Ferkans — Interactive Telecom Tutor

Phases Are Not Continuous

The hybrid architecture of Sections 20.1-20.4 treats phase shifters as continuous-valued devices: $e^{j\phi}$ for any $\phi \in [0, 2\pi)$ . Real phase shifters are digitally controlled and support only a discrete set of phases. Typical hardware at 28 GHz offers $b = 2$ to $b = 5$ bits of resolution, giving $2^b$ phase states equally spaced on the unit circle. This is a hardware reality - not a conservative assumption - driven by the cost, size, and power of high-resolution phase shifters.

The engineering question is: what is the performance penalty of $b$ -bit phase shifters relative to the continuous-phase ideal, and what is the minimum $b$ for a given target spectral efficiency? The answer guides the RFIC design for 5G and 6G mmWave arrays.

Definition:
$b$ -Bit Phase Shifter

A $b$ -bit phase shifter supports $2^b$ discrete phase states uniformly spaced on $[0, 2\pi)$ :

$\phi \in \mathcal{Q}_b = \left\{0, \frac{2\pi}{2^b}, \frac{2 \cdot 2\pi}{2^b}, \ldots, \frac{(2^b - 1) \cdot 2\pi}{2^b}\right\}.$

The quantized analog precoder has entries

$[\mathbf{F}_{\text{RF}}^{(b)}]_{m,n} = \frac{1}{\sqrt{N_t}} e^{j \widehat{\phi}_{m,n}}, \quad \widehat{\phi}_{m,n} = \arg\min_{\phi \in \mathcal{Q}_b} |\phi - \phi_{m,n}|,$

where $\phi_{m,n}$ is the continuous-phase target.

Some hardware uses phase-inverter structures ( $b = 1$ , two states $\{0, \pi\}$ ) or quadrature-hybrid networks ( $b = 2$ , four states). At higher resolution, the dominant contributors are digitally controlled transmission-line or reflection-type phase shifters with a vector modulator. Each extra bit roughly doubles the chip area and power.

Theorem: Spectral Efficiency Loss from Phase Quantization

Let $\mathbf{F}_{\text{RF}}^{\star}$ be the continuous-phase hybrid precoder solution, and let $\mathbf{F}_{\text{RF}}^{(b)}$ be obtained by rounding each phase entry to the nearest $b$ -bit grid. For large $N_t$ and random phases, the expected beamforming gain loss relative to the continuous optimum satisfies

$\eta(b) \triangleq \mathbb{E}\!\left[\frac{|\mathbf{v}^{(b)\,H} \mathbf{a}|^2}{|\mathbf{v}^{\star\,H} \mathbf{a}|^2}\right] = \text{sinc}^2\!\left(\frac{1}{2^b}\right) = \left(\frac{2^b}{\pi} \sin\!\frac{\pi}{2^b}\right)^2.$

The corresponding array-gain loss in dB is $-10\log_{10} \eta(b)$ , yielding approximately:

$b$ (bits)	Loss (dB)	$\\eta(b)$
1	3.92	0.405
2	0.91	0.811
3	0.22	0.950
4	0.056	0.987
5	0.014	0.997

Three-bit phase shifters lose less than a quarter of a dB of array gain

enough justification for the ubiquitous $b = 3$ choice in deployed mmWave hardware.

Rounding an optimal phase to the nearest discrete level introduces a uniformly distributed error in $[-\pi/2^b, \pi/2^b]$ . The resulting complex gain is modulated by $\cos(\Delta\phi)$ ; averaging over the uniform error gives $(2/\pi \cdot 2^{b-1} \sin(\pi/2^b))^2$ , which simplifies to the stated $\text{sinc}^2$ formula.

Show Hint

Start from the beamforming inner product $\sum_m e^{j(\phi_m + \Delta_m)}$ where $\Delta_m$ is the uniform quantization error.

Take expectation over $\Delta_m \sim \text{Uniform}(-\pi/2^b, \pi/2^b)$ : the cross terms vanish by independence, and the diagonal gives $|\mathbb{E}[e^{j\Delta}]|^2$ .

Compute $\mathbb{E}[e^{j\Delta}] = \text{sinc}(1/2^b)$ by direct integration.

