Channel estimation errors are the difference between the channel a receiver or transmitter believes it has and the channel that actually exists. Their impact on system performance is an SNR loss on a single link and a hard ceiling on multiuser links: with zero forcing from 8 antennas, an error of -20 dB caps four users at about 30 b/s/Hz however much power is added. Error comes from noise on the pilots, too few pilots, interpolation and channel ageing.
The channel state information page explains how CSI is estimated, and CSI estimation techniques looks at the estimators. This page starts where those stop. It takes an estimate with a known error and works out what that error costs in rate, EVM and multiuser capacity, and how many pilots are worth sending.
How channel estimation errors are modelled
Write the true channel as the estimate plus an error:
h = ĥ + e
With an MMSE estimator, the estimate and error are uncorrelated. For a unit-power channel, the estimate holds power 1 – σe2 and the error holds σe2. Here σe2 is the normalised mean squared error (NMSE), the usual figure of merit, quoted in dB.
For Np pilots at the same SNR ρ as the data, the MMSE error on a Rayleigh channel is:
σe2 = 1 / (1 + Npρ)
This covers the noise term only. Real systems add other sources:
- Pilot noise, from low SNR on the reference signal.
- Interpolation error, when pilots are spaced too widely in frequency or time. It sets a floor that pilot power does not remove.
- Channel ageing, as the user moves between the estimate and its use.
- Pilot contamination, from other users sending the same sequence.
- RF impairments, such as carrier frequency offset, phase noise and calibration drift between chains.
Impact on a single link
The receiver treats the error as extra noise, because a share of the signal power is now unknown. The effective SNR is:
SNReff = ρ(1 – σe2) / (1 + ρσe2)
Two regimes follow. When ρσe2 is well below 1, noise still dominates and the error hardly matters. When it is above 1, the error dominates, and SNReff flattens at about 1/σe2. With a fixed floor of -20 dB, the SNR saturates at 20 dB, however strong the link is. At 30 dB of raw SNR, the formula gives 19.5 dB.
Worked example: how many pilots at 15 dB SNR?
Take ρ = 15 dB (31.6) and a coherence block of 100 symbols. Perfect CSI gives log2(32.6) = 5.03 b/s/Hz. Each row applies the formulas, with spectral efficiency scaled by (1 – Np/100):
| Pilots Np | Error σe2 (dB) | SNReff (dB) | Loss (dB) | Spectral efficiency (b/s/Hz) |
|---|---|---|---|---|
| 1 | -15.1 | 11.9 | 3.1 | 4.01 |
| 2 | -18.1 | 13.2 | 1.8 | 4.36 |
| 4 | -21.1 | 14.0 | 1.0 | 4.52 |
| 8 | -24.0 | 14.5 | 0.5 | 4.47 |
| 16 | -27.0 | 14.7 | 0.3 | 4.15 |
One pilot costs 3 dB, since pilot and data then have equal SNR. Four pilots reduce the loss to 1 dB. Beyond that the gain is less than the overhead, and spectral efficiency falls. A search over integers puts the best value at 5 pilots, giving 4.53 b/s/Hz.
The best pilot count depends on both SNR and channel speed:
| Coherence block (symbols) | Best pilots, 0 dB | Best pilots, 15 dB |
|---|---|---|
| 20 | 4 | 2 |
| 100 | 11 | 5 |
| 500 | 25 | 11 |
The trend is the useful part: pilots should scale with roughly the square root of the coherence block. A fast channel gets few pilots and a noisy estimate, which is the reason high-speed links suffer twice. These numbers are for one antenna pair. A MIMO link needs orthogonal pilots for each transmit stream, so the overhead multiplies by the stream count.
