Quadrature Amplitude Modulation: A Working Engineer’s Tutorial
# Quadrature Amplitude Modulation: A Working Engineer’s Tutorial
*Written for the engineer who is new to RF — including the one who has never knowingly looked at a modulated signal.*
Last revised: September 2026
—
## A Note Before We Start
Three things you should know about how this document was made.
**First, the sourcing.** Every historical claim and every specification number below is footnoted. Where I could reach primary sources — the original IRE and AIEE papers, ITU-T Recommendations, CableLabs specifications — I used them in preference to secondary accounts.
**Second, the mathematics.** Some of what follows is arithmetic I performed myself. Every time that happens, I show the working and label it as mine, so you can check it rather than trust it. Where my arithmetic reproduces a published value, I say so, because that agreement is evidence. Where I am quoting a standard textbook result rather than deriving it, I say that too.
**Third, a confession about the subject.** QAM is one of those topics where the underlying idea is genuinely simple — simple enough to explain properly in two pages — and where nearly all the difficulty lives in the gap between the idea and a working receiver. I have tried to give both honestly. If Part III feels easy and Part IV feels hard, that is not a failure of the writing. That is the actual shape of the subject.
Read this with a piece of test equipment nearby if you can. A constellation display teaches things that no amount of prose will.
—
## Table of Contents
**Part I — What QAM Is**
1. [The Problem: Getting Digits Across a Channel](#1-the-problem-getting-digits-across-a-channel)
2. [Carriers, and the Three Things You Can Change](#2-carriers-and-the-three-things-you-can-change)
3. [The Quadrature Trick](#3-the-quadrature-trick)
4. [The Constellation](#4-the-constellation)
**Part II — The History**
5. [Foundations: Nyquist, Hartley, Shannon](#5-foundations-nyquist-hartley-shannon)
6. [Analog Quadrature Came First: Color Television](#6-analog-quadrature-came-first-color-television)
7. [1960–1962: Digital QAM Is Born](#7-19601962-digital-qam-is-born)
8. [The Modem Era](#8-the-modem-era)
9. [Broadcast, Cable, and the Standards That Followed](#9-broadcast-cable-and-the-standards-that-followed)
10. [QAM Inside OFDM](#10-qam-inside-ofdm)
**Part III — The Mathematics**
11. [The Signal, Written Down](#11-the-signal-written-down)
12. [Constellation Geometry and Power](#12-constellation-geometry-and-power)
13. [Gray Coding](#13-gray-coding)
14. [Bandwidth: Nyquist and the Raised Cosine](#14-bandwidth-nyquist-and-the-raised-cosine)
15. [Noise, Distance, and the 6 dB Rule](#15-noise-distance-and-the-6-db-rule)
16. [The Shannon Limit and How Close QAM Gets](#16-the-shannon-limit-and-how-close-qam-gets)
**Part IV — Practice**
17. [The Transmitter Chain](#17-the-transmitter-chain)
18. [The Receiver Chain](#18-the-receiver-chain)
19. [Measuring Quality: MER, EVM, CNR](#19-measuring-quality-mer-evm-cnr)
20. [What Breaks QAM](#20-what-breaks-qam)
21. [Adaptive Modulation](#21-adaptive-modulation)
22. [Traps and Misconceptions](#22-traps-and-misconceptions)
23. [Glossary](#23-glossary)
24. [References](#24-references)
—
# Part I — What QAM Is
## 1. The Problem: Getting Digits Across a Channel
Start with what you actually have.
You have a channel — a coaxial cable, a fiber, a slice of radio spectrum, a telephone pair. It occupies some band of frequencies. It attenuates what you send, adds noise, and distorts the signal in various ways. That is the whole physical reality.
You have some bits you want to move across it.
Your job is to turn bits into a physical waveform that survives the channel, and then turn the received waveform back into the same bits. Every modulation scheme ever invented is an answer to that question. QAM is a particularly good answer, which is why it is nearly everywhere.
### Why not just send the bits directly?
You could put a voltage on the wire — high for 1, low for 0. This is called **baseband** signaling, and it works fine over short distances.
It fails for most real channels for two reasons.
**The channel may not pass DC or low frequencies.** A cable television plant has amplifiers and transformers that block DC. A radio channel physically cannot radiate efficiently at low frequencies without an enormous antenna. Your square-wave baseband signal has energy right down to DC, and the channel throws that energy away.
**You want to share the channel.** If every user’s signal occupies DC to some upper limit, only one user can use the wire. Moving different users to different frequency bands is how cable carries hundreds of channels on one coax, and how radio carries thousands of stations in one atmosphere.
The fix for both is the same. Take your information and use it to alter some property of a **carrier** — a sine wave at a frequency the channel likes.
## 2. Carriers, and the Three Things You Can Change
A sine wave has exactly three properties you can vary:
$$s(t) = A \cos(2\pi f_c t + \phi)$$
– **A** — the amplitude
– **f_c** — the frequency
– **φ** — the phase
That gives you the three classical families of modulation. Vary the amplitude and you have **AM** (in digital form, amplitude-shift keying). Vary the frequency and you have **FM** (frequency-shift keying). Vary the phase and you have **PM** (phase-shift keying).
This list is exhaustive. There is nothing else about a sine wave to change.
### Why not use just one?
Each one alone leaves capability on the table.
Pure amplitude modulation puts all your constellation points on a line. To send more bits per symbol you need more amplitude levels, packed closer together, and they become progressively easier for noise to confuse — while your peak power grows.
Pure phase modulation puts your points on a circle. This is elegant, and it has a real advantage: constant envelope, which amplifiers love. But a circle has limited room. As you add points, the arc between adjacent points shrinks, and beyond about eight points, phase-shift keying becomes inefficient in its use of power.
The insight — and it took the field until 1960 to state it clearly — is that **amplitude and phase are independent**. You can vary both at once. Instead of points on a line, or points on a circle, you get points on a **plane**. A plane holds far more points at a given minimum separation than a line or a circle does.
That is QAM.
## 3. The Quadrature Trick
Here is the piece of mathematics that makes it practical, and it is worth understanding properly because everything else follows from it.
Take two carriers at the **same frequency**, one a cosine and one a sine. They are 90° apart — in **quadrature**. Amplitude-modulate each one independently, and add them together:
$$s(t) = I(t)\cos(2\pi f_c t) – Q(t)\sin(2\pi f_c t)$$
**I** is the *in-phase* component. **Q** is the *quadrature* component.
### Why the two don’t interfere
The magic is that sine and cosine at the same frequency are **orthogonal**. Over any whole number of cycles, the integral of their product is zero.
That means a receiver can separate them cleanly. Multiply the received signal by $2\cos(2\pi f_c t)$ and low-pass filter, and you recover $I(t)$ — the Q term integrates away to nothing. Multiply by $-2\sin(2\pi f_c t)$ and low-pass filter, and you recover $Q(t)$.
**Two independent data streams travel in the same frequency band at the same time, and they do not interfere.** You have doubled the information carried in a given bandwidth, using nothing but a 90° phase shift.
This is the whole idea. Everything after this is refinement.
### The equivalent view
There is a second way to see the same signal, and switching between them fluently is a mark of someone who really understands this.
Any combination of $I\cos(\omega t) – Q\sin(\omega t)$ can be rewritten as a single cosine with an amplitude and a phase:
$$s(t) = A\cos(2\pi f_c t + \phi), \quad \text{where } A = \sqrt{I^2 + Q^2} \text{ and } \phi = \arctan(Q/I)$$
So “amplitude-modulating two quadrature carriers” and “jointly modulating the amplitude and phase of one carrier” are **the same operation described two ways**. I/Q is the rectangular description; amplitude/phase is the polar description.
