2)}. This is a geometric sum (Dirichlet kernel), maximized when p/Nr=sin⁡ϕℓ/2p/N_r = \sin\phi_\ell/2 and decaying as ∣p/Nr−sin⁡ϕℓ/2∣|p/N_r - \sin\phi_\ell/2| increases.

Sparsity conclusion

For each path ℓ\ell, only the entries near row pℓ∗≈Nrsin⁡ϕℓ/2p_\ell^* \approx N_r\sin\phi_\ell/2 and column qℓ∗≈Ntsin⁡ψℓ/2q_\ell^* \approx N_t\sin\psi_\ell/2 are significant. For LL paths, approximately LL "bright spots" appear in H~\tilde{\mathbf{H}}.

For L≪min⁡(Nt,Nr)L \ll \min(N_t, N_r) (sparse scattering), H~\tilde{\mathbf{H}} is approximately LL-sparse. ■\blacksquare

Historical Note: Sayeed's Virtual Channel Representation

2002

The virtual channel representation was introduced by Akbar M. Sayeed at the University of Wisconsin–Madison in a landmark 2002 paper: "Deconstructing multiantenna fading channels." Prior to this work, MIMO channel analysis relied heavily on the full channel matrix H\mathbf{H} without a natural spatial basis.

Sayeed's key insight was that the DFT provides a canonical basis for ULA channels that simultaneously diagonalizes both the spatial covariance structure and the sparse path representation. This made angular-domain sparsity exploitable for the first time. The framework later became the foundation for beamspace processing, hybrid precoding analysis, and compressed-sensing-based channel estimation in massive MIMO.

Virtual Channel: Angular-Domain Sparsity

Visualizes ∣H~∣2|\tilde{\mathbf{H}}|^2 (normalized power in each angle bin) for a channel with LL propagation paths. Observe how increasing the number of paths fills in more angle bins, and how the virtual channel transitions from sparse (mmWave-like) to dense (rich scattering / i.i.d. Rayleigh limit). The marginal sums along rows and columns correspond to the receive-side and transmit-side angular power spectra.

Parameters
32
16
3
5

Building the Virtual Channel Path by Path

Animates the accumulation of paths in the virtual channel domain. Each frame adds one new path, showing how the energy pattern in ∣H~∣2|\tilde{\mathbf{H}}|^2 evolves from a single bright spot to a sparse set of concentrated clusters.

Parameters
16
5

Sparsity, Compressed Sensing, and Hybrid Beamforming

The sparsity of H~\tilde{\mathbf{H}} in the angular domain has three major design implications:

  1. Compressed channel estimation: Instead of estimating all Nt⋅NrN_t \cdot N_r entries of H\mathbf{H}, we need only estimate O(Llog⁡Nt)O(L \log N_t) angular coefficients. This is the basis for compressed sensing channel estimation — using far fewer pilots than the Nyquist rate would require.

  2. Hybrid beamforming: In a hybrid analog-digital architecture with NRF≪NtN_{\text{RF}} \ll N_t RF chains, the analog beamforming matrix should align with the NRFN_{\text{RF}} dominant columns of Ut\mathbf{U}_t (the DFT columns at the strongest angle bins). This is the "beamspace" approach to hybrid precoding.

  3. Pilot decontamination: Users whose angular supports (non-zero virtual channel regions) do not overlap can share the same pilot sequences without mutual estimation interference — even in multi-cell settings. This is the spatial filtering principle behind Chapter 3's covariance-based pilot decontamination.

,

Definition:

Effective Rank and Angular Resolution

The effective rank of Rt\mathbf{R}_t measures the number of statistically distinguishable spatial directions:

eff-rank(Rt)≜(tr(Rt))2tr(Rt2)=∥λ∥12∥λ∥22,\text{eff-rank}(\mathbf{R}_t) \triangleq \frac{\left(\text{tr}(\mathbf{R}_t)\right)^2}{\text{tr}(\mathbf{R}_t^2)} = \frac{\|\boldsymbol{\lambda}\|_1^2}{\|\boldsymbol{\lambda}\|_2^2},

where λ\boldsymbol{\lambda} are the eigenvalues of Rt\mathbf{R}_t. It satisfies 1≤eff-rank(Rt)≤Nt1 \leq \text{eff-rank}(\mathbf{R}_t) \leq N_t, with equality to NtN_t for i.i.d. Rayleigh (Rt=I\mathbf{R}_t = \mathbf{I}).

The angular resolution of a ULA with NtN_t half-wavelength elements is δψ=1/Nt\delta_\psi = 1/N_t in spatial frequency, corresponding to an angular resolution of δθ≈2/(Ntcos⁡θ0)\delta_\theta \approx 2/(N_t\cos\theta_0) radians at angle θ0\theta_0.

