Chapter Summary

Chapter Summary

Key Points

• 1.

  Real RIS hardware is discrete. Unit cells with $B$ control bits realize $L = 2^B$ phase levels; the feasible set is the finite discrete torus $\mathcal{M}_B$. The continuous-phase problem from Chapters 5–7 is a relaxation; the discrete problem is combinatorial and NP-hard (in fact MAX-CUT-hard at $B = 1$).

• 2.

  Projection-from-continuous is the simplest, and usually best, approach. Solve the continuous-phase problem with AO+WMMSE, project each $\phi_n^\star$ to the nearest grid point, optionally re-optimize the precoder. Adds $< 1\%$ compute overhead to the continuous pipeline. Single-user loss: $\text{sinc}^2(\pi/2^B)$, matching the hardware quantization-only loss — no extra algorithmic penalty.

• 3.

  Direct discrete BCD recovers a small edge in multi-user. Update each $\phi_n$ to the best discrete level at each sweep, tracking quantized coordinates through the iteration. Beats projection by $0.1$-$0.5\text{ dB}$ in multi-user scenarios ($K \geq 4$). For single-user, equivalent to projection. Branch-and-bound gives provable global optima but is feasible only for $N \leq 20$.

• 4.

  Sum-rate loss scales linearly with $K$; max-min loss stays per-user. For $B = 3$-bit and typical channels: negligible per-user penalty (~$0.2\text{ dB}$), $K$-fold for sum-rate aggregate. At $B = 1$-bit the penalty is $\sim 4\text{ dB}$ per user — a substantial hit worth avoiding unless hardware cost is decisive.

• 5.

  Industry sweet spot: $B = 3$-bit, $N$ as large as the panel budget allows. Production mmWave RIS panels in 2024 use 3-bit resolution with $N = 256$-$1024$. Continuous phase is reserved for research and ISAC/sensing applications where the extra precision pays back. Going below $B = 2$ rarely makes sense except for very cost-constrained IoT scenarios.