This is not a coincidence — it is the historical origin of the name. When Campopiano and Glazer introduced the square constellation in 1962, they described their system as essentially amplitude modulation of “two carriers that have the same frequency but are in quadrature with each other” — the first time combined amplitude-and-phase modulation had been framed as amplitude modulation on quadrature carriers [4].
The I/Q view is the one that dominates in practice, for a very practical reason: **it is what the hardware does.** A modern transmitter has an I DAC and a Q DAC feeding a quadrature mixer. A modern receiver has a quadrature downconverter producing I and Q samples. You will spend your career thinking in I/Q because that is what the silicon gives you.
## 4. The Constellation
Plot I on the horizontal axis and Q on the vertical. Every symbol the transmitter can send becomes a **point** on that plane. The set of all such points is the **constellation**.
For **16-QAM**, the classic square arrangement uses four levels on each axis — conventionally the odd integers −3, −1, +1, +3 — giving 4 × 4 = 16 points [32].
Count the information: 16 possible points means each symbol carries $\log_2 16 = 4$ bits.
The general rule, and it is the single most useful fact in this document:
> **A constellation with M points carries log₂(M) bits per symbol.**
| Constellation | Points (M) | Bits per symbol |
|—|—|—|
| BPSK | 2 | 1 |
| QPSK (= 4-QAM) | 4 | 2 |
| 16-QAM | 16 | 4 |
| 64-QAM | 64 | 6 |
| 256-QAM | 256 | 8 |
| 1024-QAM | 1024 | 10 |
| 4096-QAM | 4096 | 12 |
Notice how slowly the right column grows. **Doubling your bits per symbol requires squaring the number of constellation points.** Going from 64-QAM to 4096-QAM — a 64-fold increase in constellation density — only doubles your throughput. This diminishing return is the central economic fact of high-order modulation, and §15 will show you what it costs in signal quality.
### What the receiver actually does
The receiver measures I and Q, which lands it somewhere on the plane — never exactly on an ideal point, because of noise. It then asks: *which ideal constellation point is nearest?* That decision is called **slicing**, and the answer is the received symbol, which maps back to bits.
So the entire game is: **keep the received points close enough to where they should be that the nearest-neighbour decision is right.** Every impairment in Part IV is a different way of failing at that.
—
# Part II — The History
I want to give you this properly, because the sequence explains why things are the way they are, and because the field’s habit of naming things after their inventors becomes much less intimidating once you know who those people were.
## 5. Foundations: Nyquist, Hartley, Shannon
Before anyone could invent QAM, someone had to work out what a channel can carry.
**Harry Nyquist, 1924.** In “Certain Factors Affecting Telegraph Speed,” Nyquist was among the first to treat what telegraphers sent as *information*, and identified the factors setting the maximum speed of transmission [6][41].
**Harry Nyquist, 1928.** “Certain Topics in Telegraph Transmission Theory,” presented at the AIEE Winter Convention in February 1928 and published in the *Transactions of the AIEE*, volume 47, pages 617–644 [5][41]. This paper introduced the class of pulse shapes now called **Nyquist pulses** — shapes that carry symbols without the symbols smearing into each other [41]. §14 is essentially an exposition of this paper’s consequences.
**Ralph Hartley, 1928.** “Transmission of Information,” *Bell System Technical Journal*, July 1928 [7].
**Claude Shannon, 1948.** “A Mathematical Theory of Communication,” *Bell System Technical Journal*, volume 27, pages 379–423 (July) and 623–656 (October) [8]. Shannon cited Nyquist’s 1924 and 1928 papers and Hartley’s 1928 paper in his opening [8]. He followed it with “Communication in the Presence of Noise” (*Proceedings of the IRE*, 1949) [9].
Shannon established the ceiling. Everything that follows is engineering trying to reach it.
## 6. Analog Quadrature Came First: Color Television
Here is a fact that surprises people, and it is worth knowing because it corrects a common misconception about where QAM came from.
**Quadrature amplitude modulation was deployed in the United States in 1953 — seven years before anyone published a digital version of it.**
In December 1953 the FCC approved the NTSC color television standard. The problem the standard had to solve was elegant and brutal: add color to a television signal without breaking the millions of black-and-white receivers already in homes.
The solution was to encode color as two color-difference signals, conventionally called **I and Q**, and impress them on a subcarrier at 3.579545 MHz (precisely 315/88 MHz) using quadrature amplitude modulation with the carrier suppressed [14][15]. The FCC’s own signal specifications describe the chrominance components as being transmitted as “the amplitude modulation sidebands of a pair of suppressed subcarriers in quadrature” [14].
Two details are worth carrying forward.
**It is not textbook QAM.** The I and Q channels were given different bandwidths — the I component roughly 1.3 to 1.5 MHz, the Q component roughly 0.5 MHz — and the I channel’s upper sideband was truncated, making the arrangement a hybrid rather than a symmetrical double-sideband QAM [15]. This is a real engineering compromise driven by the fact that human vision resolves some color axes better than others.
**The name I and Q comes from here.** When you say “I and Q” today, you are using vocabulary that entered mass production in a 1953 television standard.
The lesson for a junior engineer: quadrature carrier multiplexing is an *analog* idea that long predates its digital application. QAM was not invented for data. It was adapted to data.
## 7. 1960–1962: Digital QAM Is Born
Three papers, all in the *IRE Transactions on Communications Systems*, over about eighteen months.
**C. R. Cahn, 1960.** “Combined Digital Phase and Amplitude Modulation Communication Systems,” volume CS-8, issue 3, September 1960, pages 150–155 [1]. Cahn is generally credited with first proposing QAM [10][11][13]. His constellations placed equal numbers of symbols on concentric circles — what is now called **Star QAM**. These are later labelled **Type I** [4].
**J. C. Hancock and R. W. Lucky, 1960.** “Performance of Combined Amplitude and Phase-Modulated Communication Systems,” volume CS-8, issue 4, 1960, pages 232–237 [2]. They observed that Cahn’s constellations were suboptimal for noise, because on concentric rings with equal point counts, the inner-ring points sit closer together than the outer-ring ones. Their fix was to put **more** symbols on the outer rings than the inner ones, balancing the distances [12]. This became **Type II**. They also concluded, candidly, that such systems were more complicated to implement [12].
**C. N. Campopiano and B. G. Glazer, 1962.** “A Coherent Digital Amplitude and Phase Modulation Scheme,” volume CS-10, issue 1, 1962, pages 90–95 [3]. This is the one that mattered. They introduced the **square QAM constellation** — **Type III** — the rectangular grid that everything from your cable modem to your Wi-Fi router uses today [4][12].
The comparison they drew is instructive and still relevant. Type II performed slightly better than Type III in noise, but Type III’s implementation and detection complexity was dramatically lower [11]. Square QAM won not because it was theoretically optimal but because it was *enormously* easier to build. Its I and Q axes decompose into two independent one-dimensional decisions, which is why a slicer can be two comparators instead of a two-dimensional nearest-neighbour search.
**That trade — a fraction of a dB of theoretical performance for a large reduction in complexity — is one you will make again and again in your career. The people who made it in 1962 made it correctly.**
One more historical nicety, from the definitive academic history of the subject: Campopiano and Glazer described their scheme in quadrature-carrier terms, but **the acronym QAM was not suggested in their paper** [4]. The name came later.
## 8. The Modem Era
QAM’s first mass deployment for data was the voiceband modem — moving digital data across a telephone channel roughly 3 kHz wide. When bandwidth is that scarce, bits per symbol is everything.
**V.29.** Used a 16-point constellation; became widely deployed on four-wire leased lines and remains the modulation used by Group 3 fax [23].
**V.32.** A 9600 bit/s full-duplex modem over a two-wire switched telephone line. It used a 1800 Hz carrier and a symbol rate of 2400 baud, with 16 constellation states uncoded, or 32 states with trellis coding [21][23].
