Make your paper interactiveCreate

Text from the interactive explanation

Coded Modulation Assisted Radial Basis Function Aided Turbo Equalization for Dispersive Rayleigh-Fading Channels

Soon Xin Ng, Mong-Suan Yee, Lajos Hanzo

Original paper

Radio spectrum is scarce, so every pulse a phone receives must carry more bits. A dense alphabet of 16 signal shapes does that, but echoes smear its shapes into one another. In the paper's computer simulations, a receiver that lists every possible echo and confers with an error-correcting decoder needed only 1.88 dB more energy per bit over a fading, echo-filled channel than an equally fast uncoded signal needs on a perfectly clean one — while doing about 36 times fewer additions than its full-strength sibling.

Watch the cloud: each pulse is one round of the receiver's conversation.

Simplified live simulation — colours show which of the 16 symbols was actually sent. Each burst of 684 symbols is a fresh random fading draw at Eb/N0 = 10 dB. Dots start where a receiver that ignores the echo would put them, then move to the receiver's best estimate after each of 8 turbo iterations; off-colour stragglers are errors. The arrows show the direct path and the echo for this burst (strength and phase); the bars count bit errors after each iteration.

A 16-letter radio alphabet

A radio carrier is a wave. The transmitter mixes two copies of it that are a quarter of a cycle (90°) apart: the in-phase wave I and the quadrature wave Q. How much of each it mixes in fixes a point on a plane, and that point is one transmitted symbol.

16QAM (16-point quadrature amplitude modulation) uses a 4 × 4 grid of points, so each symbol carries 4 bits. Neighbouring points differ in only one bit (Gray coding), so a small slip usually costs a single bit. Each point keeps its colour throughout this page: hue follows its angle, brightness its ring.

Tap a point, or focus the grid and use the arrow keys, to see its 4 bits and the wave it sends.

Type a word: it becomes bits, and every 4 bits become one coloured symbol.

More points mean more bits per pulse, but the points sit closer together, so less noise or distortion is needed to push one into a neighbour's place. The paper's coded modulation spends 1 of the 4 bits on redundancy: 3 data bits plus 1 check bit per 16QAM symbol (rate 3/4, 3 information bits per symbol).

Pick a point

The echo

In a city your phone hears the tower directly and again a moment later via a reflection. The paper's channel has exactly two paths, one symbol apart and equally strong on average: h = 0.707 + 0.707 z−1. So every symbol arrives twice, and each received sample is the current symbol plus an echo of the previous one:

rk = h0·sk + h1·sk−1 + noise

That overlap is intersymbol interference. The taps also fade randomly (Rayleigh fading, normalized Doppler frequency 3.25 × 10−5), staying constant within each transmission burst.

At what echo strength do the 16 clusters start to merge?

Drag the echo slider. (Moving it holds the fading still, with the echo in phase with the direct path, so you can see the geometry; press Play to let the channel drift again.)

In the paper the echo is, on average, as strong as the direct path, so one sample alone cannot tell the symbols apart. But the echo is not random noise: it is a delayed copy of information the receiver can use.

Illustration: noiseless channel states plus light noise; Play lets both taps fade (strength and phase drift), which twists and breathes the cloud.

Every burst carries a known training sequence in its middle (burst structure adapted from Fig. 3, the FRAMES FMA1 non-spread burst):

A TDMA frame holds 16 such slots per 4.615 ms. The paper's interleaver spans 8 frames (8 × 684 = 5472 symbols, about 37 ms of delay).

The receiver guesses the two taps, predicts each training sample, and nudges its guess in proportion to the error: the least-mean-squares (LMS) algorithm. The paper used a step size of 0.05 in the first pass and 0.01 in later turbo iterations.

Compare step sizes: a big step learns fast but jitters, a small one is smooth but slow.

List every possibility, then sort

A conventional equalizer tries to undo the channel: it filters the received signal to mimic the channel's inverse. The radial basis function (RBF) equalizer works the other way round. From its estimate of the two taps it lists every noiseless sample the channel could produce — each combination of current and previous symbol, 16 × 16 = 256 "channel states" — and places a soft bump on each.

A radial basis function's height depends only on the distance from its centre, exp(−|v−r|²/2σ²). To classify a received sample, the receiver adds up the bump heights belonging to each candidate symbol: Bayesian classification.

Drag the white probe (a received sample) around the plane, or focus the plane and use the arrow keys. The bars show how likely each of the 16 symbols is.

Find a spot where two symbols tie.

