Ferkans — Interactive Telecom Tutor

From $d_{\rm free}^2$ to $P_e$ : The Engineering Bottom Line

The design rules of TUngerboeck Design Rules maximize free Euclidean distance $d_{\rm free}^2$ ; the free-distance lower bound of TLower Bound on Free Euclidean Distance via Partition Levels explains why set partitioning helps. But an engineer ultimately cares about the bit-error rate $P_e$ at a given $E_s/N_0$ , and about whether the coding gain survives the transition from theory to real hardware.

This section closes the loop: we derive the asymptotic coding-gain formula $\gamma_c = d_{\rm free}^2 / d_{\rm uncoded}^2$ , apply it to the canonical 8-PSK and 16-QAM TCM families (finding the celebrated 3–6 dB figure), and then switch to engineering mode — the V.32 and V.34 modem standards, which turned TCM from a theoretical result into a line of products that carried much of the global internet traffic through the 1990s. We close with an honest comparison to BICM (Chapter 5), which eventually replaced TCM in wireless standards.

Definition:
Asymptotic Coding Gain

The asymptotic coding gain of a TCM scheme over an uncoded reference at the same spectral efficiency is

$\gamma_c \;\triangleq\; \frac{d_{\rm free}^2 / E_s^{\rm TCM}}{d_{\rm uncoded}^2 / E_s^{\rm unc}},$

where $E_s^{\rm TCM}$ and $E_s^{\rm unc}$ are the average symbol energies of the TCM constellation and the uncoded baseline, respectively. When both constellations are normalized to unit average energy, this simplifies to

$\gamma_c = d_{\rm free}^2 / d_{\rm uncoded}^2.$

Expressed in dB, $\gamma_c^{\rm dB} = 10 \log_{10}(d_{\rm free}^2 / d_{\rm uncoded}^2)$ .

"Asymptotic" means "at high SNR, ignoring the error-coefficient prefactor in the union bound." At moderate SNR the actual gain can be $0.1$ – $0.5$ dB smaller, because the multiplicity $N_{\rm free}$ of the dominant error event shifts the BER curve left by a constant amount.

,

Definition:
Shaping Gain (Preview)

The shaping gain $\gamma_s$ is a separate contribution to the gap to Shannon capacity, arising from using a non-uniform constellation (or equivalently a uniform constellation on a non-square region of signal space). For 2D constellations the ultimate shaping gain is $\gamma_s^{\rm max} = \pi e/6 \approx 1.53$ dB, achieved by circular or Voronoi-shaped constellations.

In TCM, the shaping gain is separate from the coding gain: a TCM scheme built on an Ungerboeck partition of a square 16-QAM has $\gamma_s = 0$ dB; to harvest the shaping gain one must additionally apply constellation shaping, as V.34 does via shell mapping.

This chapter treats only the coding gain. Chapter 4 (coset codes and lattice-based coded modulation) covers shaping in depth. For now, the reader should remember that the 3–6 dB coding gain of TCM and the $\leq 1.53$ dB shaping gain add (approximately) at the system level.

Asymptotic Coding Gain

The ratio of free Euclidean distance of a coded scheme to the minimum distance of an uncoded baseline at the same spectral efficiency, measured in dB. Governs the horizontal shift of the BER curve at high SNR.

Theorem: Asymptotic Coding Gain Formula

On AWGN with one-sided noise PSD $N_0$ , let a TCM scheme with free squared Euclidean distance $d_{\rm free}^2$ and multiplicity $N_{\rm free}$ be compared to an uncoded reference with minimum squared distance $d_{\rm uncoded}^2$ at the same spectral efficiency and average symbol energy. At high $E_s/N_0$ , the bit-error probability satisfies