Do the arithmetic yourself, because it is a good check on §4: 2400 symbols/s × 4 bits/symbol = 9600 bit/s. That is where the number comes from.
**Gottfried Ungerboeck, 1982.** “Channel Coding with Multilevel/Phase Signals,” *IEEE Transactions on Information Theory*, volume IT-28, number 1, January 1982, pages 55–67 [16][16].
**Trellis-coded modulation** is one of the genuinely great ideas in communications, and it is worth understanding the shape of it even if you never implement one. Ungerboeck’s insight was that in a bandwidth-limited channel, the redundancy needed for error correction should be obtained by **expanding the constellation while keeping the bandwidth fixed**, rather than by expanding the bandwidth while keeping the constellation fixed [18].
Concretely: instead of using 16-QAM uncoded, use 32-QAM and spend the extra bit on a convolutional code. You lose some minimum distance by packing the constellation tighter, but the code buys back much more than you lost. Ungerboeck showed that doubling the constellation captures most of the available coding gain [18][18].
The reported results: V.32 used an 8-state trellis code for a coding gain of about 3.5 dB; V.34 used 16- to 64-state codes for gains of 4.0 to 4.5 dB [17]. The 8-state two-dimensional code used in V.32 was due to Lee-Fang Wei, with a nonlinear modification making it rotationally invariant [18].
Three to four and a half decibels sounds modest. It is not. §15 will show you that 6 dB buys you two more bits per symbol. Ungerboeck’s codes bought most of that for free.
**V.32bis** raised the rate to 14,400 bit/s [20]. **V.34** (1994) went to 28,800 and later 33,600 bit/s, specifying QAM with selectable symbol rates — mandatory rates of 2400, 3000 and 3200 symbols/s plus optional rates of 2743, 2800 and 3429 symbols/s [20]. V.34 also introduced **nonlinear coding**, spacing constellation points unequally so that high-amplitude points sit farther apart, and **precoding** to combat intersymbol interference [22].
If you ever wondered why V.34 modems made that long negotiation noise: they were probing the channel from 150 Hz to 3750 Hz with test tones to decide which symbol rate and constellation the line could support [22]. That is adaptive modulation (§21), thirty years ago, in a consumer product.
## 9. Broadcast, Cable, and the Standards That Followed
QAM’s second mass deployment was digital television, and this is where the cable industry enters.
**ITU-T Recommendation J.83** defines digital multi-programme television distribution by cable. It has regional annexes.
**Annex A** — the basis of **DVB-C**, used in Europe and much of the world. The 1995 edition allowed 16-, 32- and 64-QAM constellations, with provision for extension to 128- and 256-QAM [26]. Typical deployment uses an 8 MHz channel with a roll-off factor of 0.15; the theoretical maximum symbol rate in 8 MHz is then about 6.96 Mbaud [28].
**Annex B** — the North American cable variant, based on testing and characterization of North American cable systems [27]. It specifies 64-QAM and 256-QAM in two modes: Mode 1 at a symbol rate of 5.056941 Msym/s, typically used for 64-QAM, and Mode 2 at 5.360537 Msym/s, typically used for 256-QAM [24][25][29].
Those two symbol rates are worth memorizing if you work in cable. They appear directly in the DOCSIS specifications, which require the CMTS or Edge QAM downstream to operate per J.83 Annex B [30].
**Annex C** — the Japanese variant. Same framing and channel coding as Annex A, with the symbol rate and roll-off optimized for a different channel plan [24].
Note the FEC: J.83 uses **concatenated coding** — an outer Reed-Solomon block code with an inner trellis code — chosen because it produces high coding gain at moderate complexity and overhead [25]. The system is designed for **quasi error-free** operation, defined in Annex B as about one error event per 15 minutes [25].
This is the pattern you should notice: **by the mid-1990s nobody deployed QAM without forward error correction.** Uncoded QAM is a teaching device. Real systems are QAM plus a code, and the code is not an afterthought — it is what makes the modulation order achievable.
## 10. QAM Inside OFDM
The last chapter of the story, and the one that matters most for current work.
Modern systems rarely use one wide QAM carrier. Instead they use **OFDM** — Orthogonal Frequency Division Multiplexing — which divides the channel into many narrow subcarriers, and modulates **each subcarrier with its own QAM constellation**.
This does not replace QAM. It multiplies it. An OFDM symbol is thousands of simultaneous QAM symbols.
The advantage is that a narrow subcarrier sees an almost flat channel, so frequency-selective distortion becomes a per-subcarrier gain rather than intersymbol interference. And critically, **different subcarriers can carry different constellation orders**, matched to the local signal quality.
Where this stands today:
– **DOCSIS 3.1 and 4.0.** Downstream OFDM requires modulation orders up to 4096-QAM with optional support to 16384-QAM; upstream OFDMA requires up to 1024-QAM with optional support up to 4096-QAM [34]. Legacy single-carrier DOCSIS uses 64-QAM and 256-QAM downstream per J.83 Annex B [30].
– **Wi-Fi.** 802.11ac (Wi-Fi 5) supports up to 256-QAM; 802.11ax (Wi-Fi 6) up to 1024-QAM, carrying 10 bits per symbol; 802.11be (Wi-Fi 7) up to 4096-QAM, carrying 12 bits, for a nominal 20% rate increase at the same coding rate [35][36][37][38].
– **Cellular.** Sources I could verify describe 5G NR as supporting a maximum of 256-QAM [39]. Higher orders have been discussed in later 3GPP work; I have not verified the current position, and you should check a current 3GPP specification rather than take a number from me.
One detail from the Wi-Fi 7 work that illustrates the cost of high-order QAM beautifully: the IEEE standard sets a transmitter EVM requirement of **−38 dB for 4096-QAM**, a 3 dB tightening relative to Wi-Fi 6 [37]. Getting 12 bits per symbol is not a software change. It is a demand on every analog component in the chain.
—
# Part III — The Mathematics
## 11. The Signal, Written Down
The transmitted QAM signal:
$$s(t) = I(t)\cos(2\pi f_c t) – Q(t)\sin(2\pi f_c t)$$
where $I(t)$ and $Q(t)$ are baseband waveforms carrying the symbol values, shaped by the pulse filter of §14.
Equivalently, in complex baseband notation, which is how every modern DSP implementation actually represents it:
$$s(t) = \Re\left\{ \tilde{s}(t)\, e^{j2\pi f_c t} \right\}, \quad \tilde{s}(t) = I(t) + jQ(t)$$
$\tilde{s}(t)$ is the **complex envelope**. Almost all analysis and all simulation happens here, with the carrier stripped away. This is why MER is described as a *complex baseband* signal-to-noise ratio [32] — the measurement lives in this domain.
### Symbol rate versus bit rate
Two rates, constantly confused:
– **Symbol rate** (or **baud**), $R_s$ — symbols per second.
– **Bit rate**, $R_b = R_s \times \log_2 M$ — bits per second.
They are equal only for binary signaling. A 5.360537 Msym/s 256-QAM cable channel carries 5.360537 × 8 ≈ 42.9 Mbit/s of raw symbol payload before FEC overhead [30].
**”Baud” and “bits per second” are not synonyms.** Using them interchangeably marks you out immediately.
## 12. Constellation Geometry and Power
### Square constellations
For a square M-QAM constellation placed on an integer grid with points spaced by 2 on each axis — so the coordinates are the odd integers ±1, ±3, … ±(√M − 1) — two quantities matter.
**Peak constellation power**, the squared magnitude of a corner point:
$$P_{\text{peak}} = 2(\sqrt{M} – 1)^2$$
**Average constellation power**, averaged equally over all points:
$$P_{\text{avg}} = \tfrac{2}{3}(M – 1)$$
Both formulas are from Hranac and Currivan [32]. Check them on 16-QAM: coordinates come from {−3, −1, 1, 3}, so peak power is 3² + 3² = 18, and the formula gives 2(4−1)² = 18 ✓. Average power is (2/3)(15) = 10 ✓ — which, as they note, happens to equal the power of the point (3,1), since 3² + 1² = 10 [32].
