Make your paper interactiveCreate

Text from the interactive explanation

Soft-Decision Star-QAM Aided BICM-ID

Dandan Liang, Soon Xin Ng, Lajos Hanzo

Original paper

Liang, Ng & Hanzo · IEEE Signal Processing Letters · 2011

Could a phone pass along someone else’s message without ever measuring the channel it travels through?

Channel gain now: —

Left, the brass trails show where the storm has dragged each of the 16 points; the brass hand is the channel itself. Right, the same symbols compared pairwise snap onto a fixed star — no measurement of the channel needed. This letter’s contribution is teaching that comparing receiver to say how sure it is about every bit, so that a powerful error-correcting code can use it. In the authors’ simulations, that needs about 25 times less energy per bit than a comparable older scheme.

Everything moving on this page is this page’s own simulation of the set-up described in the paper; numbers attributed to the paper are labelled as such.

Plate I

Every message is a point on a dial

Digital radio speaks with a tiny vocabulary. Transmitter and receiver agree on a pattern of points — engineers call it a constellation . Each point sits at a certain distance from the centre (its amplitude) and a certain angle (its phase). Sending a symbol means picking one point.

With 16 points, every pick carries 4 bits , because 2 × 2 × 2 × 2 = 16. Here are three ways to arrange 16 points with the same average energy. Tap a point to see the bits it stands for.

The paper’s shape is the star : 16-level Star-QAM (“16-StQAM”), two rings of eight points. The inner ring has radius 0.63, the outer 1.26 — twice as far out.

Watch the “closest neighbours” readout. Squeeze all 16 points onto one ring (16PSK) and neighbours crowd together, so a little noise is enough to confuse them. That crowding is what later gives the single-ring scheme its error floor in the paper.

Tap a point.

Closest neighbours: —

All three shapes have average energy 1. Bit labels shown here mark positions; in Star-QAM, as you will see in Plate III, the bits really describe moves.

Plate II

The storm

Now let the phone move. Its signal reaches the mast along many reflected paths that add up and cancel as distances change. Engineers call this Rayleigh fading , and its effect on the dial is simple and brutal: every symbol gets multiplied by one unknown complex number, h.

Multiplying by h rotates the whole constellation and shrinks or stretches it. And h keeps drifting, sometimes collapsing towards zero in a deep fade.

Read the dial blind. Each round, one of the 16 brass points is sent through a fresh storm. The cyan dot is what arrives. Tap the point you think was sent.

A conventional, coherent receiver solves this by estimating h — its channel state information — before it reads any absolute position. That costs effort: known “pilot” symbols that carry no data, tracking circuitry, and the risk of locking onto the wrong rotation, which the paper calls false-phase locking.

The authors point out why this matters more and more: in cooperative networks, a mobile phone may relay someone else’s signal, and “it might be unrealistic to expect from a relay station constituted by a cooperating mobile phone to estimate the channel of the link it is relaying.”

Score 0 / 0

Round 1 of 6 — which brass point was sent?

Each round draws a random channel and a little noise (this page’s illustration). With the channel given, the rotated, rescaled star is drawn for you — that is what perfect channel knowledge buys.

Plate III

Say it with changes

Star-QAM’s way out is to stop sending positions and send moves. Each group of 4 bits says how to get from the previous point to the next:

The first bit, b3 : stay on the same ring (0) or hop to the other ring (1).

The other three, b2 b1 b0 : how far to turn — one of eight steps of 45°.

In the paper’s notation the new direction is the old one turned by the step, vk = vk−1 · wk, and the point sent is that direction at the chosen ring size, xk = ak · vk. Every frame begins from an agreed reference point on the inner ring.

Set four bits and press Send , or type a short word and watch it become a dance of moves.

Start: reference point, inner ring, 0°.

Turn steps use a Gray labelling of the eight 45° turns, so neighbouring turns differ in one bit. Each letter is 8 bits, i.e. two moves.

Divide out the storm

At the receiver, divide each symbol by the one before it. If the channel h hardly changed in between, it appears on top and bottom — and cancels:

yk ÷ yk−1 ≈ (ak ÷ ak−1) × wk

What’s left is the move itself: a size change of ×1 , ×0.5 or ×2 , and a turn that is a multiple of 45°. Three rings, eight spokes, 24 possible landing spots — and no channel estimate needed.

It is a navigator’s trick: the angle between two stars doesn’t depend on how the ship rolls.

The catch is in that “≈”. Slide the channel speed up: the paper assumes the channel barely changes between neighbouring symbols (its simulations use a normalised Doppler frequency of 0.01). When the storm moves faster, the comparison itself smears.

Moves misread: —

This page’s simulation at Eb/N0 = 25 dB with simple hard decisions; the readout counts 40,000 simulated moves at the chosen speed.

Plate IV

Three rings, three blurs

Before you look: the receiver compares each symbol with a previous symbol that carries noise of its own. Which kind of move will it see most blurrily?

