Synchronization: Time, Frequency, and Phase

Synchronization Is the Other Half of Coherence

Reciprocity calibration guarantees that the downlink precoder matches the uplink channel. But even a perfectly calibrated array fails to form a coherent beam if the samples feeding the DACs are not aligned in time, in frequency, and in phase. Synchronization is the second half of the coherence problem: it fixes the axis along which the signal travels through the pipeline, before calibration adjusts the per-antenna amplitudes.

A real testbed has to reconcile three clock disciplines — the slot clock, the carrier oscillator, and the sample clock — across all antennas. In a centralized array a single reference clock serves them all; in a cell-free system each AP has its own oscillator and each must be disciplined against a shared reference. This section walks through the three synchronization layers, derives the BER penalty of residual offsets, and closes with a look at GPS-assisted time distribution in cell-free testbeds.

Definition:

Three Synchronization Layers

A 5G NR link maintains synchronization at three layers:

  1. Frame-level — the BS and UE must agree on which slot is which. Handled by the Primary and Secondary Synchronization Signals (PSS, SSS), which occupy specific OFDM symbols in the SS burst block. Correlating against the PSS/SSS templates yields the frame boundary to within an OFDM symbol.

  2. Symbol-level — within a frame, the UE must align its FFT window with the BS's OFDM symbols. Residual timing offsets of a fraction of the cyclic prefix are tolerable; larger offsets leak inter-symbol interference into the frequency domain.

  3. Frequency/phase-level — the UE's local oscillator must match the BS's carrier to within a small fraction of the subcarrier spacing. Residual carrier frequency offsets produce both a per-symbol phase rotation and, more perniciously, intercarrier interference proportional to εf/Δf\varepsilon_f / \Delta f.

In massive MIMO the cost of an offset at any of these layers is amplified by the array: a misaligned per-antenna phase does not only hurt a single link but tears apart the coherent beam.

Definition:

Primary and Secondary Synchronization Signals (PSS/SSS)

In 5G NR, the PSS is a length-127 sequence modulated from a BPSK mm-sequence, occupying one OFDM symbol at a fixed location in the SS burst. It encodes the cell ID modulo 3. The SSS is another length-127 sequence encoding the cell ID fully. A UE correlates the received signal against the three PSS candidates, detects the coarse frame boundary, and then disambiguates the full cell ID from the SSS. The combined search yields:

  • Coarse frame timing to within one OFDM symbol;
  • A fractional carrier frequency offset estimate from the phase of the PSS auto-correlation peak;
  • The cell ID from the SSS decoding.

Refinement to sub-symbol timing accuracy is done afterwards from the DMRS of the first data slot.

Theorem: BER Penalty of Residual Carrier Frequency Offset

Consider an OFDM-massive-MIMO receiver with subcarrier spacing Δf\Delta f, FFT size NN, and residual carrier frequency offset εf\varepsilon_f. After FFT, the received signal on subcarrier kk is Yk=α(ε)Xk+Ik(ε)+Wk,Y_k = \alpha(\varepsilon) X_k + I_k(\varepsilon) + W_k, where α(ε)=sinc(ε)\alpha(\varepsilon) = \mathrm{sinc}(\varepsilon) with ε=εf/Δf\varepsilon = \varepsilon_f / \Delta f, and the intercarrier interference power is σI2(ε)≈(πε)23Ps,\sigma_I^2(\varepsilon) \approx \frac{(\pi\varepsilon)^2}{3} P_s, with PsP_s the signal power. The effective SINR after receive beamforming on NtN_t antennas degrades as SNReff(ε)=∣α(ε)∣2NtSNR1+NtSNR⋅(πε)2/3.\text{SNR}_{\rm eff}(\varepsilon) = \frac{|\alpha(\varepsilon)|^2 N_t \text{SNR}}{1 + N_t\text{SNR}\cdot (\pi\varepsilon)^2/3}.