### Maximum-to-average ratio
The ratio of these two, the **MTA**, tells you how much of your power budget is consumed by the outermost points:
$$\text{MTA} = \frac{2(\sqrt{M}-1)^2}{\frac{2}{3}(M-1)} = \frac{3(\sqrt{M}-1)}{\sqrt{M}+1}$$
*(The simplification is mine: factor M − 1 = (√M − 1)(√M + 1) and cancel.)*
As M grows, this approaches 3, or 4.77 dB [32]. Working the values out:
| Constellation | MTA (linear) | MTA (dB) |
|—|—|—|
| QPSK (4-QAM) | 1.000 | 0.00 |
| 16-QAM | 1.800 | 2.55 |
| 64-QAM | 2.333 | 3.68 |
| 256-QAM | 2.647 | 4.23 |
| 1024-QAM | 2.818 | 4.50 |
| 4096-QAM | 2.908 | 4.64 |
*(Values computed by me from the formula above. The QPSK result of 0 dB is a useful sanity check — all four QPSK points have equal magnitude, so peak equals average exactly.)*
### A distinction that catches people
**MTA is not the same as the peak-to-average ratio of the transmitted RF signal.** MTA accounts only for the distribution of ideal constellation symbols. The actual PAR of a modulated carrier is affected by pulse shaping and filtering, and typically lies in the range of 6 to 13 dB or more; measured values of 6.3 to 7.3 dB for 64-QAM and 6.5 to 7.5 dB for 256-QAM have been reported [32].
Why does this matter? Because your power amplifier has to handle those peaks linearly. QAM’s amplitude variation is precisely why QAM signals cannot be amplified as efficiently as constant-envelope signals.
### Non-square constellations
Square constellations need an **even** number of bits per symbol, because $\sqrt{M}$ must be an integer. For 32-QAM, 128-QAM, 512-QAM — odd bit counts — you need something else.
Two families are used:
**Cross constellations**, which take a square and move the corner points inward, producing a plus-sign shape.
**Double-square constellations**, defined as a subset consisting of half the points of the next-higher square constellation, arranged like the black squares on a checkerboard, having the same peak and average values as that next-higher square [32].
DOCSIS uses 64-QAM and 256-QAM square constellations downstream, and both square and double-square constellations from QPSK through 128-QAM upstream [32].
## 13. Gray Coding
A small idea with a large payoff.
When noise causes a symbol error, the received point is almost always a **nearest neighbour** of the correct one — errors to distant points are far less likely. So arrange your bit-to-symbol mapping such that **adjacent constellation points differ in exactly one bit**. That is **Gray coding**.
The consequence: the overwhelming majority of symbol errors produce exactly one bit error, rather than an average of half the bits in the symbol.
For a Gray-coded square constellation, the bit error rate is approximately the symbol error rate divided by the number of bits per symbol:
$$\text{BER} \approx \frac{\text{SER}}{\log_2 M}$$
*(This is the standard relationship given in the digital communications literature [40]; it holds at moderate-to-high SNR where nearest-neighbour errors dominate.)*
Without Gray coding you would lose several dB of effective performance for no reason at all. It costs nothing but a lookup table.
## 14. Bandwidth: Nyquist and the Raised Cosine
### The fundamental limit
Nyquist’s 1928 result [5] gives the floor: to transmit $R_s$ symbols per second without **intersymbol interference** (ISI — symbols smearing into their neighbours), you need a minimum bandwidth of $R_s/2$ at baseband, or $R_s$ at passband.
The ideal pulse that achieves this is the sinc function. It has a problem: it decays slowly, so timing errors cause large ISI, and it is not realizable [41].
### The raised cosine compromise
The practical answer is the **raised-cosine** family, parameterized by a **roll-off factor** α between 0 and 1. Occupied bandwidth becomes:
$$B = R_s (1 + \alpha)$$
α is the price you pay in excess bandwidth for a filter you can actually build and a receiver that tolerates timing error.
In practice the filtering is split: a **root-raised-cosine** filter in the transmitter and a matching one in the receiver, so the cascade is raised-cosine and the receiver filter is simultaneously the matched filter for the pulse. J.83 requires exactly this — the I and Q signals are square-root raised cosine filtered prior to modulation [24].
### Worked example: where 6 MHz comes from
This is my own arithmetic, and it is worth doing because it connects three numbers you will meet constantly.
J.83 Annex B Mode 2 specifies 5.360537 Msym/s for 256-QAM in a 6 MHz channel [24][29]. What roll-off does that imply?
$$\alpha = \frac{6.000}{5.360537} – 1 = 0.1193$$
And for Mode 1, 64-QAM at 5.056941 Msym/s:
$$\alpha = \frac{6.000}{5.056941} – 1 = 0.1865$$
Those come out at approximately 0.12 and 0.18 — the roll-off values Annex B specifies for 256-QAM and 64-QAM respectively. My arithmetic reproduces the standard’s design, which is good evidence that I have the relationship right and that you now understand where a cable channel’s symbol rate comes from: it is the largest rate that fits in 6 MHz once you have chosen a realizable filter.
Sanity-check the same relation against DVB-C: 6.96 Mbaud × (1 + 0.15) = 8.004 MHz, filling an 8 MHz channel [28]. ✓
### The point that matters
**Bandwidth is set by symbol rate and roll-off. It is not affected by constellation order.**
64-QAM and 256-QAM at the same symbol rate occupy the same bandwidth. The 256-QAM signal carries 33% more data in that bandwidth. It does not get that for free — it pays in required signal quality, which is the next section.
## 15. Noise, Distance, and the 6 dB Rule
### Minimum distance is everything
For a receiver making nearest-neighbour decisions in Gaussian noise, error probability is governed almost entirely by the **minimum Euclidean distance** between constellation points, relative to the noise.
For square M-QAM at a fixed average symbol energy $E_s$, using the grid of §12 where the minimum distance is $d = 2$ and the average power is $\frac{2}{3}(M-1)$:
$$\frac{d^2}{E_s} = \frac{4}{\frac{2}{3}(M-1)} = \frac{6}{M-1} \quad \Longrightarrow \quad d = \sqrt{\frac{6E_s}{M-1}}$$
*(My derivation, from the power formulas of Hranac and Currivan [32].)*
### The 6 dB rule
Now the consequence. Go from M to 4M points — two more bits per symbol — holding average power constant. The minimum distance shrinks by:
$$\frac{d_{4M}}{d_M} = \sqrt{\frac{M-1}{4M-1}}$$
Evaluating:
| Step | Distance ratio | In dB |
|—|—|—|
| 16 → 64 | 0.4880 | −6.23 |
| 64 → 256 | 0.4971 | −6.07 |
| 256 → 1024 | 0.4993 | −6.03 |
| 1024 → 4096 | 0.4998 | −6.02 |
*(Computed by me from the expression above.)*
So the working rule, which you should commit to memory:
> **Every two extra bits per symbol costs approximately 6 dB of required signal quality.** Equivalently, about 3 dB per bit per symbol.
The convergence to exactly 6 dB is fast; treat it as 6 dB for anything above 16-QAM.
This single rule explains an enormous amount about real networks. It explains why the jump from 256-QAM to 4096-QAM in DOCSIS 3.1 required plant improvement rather than just new modems. It explains why Wi-Fi 7’s 4096-QAM needs beamforming to work at all. It explains why your modem falls back to a lower order when the plant degrades.