Impact on EVM and modulation
Estimation error appears directly in EVM. After equalisation with a noisy estimate, to first order:
EVM2 ≈ σe2 + 1/ρ
So a link with high SNR still shows residual EVM from its reference signals. Allowing the error half of the budget, this table shows the requirement for typical base station EVM limits:
| Modulation | Approximate EVM limit | Limit (dB) | Largest σe2 (dB) | Pilot SNR, 4 pilots (dB) |
|---|---|---|---|---|
| 16-QAM | 12.5% | -18.1 | -21.1 | 15.1 |
| 64-QAM | 8% | -21.9 | -24.9 | 18.9 |
| 256-QAM | 3.5% | -29.1 | -32.1 | 26.1 |
This explains why high-order QAM needs clean reference signals. In practice link adaptation sees the effect as a CQI report that stops rising. Other contributors are covered in modulation, EVM and PAPR.
Impact on multiuser MIMO
Multiuser links suffer far more than single links. Zero-forcing precoding places nulls on other users, and those nulls are only as accurate as the estimate. With an error, interference scales with transmit power, so raising power raises signal and interference together.
For unit-norm ZF beams with M antennas, K users and equal power:
SINR ≈ (ρ/K)(M – K + 1)(1 – σe2) / (1 + ρσe2(K – 1)/K)
At high SNR, this tends to the floor:
SINRmax ≈ (M – K + 1)(1 – σe2) / (σe2(K – 1))
Take M = 8, K = 4. An error of -20 dB gives a floor of 5 × 0.99/(0.01 × 3) = 165, or 22.2 dB per user, which predicts 4 log2(166) = 29.5 b/s/Hz. An error of -13 dB gives 15.0 dB per user and 20.1 b/s/Hz. A Monte Carlo simulation of the link gives 30.0 and 20.7 b/s/Hz at 40 dB SNR, close to these approximations.
| Estimation error | Sum rate at 20 dB SNR | Sum rate at 30 dB SNR | Sum rate at 40 dB SNR |
|---|---|---|---|
| Perfect | 27.3 | 40.6 | 53.8 |
| -30 dB | 27.0 | 37.6 | 42.4 |
| -20 dB | 24.2 | 29.0 | 30.0 |
| -13 dB | 18.9 | 20.5 | 20.7 |
| -10 dB | 15.7 | 16.6 | 16.7 |
Compare this with the single-link result. At 20 dB, an error of -20 dB costs a single link about 3 dB. Here it costs 11% of sum rate at 20 dB and 44% at 40 dB. Multiuser gain is the first thing lost when CSI degrades, which is why schedulers fall back to fewer streams when channel quality reports worsen.
Reducing the impact of estimation errors
The remedies fall into three groups:
- Better estimates. Use LMMSE in place of least squares, filter across time and frequency, and send more pilots on slow channels.
- Fresher estimates. Shorten the report period for fast users, or predict the channel forward.
- Tolerant processing. Regularised ZF and MMSE precoding back off from exact nulls and degrade more gently. Rank adaptation drops streams before interference takes over. Link adaptation holds a margin.
Where feedback is involved, quantisation adds its own error. See CSI quantization and deep learning CSI feedback.
Frequently asked questions
What causes channel estimation errors?
Noise on the pilots, too few pilots, interpolation between them, channel change over time and interference from other users’ pilots. Hardware effects such as frequency offset and phase noise add more.
How much SNR is lost to channel estimation error?
With one pilot at the data SNR, about 3 dB at high SNR. With four pilots, about 1 dB. The loss then falls by half each time the pilots double, until a fixed floor such as interpolation error takes over.
What is the impact of imperfect CSI on MIMO capacity?
Single-stream links lose a few decibels. Multiuser and high-rank links lose much more, because interference that should have been cancelled remains. In the example above, an error of -20 dB caps a four-user link at about 30 b/s/Hz.
How many pilots should be used?
Enough that estimation error sits a little below the noise, and no more. In the worked example the optimum was 5 pilots in 100 symbols at 15 dB. It rises at low SNR and in slow channels.
Related topics
The wireless channel fading article explains how fast the channel changes, and massive MIMO spectral efficiency applies these ideas at scale. MIMO detection algorithms shows how the receiver uses the estimate.