Proof

Quantization-error model

Let $\phi_m = \phi_m^{\star} + \Delta_m$ where $\Delta_m \sim \text{Uniform}(-\pi/2^b, \pi/2^b)$ and $\{\Delta_m\}_{m=1}^{N_t}$ are independent (a good approximation for random target phases). The quantized beamformer is $[\mathbf{v}^{(b)}]_m = \frac{1}{\sqrt{N_t}} e^{j(\phi_m^{\star} + \Delta_m)}$ .

Expected inner product

The inner product with the target steering vector is $\mathbf{v}^{(b)\,H} \mathbf{a} = \frac{1}{N_t} \sum_m e^{j\Delta_m} \cdot |[\mathbf{v}^{\star\,H} \mathbf{a}]_m^{\text{effective}}|$ . Taking expectations, the magnitude squared of the expected inner product factors as $|\mathbb{E}[e^{j\Delta}]|^2 \cdot |\mathbf{v}^{\star\,H} \mathbf{a}|^2$ .

Compute $\mathbb{E}[e^{j\Delta}]$

By direct integration: $\mathbb{E}[e^{j\Delta}] = \frac{2^b}{2\pi} \int_{-\pi/2^b}^{\pi/2^b} e^{j\Delta}\,d\Delta = \frac{2^b}{2\pi} \cdot \frac{2\sin(\pi/2^b)}{1} = \frac{2^b}{\pi}\sin\!\frac{\pi}{2^b} = \text{sinc}\!\frac{1}{2^b}.$

Combine

Squaring yields $\eta(b) = \text{sinc}^2(1/2^b)$ . Substituting $b = 1, 2, 3, 4, 5$ produces the table. $\blacksquare$

,

Array-Gain Loss vs. Phase-Shifter Resolution

Plot the spectral efficiency of a hybrid-beamformed mmWave link versus SNR for different phase-shifter resolutions $b \in \{1, 2, 3, 4, \infty\}$ . Also shows the asymptotic quantization loss formula $\eta(b) = \text{sinc}^2(1/2^b)$ .

Parameters

N_t

64

Max SNR (dB)25

Example: Choosing $b$ for a 1 dB Budget

A 28 GHz deployment tolerates at most 1 dB of beamforming-gain loss due to phase quantization. What is the minimum phase-shifter resolution $b$ that satisfies the budget? Give the realized loss.

Solution

Solve $\text{sinc}^2(1/2^b) \geq 10^{-0.1}$

$10^{-0.1} \approx 0.794$ . From the table in Theorem TSpectral Efficiency Loss from Phase Quantization, $b = 2$ gives $\eta(2) = 0.811 \geq 0.794$ . $b = 1$ gives $\eta(1) = 0.405 < 0.794$ .

Verify $b = 2$

$\eta(2) = 0.811 \Rightarrow -10\log_{10}(0.811) = 0.91$ dB $\leq 1$ dB. OK.

Conclusion

$b = 2$ bits (4 phase states) is the minimum resolution. Realized loss is 0.91 dB, using up 91 % of the budget. Going to $b = 3$ bits drops the loss to 0.22 dB, leaving substantial margin at a modest hardware cost. Most production mmWave arrays pick $b = 3$ for this reason. $\blacksquare$

⚠️Engineering Note

Realistic Phase Errors Exceed the Theoretical Floor

Theorem TSpectral Efficiency Loss from Phase Quantization assumes perfect phase control within each quantization level. In reality, additional error sources degrade the array: (1) phase-shifter non-linearity: frequency-dependent insertion phase ripple adds $\pm 3$ - $5^\circ$ of RMS error even in 4-bit devices; (2) element-to-element amplitude variation: $\pm 1$ dB per element from MMIC process variation; (3) calibration drift: phase-shifter settings drift by $\sim 2^\circ$ per 10 K temperature change. The net effect is that a nominal 4-bit (0.06 dB theoretical loss) array in practice shows 1-2 dB of realized loss at the beam peak. Periodic calibration with built-in self-test is therefore a standard feature of deployed mmWave base stations.