The paper's equalizer looks at 3 successive samples (feedforward order m = 3, decision delay τ = 2) and feeds back 1 past decision (n = 1), which leaves 163 = 4096 bumps (RBF nodes) for every decision . That is the cost problem section 5 solves.

fi(v) = Σj wji · exp(−‖v − cji‖² / 2σ²)

P(symbol i | v) = fi(v) / Σl fl(v)

The centres c are the channel states (the channel taps applied to every possible combination of symbols), σ² is the noise variance, and the weights w are the prior probabilities of each channel state. Without any other knowledge all weights are equal; in turbo equalization the decoder supplies them (section 4).

One illustrative channel (echo 0.55× the direct path, rotated 35°). Each bar is one symbol, in its colour.

Two experts that talk

An error-correcting code adds structured redundancy: only a small fraction of all possible bit sequences are legal codewords, so the decoder can judge which sequences are plausible. In turbo equalization the soft-in/soft-out equalizer and the decoder trade soft beliefs (log-probabilities) for several rounds.

The rule: pass on only what's new . Each expert subtracts what it was told from what it now believes and passes on just this extrinsic information (Eq. 1), so neither is fed its own opinion back. The decoder's beliefs become the RBF weights — the prior probabilities of the channel states — so unlikely neighbours fade. An interleaver (Π) shuffles the order in between, which makes neighbouring symbols statistically independent.

Turbo lab

What happens if the two experts share everything they believe, instead of only what's new?

16 bumps, one per possible previous symbol. Their size and brightness are the decoder's beliefs; the cross is the received sample.

In the paper, gains became marginal after about 3 iterations for the full RBF turbo equalizer.

Split the problem in two

The channel taps are complex numbers: they rotate the signal as well as scale it. That couples the two halves, so the received I part depends on both the sent I and the sent Q, and so does the received Q part (Eq. 12).

Move only the Q lever and watch the received I part move too.

The trick (Eqs. 11–13, Fig. 6): the decoder's current beliefs give estimates ŝI and ŝQ. Subtract the estimated cross-coupled part, and what is left of the I half, r′I = sI∗h + noise, depends only on the 4 possible I levels; the same holds for Q. Two small RBF equalizers then replace one big one. The first iteration has no decoder beliefs yet, so the paper uses a cheap conventional decision-feedback equalizer (DFE) there. Early decoupling is imperfect because the estimates are unreliable, and later iterations compensate.

Raise the quality of the decoder's guesses. With good guesses, the I-arm samples collapse onto just 16 centres (4 current I levels × 4 previous I levels), coloured by the full symbol that was sent. With poor guesses the leftover Q part keeps them smeared.

This is why I/Q needs more iterations: in the paper, gains became marginal after about 6 iterations for RBF-I/Q-TEQ versus 3 for full RBF-TEQ. In Fig. 10, RBF-I/Q-TEQ-TTCM with 8 iterations matched full RBF-TEQ-TTCM with 4 iterations (for BER below 10−4).

Bonus: estimating the channel costs I/Q almost nothing. It has fewer centres to pick through decision feedback (Mn → Mn/2), so wrong decisions propagate less, and LMS channel estimation lost only marginally against perfect channel knowledge (Fig. 7).

How much arithmetic does the split save?

Nodes ns,i = M̄L+m−n, scalar states ns,f = M̄L+1; per iteration, additions ns,i(m+2) − 2M̄ + ns,f and multiplications 2ns,f, with M̄ = M for the full equalizer and √M for I/Q (m = 3, n = 1). The paper's formula (Sec. V-B) applied to other settings; the paper simulates 16QAM with L = 1.

The race (paper's 16QAM, L = 1 numbers)

To equal performance: full RBF-TEQ for 4 iterations against RBF-I/Q-TEQ for 8 iterations, whose first pass is a cheap DFE of negligible cost, so 7 RBF-I/Q iterations count. Both work at the same pace per operation.

The race starts when it scrolls into view.

What the simulations showed

All results in the paper come from computer simulation. Energy is measured as Eb/N0, the energy per bit relative to the noise, in decibels: +10 dB means 10 times the energy, +3 dB roughly double. Every marker below is the value needed for a bit error ratio (BER) of 10−5, one wrong bit in 100,000, at 3 bits per symbol.

Set an energy budget. Which receivers reach the target?

14.85 − 12.97 = 1.88 dB ≈ 1.54× the energy per bit of a clean channel. 31.63 − 14.85 = 16.78 dB of coding gain ≈ 47.6× less energy per bit than the conventional receiver.

Four coded-modulation schemes

TCM (trellis-coded modulation), TTCM (turbo TCM), BICM (bit-interleaved coded modulation) and BICM-ID (BICM with iterative decoding), all with similar decoding complexity. TTCM, BICM and BICM-ID with RBF-TEQ came within about 2 dB of the clean-channel benchmark (Fig. 4).

TCM trails in both rankings: some of its bits are left unprotected and it has no internal iterations.