Common Mistake: DFT Grid ≠ Physical Angle Grid

Mistake:

Assuming that the DFT angle bins at ψq=q/N\psi_q = q/N for q=0,…,N−1q = 0, \ldots, N-1 are uniformly spaced in physical angle θ\theta.

Correction:

The DFT bins are uniformly spaced in spatial frequency ψ=sin⁡θ/2∈[−1/2,1/2]\psi = \sin\theta/2 \in [-1/2, 1/2], but nonuniformly spaced in physical angle θ=arcsin⁡(2ψ)\theta = \arcsin(2\psi):

Δθq=arcsin⁡(2(q+1)/N)−arcsin⁡(2q/N)≈2/N1−(2q/N)2.\Delta\theta_q = \arcsin(2(q+1)/N) - \arcsin(2q/N) \approx \frac{2/N}{\sqrt{1 - (2q/N)^2}}.

At broadside (q=0q = 0): Δθ≈2/N\Delta\theta \approx 2/N radians. At endfire (q≈N/2q \approx N/2): Δθ→∞\Delta\theta \to \infty. This angular compression toward endfire means that the DFT provides finer resolution at broadside and coarser resolution near endfire. Algorithms that treat all bins as equal in angle will have biased performance near endfire.

🔧Engineering Note

Extension to 2D Arrays and Azimuth-Elevation Domains

The virtual channel framework extends naturally to 2D planar arrays (UPAs). For an MH×MVM_H \times M_V UPA (horizontal × vertical elements), the 2D DFT UMH⊗UMV\mathbf{U}_{M_H} \otimes \mathbf{U}_{M_V} transforms to the joint azimuth-elevation angle domain. The virtual channel H~\tilde{\mathbf{H}} is now indexed by (azimuth bin, elevation bin) at each end.

In practice, 5G NR base stations use 2D arrays (e.g., 8×4 = 32 elements with dual polarization → 64 ports). The 3GPP channel models (TR 38.901) generate channels in the azimuth-elevation domain and then project to the array response vectors. Full 3D beamforming ("FD-MIMO") exploits both azimuth and elevation dimensions.

Practical Constraints
  • •

    5G NR supports up to 256 antenna ports (3GPP TS 38.211)

  • •

    Type II CSI feedback reports precoder coefficients per sub-band — implicitly captures angular structure

  • •

    2D DFT for MH×MVM_H \times M_V UPA requires F=UMH⊗UMV\mathbf{F} = \mathbf{U}_{M_H} \otimes \mathbf{U}_{M_V}, a separable transform

📋 Ref: 3GPP TS 38.211, Section 7.3.1.4

Key Takeaway

The DFT maps space to angle, sparsity to structure. The virtual channel H~=UrHHUt\tilde{\mathbf{H}} = \mathbf{U}_r^H \mathbf{H} \mathbf{U}_t concentrates energy in O(L)O(L) angle-bin pairs for LL-path channels. This sparsity enables compressed estimation (Chapter 3), structured precoding (Chapter 7), and efficient hybrid beamforming (Chapter 20). For dense scattering environments, the virtual channel fills uniformly and the structure collapses to i.i.d. — but even then, the angular-domain view clarifies what "uniform" really means geometrically.

Quick Check

For a single-path channel H=α a^(ϕ)a(ψ)H\mathbf{H} = \alpha \, \hat{\mathbf{a}}(\phi) \mathbf{a}(\psi)^H with Nt=Nr=NN_t = N_r = N and on-grid angles (Nsin⁡ϕ/2=p∗N\sin\phi/2 = p^*, Nsin⁡ψ/2=q∗N\sin\psi/2 = q^* for integers p∗,q∗p^*, q^*), what does H~\tilde{\mathbf{H}} look like?

Full matrix with all entries equal to ∣α∣2/N2|\alpha|^2/N^2

A matrix with only entry (p∗,q∗)(p^*, q^*) nonzero, equal to α\alpha

A rank-1 matrix with a sinc-shaped envelope around (p∗,q∗)(p^*, q^*)

The identity matrix scaled by α\alpha

Correction:
A matrix with only entry (p∗,q∗)(p^*, q^*) nonzero, equal to α\alpha

For an on-grid path, UNHa(ψ)=eq∗\mathbf{U}_N^H \mathbf{a}(\psi) = \mathbf{e}_{q^*} exactly, so H~=αep∗eq∗H\tilde{\mathbf{H}} = \alpha \mathbf{e}_{p^*} \mathbf{e}_{q^*}^H — a rank-1 matrix with one nonzero entry.

Spatially Correlated Channels3GPP Spatial Channel Models