The histogram counts the size ratio |yk| ÷ |yk−1| for a long stream of simulated symbols — this page’s re-creation of the paper’s Fig. 2. All symbols suffer the same radio noise, yet the three peaks come out very different.

The reason is in the paper’s Eq. 4: the old symbol’s noise is multiplied by the jump p before it lands on the comparison. So each move carries its own effective noise, Ñ0 = N0(1 + p²) (Eq. 8):

A grow move also starts from the small inner ring, so dividing by it magnifies the blur further. Any receiver that wants honest probabilities must account for these different blurs — the heart of the next plate.

Peak heights —

Bars: this page’s simulation (correlated Rayleigh fading, normalised Doppler 0.01), scaled so the ×1 peak matches the paper’s 1.25. Brass ticks : peak heights read from the published Fig. 2 at 25 dB (≈1.05, 1.25, 0.27).

Plate V · the paper’s contribution

Don’t decide — weigh the evidence

Earlier Star-QAM receivers made a hard decision : find the nearest star, output its 4 bits, move on. That throws information away. Modern error-correcting codes work far better when the receiver also says how sure it is about every bit. The paper’s central contribution is the formula for those probabilities: a soft demapper for 16-StQAM, which needs no channel estimate.

Drag the cyan point (or focus the plate and use the arrow keys) — it is the new received symbol. The hollow brass ring is the previous one. Around it sit the 24 places the new symbol would land for each possible move, each with a halo showing how far noise can push it.

—

—

Computed live with the paper’s Eqs. 9–12, no prior hints from a decoder. The numbers after each percentage are log-likelihood ratios (LLRs): positive means “probably 1”, and each +1 multiplies the odds by about 2.7. Halo radius ∝ √(N0(1+R²)).

Score every candidate. For each of the eight turns and each size change, ask how likely the new symbol would land where it did. The answer is a bell curve around the candidate, with width set by that move’s own effective noise — 2 N₀ for stay, 1.25 N₀ or 5 N₀ for the two kinds of switch (Eq. 9, Eq. 10).

Switch means two possibilities. Without a channel estimate the receiver can’t be sure which ring the previous symbol was on, so “switch ring” adds the ×0.5 and ×2 candidates together (Eq. 10).

Add up per bit. A bit’s probability of being 1 is the total score of all candidates whose label has that bit set, against the rest (Eqs. 11–12). Those probabilities, as LLRs, go to the decoder.

Try the outer ring’s edge and flip the noise model. If the receiver wrongly assumes every move is equally blurred, a grow ×2 landing a little off-target looks implausible, and the ring bit b₃ swings with false confidence. The ring-aware model knows ×2 jumps are blurry and stays honest.

Plate VI

Two experts in conversation

The soft demapper’s probabilities go to the decoder of an error-correcting code. The paper uses a rate-3/4 convolutional code : every 3 data bits become 4 coded bits, which is exactly one 16-StQAM symbol. Before mapping, the coded bits are shuffled by interleavers, so bits that travel together don’t fade together. This is BICM , bit-interleaved coded modulation.

In the iterative version, BICM-ID , the two experts talk. The decoder sends back what it has worked out about each bit; the demapper re-weighs the evidence using those hints (the “a-priori” probabilities in Eqs. 11–12); and so on. One iteration means each expert has spoken once.

Redrawn after the paper’s Fig. 1 (interleavers omitted there, as here). Swipe sideways to follow the whole chain.

Why the labels on the eight turns matter

How much a hint helps depends on which bit pattern is written on which turn. Set the decoder’s hints about b₀ and b₁ and watch the demapper’s confidence in the remaining bit b₂ for one noisy received comparison.

—

turns with b₂ = 1turns with b₂ = 0received comparison

This page’s illustration at Eb/N0 = 10 dB; a hint is an a-priori probability of 98%. Faded turns are ruled out by the hints.

With Gray labels, neighbouring turns differ in only one bit — ideal when nobody is helping, which is why Gray-labelled BICM wins on a single pass. With set partitioning (SP), once the decoder vouches for b₀ and b₁, the two remaining candidates sit 180° apart and are almost impossible to confuse. That is why the paper finds SP-based BICM-ID worse than Gray BICM on the first iteration, but about 2 dB better after the second.

Plate VII · the paper’s results

The race to one error in a million

The paper’s Fig. 3, redrawn. Down is better: the bit error ratio is the fraction of data bits that arrive wrong. Right means more energy per bit, Eb/N0; every 10 dB is ten times more. These are the authors’ Monte-Carlo simulations over correlated Rayleigh fading — 20,000 frames of 1,200 symbols each. Drag across the chart or use the power dial.

Markers and lines: values read approximately from the published Fig. 3 (solid = Gray-labelled BICM, dotted = SP-based BICM-ID). Glowing dots: your own short runs from the lab below.