$P_e^{\rm TCM} \;\sim\; N_{\rm free} \cdot Q\!\left(\sqrt{\frac{d_{\rm free}^2}{2 N_0}}\right), \qquad P_e^{\rm unc} \;\sim\; N_{\rm unc} \cdot Q\!\left(\sqrt{\frac{d_{\rm uncoded}^2}{2 N_0}}\right),$

and the horizontal shift between the two curves at any fixed target BER is

$\Delta \!\left(\frac{E_s}{N_0}\right) \;=\; 10 \log_{10}\!\frac{d_{\rm free}^2}{d_{\rm uncoded}^2} \;=\; \gamma_c^{\rm dB} \quad \text{(plus a small $\mathcal{O}(\log N_{\rm free})$ correction)}.$

Both BER curves are $Q$ -functions with arguments $\sqrt{d^2/(2N_0)}$ . At high SNR, the $Q$ -function behaves like $Q(x) \sim e^{-x^2/2}$ , so a change of $d^2$ by a factor $\gamma_c$ shifts the required $E_s/N_0$ by $10 \log_{10} \gamma_c$ dB. The multiplicity prefactors $N_{\rm free}$ and $N_{\rm unc}$ contribute an additional sub-dB correction that becomes negligible as SNR grows.

Show Hint

At high SNR, the union bound on $P_e$ is dominated by a single term: $P_e \approx N_{\rm free} Q(\sqrt{d_{\rm free}^2/(2N_0)})$ .

Solve $P_e^{\rm TCM}(SNR_c) = P_e^{\rm unc}(SNR_u)$ for the shift $SNR_u - SNR_c$ at any fixed target $P_e$ .

Proof

Union bound at high SNR

The exact bit-error probability for a TCM code on AWGN can be written as a sum over error events $P_e = \sum_{\text{events } e} N_e \cdot Q\!\left(\sqrt{d_e^2/(2N_0)}\right),$ where $d_e^2$ is the squared Euclidean distance of event $e$ and $N_e$ its frequency. At high SNR, the smallest $d_e^2 = d_{\rm free}^2$ dominates because $Q(x)$ decays faster than any polynomial — the term with the smallest distance is exponentially larger than all others. Hence $P_e^{\rm TCM} \;\sim\; N_{\rm free} \cdot Q\!\left(\sqrt{d_{\rm free}^2/(2N_0)}\right).$ The same argument applied to the uncoded reference (where the dominant "error event" is a single nearest-neighbour swap) gives $P_e^{\rm unc} \sim N_{\rm unc} \cdot Q(\sqrt{d_{\rm uncoded}^2/(2N_0)})$ .

Solve for the SNR shift

Set $P_e^{\rm TCM}(SNR_c) = P_e^{\rm unc}(SNR_u)$ at a fixed target BER. Using $Q(x) \sim e^{-x^2/2}/(\sqrt{2\pi} x)$ and taking $\log$ s, $\frac{d_{\rm free}^2}{2N_0^{\rm TCM}} = \frac{d_{\rm uncoded}^2}{2 N_0^{\rm unc}} + \log\!\frac{N_{\rm free}}{N_{\rm unc}} + \mathcal{O}(\log SNR).$ For equal $N_0$ , this gives $E_s^{\rm unc} = E_s^{\rm TCM} \cdot d_{\rm uncoded}^2 / d_{\rm free}^2$ ; the dB-scale shift is $\gamma_c^{\rm dB} = 10 \log_{10} \frac{d_{\rm free}^2}{d_{\rm uncoded}^2}. \qquad \blacksquare$

Multiplicity correction

The $\log(N_{\rm free}/N_{\rm unc})$ term gives an extra SNR penalty of about $0.1$ – $0.3$ dB for typical TCM codes (with $N_{\rm free}$ between $4$ and $32$ ). This is why Ungerboeck's Table I reports coding gains slightly smaller than $10 \log_{10} d_{\rm free}^2 / d_{\rm uncoded}^2$ : for the 4-state 8-PSK code, the asymptotic formula gives $3.01$ dB but the actual BER-curve shift at $10^{-6}$ is closer to $2.7$ dB.