A residual CFO has two effects: it rotates the desired subcarrier symbol (the α\alpha factor), and it smears energy from neighboring subcarriers onto the target (the ICI). The first hurts the signal coherently and is partially recovered by per-subcarrier phase tracking; the second is additive noise. In massive MIMO the array gain amplifies both the desired signal and the ICI, so the effective SINR has a fixed asymptotic ceiling set by 1/(πε)21/(\pi\varepsilon)^2.

Show Hint

Expand the received time-domain sample as x~nej2πεfnTs\tilde x_n e^{j 2\pi \varepsilon_f n T_s} before the FFT.

Apply the DFT to ej2πεn/Ne^{j 2\pi \varepsilon n / N} and identify the attenuation factor on the target subcarrier and the leakage to neighbors.

The attenuation is sin⁡(πε)Nsin⁡(πε/N)≈sinc(ε)\frac{\sin(\pi\varepsilon)}{N \sin(\pi\varepsilon/N)} \approx \mathrm{sinc}(\varepsilon) for small ε\varepsilon and large NN.

The leakage sum gives σI2≈(πε)2/3\sigma_I^2 \approx (\pi\varepsilon)^2/3 to leading order; combine with the massive-MIMO array gain to get the final SINR expression.

Proof
Time-domain CFO phase ramp

The received time-domain signal, after frequency downconversion with error εf\varepsilon_f, is rn=x~nej2πεfnTs+wnr_n = \tilde x_n e^{j 2\pi \varepsilon_f n T_s} + w_n, where Ts=1/(NΔf)T_s = 1/(N\Delta f) is the sample period. Writing ε=εf/Δf\varepsilon = \varepsilon_f / \Delta f gives a phase ramp ej2πεn/Ne^{j 2\pi\varepsilon n/N} over the FFT window.

Apply DFT and identify attenuation

The FFT of the product x~nej2πεn/N\tilde x_n e^{j 2\pi\varepsilon n/N} has the target subcarrier kk multiplied by αN(ε)=1N∑n=0N−1ej2πεn/N=sin⁡(πε)Nsin⁡(πε/N).\alpha_N(\varepsilon) = \frac{1}{N}\sum_{n=0}^{N-1} e^{j 2\pi\varepsilon n/N} = \frac{\sin(\pi\varepsilon)}{N\sin(\pi\varepsilon/N)}. For large NN and ∣ε∣≪1|\varepsilon| \ll 1 this simplifies to sinc(ε)=sin⁡(πε)/(πε)\mathrm{sinc}(\varepsilon) = \sin(\pi\varepsilon)/(\pi\varepsilon).

Intercarrier leakage

The leakage to subcarrier k+ℓk + \ell for ℓ≠0\ell \neq 0 is αN(ε−ℓ)≈ε/ℓ\alpha_N(\varepsilon - \ell) \approx \varepsilon/\ell (for small ε\varepsilon). Summing the squared magnitudes over all ℓ≠0\ell \neq 0 gives σI2≈(πε)2/3⋅Ps\sigma_I^2 \approx (\pi\varepsilon)^2/3 \cdot P_s, assuming the Xk+ℓX_{k+\ell} are i.i.d. with power PsP_s.

Array gain

With NtN_t receive antennas and matched-filter combining, both the desired signal and the noise-plus-ICI scale by NtN_t. Hence the effective SINR is ∣α(ε)∣2NtSNR/(1+NtSNR⋅(πε)2/3)|\alpha(\varepsilon)|^2 N_t\text{SNR} / (1 + N_t\text{SNR}\cdot(\pi\varepsilon)^2/3), which has a finite ceiling as SNR→∞\text{SNR}\to\infty. ■\blacksquare

Example: Sizing the CFO Target for a 30 kHz Numerology

A 5G NR cell with Δf=30\Delta f = 30 kHz and Nt=64N_t = 64 antennas wants to hold the effective SINR within 0.2 dB of the no-CFO reference at an operating point of SNR=10\text{SNR} = 10 dB. What residual carrier frequency offset εf\varepsilon_f does this demand?