### Error probability
For completeness, the standard result. For square M-QAM in AWGN, the symbol error rate is
$$\text{SER} = 1 – \left(1 – 2\left(1 – \tfrac{1}{\sqrt{M}}\right) Q\!\left(\sqrt{\tfrac{3E_s}{(M-1)N_0}}\right)\right)^2$$
where $Q(\cdot)$ is the Gaussian tail probability. With Gray coding, BER ≈ SER / log₂M (§13).
*(This is the standard textbook expression for square QAM [40]; I am quoting it, not deriving it. The structure is worth reading even if the algebra is not: the inner term is the error probability of one √M-ary amplitude decision on one axis, and the square reflects the fact that a square constellation decomposes into two independent one-dimensional decisions. That decomposition is exactly the implementation advantage that made square QAM win in 1962.)*
### Real-world thresholds
Theory gives you the shape. Operators use empirical numbers with margin built in.
Hranac and Currivan report that many cable operators use these minimum acceptable operational unequalized RxMER values [32]:
| Modulation | Minimum operational RxMER |
|—|—|
| QPSK | ~18 dB |
| 16-QAM | ~24 dB |
| 64-QAM | ~27 dB |
| 256-QAM | ~31 dB |
They further advise keeping unequalized RxMER at least 3 to 6 dB above the lower Es/N₀ threshold for the modulation in use, to accommodate temperature-related level variation, amplifier and optoelectronics misalignment, test equipment calibration, and similar factors [32].
Two observations for you. First, notice that the gaps between these numbers are smaller than 6 dB — because these are practical operating minima including FEC, not uncoded theoretical thresholds. Second, notice that **margin is explicitly designed in**. A plant that meets threshold exactly is a plant that fails on a hot afternoon.
## 16. The Shannon Limit and How Close QAM Gets
Shannon’s capacity for a band-limited channel with additive white Gaussian noise:
$$C = B \log_2\left(1 + \frac{S}{N}\right)$$
where C is capacity in bits per second, B is bandwidth in hertz, and S/N is the signal-to-noise power ratio [8][9].
This is the ceiling. No modulation, no code, no cleverness exceeds it.
### The gap
QAM with uniformly-spaced points and equally-likely symbols does not quite reach capacity even with perfect coding. The reason is that capacity is achieved with Gaussian-distributed inputs, and a uniform QAM constellation is not Gaussian. The resulting loss — the **shaping gain** available in principle — is at most 1.53 dB [19].
*(That figure is πe/6 expressed in decibels; I have verified the arithmetic gives 1.53 dB, and the result is standard in the coding literature [19].)*
Recovering some of that gap is what **constellation shaping** does, and it is why probabilistic shaping has become an active area in high-speed optical and wireless systems.
### The practical reading
Put the pieces together and you get the honest engineering picture.
Shannon says capacity rises with the **logarithm** of SNR. §15 says every 6 dB of SNR buys 2 bits per symbol. These are the same statement.
**There is no cliff, and there is no free lunch.** Every additional bit per symbol costs roughly 3 dB, forever. That is why modulation order climbs slowly across technology generations while bandwidth and spatial multiplexing climb fast — spectrum and antennas scale linearly, modulation order scales logarithmically.
—
# Part IV — Practice
## 17. The Transmitter Chain
The stages, in order:
1. **FEC encoding.** Redundancy is added — Reed-Solomon, convolutional, trellis, LDPC, depending on the standard. As §9 noted, real QAM systems are always coded.
2. **Bit-to-symbol mapping.** Groups of log₂M bits become constellation points, Gray-coded (§13).
3. **Pulse shaping.** Symbols are filtered by a root-raised-cosine filter, setting the occupied bandwidth (§14).
4. **Quadrature upconversion.** I and Q modulate the cosine and sine carriers and are summed.
5. **Amplification.** Linearly, with enough back-off to handle the peaks discussed in §12.
Modern implementations do steps 1–3 and often step 4 digitally, with the DAC placed as late in the chain as the sample rate allows.
## 18. The Receiver Chain
This is where the real engineering lives. The transmitter knows everything; the receiver knows nothing and must work it all out from the signal.
**Automatic gain control.** Set the signal amplitude to suit the ADC. Constellation decisions are amplitude-sensitive, so gain errors directly scale the constellation.
**Quadrature downconversion.** Recover I and Q.
**Timing recovery.** Determine when, within each symbol period, to sample. Sampling at the wrong instant means sampling a pulse that has not fully settled — instant ISI.
**Carrier recovery.** Determine the exact carrier frequency and phase. This one deserves emphasis. **A phase error rotates the entire constellation.** A frequency error makes it spin. Since the decision boundaries are fixed, a rotated constellation produces errors even with zero noise. This is why QAM demands *coherent* detection and why phase noise (§20) is so damaging at high orders.
**Equalization.** Compensate the channel’s frequency response and remove ISI. Two flavours you will meet in cable:
– **Adaptive receive equalization**, where the receiver adapts a filter to invert the channel.
– **Pre-equalization**, where the *transmitter* predistorts. DOCSIS does this upstream: each modem contains a pre-equalizer, defined in DOCSIS 2.0 as a 24-tap symbol-spaced FIR filter. The CMTS adapts its burst-receiver equalizer on a ranging burst, then sends the coefficients back to the modem, which loads them into its pre-equalizer [32].
A parameter worth knowing: the equalizer **span** — (number of taps − 1) × tap spacing. At 5.12 Msym/s, 24 taps gives (24−1)/5.12 = 4.5 µs [32]. **If an echo arrives later than the span, the equalizer simply cannot correct it.** That is a hard architectural limit, not a tuning problem.
**Slicing.** Nearest-neighbour decision (§4), then symbol-to-bit mapping and FEC decoding.
## 19. Measuring Quality: MER, EVM, CNR
You will be asked to interpret these numbers constantly, and they are routinely confused. This section is the highest-value practical material in the document, drawn from Hranac and Currivan’s treatment [32].
### MER
**Modulation Error Ratio** is the digital complex baseband signal-to-noise ratio:
$$\text{MER (dB)} = 10\log_{10}\left(\frac{\text{average symbol power}}{\text{average error power}}\right)$$
The **error vector** is the vector difference between the ideal target symbol and the received symbol. Plotted over time, symbol landing points form a cloud rather than a point, and **MER is a measure of how fuzzy that cloud is** [32].
Higher is better. MER is a direct measure of modulation quality with linkage to bit error rate [32].
### EVM
**Error Vector Magnitude** is the same underlying quantity, reported differently:
$$\text{EVM}_\% = \frac{E_{\text{RMS}}}{S_{\max}} \times 100$$
where $E_{\text{RMS}}$ is the RMS error magnitude and $S_{\max}$ is the maximum symbol magnitude [32]. **Lower is better.** EVM is linear and in percent; MER is logarithmic and in dB.
The conversion between them requires the MTA of §12, because the two are referenced to different quantities — MER to average power, EVM to peak [32]:
$$\text{EVM}_\% = 100 \times 10^{-(\text{MER}_{dB} + \text{MTA}_{dB})/20}$$
For scale: a downstream 256-QAM signal measured in a headend or at a node output might show an EVM of about 0.9% [32].
### CNR — and the confusion
**Carrier-to-noise ratio** is the RF measurement you make on a spectrum analyzer.
In an ideal system with no impairment other than AWGN and full traffic loading, SNR, CNR, Es/N₀, RxMER and CINR are all equal [32]. In the real world they diverge, **and the divergence is the diagnostic information**.
The rule to remember: **MER is less than or at best equal to CNR, but never greater** [33]. MER captures everything CNR does *plus* transmitter imperfections, linear distortion, phase noise and receiver implementation loss.
### The single most useful practical fact in this document
When a cable modem or set-top box reports “SNR”, **it is not reporting CNR.** It is reporting equalized downstream RxMER. Likewise, the “SNR” a CMTS reports for the upstream is equalized (or in some implementations unequalized) RxMER, as defined in the DOCSIS MIB [32].