, ,

Example: Asymptotic Coding Gain of 8-State 16-QAM TCM over Uncoded 8-PSK

The 8-state 16-QAM TCM ( $m = 3$ , $\tilde{m} = 2$ , $\nu = 3$ ) has $d_{\rm free}^2 = 20$ (from Ungerboeck Table II) when the 16-QAM is normalized to unit inter-point spacing ( $\Delta_0^2 = 4$ ). The uncoded reference at 3 bit/symbol is 8-PSK; when normalized to unit average energy it has $d_{\rm uncoded}^2 = 2 - \sqrt{2} \approx 0.586$ .

Compute the asymptotic coding gain, first in the normalization of Ungerboeck's table and then with both constellations rescaled to the same average symbol energy.

Solution

Compute average symbol energies

16-QAM in the $\pm 1, \pm 3$ normalization has $E_s^{\rm 16QAM} = \tfrac{1}{16}\sum_{a, b \in \{\pm 1, \pm 3\}} (a^2 + b^2) = 10.$ Unit-energy 8-PSK has $E_s^{\rm 8PSK} = 1$ .

Normalize distances to equal energy

Rescale 16-QAM so its average energy is $1$ : multiply every point by $1/\sqrt{10}$ . The free distance rescales as $d_{\rm free}^2 \to 20/10 = 2$ . Keep 8-PSK at unit energy: $d_{\rm uncoded}^2 = 2 - \sqrt{2} \approx 0.586$ .

Compute the asymptotic coding gain

$\gamma_c = \frac{d_{\rm free}^2}{d_{\rm uncoded}^2} = \frac{2}{2 - \sqrt{2}} = 2 + \sqrt{2} \approx 3.414.KATEXPLACEHOLDER0END\gamma_c^{\rm dB} = 10 \log_{10}(2 + \sqrt{2}) \approx 5.33 \text{ dB}.$ $Subtract the multiplicity correction (about$ 0.3 $dB for this code) to get the BER-curve shift at$ 10^{-6} $:$ \sim 5.0$ dB.

Comment

This is one of the celebrated 5–6 dB gains of TCM. The 64-state 16-QAM TCM from Table II (reached with $\tilde{m} = 2$ , $\nu = 6$ , carefully chosen generator polynomials) reaches $d_{\rm free}^2 = 24$ at unit energy — $\gamma_c^{\rm dB} = 10 \log_{10}(24/10 / 0.586) = 6.12$ dB. This is the "6 dB ceiling" that Ungerboeck advertised: the best 2D 16-QAM TCM with $\tilde{m} = 2$ coding cannot do better than about 6 dB, and reaching the upper end requires moving to 4D constellations (as V.34 did) or to BICM with modern LDPC codes (Chapter 9).

,

Asymptotic Coding Gain vs. Number of Trellis States

Asymptotic coding gain $\gamma_c$ (dB) of Ungerboeck TCM over the uncoded baseline at the same spectral efficiency, as a function of the number of trellis states $N_s \in \{4, 8, 16, 32, 64, 128, 256\}$ . The 8-PSK curve (2 bit/symbol, reference = QPSK) saturates near 6 dB; toggle the 16-QAM curve to see the 3 bit/symbol design (reference = 8-PSK) reaching similar gains. The curves are flat at the right because parallel transitions cap the gain unless $\tilde{m}$ is also increased — a key engineering insight from TLower Bound on Free Euclidean Distance via Partition Levels.

Parameters

Show 16-QAM TCM curve

⚠️Engineering Note

The V.32 Modem (1984): TCM's First Commercial Win

The V.32 dial-up modem standard (ITU-T Recommendation V.32, 1984) transmits 9600 bit/s over a 2400 baud ( $T_s^{-1} = 2400$ Hz) telephone line using an 8-state 32-CROSS TCM. The 32-CROSS constellation is a rotationally-invariant 32-point 2D constellation (selected by L.-F. Wei of AT&T to avoid phase-ambiguity problems during carrier recovery). The effective spectral efficiency is $9600 / 2400 = 4$ bit/symbol.

