Make your paper interactiveCreate

Text from the interactive explanation

Near-Capacity Turbo Coded Soft-Decision Aided DAPSK/Star-QAM for Amplify-and-Forward Based Cooperative Communications

Dandan Liang, Soon Xin Ng, Lajos Hanzo

Original paper

Send a message

Why it twists

Read the change

One ring or four?

Weigh, don't decide

Two experts

Closer to the limit

A relay that doesn't think

The answer

Liang, Ng & Hanzo · IEEE Trans. Commun. · 2013

Can a message get through a radio channel that nobody ever measures — even when a helper simply shouts it onward — and how close to the theoretical limit can it get? Start by sending one. Each character becomes one point of a 64-point star (6 bits), and two writers send it through the same twisting channel.

The star as the channel twists it now

Position reader's measured map

Channel twist h

Received points; the turn + hop between them

Illustrative, uncoded simulation built from the paper's equations: correlated Rayleigh fading, with the same channel and the same noise for both writers. The change reader picks the most likely change with the paper's soft demapper (Eq. 12) and never estimates the channel.

Chapter 1

Why the signal twists

A radio signal reaches a phone along many paths at once: straight from the mast and bounced off buildings. Each echo arrives as a small wave with its own timing. Add the echoes like arrows, head to tail, and you get the signal the phone actually receives — the magenta arrow h.

As the phone moves, every echo's arrow turns at its own rate. So h keeps rotating (the twist ) and growing or shrinking (the strength ). When the arrows happen to cancel, the signal all but vanishes: a deep fade .

Try it. One echo: steady strength, only the twist turns. Two echoes: regular dips. Six or 24: random deep fades — Rayleigh fading, the channel the paper studies.

—

Watch the last number: from one symbol to the next, the twist barely changes. The paper's channel changes slowly from symbol to symbol (normalised Doppler frequency fd = 0.01). That small loophole is the whole idea.

Chapter 2

Read the change, not the position

Most receivers read the position of each received point. To do that they must first measure the twist — "channel estimation" — which costs effort and fails when things change fast.

Here the transmitter uses an 8-position dial. You see the previous symbol (hollow) and the current one (filled), but the channel has turned the whole picture by an amount you don't know. Answer both questions.

Start with either question.

Positions 0/5 · Turns 0/5 — 0 of 5 rounds played

(a) Where on the transmitter's dial was the current symbol sent?

(b) How far did it turn since the previous one?

Reading the change works because both points suffered almost the same twist. The paper assumes exactly this: hk ≈ hk−1, which it notes is needed to avoid an "error floor" in differential detection. Turn on the faster world and even the change starts to fail.

Chapter 3

One ring or four rings?

To carry 6 bits per symbol, the receiver must tell apart 64 different changes. 64-DPSK puts all 64 directions on one ring and uses only the turn: neighbouring turns are just 5.6° apart. 64-DAPSK(4,16) , also called Star-QAM , uses 4 rings × 16 directions: each symbol hops between rings (2 bits) and turns (4 bits), so turns can be 22.5° apart.

Each dot below is one received change, z = yk / yk−1: its angle is the turn, its distance from the centre is the ring ratio (on a log scale, so ×1.4, ×1.96 … are evenly spaced). Faint lines are decision boundaries. Cyan = read correctly, red = mistake.

Same channel, same noise, same 6 bits per symbol. Which will make fewer mistakes?

Symbol errors: ?

Symbol errors: ?

The wrap-around. The ring index moves in a cycle: bk = (bk−1 + fk) mod 4. So "one ring out" from the outermost ring lands on the innermost one. The same 2-bit message, +1 ring, can arrive as ×1.4 or as ÷2.74 (that is, 1.4³). A good receiver has to consider both — which is exactly what the paper's formula does in the next chapter.

These are uncoded symbol errors from an illustrative simulation (correlated Rayleigh fading, simple decisions on the change plane): raw mistakes before any error correction. Push the signal strength up and 64-DPSK hits an error floor — the twist's own drift is enough to cross its narrow 5.6° boundaries. The paper adds a turbo code; that comes next.

Chapter 4 · the heart of the paper

Weigh, don't decide

An error-correcting code can fix mistakes — but only if the receiver tells it how sure it is about every bit. This is the paper's core derivation. Start from the previous received point yk−1. Every possible change predicts where the current point should land: yk−1 × ring ratio × turn. Because of the wrap-around there are 112 predictions (7 ring ratios × 16 turns) for 64 messages, so each colour except white appears on two rings. Each prediction is fuzzy, and outer ones are fuzzier: the reference yk−1 is itself noisy, and scaling it up scales its noise too (Eq. 10).