Practical Constraints

•
RFIC process tolerance: $\pm 1^\circ$ to $\pm 5^\circ$ phase error
•
Frequency dependence: $\sim 1^\circ$ /GHz phase slope per element
•
Calibration period: every 1-10 minutes depending on temperature swing

Beampattern with Quantized Phases — Beampatterns of a 64-antenna ULA with $b \in \{1, 2, 3, \infty\}$ -bit phase shifters, all steered to broadside. The 1-bit pattern exhibits visible quantization lobes and a reduced peak; the 3-bit pattern is nearly indistinguishable from the continuous-phase ideal.

Historical Note: A Brief History of Digital Phase Shifters

1962-present

Digital phase shifters for radar and satellite phased arrays date to the 1960s. The earliest implementations used ferrite toroids switched between magnetic states (White, 1962), giving typically 3- to 5-bit resolution at L- and S-band. The 1980s saw the transition to PIN-diode phase shifters (smaller, lossier, but faster-switching), still in 3- to 5-bit designs. Modern mmWave phase shifters use vector modulators based on quadrature mixers or variable attenuators, achieving 5- to 6-bit resolution at cellular mmWave frequencies. The "3-bit rule" - that three bits are enough for array-gain purposes - was empirically established in the radar community decades before it appeared in the mmWave 5G literature.

Digital Phase Shifter

An RF component whose insertion phase can be set to one of $2^b$ discrete values selected by a $b$ -bit control word. Common mmWave implementations include switched-line (low loss, large area), loaded-line (moderate loss, narrow band), reflection-type (moderate loss, wide band), and vector modulator (fine resolution, higher power). Typical 5G mmWave devices use $b = 4$ to $b = 6$ bits.

Common Mistake: Quantization Loss Is Not Additive in SNR

Mistake:

Some analyses treat phase quantization as an independent noise term to be added to the SNR denominator, concluding that low-resolution shifters "destroy" high-SNR performance.

Correction:

Phase-quantization loss is multiplicative on the beamforming gain, not additive on the noise. The effective SNR is $\\eta(b) \\cdot \N_t \\cdot \\text{SNR}$ , which shifts the rate-vs-SNR curve by $-10\\log_{10}\\eta(b)$ dB - a constant offset at all SNRs. The curve shape is unchanged. This is why $b = 3$ bits suffice: the 0.22 dB shift is negligible at any operating point.

Quick Check

A deployment budget tolerates 0.3 dB of array-gain loss from phase-shifter quantization. What is the minimum $b$ ?

$b = 1$

$b = 2$

$b = 3$

$b = 4$

Correction:

b = 3

$\eta(3) = \text{sinc}^2(1/8) \approx 0.950$ gives 0.22 dB loss, just under the 0.3 dB budget. This is why 3-bit phase shifters are the practical sweet spot.

Key Takeaway

Phase-shifter quantization introduces a $\text{sinc}^2(1/2^b)$ array-gain penalty, which vanishes rapidly with $b$ : 0.91 dB at $b = 2$ , 0.22 dB at $b = 3$ , and under 0.1 dB for $b \geq 4$ . Three-bit phase shifters are the pragmatic sweet spot and are what deployed 5G mmWave hardware uses. Calibration, process variation, and temperature drift typically dominate the theoretical quantization loss in real devices.

Phase Shifter Quantization

Phases Are Not Continuous

Definition: bbb-Bit Phase Shifter

Theorem: Spectral Efficiency Loss from Phase Quantization

Quantization-error model

Expected inner product

Compute $\mathbb{E}[e^{j\Delta}]$

Combine

Array-Gain Loss vs. Phase-Shifter Resolution

Parameters

Example: Choosing bbb for a 1 dB Budget

Solve $\text{sinc}^2(1/2^b) \geq 10^{-0.1}$

Verify $b = 2$

Conclusion

Realistic Phase Errors Exceed the Theoretical Floor

Beampattern with Quantized Phases

Historical Note: A Brief History of Digital Phase Shifters

Digital Phase Shifter

Common Mistake: Quantization Loss Is Not Additive in SNR

Quick Check

Key Takeaway

Definition:
$b$ -Bit Phase Shifter

Example: Choosing $b$ for a 1 dB Budget