Over a typical 30 dB SNR phone line, V.32 achieved a raw BER around $10^{-5}$ — adequate for reliable data transfer under a protocol with ARQ retransmission. The coding gain over uncoded 16-QAM was about $4$ dB, which translated directly into robustness on noisy lines. Before V.32, the V.29 modem (1976) operated at the same 9600 bit/s but relied on a 16-QAM constellation with no coding — it often failed on marginal lines where V.32 succeeded. This was the first large-scale demonstration that coded modulation was a commercial necessity, not a theoretical curiosity.

Practical Constraints

•
Symbol rate $1/T_s = 2400$ baud, fixed by telephone-channel bandwidth
•
Constellation: 32-CROSS (rotationally invariant, 4 bit/symbol)
•
Trellis: 8-state TCM (Wei 1984 rotationally-invariant construction)
•
Coding gain over uncoded 16-QAM: $\sim 4$ dB

📋 Ref: ITU-T V.32 (1984)

,

🚨Critical Engineering Note

The V.34 Modem (1994): 33.6 kbps Through TCM + Shell Mapping

V.34 (ITU-T Recommendation V.34, 1994) pushed the same telephone line to 33.6 kbit/s — roughly the Shannon limit of the public switched telephone network at typical SNR levels (about 35 dB). Achieving this required combining several techniques into a single system:

4D TCM. A 4-dimensional constellation (two consecutive 2D symbols treated as one 4D point) with a 16-state trellis code gave about 5 dB of coding gain.
Shell mapping. A constellation-shaping algorithm (Laroia, Farvardin, Tretter, 1994) reshaped the 4D constellation's amplitude distribution toward a Maxwell-Boltzmann density, harvesting about 1 dB of the 1.53 dB shaping gain ceiling.
Non-linear precoding. For the highest-rate modes, a non-linear precoder compensated for the telephone-line echo path.
Adaptive rate selection. V.34 supports 28 symbol rates between 2400 and 3429 baud and 59 constellation sizes, chosen per-call via a training-based MCS (modulation and coding scheme) selection.

V.34 modems dominated internet access from 1994 until the DSL and cable-modem rollout of the late 1990s. For most users, this was the fastest digital communications service in their lives at the time — and TCM was the single largest contributor to the gain over the earlier V.29/V.32 modems.

Practical Constraints

•
Line bandwidth: $\sim 3$ kHz, SNR $\sim 35$ dB
•
Shannon capacity at these conditions: $\sim 35$ kbit/s (V.34 reaches 33.6 kbit/s, within 1 dB)
•
Trellis: 16-state 4D TCM (Wei 1987 construction)
•
Shaping: 16-dimensional shell mapping (~1 dB of shaping gain)

📋 Ref: ITU-T V.34 (1994) and revised 1998

,

Historical Note: From TCM to BICM: A Change of Channel

1982–1998

TCM was built for AWGN-dominant channels — the telephone line, the satellite downlink, point-to-point microwave. Its coding gain depends on Euclidean distance in signal space, which is exactly the right currency when the channel is (approximately) AWGN and static.

The wireless channel, however, is neither AWGN nor static: Rayleigh fading makes the received amplitude a random variable, and on a time-varying mobile channel each transmitted symbol experiences a different channel realization. In this setting, Euclidean distance is no longer the right figure of merit; what matters is the diversity order, the number of independent fading coefficients that must all fade simultaneously for the code to fail.

Caire, Taricco, and Biglieri's 1998 paper "Bit-Interleaved Coded Modulation" (IEEE Transactions on Information Theory) showed that on fading channels a binary convolutional code + a bit interleaver + a Gray-labelled constellation mapper outperforms TCM, because the bit interleaver maximizes the effective diversity order of the code. This single paper, with over 4000 citations today, shifted the wireless standards community from TCM-based designs (GSM's 1990s era) to BICM-based designs (everything from W-CDMA onward through 5G NR).

TCM is not obsolete — it still wins on AWGN-dominant links where the fading assumption is not operative (satellite, cable, deep-space probes, even probabilistic-shaping systems that reuse the set-partitioning idea) — but for wireless, BICM supplanted it in the late 1990s. The irony: Caire's first major result was to kill the dominant paradigm of his advisor's generation. We will cover BICM in detail in Chapter 5.

