How Can Hybrid Beamforming Pave the Way for Sub‑THz Communications?

Imagine trying to talk to a friend miles away using only light. That’s the core idea behind sub‑THz communications, where frequencies are so high they behave almost like light beams. To send data at breakneck speeds, however, you need a “brain” that can steer those beams precisely.

The new arXiv paper proposes a hybrid architecture: the antenna is split into “tiles,” each a small phased‑array that forms an analog beam. After the beam, the signals go to a digital signal processor (DSP) that combines and cleans them. This reduces hardware complexity compared to a fully‑digital array, where every antenna would need its own circuit.

Scaling both array size and bandwidth brings two challenges. First, “beam squint”: at different frequencies the beam points elsewhere, like a flashlight changing direction when you switch colors. Second, power consumption, which rises with the number of antennas.

The authors compare several tricks to tame squint, such as widening the beam or sharing tiles among users. They use theoretical models and OFDM simulations to estimate spectral efficiency, pitting a linear per‑subcarrier receiver against an ideal, unconstrained one.

The takeaway? With the right tile size and DSP algorithms, you can keep ultra‑high data rates without blowing up power use, opening the door to ultra‑fast fixed links for homes and industry.

The work sits within the sub‑THz communication regime (300 GHz‑1 THz), where available bandwidth exceeds that of traditional 5G bands by an order of magnitude. In this regime, array element density rises sharply, making a fully‑digital approach impractical: each element would need a high‑speed A/D converter, power amplifiers, and feed lines, leading to power consumption exceeding 10 W per cm² of antenna surface.

The proposed hybrid architecture partitions the array into N_tile tiles, each with M elementary antennas. Every tile performs analog beamforming via phase or attenuation networks, creating a radiation pattern directed toward a spatial sector. The combined tile signal is then sampled at baseband and processed digitally. This reduces the number of RF‑DSP chains from N_total = N_tile·M to N_tile, yielding a linear power saving.

Beam squint is modeled as a frequency‑dependent phase response θ(f) = θ_0 + 2π·d·sin(φ)·(f‑f_c)/c, where d is the inter‑antenna spacing, φ the steering angle, and f_c the center frequency. For bandwidths exceeding 10 % of f_c, pointing deviation can exceed several degrees, degrading inter‑user interference. Mitigation strategies include: (i) beam broadening via non‑uniform amplitude weighting (tapering), (ii) using multiple sub‑arrays with slightly offset pointings (multi‑beam tiling), and (iii) dynamic per‑sub‑carrier phase coefficient adaptation (per‑subcarrier RF beamforming), albeit with more complex hardware.

From a digital processing perspective, the receiver adopts an MIMO‑OFDM model with H_k ∈ ℂ^{U×N_tile} for the k‑th sub‑carrier, where U is the number of users. The theoretical capacity is C = (1/K)∑_{k=1}^{K} log₂ det(I_U + (ρ/σ²) H_k W_k W_k^H H_k^H), with ρ transmit power, σ² noise, and W_k the digital weight matrix. The paper compares this capacity with that of a per‑sub‑carrier linear receiver (ZF or MMSE) and an ideal benchmark (ML) to assess the gap introduced by hardware constraints.

Results show that, for tile sizes of 8×8 or 16×16, an spectral efficiency of 70‑80 % of the ideal limit is achievable, provided beam broadening and tile sharing among users with similar arrival angles are employed. Energy analysis indicates total power can stay below 5 W per 1 m² of antenna surface, making the approach viable for ultra‑high‑bandwidth fixed access scenarios.