Drag the amber point (or focus the picture and use the arrow keys). Brightness shows how likely each prediction is.

same ring

+1 ring

+2 rings

+3 rings

Most likely change

—

In hard mode the receiver keeps only its single best guess, so every bit looks certain — even a coin flip. That lost confidence is exactly what the turbo code needs. Notice the "torn" case: thanks to Gray labelling, neighbouring turns differ in just one bit, so only one bar is unsure.

Earlier soft receivers either made hard decisions or used a metric called BMIAD that assumes the noise is negligible — fine at high signal strength, but near-capacity systems work where noise matters. The authors extended an exact-noise soft demapper from two rings to any number of rings. In their simulations of 16-DAPSK(2,8) it beat BMIAD by 0.3 dB (Fig. 3), and a second round of demapper–decoder exchange added about 0.2 dB (Fig. 4).

yk ≈ (ak/ak−1) · yk−1 · wk + ñk (7)

The current point is the previous one, scaled by the ring ratio and turned by wk, plus a combined noise ñk. No channel term appears: it cancelled because hk ≈ hk−1.

N0(q) = N0 (1 + α2q) (10)

The combined noise grows with the ring ratio αq, because the noisy reference is scaled by it too.

P(yk | dk, fk) ≈ Σq : q mod 4 = fk exp( −|yk − yk−1 αq μ(dk)|² / N0(q) ) (12)

For each message (ring step f, turn d), add up every way it could show up — for example ×1.4 or ÷2.74 for "+1 ring" — each judged with its own noise. The bit confidences are log-likelihood ratios, L = ln( P(bit = 0) / P(bit = 1) ), summed over the 64 messages. The picture above computes exactly this.

Chapter 5

Two experts in conversation

The receiver has two experts. The demapper reads the received changes. The turbo decoder knows the rules of the code — a half-rate code, so each symbol's 6 coded bits carry 3 message bits. They take turns, each passing on only what is new to the other ("extrinsic" information).

An EXIT chart predicts whether that conversation reaches certainty. The amber curve is the demapper: how much it can tell back (up) given what it already knows (across). The white curve is the decoder, drawn with its axes swapped. The staircase is the conversation. If a tunnel stays open between the curves, the staircase climbs to the top-right corner and errors vanish. If the curves meet, it gets stuck.

The shaded area under the demapper's curve is roughly the capacity of the signal set. A narrow tunnel means the code is closely matched to it — working near capacity.

Demapper (inner)

Decoder (outer)

—

The paper's reading of Fig. 5 at 16 dB: the tunnel is open for the turbo code with Gray-labelled 64-DAPSK, closed for 64-DPSK; set-partitioning labelling makes the curves intersect; and a convolutional code of the same complexity does not match the DAPSK curve. 64-QAM gets through with room to spare — with perfect channel knowledge it is far from its limit at this SNR.

Chapter 6

Closer to the limit

These are the paper's computer-simulation results for the full system — turbo code, soft demapper and iterations — over correlated Rayleigh fading with fd = 0.01. The target is a bit error rate of 10⁻⁵: one wrong bit in 100 000. Further left means less power needed; +3 dB is roughly double the power.

64-QAM — needs perfect channel knowledge

64-DAPSK (4,16) — no channel estimate

64-DPSK — no channel estimate

—

2.3 dB 64-DAPSK beats 64-DPSK — same throughput, neither measures the channel.

4.6 dB Coherent 64-QAM, which needs perfect channel knowledge, is ahead at a 400-symbol block: the price of never measuring.

5.25 dB gained by 64-DAPSK when the turbo block grows from 400 to 40 000 symbols.

How far from the limit?

Here capacity means the most bits per symbol this signal set could carry at a given SNR with this kind of receiver (the paper's "DCMC" capacity) — not an unconstrained Shannon limit. Each cross marks the SNR the turbo-coded scheme needed to reach 10⁻⁵ at its throughput: half the bits per symbol, because the code is half-rate.

—

Why longer blocks help. A deep fade wipes out a run of neighbouring symbols. In a short block, one fade can hit a big share of everything the code has to work with. In a long block the damage is a small fraction, spread (by the turbo code's interleaver) among many healthy symbols the code can lean on.

2.25 dB is about 1.7× the minimum power this signal set would need; 7.56 dB is about 5.7×.