Solution
Translate the 0.2 dB tolerance

A 0.2 dB degradation corresponds to a ratio SNReff/SNRref≥0.955\text{SNR}_{\rm eff} / \text{SNR}_{\rm ref} \geq 0.955. Using the theorem with NtSNR=64⋅10=640N_t\text{SNR} = 64 \cdot 10 = 640: ∣α(ε)∣21+640⋅(πε)2/3≥0.955.\frac{|\alpha(\varepsilon)|^2}{1 + 640 \cdot (\pi\varepsilon)^2/3} \geq 0.955.

Expand for small $\varepsilon$

For small ε\varepsilon, ∣α(ε)∣2≈1−(πε)2/3|\alpha(\varepsilon)|^2 \approx 1 - (\pi\varepsilon)^2/3. Combining, at leading order: (πε)2(1+640)/3≤0.045,(\pi\varepsilon)^2 (1 + 640)/3 \leq 0.045, so (πε)2≤2.1×10−4(\pi\varepsilon)^2 \leq 2.1 \times 10^{-4} and ∣ε∣≤4.6×10−3|\varepsilon| \leq 4.6 \times 10^{-3}.

Convert to Hz

∣εf∣=∣ε∣⋅Δf≤4.6×10−3⋅30 kHz=138 Hz|\varepsilon_f| = |\varepsilon| \cdot \Delta f \leq 4.6 \times 10^{-3} \cdot 30~\mathrm{kHz} = 138~\mathrm{Hz}. At carrier f0=3.5f_0 = 3.5 GHz this is an oscillator stability of ≈40\approx 40 parts per billion — well within standard TCXO specifications. The array amplification of ICI means that at higher NtN_t the CFO budget tightens further.

BER Degradation under Residual Timing and Frequency Offsets

Scan the residual CFO (as a fraction of subcarrier spacing) and the residual timing offset (as a fraction of the cyclic prefix). Observe how the post-FFT SINR falls as both offsets grow, and how the massive-MIMO array gain amplifies rather than mitigates the ICI floor.

Parameters
10
64

Definition:

Local-Oscillator Phase Noise

The local oscillator driving the RF front end has a non-ideal spectrum; its instantaneous phase ϕ(t)\phi(t) is a random process with power spectral density Sϕ(f)S_\phi(f), typically modeled as a Wiener process with rate σϕ2\sigma_\phi^2 or as a filtered shape in 1/f21/f^2 and 1/f31/f^3 regions. The effect on the baseband is a multiplicative complex rotation ejϕ(t)e^{j\phi(t)} applied to every sample of every antenna sharing that LO.

Phase noise has two distinct effects on the receive signal:

  1. Common phase error (CPE): a constant phase rotation per OFDM symbol that is the same for all subcarriers in the symbol. Correctable with a one-tap tracking loop using DMRS.

  2. Intercarrier interference: a per-subcarrier leakage that is not correctable — it acts exactly like the ICI of residual CFO, but with variance set by the integrated phase noise over the OFDM symbol.

FR2 (mmWave) testbeds are dominated by phase noise because the carrier-frequency-proportional noise floor is 30 dB worse than at sub-6 GHz.

⚠️Engineering Note

GPS-Assisted Synchronization for Cell-Free Testbeds

In a centralized massive MIMO testbed a single reference oscillator drives every RF chain, so sub-nanosecond relative timing is trivial. In a cell-free testbed each access point has its own oscillator, and the entire coherence claim depends on disciplining those oscillators to a shared reference. The three options:

  • GPS/GNSS disciplined oscillators (GPSDO). Each AP has a GPS receiver that generates a 1PPS signal referenced to GPS time. Pulse-per-second accuracy is ∼30\sim 30 ns outdoors — sufficient for the guard-interval budget of 5G NR but not for sub-gridpoint phase alignment. Phase is bootstrapped from a scheduled calibration sequence between APs.

  • White Rabbit over fiber. A deterministic Ethernet extension that achieves sub-nanosecond timing and picoseconds-level jitter. Used in high-end scientific testbeds.

  • Shared reference clock distribution. A central BS distributes a 10 MHz reference over coax or fiber to each AP. Works well over small campus deployments; fails over wide areas.