,

Why This Matters: Where TCM Ideas Live On in Modern Wireless

Although BICM (Chapter 5) dominates wireless PHY layers, Ungerboeck's set-partitioning idea has a second life in probabilistic amplitude shaping (PAS) — the technique used by all high-SE modern DVB-S2X and optical fiber transponders to harvest the shaping gain on top of a BICM + LDPC system.

PAS works as follows: a distribution matcher maps uniform bits to non-uniform amplitudes (Maxwell-Boltzmann distribution), a binary code protects the sign bits, and the result is mapped onto a QAM constellation via a set-partitioning-style labeling — high-order bits are uncoded and selected by the shaping source; low-order bits are coded and determine which high-MSED coset. This is structurally the same split as Ungerboeck's rule (R1)/(R2), redirected from coding-gain harvesting to shaping-gain harvesting.

The takeaway: Ungerboeck's partition idea is still alive; the trellis code + Viterbi side of TCM has largely been replaced by LDPC

sum-product decoding, but it survives in legacy systems and as a pedagogical benchmark.

See full treatment in Chapter 5

Forward Reference: BICM (Chapter 5)

In Chapter 5 we will derive the BICM (Bit-Interleaved Coded Modulation) framework of Caire, Taricco, and Biglieri (1998) — the CommIT group's foundational contribution to coded modulation theory. BICM replaces the joint code-and-mapping design of Ungerboeck with a binary code, a bit interleaver, and a Gray-labelled mapper. On AWGN, BICM with Gray labeling is slightly worse than TCM (about 0.2–0.5 dB gap in BICM capacity vs. CM capacity at moderate $\text{SNR}$ ); on fading channels, BICM substantially outperforms TCM thanks to the interleaver's diversity-exploiting action.

For this chapter, however, the reader should leave with a crisp picture of Ungerboeck-style TCM: partition, label, code the coset bits, Viterbi-decode. Every idea in Chapters 3–9 either builds on it (MLC, coset codes) or explicitly departs from it (BICM).

Common Mistake: Overselling 'No Bandwidth Expansion'

Mistake:

Reading Ungerboeck's "coding gain without bandwidth expansion" slogan as implying that TCM is strictly better than any binary-coded scheme.

Correction:

The slogan is correct but narrow in scope. It says TCM achieves a 3–6 dB gain at the same baud rate and bit rate. It does not say:

TCM is better than BICM on fading channels (it is not).
TCM is better than modern LDPC+QAM at spectral efficiencies $\geq 8$ bit/symbol (it is not; LDPC approaches Shannon capacity to within $< 0.5$ dB at those rates).
TCM is optimal for short blocklengths (it is not; short-blocklength coding is its own problem — see Chapter 22).

TCM was the right answer for the bandwidth-limited AWGN channels of 1982–1994. Later channels and better codes changed the answer.

Quick Check

A TCM code with $d_{\rm free}^2 = 12$ is compared to an uncoded constellation with $d_{\rm uncoded}^2 = 3$ at the same average symbol energy and the same spectral efficiency. What is the asymptotic coding gain in dB?

$3$ dB

$6$ dB

$10$ dB

$12/3 = 4$ dB

Correction:

6

dB

$\gamma_c = d_{\rm free}^2 / d_{\rm uncoded}^2 = 12/3 = 4 \;\longrightarrow\; \gamma_c^{\rm dB} = 10 \log_{10} 4 \approx 6.02$ dB. This is at the upper end of what 2D TCM typically achieves and corresponds to a carefully-designed 64-state 16-QAM TCM.

Quick Check

V.34 achieves 33.6 kbit/s over a 3 kHz, 35 dB SNR telephone line. The Shannon capacity of this channel is $W \log_2(1 + \text{SNR}) = 3000 \log_2(1 + 10^{3.5}) \approx 34.9$ kbit/s. How far is V.34 from Shannon, and where does the remaining gap come from?

About 10 dB — the rest is the uncoded-modulation loss.

About 1 dB — the remaining gap comes from the finite-blocklength penalty, shaping incompleteness, and decoder suboptimality.