Study settings (all simulated)

Alphabet 16QAM, 3 information bits per symbol

Channel equal-weight two-path, symbol-spaced Rayleigh fading

Fading normalized Doppler 3.25 × 10−5, constant within a burst

Channel estimate LMS on the 49-symbol training sequence

RBF-DFE τ = 2, m = 3, n = 1

Interleaver 5472 symbols (8 frames)

Codes memory 3 components; TTCM 4 internal iterations, BICM-ID 8

Other comparisons

Against a conventional DFE with uncoded 8PSK, coded modulation with a conventional DFE gained 4–7 dB; with RBF-TEQ after 3 iterations it gained 10–17 dB (Fig. 5).

With LMS channel estimation, full RBF-TEQ lost some performance against perfect channel knowledge, and that loss shrank with iterations; RBF-I/Q-TEQ lost only marginally (Fig. 7).

With 6 iterations, RBF-I/Q-TEQ-CM performed like RBF-TEQ-CM with 3, except TCM, whose frame error ratio was about 1 dB worse (Fig. 9).

A trellis-based turbo equalizer would need on the order of 16² operations: comparable to RBF-I/Q-TEQ and 16 times lower than full RBF-TEQ. The paper did not simulate trellis-based schemes with coded modulation.

Back to the echo

The smear. A new burst blooms where you would put each sample if you ignored the echo: every point is spread into a 4 × 4 mini-grid (section 2).

The bumps. The first pass scores every sample against all channel states and sorts it by summed bump heights (section 3).

The conversation. Each pulse is one turbo iteration: the decoder's new beliefs reweight the bumps, and the clusters sharpen (section 4).

The split. Switch to the I/Q receiver: two small equalizers start with no estimates and need more rounds, but reach the same destination (section 5).

The authors frame their result carefully. In their simulations, coded modulation with RBF-aided turbo equalization came within about 2 dB of an identical-throughput benchmark on a clean channel, and the I/Q version kept that performance while cutting additions about 36-fold and multiplications about 9-fold. Dense alphabets over hostile, fading channels at close to clean-channel energy, with modest arithmetic: that is what this receiver design brings closer.

Limits worth keeping in mind: every result is a computer simulation of one channel model (an equal-weight two-path Rayleigh channel), and trellis-based turbo equalizers with coded modulation were not simulated for comparison.

Source and notes

Soon Xin Ng, Mong-Suan Yee and Lajos Hanzo, “Coded Modulation Assisted Radial Basis Function Aided Turbo Equalization for Dispersive Rayleigh-Fading Channels,” IEEE Transactions on Wireless Communications, vol. 3, no. 6, pp. 2198–2206, November 2004. DOI: 10.1109/TWC.2004.837410. Author-hosted copy: southampton.ac.uk/~sxn/paper/35-tow-04.pdf.

Code: a simple rate-1/2, 4-state convolutional code (2 information bits per 16QAM symbol, Gray labelling, bit interleaving), instead of the paper's rate-3/4 TCM, TTCM, BICM and BICM-ID schemes.

One burst of 684 symbols per interleaver (the paper interleaves over 8 frames, 5472 symbols).

The equalizer looks at the two samples that contain each symbol (rk and rk+1), without decision feedback; the paper's RBF-DFE uses m = 3, n = 1, τ = 2.

Perfect channel knowledge; the paper estimates the channel with LMS.

For its first pass the paper's I/Q receiver uses a conventional DFE; our simplified I/Q receiver begins with no estimates at all and treats the unknown other half as noise.

Eb/N0 values in the lab refer to this toy system and are not comparable with the paper's dB figures. Deep fades happen in our bursts too: some cannot be rescued at any number of iterations.

All visuals are original and computed live; the block diagrams are adapted from Figs. 1, 3 and 6 of the paper. The paper's result curves are not reproduced; quoted numbers come from its text.

Topics in this iPaper

Engineering
channel equalization is invoked for mitigating the effects of intersymbol interference (ISI) in the context of single carrier modulation

Section I, page 2198

Computer science
formulates the channel equalization procedure as a classification problem

Section I, page 2198

turbo equalization
radial basis function (RBF)-based turbo equalization (TEQ) schemes are investigated

Abstract, page 2198

radial basis function network
The RBF network-based equalizer is capable of utilising the a priori information

Section III, page 2200

coded modulation
coded modulation (CM) schemes, which are based on combining the functions of channel coding and modulation

Section I, page 2198

rayleigh fading
when communicating over dispersive Rayleigh-fading channels

Abstract, page 2198

intersymbol interference
mitigating the effects of intersymbol interference (ISI)

Section I, page 2198

quadrature amplitude modulation
16 quadrature amplitude modulation-based trellis coded modulation (TCM)

Abstract, page 2198

Coded Modulation Assisted Radial Basis Function Aided Turbo Equalization for Dispersive Rayleigh-Fading Channels — Interactive explanation | iPaper.ai