The single-ring 16DPSK-BICM stalls on an error floor: its 16 points are too crowded.

16-StQAM-BICM with the new soft demapper reaches one error in a million about 12 dB sooner.

BICM-ID starts worse on its first iteration, then after the second wins about 2 dB more: roughly 14 dB in all, about 25 times less energy per bit.

Coherent 16QAM and 16PSK , given perfect channel knowledge, remain ahead — the price of never measuring the channel.

Run it yourself

The lab re-simulates the chain from the paper’s Table I in your browser: random data bits, the rate-3/4 code, four interleavers, the differential mapper, correlated Rayleigh fading (normalised Doppler 0.01), the soft demapper of Eqs. 9–12 and an iterative decoder, 1,200 symbols per frame. Each result is plotted as a glowing dot on the chart above.

A short re-simulation in your browser, following Table I. The paper’s curves used 20,000 frames per point; short runs cannot reach the last decades near one error in a million, and errors arrive in bursts during deep fades, so early estimates wobble. Runs usually land within about a dB of the published curves. Details the paper leaves open (which coded bit goes to which symbol bit, the fading generator) are this page’s choices.

Plate VIII

Back to the relay

We began with a phone that might pass along someone else’s message without measuring the channel. This letter doesn’t build such a relay — it supplies a missing piece for one. A receiver that only compares neighbouring symbols can now hand honest probabilities to a powerful iterative decoder, instead of bare yes-or-no guesses.

Soft decisions make channel-estimation-free 16-StQAM compatible with BICM and BICM-ID. In simulation: about 12 dB over 16DPSK-BICM, about 14 dB with two decoding iterations, at one error in a million.

Coherent 16QAM and 16PSK with perfect channel knowledge still do better. And the comparison trick assumes the channel barely changes between neighbouring symbols; the results are simulations, not field measurements.

The authors note the same soft-decision approach can be extended to DAPSK schemes with more than two rings — more stars in the sky.

Sources & credits

Read the original

Dandan Liang, Soon Xin Ng and Lajos Hanzo, “Soft-Decision Star-QAM Aided BICM-ID,” IEEE Signal Processing Letters, vol. 18, no. 3, pp. 169–172, March 2011.

DOI: 10.1109/LSP.2011.2104951 · Author-hosted PDF: southampton.ac.uk

© 2011 IEEE. All visuals on this page are original redraws. Fig. 3 values are approximate readings of the published figure; Fig. 2 peak heights likewise. The live simulations are this page’s re-implementations of the set-up described in the paper, not the authors’ code.

Received signal (Eq. 3). yk = hk xk + nk The sent point is multiplied by the fading channel and noise of variance N0/2 per dimension is added.

Comparison (Eq. 4). If the channel barely changes, yk = pk yk−1 wk + ñk, ñk = nk − pk nk−1 wk with ring ratio pk ∈ {1, 0.5, 2}.

Effective noise (Eq. 8). Ñ0 = N0 (1 + |pk|²)

Likelihoods (Eqs. 9–10). Stay: P(yk | w, b3=0) = exp(−|yk − yk−1 w|² / 2N0) / (π · 2N0) Switch: the sum of the same expression for ×0.5 (noise 1.25 N0) and ×2 (noise 5 N0).

Bit probabilities (Eqs. 11–12). For a turn bit, add the likelihoods of every candidate turn with that bit set to 1 (both ring options), each weighted by the decoder’s hints about the other bits; compare with the candidates where it is 0. For the ring bit, add over all eight turns for “switch” versus “stay”.

Simulation set-up (Table I). Rate-3/4 code, memory 3, octal generators [4 4 4 4; 0 6 2 4; 0 2 5 6]; approximate Log-MAP decoding; 1,200 symbols per frame; 20,000 frames; correlated Rayleigh fading with normalised Doppler 0.01; Gray labelling for BICM (1 iteration), set partitioning for BICM-ID (1, 2, 4 iterations).

Topics in this iPaper

Engineering
Differentially detected noncoherent Star Quadrature Amplitude Modulation (Star-QAM) is ideal for low-complexity wireless communications

Abstract, page 169

quadrature amplitude modulation
16-level Star-QAM (16-StQAM) is based on two concentric 8-level Phase-Shift Keying (8PSK) constellations having two different amplitudes

Section I, page 169

noncoherent detection
differentially detected noncoherent Star-QAM was proposed in [5] in order to dispense with high-complexity CSI estimation

Section I, page 169

bit interleaved coded modulation
A sequence of 3-bit information symbols is encoded by a rate-3/4 BICM encoder

Section II, page 169

iterative decoding
Iterative detection is then carried out by exchanging extrinsic information between the 16-StQAM soft demapper and BICM decoder

Section II, page 169

rayleigh fading
when communicating over correlated Rayleigh fading channels

Section III, page 171

Soft-Decision Star-QAM Aided BICM-ID — Interactive explanation | iPaper.ai