How It Works
From electromagnetic wave propagation to a computable instability boundary — the physics that explain why wireless links fail and how WaveLynk predicts it.
Why high-frequency links fail without warning
As wireless systems push to 6 GHz (Wi-Fi 7) and beyond, a fundamental fragility emerges in the most commonly used beamforming strategy.
Zero-Forcing (ZF) beamforming
ZF cancels interference between users by inverting the wireless channel matrix H. With accurate, up-to-date Channel State Information (CSI), this produces near-perfect spatial multiplexing — each user's signal is cleanly separated.
WZF = precoder matrix, H = channel matrix, H† = conjugate transpose.
The hidden vulnerability
The CSI used for beamforming is always slightly out of date. The channel is measured, fed back, and then used — all within milliseconds. But at high carrier frequencies (6 GHz+), even a few milliseconds of user movement causes the channel to change significantly.
When the CSI estimate becomes stale, ZF doesn't degrade gracefully — matrix inversion amplifies the estimation error, causing a sudden, catastrophic jump in interference. This is the Coherence Cliff.
The wireless channel as a matrix
In a MIMO system, the channel between transmitter and receiver is described by a complex matrix H. Each element Hij captures the complex gain — amplitude and phase — from transmit antenna j to receive antenna i.
y = received signal, x = transmitted signal, n = additive noise.
The base station periodically estimates H — this is Channel State Information (CSI). But by the time this estimate is used for beamforming, some time τ has elapsed. Whether τ is "too long" depends on how fast the channel is changing — its coherence time Tc.
Key parameters
- τ — CSI feedback delay (how old the estimate is)
- Tc — Channel coherence time (how long the channel stays constant)
- fD = v·fc/c — Doppler frequency (proportional to user speed and carrier frequency)
- κ(H) — Condition number of H (sensitivity to perturbation)
- SINR — Signal-to-interference-plus-noise ratio
How CSI becomes stale: Jake's autocorrelation model
The statistical correlation between the true channel H(t) and an estimate made τ seconds earlier follows the Jake's model:
J₀ = Bessel function of the first kind, order zero. fD = Doppler frequency.
When τ is small relative to Tc, ρ ≈ 1 — the estimate is still accurate. As τ increases, ρ decays. The critical insight is that this decay is nonlinear: it has a sharp knee at τ ≈ 0.3·Tc, after which correlation drops rapidly.
Below this knee, ZF works. Past it, errors begin amplifying exponentially through matrix inversion.
| Condition | Correlation ρ | ZF Status |
|---|---|---|
| τ ≪ Tc | ρ ≈ 1.0 | Stable |
| τ ≈ 0.3·Tc | ρ ≈ 0.8 | Transition |
| τ ≈ 0.5·Tc | ρ ≈ 0.6 | Switch to MRT |
| τ ≥ Tc | ρ ≈ 0 | Cliff — ZF fails |
Condition number makes it worse
The condition number κ(H) = σmax/σmin measures how sensitive the channel matrix is to small perturbations. For ZF, which requires matrix inversion, this matters critically:
The relative error in beamforming weights W is bounded by κ times the relative channel estimation error.
This is why the instability is nonlinear: it's not just that CSI ages — the conditioning of the channel amplifies that aging. A channel with κ = 10 turns a 5% CSI error into a 50% beamforming weight error.
Error amplification by condition number
| κ(H) | CSI Error 5% | BF Weight Error |
|---|---|---|
| 1 (ideal) | 5% | 5% |
| 3 | 5% | 15% |
| 10 | 5% | 50% |
| 20 | 5% | 100% → collapse |
Building the Conditioned Coherence Index
Combining Doppler decorrelation, matrix conditioning, signal quality, and CSI aging into a single predictive metric.
Step 1: Channel autocorrelation decay
Using the Jake's model for isotropic Rayleigh fading, the temporal autocorrelation of the channel is:
For small arguments, J₀(x) ≈ 1 − x²/4, so ρ ≈ 1 − (π·fD·τ)². The ratio τ/Tc is the key dimensionless parameter controlling when this decay becomes significant. When it exceeds ~0.3, correlation drops below 0.8 — the onset of instability.
Step 2: ZF weight perturbation bound
ZF weights are computed as W = H†(HH†)−1 (pseudoinverse). Standard matrix perturbation theory gives:
This bound grows super-linearly with κ. When the denominator approaches zero — i.e., when κ(H)·‖ΔH‖/‖H‖ → 1 — the error is unbounded. This is the coherence cliff.
Step 3: Assembling the CCI
The CSI error ‖ΔH‖/‖H‖ is proportional to 1 − ρ(τ), which itself scales with the Bessel function. Combining all four physical drivers:
Each term is independently measurable from standard Wi-Fi CSI feedback. Their product gives a single number capturing all four drivers of beamforming instability.
Step 4: Deriving the threshold γ = 0.6
At a typical feedback delay τ ≈ 0.3·Tc, the Bessel argument becomes 2π·fD·0.3·Tc = 2π·0.3·0.423 ≈ 0.80.
Multiplying by typical SINR weight (≈ 0.95 at 20 dB) and aging factor (e−0.3 ≈ 0.74) gives a combined baseline product of ≈ 0.60. This is the operating point at which any ill-conditioning (κ > 1) pushes the system into the instability zone.
Therefore γ = 0.6 is the natural stability boundary — below this, ZF is safe; at or above this, ZF errors grow and MRT provides better performance.
ZF vs. MRT — and why neither alone is optimal
Zero-Forcing (ZF)
Cancels multi-user interference by inverting the channel matrix. Peak performance when CSI is fresh.
| Property | Characteristic |
|---|---|
| Peak throughput | High — full spatial multiplexing |
| Interference | Cancelled (when CSI is accurate) |
| CSI sensitivity | Very high — fails with stale CSI |
| Failure mode | Sudden collapse at the cliff |
Maximum Ratio Transmission (MRT)
Maximizes received signal strength per user. Robust to stale CSI, lower peak throughput.
| Property | Characteristic |
|---|---|
| Peak throughput | Moderate — no interference cancellation |
| Interference | Present but bounded |
| CSI sensitivity | Low — graceful degradation |
| Failure mode | Gradual degradation only |
See it validated
The theory is backed by 100 hardware trials and a full Monte Carlo simulation sweep.