Chapter 7

A relay that doesn't think

Now add a helper halfway between source and destination. In time slot 1 the source broadcasts and both the relay and the destination listen. In slot 2 the relay simply amplifies what it heard — signal and noise alike — and forwards it. It doesn't decode, and it doesn't measure the channel: its amplification factor, β = 1/√(N0(γ̄sr + 1)), needs only the average signal strength γ̄sr.

Each hop is half the distance, so in the paper's model it gets 4× the power (about +6 dB). The destination sees the two hops' twists multiplied together — still a slowly changing twist, so reading the change still works. Its demapper just allows for more noise (Eq. 23 adds up four noise terms). The paper does not rely on the relay knowing anything about the channel — it argues it is unrealistic to expect a relay to "altruistically" estimate it.

Star insets: illustrative simulation of what each node receives (64-DAPSK points through independent, slowly fading hops). The chart uses the paper's results.

—

In the paper's simulations, with a 40 000-symbol turbo block, the amplify-and-forward relay saved about 4.5 dB (Fig. 8): the link works with roughly a third of the power.

Chapter 8

Back to the opening question

TwistFading twists every symbol, but neighbours share almost the same twist (the paper simulates fd = 0.01).

ChangeSo write the message in the change: a hop between 4 rings plus a turn — 64-DAPSK(4,16).

WeighWeigh every possible change instead of deciding (Eqs. 10–12): 0.3 dB better than the earlier BMIAD metric for 16-DAPSK(2,8) (Fig. 3).

TalkLet the demapper and a turbo decoder talk: the EXIT tunnel is open at 16 dB (Fig. 5), and 64-DAPSK beats 64-DPSK by about 2.3 dB (Fig. 6).

LimitWith 40 000-symbol blocks it runs within 2.25 dB of this signal set's capacity (Fig. 7) — about 1.7× the minimum power.

RelayA relay that only amplifies saves about another 4.5 dB (Fig. 8) — roughly a third of the power.

So — in the paper's simulations — yes: a message gets through a channel that nobody measures, within about 2.25 dB of what this signal set allows, and a relay that just shouts it onward cuts the power needed by another 4.5 dB. The price is visible too: coherent 64-QAM with perfect channel knowledge needs 4.6 dB less power at short blocks.

What it brings closer: links and relays — possibly other people's devices — that never measure the channel, yet run near their signal set's capacity. That matters most where channels change too fast to measure, or where a relay can't or won't do the work. The authors' stated next step is to reduce the receiver's complexity.

Sources & credits

Dandan Liang, Soon Xin Ng and Lajos Hanzo, "Near-Capacity Turbo Coded Soft-Decision Aided DAPSK/Star-QAM for Amplify-and-Forward Based Cooperative Communications," IEEE Transactions on Communications, vol. 61, no. 3, pp. 1080–1087, 2013. DOI: 10.1109/TCOMM.2012.122712.110786. Author-hosted copy: southampton.ac.uk/~sxn/paper/tcom2013-liang.pdf.

Charts redrawn from approximate values read off Figs. 5–8 of the paper (original figures © 2013 IEEE). All published results are computer simulations.

Interactive simulations are simplified, uncoded illustrations built from the paper's equations; the published results come from the authors' turbo-coded simulations.

Fonts: Fraunces, Space Grotesk and JetBrains Mono — SIL Open Font License.

Topics in this iPaper

Engineering
Index Terms—Soft-decision, iterative detection, DAPSK, turbo coding, amplify-and-forward (AF) relaying protocol, correlated Rayleigh fading channel, near-capacity transceivers.

Abstract and Index Terms, page 1080

differential modulation
Multilevel Differential Amplitude and Phase-Shift Keying (DAPSK) schemes do not require any channel estimation, which results in low complexity.

Abstract, page 1080

turbo codes
Turbo Coding (TC) [12] is employed in our DAPSK scheme in order to achieve further improvements.

Section I Introduction, page 1080

iterative detection
we exchange extrinsic information between the M-DAPSK soft demapper and the TC decoder to accomplish iterative detection.

Section II, page 1081

extrinsic information transfer chart
an open EXIT chart tunnel implies having an infinitesimally low BER

Section III and Fig. 5, page 1085

cooperative communications
According to the AF relaying protocol, the Relay Node (RN) amplifies the signal received from the Source Node (SN) and forwards it to the Destination Node (DN)

Section II-C Amplify-and-Forward, page 1083

rayleigh fading
Correlated Rayleigh fading channel having a normalised Doppler frequency of 0.01

Table I, page 1084

Near-Capacity Turbo Coded Soft-Decision Aided DAPSK/Star-QAM for Amplify-and-Forward Based Cooperative Communications — Interactive explanation | iPaper.ai