Channel Equalisation Calculator
Design zero-forcing, MMSE and decision feedback equalisers for a multipath channel and compare the output SNR each one leaves.
Free to use. Sign up free to save, download and print.
Worked example: a 0.7 echo at 25 dB SNR
Zero-forcing output SNR: 22.08 dB
Loading the Channel Equalisation Calculator…
How it works
Zero-forcing and MMSE equaliser formulas
An equaliser is a filter at the receiver that undoes multipath. Each echo smears a symbol into the symbols after it, which is inter-symbol interference, and how well an equaliser works is the SNR left at its output: the wanted symbol against the interference it did not remove plus the noise it amplified.
A zero-forcing equaliser aims to remove all the interference and ignores the noise, so it amplifies the noise wherever the channel is weak. An MMSE equaliser minimises interference and noise together, and an MMSE decision feedback equaliser also subtracts the echoes of symbols already decided (Al-Dhahir and Cioffi, 1995). The taps solve the Wiener-Hopf equations at the decision delay with the least error, and the limit that unlimited taps would reach is worked out from the channel's frequency response.
R v = p, with Rₖⱼ = Σ h[a] h*[a + k − j] + σ² δₖⱼ and pₖ = h[D − k]
SINR = |c[D]|² / (ISI + σ² Σ |w[k]|²), with c = h ∗ w and ISI = Σ |c[n]|² for n ≠ D
SNR_ZF = SNR / mean(1 / |H(f)|²)
SNR_MMSE = 1 / mean(1 / (1 + SNR |H(f)|²)) − 1
SNR_DFE = exp(mean ln(1 + SNR |H(f)|²)) − 1
- h[n]
- channel impulse response, one sample per symbol
- R, p
- correlation of the received samples, and of the received samples with the wanted symbol; for the MMSE-DFE, R leaves out the echoes the feedback cancels
- v, w
- the solution, and the equaliser taps, its complex conjugate
- D
- decision delay, in symbols, chosen for the least error
- c
- channel and equaliser combined; c[D] is the wanted symbol, and for the MMSE-DFE the echoes its feedback cancels are left out of the ISI
- σ²
- noise variance, 1 / SNR for unit-power symbols; 0 for zero-forcing
- H(f)
- channel frequency response, with f in cycles per symbol; mean() is the average across the band
- SNR, SINR
- input Es/N0, and the output signal to interference plus noise ratio, as ratios here and in dB in the example
Standards
- ITU-R M.1225 pedestrian A channel delays and powers
- 3GPP TS 36.211 and TS 38.211 cyclic prefix lengths
Worked example
Worked example: a 0.7 echo at 25 dB SNR
A direct path and an in-phase echo at 0.7 of its amplitude one symbol later, at an input SNR of 25 dB, with a zero-forcing equaliser of unlimited length.
| Step | Working | Result |
|---|---|---|
| Channel power response | |1 + 0.7 e^(−j2πf)|² = 1.49 + 1.4 cos 2πf | 0.09 to 2.89 |
| Mean of 1/|H(f)|² | 1 / (1 − 0.7²) | 1.96 |
| Noise enhancement | 10 log₁₀(1.96) | 2.92 dB |
| Zero-forcing output SNR | 25 − 2.92 | 22.08 dB |
With unlimited taps, an MMSE equaliser reaches 22.13 dB, little better at this SNR, and an MMSE decision feedback equaliser 25.01 dB, since it subtracts the echo of the symbol already decided, assuming that decision was right. A real filter has a set number of taps: with five, the zero-forcing equaliser reaches 16.91 dB, because the echo it cannot cancel is left 18.29 dB below the symbol, and with eight it reaches 21.10 dB.
FAQ
Frequently asked questions
Not here? Ask us
Calculations run in your browser and nothing you enter is sent to a server.
When should I use zero forcing, MMSE or decision feedback?
Zero forcing aims to invert the channel and ignores the noise, which works at high SNR on a channel without nulls but amplifies noise badly near a null. MMSE balances interference cancellation against noise and is the usual choice. Decision feedback adds a filter that cancels interference from past decisions, which helps on severe multipath, although a wrong decision can propagate; the calculator assumes past decisions are correct. All three are compared on the same channel.
What is noise enhancement in a zero-forcing equaliser?
A zero-forcing equaliser boosts each frequency by the inverse of the channel there, and the noise at that frequency is boosted with it. The noise enhancement is that boost averaged across the band, the mean of 1/|H(f)|². With an echo at 0.7 of the direct path one symbol later, the channel power falls to 0.09 at its deepest dip, where the noise is boosted by 10.46 dB, and the average is 2.92 dB. If the echo is as strong as the direct path, the channel has exact nulls and no zero-forcing equaliser can invert it, however many taps it has. An MMSE equaliser holds back where the noise would dominate: with an equal echo in opposite phase three symbols later, at 30 dB, its limit is 17.94 dB.
How is the channel modelled?
The channel is a set of complex multipath taps, each with a delay in symbol periods, a magnitude and a phase, so the phase of an echo matters. Fractional delays are evaluated exactly in the frequency domain and sampled from a band-limited pulse for the equaliser design. Presets give starting points, named by their taps.
What is the difference between LMS, RLS and decision directed?
LMS nudges the equaliser taps a small step along the error gradient on each symbol, which is simple but converges slowly. RLS tracks the inverse correlation matrix and converges much faster, at a higher computing cost per symbol. Decision directed mode runs LMS on known symbols for the first 60 symbols, then uses the receiver's own decisions as the reference, which suits tracking a slowly changing channel after training ends.
How does OFDM equalisation differ from single carrier equalisation?
In OFDM the cyclic prefix removes interference between symbols, so the equaliser only needs to correct the amplitude and phase of each subcarrier. This makes per subcarrier equalisation much simpler than a time domain equaliser on a wideband channel, which is why LTE, 5G NR and Wi-Fi use it. The cost is the cyclic prefix overhead, which reduces spectral efficiency.
Is the bit error rate table a compliance test?
No. It lists the uncoded bit error rate each modulation would have at the equaliser output in AWGN, with residual interference treated as Gaussian. Coding and fading are not included, and it is not a conformance test against any specification.
Sign up free to keep your work.
Free to use. Saving, downloading and printing need a free account.