Hranac and Currivan give a field example that I would like every junior engineer to read.
An upstream plant showed excellent CNR of about 36 dB on a spectrum analyzer — no visible ingress, no common path distortion, no impulse noise near the carrier. QPSK worked fine. Unequalized 16-QAM was unusable. The CMTS reported “SNR” of 21 dB, which third-party equipment confirmed as unequalized MER.
The cause was a severe impedance mismatch about 1,100 feet from the node, producing a micro-reflection and consequent amplitude and group delay ripple. **These distortions were invisible on the spectrum analyzer, yet degraded RxMER enough to break 16-QAM** [32].
The lesson: a clean spectrum display does not mean a healthy channel. Linear distortion is invisible to a spectrum analyzer and lethal to high-order QAM. When CNR looks good and MER looks bad, suspect reflections, group delay, or something else the analyzer cannot show you.
### When not to trust a MER reading
Hranac and Currivan catalogue the ways MER measurements mislead [32]. The ones I would want you to know:
**Implementation-loss ceiling.** Even with a perfect input, the reading saturates at the receiver’s own noise floor. It is unusual for a 256-QAM receiver to read much above 40 to 45 dB even with no input noise at all. A reading of 44 dB may mean “the channel is excellent”; it does not mean “the channel is 44 dB.”
**Symbol-error floor.** Below roughly 1% pre-FEC symbol error rate, the measurement becomes invalid. When the slicer picks the wrong point, the error vector is measured to that wrong point, which is nearer — so **the MER reads better than reality precisely when things are worst.**
**Sample count.** Standard deviation of the measurement scales as 1/√N. Averaging over 10,000 samples gives ten times smaller standard deviation than 100 samples. But fewer samples lets you *see* transients from burst noise and pulsed ingress — analogous to video averaging on a spectrum analyzer.
**Measure the constellation you will use.** In one return-path system with nonlinearities present, equalized RxMER measured 38.0 dB with QPSK, 31.9 dB with 16-QAM, and 30.2 dB with 256-QAM [32]. The QPSK figure promised headroom that did not exist, because nonlinearities hit outer constellation points harder than inner ones. **A good MER at low order does not guarantee a good MER at high order.**
## 20. What Breaks QAM
A catalogue, organized by what it does to the constellation. Learning to read the failure mode off the display is a genuinely useful skill.
**Thermal noise / AWGN** — symmetric fuzzy cloud around every point, uniform across the constellation.
**Phase noise** — angular smearing. Corner points, being farthest from the origin, smear most in the tangential direction. The receiver’s carrier tracking loop removes the low-frequency portion; the DOCSIS DRFI specification separates phase noise below 50 kHz from the TxMER requirement on the assumption that the receiver’s carrier loop removes it [32]. **Effects become more critical as modulation order increases** [32].
**Micro-reflections and impedance mismatches** — linear distortion producing amplitude and group delay ripple. Correctable by equalization *if* within the equalizer span (§18). Invisible on a spectrum analyzer.
**Nonlinear distortion** — laser clipping, amplifier compression. Affects outer constellation points more severely than inner ones [32]. Produces a constellation that looks fine in the middle and smeared at the edges.
**Ingress** — narrowband interference entering the plant. Modern CMTS burst receivers include ingress cancellers; after cancellation, upstream RxMER can be much higher than the input carrier-to-interference-plus-noise ratio, because the interference has been removed [32].
**Burst and impulse noise** — short strong hits. May barely move an averaged MER reading while destroying packets. This is exactly why MER alone is insufficient and you also watch codeword error counts.
**Gain and I/Q imbalance** — constellation scaled, or stretched along one axis.
**Carrier phase error** — the whole constellation rotated.
A note on interpreting displays: it is generally not possible to determine the *cause* of low MER just by looking at a constellation [32]. The display tells you the symptom. Diagnosis requires correlating with spectrum, with equalizer coefficients, with time-domain captures, and with what else on the plant is affected.
## 21. Adaptive Modulation
Everything above leads here.
If the required signal quality depends on modulation order (§15), and the achievable signal quality varies with plant condition, frequency and subscriber location — then the sensible thing is to *choose the order to match the conditions* rather than picking one order for everybody.
V.34 modems did this in 1994 by probing the line with test tones before selecting symbol rate and constellation [22].
Modern OFDM systems do it per-subcarrier. Because each subcarrier can carry a different constellation (§10), a system can assign aggressive bit loading where the channel is clean and back off where it is not. In DOCSIS this is managed through modulation **profiles**, with a Profile Management Application computing optimized per-subcarrier bit loading from measured per-subcarrier RxMER.
This is the modern shape of the field: **QAM order is no longer a design-time constant. It is an operational variable, measured and adjusted continuously.** The mathematics of §15 has become a control loop.
## 22. Traps and Misconceptions
**”Baud equals bits per second.”** Only for binary signaling. Bit rate = symbol rate × log₂M (§11).
**”Higher QAM needs more bandwidth.”** No. Bandwidth is set by symbol rate and roll-off (§14). Higher order needs more *signal quality*, not more spectrum.
**”QAM was invented for data.”** Quadrature amplitude modulation of a suppressed subcarrier was deployed in NTSC color television in 1953 [14][15], seven years before Cahn’s 1960 paper on the digital form [1].
**”Cahn invented the square constellation.”** Cahn proposed concentric-ring (Star) constellations [1][12]. The square constellation came from Campopiano and Glazer in 1962 [3][4].
**”Square QAM is optimal.”** It is not, quite. Hancock and Lucky’s Type II performed slightly better in noise; square QAM won on implementation and detection complexity [11]. Optimal and deployable are different criteria.
**”The modem’s SNR reading is the CNR.”** It is equalized RxMER [32]. This confusion causes an enormous amount of wasted troubleshooting.
**”Good CNR means the channel is fine.”** Linear distortion is invisible on a spectrum analyzer and can break 16-QAM on a plant showing 36 dB CNR [32].
**”A high MER reading means a high-quality channel.”** It may mean you have hit the receiver’s implementation-loss ceiling — typically 40 to 45 dB for 256-QAM [32].
**”MER is reliable when things are bad.”** It is least reliable exactly then. Below roughly 1% pre-FEC symbol error rate the measurement reads optimistically, because error vectors are measured to the wrong, nearer, constellation point [32].
**”MTA is the peak-to-average ratio.”** MTA covers only the ideal constellation. Actual signal PAR after pulse shaping is typically 6 to 13 dB [32].
**”Uncoded QAM performance is what I should design to.”** No real system since the early 1990s deploys QAM without FEC [25]. Uncoded analysis is a teaching tool.