Massive Beams uses a hybrid of GPSDO for coarse timing and an O-RAN LLS-S plane (Low-Layer Split, S-plane) synchronization protocol for fine alignment. The result is sufficient for cell-free precoding across a 50-AP campus.

Practical Constraints
  • •

    GPSDO alone gives ∼30\sim 30 ns timing and requires scheduled phase calibration for cell-free

  • •

    White Rabbit gives sub-ns but needs dedicated dark fiber

  • •

    Cell-free phase coherence requires inter-AP pilot exchanges — the same idea as intra-array argos calibration, extended across APs

,

Common Mistake: Ignoring Phase Noise at FR2

Mistake:

Running theoretical capacity curves at 28 or 39 GHz as if the LO were noise-free, then being surprised when the mmWave testbed hits an SINR ceiling well below the Shannon limit.

Correction:

Phase noise power scales as fc2f_c^2, so moving from 3.5 GHz to 28 GHz multiplies the integrated phase noise by roughly 60 dB. A TCXO that gives adequate performance at sub-6 GHz is catastrophic at FR2. Use OCXOs or dedicated PLL chains with integrated phase noise budgets, and include the phase-noise-induced ICI in every link-budget analysis. The asymptotic SINR ceiling in FR2 is set not by thermal noise but by phase noise.

Common Mistake: Per-Antenna Sample Clocks in Cell-Free

Mistake:

Treating each AP in a cell-free testbed as an independent receiver that just ships its samples back to the CU. Without a shared sample clock, the samples from different APs drift against each other, and the coherent combining at the CU devolves into incoherent.

Correction:

Every AP in a coherent cell-free cluster must be locked to the same sample clock and have its sampling-phase alignment tracked. The standard approach is GPSDO for 1PPS timing, a shared 10 MHz reference for the sample clock, and a scheduled argos-like calibration between APs to equalize sampling phase. This is exactly the architectural challenge Massive Beams has engineered around.

Historical Note: Sync in Radio Astronomy and Its Echo in Cell-Free

1960s--present

The techniques for disciplining distributed apertures to a shared timebase were developed first by the radio astronomy community. Very-long-baseline interferometry (VLBI) demanded coherence across continents and motivated hydrogen-maser clocks and post-hoc delay calibration against astronomical references. Cell-free massive MIMO inherits the same structural problem: many apertures that must behave as one. The engineering answers differ — we use GPS and O-RAN S-plane rather than masers — but the conceptual framing is borrowed intact. Massive Beams engineers regularly cite VLBI literature when designing cell-free sync subsystems.

Key Takeaway

Synchronization recap. A massive MIMO testbed must synchronize (1) the frame boundary via PSS/SSS, (2) the symbol-level timing within the cyclic prefix, and (3) the oscillator frequency and phase to a fraction of the subcarrier spacing. Residual frequency offset and phase noise cause attenuation and ICI that the array amplifies rather than averages out; the resulting SINR ceiling is the dominant impairment at FR2. Cell-free deployments require GPSDO plus scheduled inter-AP calibration to achieve coherent joint transmission across a cluster.

Quick Check

Why does massive MIMO amplify the ICI penalty of a residual carrier frequency offset rather than average it out?

The array gain scales both the desired signal and the ICI equally, so the SINR ratio is unchanged and the CFO appears as a ceiling

The array averages out the ICI, so massive MIMO mitigates CFO sensitivity

CFO becomes irrelevant in massive MIMO because channel hardening takes over

Only FR2 is affected; sub-6 GHz testbeds are immune

Correction:
The array gain scales both the desired signal and the ICI equally, so the SINR ratio is unchanged and the CFO appears as a ceiling

The ICI has the same spatial signature as the desired signal — both are processed coherently by the receive beamformer. The array amplifies the signal and the interference by the same NtN_t, so the SINR asymptote is set by the CFO-induced leakage and cannot be overcome by growing the array.

Reciprocity CalibrationOver-the-Air Measurements and Dataset Curation