V.34 exceeds Shannon — the capacity formula does not apply to non-Gaussian channels.

About 5 dB — this is the TCM coding gain, which has not yet been reached by V.34.

Correction:

About 1 dB — the remaining gap comes from the finite-blocklength penalty, shaping incompleteness, and decoder suboptimality.

V.34 reaches $\sim 33.6/34.9 = 96\%$ of capacity in bit rate, or equivalently about 1 dB in required SNR. The gap comes from: (i) the shell-mapping shaping gain is $\sim 1$ dB, not the full $1.53$ dB ultimate; (ii) 4D TCM's 5 dB coding gain is about 1 dB short of the best possible with ML coding at the same complexity; (iii) finite-blocklength + synchronization overhead. V.34 was considered a minor miracle at the time.

Quick Check

Increasing the number of states of an 8-PSK TCM from 64 to 256 yields less than 0.5 dB of additional coding gain. Why does the return on adding states diminish?

Because 8-PSK is inherently bandwidth-limited.

Because parallel transitions cap $d_{\rm free}^2$ at $\Delta_{\tilde{m}+1}^2$ , which does not scale with the number of states.

Because the Viterbi algorithm becomes unstable at high state counts.

Because the error-coefficient $N_{\rm free}$ grows, canceling the distance increase.

Correction:

Because parallel transitions cap

d_{\rm free}^2

at

\Delta_{\tilde{m}+1}^2

, which does not scale with the number of states.

Once $d_{\rm free}^{(H)} \cdot \Delta_1^2 > \Delta_2^2$ (which happens around 32 states for 8-PSK), parallel transitions become the bottleneck: $d_{\rm free}^2 = \Delta_2^2 = 4$ , a constant. Further increasing the states helps the longer-event term only, which is already above the constant ceiling. To break out of the ceiling one must use $\tilde{m} = 2$ (deeper partition level) or a larger constellation — but that changes the design, not just the state count.

Key Takeaway

TCM in one paragraph. Ungerboeck's 1982 idea was to combine a convolutional code and a higher-order modulation into a joint trellis code whose coding gain comes from Euclidean-distance partitioning, not from Hamming-distance coding. The result is 3–6 dB of gain on AWGN without bandwidth expansion, confirmed by the V.32/V.34 modem standards that dominated dial-up internet for a decade. On fading (wireless) channels, BICM (Chapter 5) supplanted TCM because the bit interleaver converts the code's Hamming distance into diversity order. On modern AWGN-dominant links at very high spectral efficiency, LDPC + probabilistic shaping has closed the remaining 1–2 dB gap to Shannon. But the partition idea — split bits into coded and uncoded by their geometric robustness — lives on, now inside probabilistic amplitude shaping in DVB-S2X and optical fiber transponders.

Performance Analysis and the V.32/V.34 Modem Standards

From dfree2d_{\rm free}^2dfree2​ to PeP_ePe​: The Engineering Bottom Line

Definition: Asymptotic Coding Gain

Definition: Shaping Gain (Preview)

Asymptotic Coding Gain

Theorem: Asymptotic Coding Gain Formula

Union bound at high SNR

Solve for the SNR shift

Multiplicity correction

Example: Asymptotic Coding Gain of 8-State 16-QAM TCM over Uncoded 8-PSK

Compute average symbol energies

Normalize distances to equal energy

Compute the asymptotic coding gain

Comment

Asymptotic Coding Gain vs. Number of Trellis States

Parameters

The V.32 Modem (1984): TCM's First Commercial Win

The V.34 Modem (1994): 33.6 kbps Through TCM + Shell Mapping

Historical Note: From TCM to BICM: A Change of Channel

Why This Matters: Where TCM Ideas Live On in Modern Wireless

Forward Reference: BICM (Chapter 5)

Common Mistake: Overselling 'No Bandwidth Expansion'

Quick Check

Quick Check

Quick Check

Key Takeaway

From $d_{\rm free}^2$ to $P_e$ : The Engineering Bottom Line

Definition:
Asymptotic Coding Gain

Definition:
Shaping Gain (Preview)