**”Each extra bit per symbol is roughly free.”** Roughly 3 dB each, forever (§15).
## 23. Glossary
| Term | Meaning |
|—|—|
| **AWGN** | Additive White Gaussian Noise — the standard channel noise model |
| **Baud** | Symbols per second; equal to bit rate only for binary signaling |
| **BER / SER** | Bit Error Rate / Symbol Error Rate |
| **CNR** | Carrier-to-Noise Ratio — an RF measurement made on a spectrum analyzer |
| **CINR** | Carrier-to-Interference-plus-Noise Ratio |
| **Complex envelope** | The baseband representation I + jQ, with the carrier removed |
| **Constellation** | The set of all points a modulator can transmit, plotted on the I/Q plane |
| **Cross constellation** | Non-square constellation for odd bit counts, made by moving a square’s corners inward |
| **Double-square** | Half the points of the next-higher square constellation, in a checkerboard pattern |
| **Es/N₀** | Ratio of average energy per symbol to noise power spectral density |
| **EVM** | Error Vector Magnitude — error expressed as a percentage of peak symbol magnitude |
| **FEC** | Forward Error Correction |
| **Gray coding** | Bit mapping in which adjacent constellation points differ in exactly one bit |
| **I / Q** | In-phase and Quadrature components |
| **ISI** | Intersymbol Interference — symbols smearing into their neighbours |
| **MER** | Modulation Error Ratio — ratio of average symbol power to average error power, in dB |
| **MTA** | Maximum-to-Average constellation power ratio; approaches 4.77 dB at high order |
| **Nyquist pulse** | A pulse shape producing no ISI at the sampling instants |
| **OFDM / OFDMA** | Orthogonal Frequency Division Multiplexing / Multiple Access |
| **PAR** | Peak-to-Average Ratio of the actual transmitted signal |
| **Pre-equalization** | Transmitter-side predistortion to compensate the channel |
| **QAM** | Quadrature Amplitude Modulation |
| **Roll-off (α)** | Excess-bandwidth factor of a raised-cosine filter; occupied bandwidth = R_s(1+α) |
| **RRC** | Root Raised Cosine — the split-filter form used at transmitter and receiver |
| **RxMER** | Receive MER, as reported by a modem or CMTS — often mislabelled “SNR” |
| **SC-QAM** | Single-Carrier QAM, as distinct from QAM carried on OFDM subcarriers |
| **Slicer** | The receiver stage making the nearest-constellation-point decision |
| **TCM** | Trellis-Coded Modulation |
## 24. References
### Foundational papers
[1] C. R. Cahn, “Combined Digital Phase and Amplitude Modulation Communication Systems,” *IRE Transactions on Communications Systems*, vol. CS-8, no. 3, pp. 150–155, September 1960. DOI: 10.1109/TCOM.1960.1097623
[2] J. C. Hancock and R. W. Lucky, “Performance of Combined Amplitude and Phase-Modulated Communication Systems,” *IRE Transactions on Communications Systems*, vol. CS-8, no. 4, pp. 232–237, 1960.
[3] C. N. Campopiano and B. G. Glazer, “A Coherent Digital Amplitude and Phase Modulation Scheme,” *IRE Transactions on Communications Systems*, vol. CS-10, no. 1, pp. 90–95, 1962.
[4] L. Hanzo et al., *Quadrature Amplitude Modulation: From Basics to Adaptive Trellis-Coded, Turbo-Equalised and Space-Time Coded OFDM, CDMA and MC-CDMA Systems*, Chapter 1, “History of QAM.” University of Southampton ePrints. *(Source of the Type I / II / III framing, the quoted Campopiano and Glazer description of quadrature carriers, and the observation that the acronym QAM was not itself proposed in their paper.)* — https://eprints.soton.ac.uk/260296/1/QAM3-chaps_1-22-24.pdf
[5] H. Nyquist, “Certain Topics in Telegraph Transmission Theory,” *Transactions of the AIEE*, vol. 47, no. 2, pp. 617–644, April 1928. Presented at the AIEE Winter Convention, New York, February 13–17, 1928. — https://bayes.wustl.edu/Manual/CertainTopicsInTelegraphTransmissionTheory.pdf
[6] H. Nyquist, “Certain Factors Affecting Telegraph Speed,” *Bell System Technical Journal*, April 1924, p. 324. (Cited by Shannon in [8].)
[7] R. V. L. Hartley, “Transmission of Information,” *Bell System Technical Journal*, July 1928, p. 535. (Cited by Shannon in [8].)
[8] C. E. Shannon, “A Mathematical Theory of Communication,” *Bell System Technical Journal*, vol. 27, pp. 379–423 (July) and pp. 623–656 (October), 1948. — https://marco-dalai.unibs.it/teach/IT/shannon_1948.pdf
[9] C. E. Shannon, “Communication in the Presence of Noise,” *Proceedings of the IRE*, 1949; reprinted in *Proceedings of the IEEE*, vol. 86, no. 2, February 1998. — https://webusers.imj-prg.fr/~antoine.chambert-loir/enseignement/2020-21/shannon/shannon1949.pdf
### Corroborating histories of QAM’s origin
[10] “SEP performance of triangular QAM with MRC spatial diversity over fading channels,” *EURASIP Journal on Wireless Communications and Networking*, 2016. *(Cahn 1960 as first proposal; Hancock and Lucky’s concentric-ring refinement; Campopiano and Glazer’s square constellation.)* — https://jwcn-eurasipjournals.springeropen.com/articles/10.1186/s13638-015-0511-2
[11] “A Survey on Design and Performance of Higher-Order QAM Constellations,” arXiv:2004.14708. *(Type I / II / III naming; the finding that Type II performs slightly better in noise while Type III has significantly lower implementation and detection complexity.)* — https://arxiv.org/pdf/2004.14708
[12] *High-Order Modulation for Optical Fiber Transmission*, Springer, Chapter 2 excerpt. *(Star QAM per Cahn; Hancock and Lucky placing more symbols on outer rings for balanced Euclidean distance, and their conclusion on implementation complexity; square QAM introduced by Campopiano and Glazer in 1962.)* — https://beckassets.blob.core.windows.net/product/readingsample/577661/9783540937708_excerpt_001.pdf
[13] Mini-Circuits, “A Primer on Quadrature Amplitude Modulation (QAM).” — https://blog.minicircuits.com/a-primer-on-quadrature-amplitude-modulation-qam/
### Analog quadrature: NTSC color television
[14] FCC, *NTSC Signal Specifications*. *(Chrominance components transmitted as the amplitude modulation sidebands of a pair of suppressed subcarriers in quadrature, at +3.579545 Mc relative to the picture carrier.)* — https://www.earlytelevision.org/pdf/ntsc_signal_specifications.pdf
[15] Descriptions of the NTSC color subcarrier at 315/88 MHz, its quadrature amplitude modulation with suppressed carrier, and the I/Q bandwidth asymmetry that makes the arrangement a hybrid rather than a symmetrical double-sideband QAM. US Patents 5,550,579 (background section, “hybrid QAM” and I/Q sideband treatment) and 4,240,105 (background section, I at 1.3 MHz and Q at 500 kHz). — https://image-ppubs.uspto.gov/dirsearch-public/print/downloadPdf/5550579 and https://image-ppubs.uspto.gov/dirsearch-public/print/downloadPdf/4240105
### Coded modulation
[16] G. Ungerboeck, “Channel Coding with Multilevel/Phase Signals,” *IEEE Transactions on Information Theory*, vol. IT-28, no. 1, pp. 55–67, January 1982.
[17] G. D. Forney, Jr., “The Viterbi Algorithm: A Personal History,” arXiv:cs/0504020. *(V.32 8-state trellis code, coding gain about 3.5 dB; V.34 16- to 64-state codes, 4.0 to 4.5 dB.)* — https://arxiv.org/pdf/cs/0504020
[18] G. D. Forney, Jr., “Channel Coding: The Road to Channel Capacity,” arXiv:cs/0611112. *(The constellation-expansion principle; doubling the constellation captures most of the available coding gain; the Lee-Fang Wei codes used in V.32 and V.34.)* — https://arxiv.org/pdf/cs/0611112
[19] G. D. Forney, Jr. and G. Ungerboeck, “Modulation and Coding for Linear Gaussian Channels,” *IEEE Transactions on Information Theory*, vol. 44, no. 6, pp. 2384–2415, October 1998. *(Shaping gain and the gap between uniform QAM and Gaussian-input capacity.)*
### Voiceband modems
[20] ITU-T Recommendation V.34, *A modem operating at data signalling rates of up to 33 600 bit/s for use on the general switched telephone network and on leased point-to-point 2-wire telephone-type circuits*, February 1998. *(QAM with mandatory symbol rates of 2400, 3000 and 3200 symbols/s and optional rates of 2743, 2800 and 3429.)* — https://www.itu.int/rec/dologin_pub.asp?lang=e&id=T-REC-V.34-199802-I!!PDF-E&type=items
[21] 3am Systems, “ITU-T V.32 / V.32bis modems.” *(V.32: 1800 Hz carrier, 2400 baud, 16 states uncoded or 32 with trellis coding, 9600 bit/s.)* — https://www.3amsystems.com/Technical_library/V32
[22] V.34 nonlinear coding, precoding, trellis code selection and channel probing from 150 Hz to 3750 Hz, as described in the background section of US Patent 6,263,016. — https://image-ppubs.uspto.gov/dirsearch-public/print/downloadPdf/6263016
[23] “The Telecom Corner: Modems,” Telecom Tidbits. *(V.29 in Group 3 fax and on four-wire leased lines.)* **Note on a source conflict:** this page dates V.32 to a 1984 CCITT definition, while [17] dates the deployed trellis-coded V.32 to 1986, and ITU lists a 1993 revision. I have therefore avoided asserting a single date for V.32 in the text. — https://www.jaredsec.com/modem/
### Television and cable standards
[24] ITU-T Recommendation J.83 (12/2007), *Digital multi-programme systems for television, sound and data services for cable distribution*. *(Annex B Mode 1 at 5.057 Msym/s and Mode 2 at 5.361 Msym/s; Annex C; the requirement that I and Q be square-root raised cosine filtered prior to modulation.)* — https://www.itu.int/rec/dologin_pub.asp?lang=e&id=T-REC-J.83-200712-I!!PDF-E&type=items
[25] ITU-T Recommendation J.83 (04/97). *(Annex B concatenated coding rationale; quasi-error-free defined as one error event per 15 minutes; 64-QAM and 256-QAM constellation mapping.)* — https://www.itu.int/rec/dologin_pub.asp?lang=e&id=T-REC-J.83-199704-S!!PDF-E&type=items
[26] ITU-T Recommendation J.83 (10/95). *(Annex A: 16-, 32- and 64-QAM with provision for extension to 128- and 256-QAM; quasi-error-free operation.)* — https://www.itu.int/rec/dologin_pub.asp?lang=e&id=T-REC-J.83-199510-S!!PDF-E&type=items
[27] Keysight, “Carrier for J.83 Annex B.” *(Annex B origin in testing and characterization of North American cable systems; level 1 and level 2 interleaving.)* — https://rfmw.em.keysight.com/wireless/helpfiles/n7623/Content/Main/Carrier_for_J.83_Annex_B.htm
[28] Keysight, “Carrier for DVB-C.” *(DVB-C constellations from 16- to 256-QAM; roll-off 0.15; theoretical maximum symbol rate of 6.96 MBaud in an 8 MHz channel.)* — https://rfmw.em.keysight.com/wireless/helpfiles/n7623b/Content/Main/Carrier_for_DVB-C.htm
[29] Rohde & Schwarz, *Test Receiver EFA — ATSC/8VSB, ITU-T J.83/B, M/N Analog TV* data sheet. *(Fixed symbol rates 5.056941 Msym/s for 64-QAM and 5.360537 Msym/s for 256-QAM.)* — https://www.testmart.com/webdata/mfr_pdfs/RS/EFA_ATSC_en.pdf
[30] Cable Television Laboratories, Inc., *Data-Over-Cable Service Interface Specifications*, downstream PMD sublayer requirements referencing ITU-T J.83 Annex B, including the 64-QAM / 5.056941 Msym/s / 6 bits-per-symbol and 256-QAM / 5.360537 Msym/s / 8 bits-per-symbol entries. — https://account.cablelabs.com/server/alfresco/54bd3432-9bb8-4331-84e8-9549b07f07c3
[31] Cable Television Laboratories, Inc., *DOCSIS 4.0 Physical Layer Specification*, CM-SP-PHYv4.0-I03-201202. — https://account.cablelabs.com/server/alfresco/553828be-22fe-4566-9d9d-85b79fb94a7e
### Measurement — the essential practical source
[32] R. Hranac (Cisco Systems) and B. Currivan (Broadcom), “Digital Transmission, Part 2,” *Communications Technology*, July 2007; presented in full at SCTE Cable-Tec Expo 2007. **The single most useful reference in this document for practical work.** *(MER and EVM definitions; peak and average constellation power formulas; MTA and its 4.77 dB asymptote; MER-to-EVM conversion; operational minimum RxMER values; the 3 to 6 dB margin guidance; implementation-loss ceiling; symbol-error floor; sample-count effects; nonlinearity effects measured across QPSK, 16-QAM and 256-QAM; MTA versus signal PAR; DOCSIS 24-tap pre-equalizer and equalizer span; the 36 dB CNR versus 21 dB MER micro-reflection case study; phase noise and the DRFI 50 kHz treatment; ingress cancellation; the point that a modem’s reported “SNR” is equalized RxMER.)* — https://www.cablefax.com/archives/digital-transmission-part-2
[33] R. Hranac’s guidance that MER is less than or at best equal to CNR but never greater, and that unequalized MER should be kept 3 to 6 dB above the failure threshold for the modulation in use, as quoted at length in the background section of US Patent 8,650,606. — https://image-ppubs.uspto.gov/dirsearch-public/print/downloadPdf/8650606
### Modern QAM inside OFDM
[34] B. Volpe, “OFDM and OFDMA,” Broadband Library. *(Downstream OFDM requires modulation orders up to 4096-QAM with optional support to 16384-QAM; upstream OFDMA requires up to 1024-QAM with optional support up to 4096-QAM.)* — https://broadbandlibrary.com/ofdma/
[35] Huawei, “What Is Wi-Fi 7? Wi-Fi 7 vs. Wi-Fi 6.” *(Wi-Fi 6 at 1024-QAM carrying 10 bits per symbol; Wi-Fi 7 introducing 4096-QAM carrying 12 bits, a 20% rate increase at the same coding rate.)* — https://info.support.huawei.com/info-finder/encyclopedia/en/WiFi+7.html
[36] “IEEE 802.11be Wi-Fi 7: Feature Summary and Performance Evaluation,” arXiv:2309.15951. *(4096-QAM rate gain, and the high SNR, EVM, cost and complexity demands it imposes.)* — https://arxiv.org/pdf/2309.15951
[37] RCR Wireless News, “1K QAM vs. 4K QAM: Why it matters for Wi-Fi 7.” *(IEEE transmitter target EVM of −38 dB for 4096-QAM, 3 dB tighter than 802.11ax.)* — https://www.rcrwireless.com/20220315/network-infrastructure/wi-fi/1k-qam-vs-4k-qam-why-it-matters-for-wi-fi-7
[38] HPE Aruba Networking, “Wi-Fi 6” technical documentation. *(802.11ac up to 256-QAM; 802.11ax up to 1024-QAM; 8 to 10 bits per subcarrier.)* — https://arubanetworking.hpe.com/techdocs/aos/wifi-design-deploy/generations/wifi6/
[39] “Revisiting Wireless Internet Connectivity: 5G vs Wi-Fi 6,” arXiv:2010.11601. *(States that 5G supports a maximum of 256-QAM. This is a 2020 source; see the caveat in §10 — verify against a current 3GPP specification before relying on it.)* — https://arxiv.org/pdf/2010.11601
### Standard texts and commentary
[40] J. G. Proakis, *Digital Communications*, 5th ed., McGraw-Hill, 2008. *(Standard reference for the square-QAM symbol error rate expression and the Gray-coded BER relationship.)*
[41] Commentary on Nyquist’s 1928 paper and the class of ISI-free Nyquist pulses: “Introduction to ‘Certain topics in telegraph transmission theory’,” *Proceedings of the IEEE*; and the summary of Nyquist pulse shaping in the background section of US Patent 7,142,613. — https://ieeexplore.ieee.org/iel5/5/21333/00989874.pdf
—
## A Closing Word
Six things to carry with you.
1. **QAM is amplitude modulation on two orthogonal carriers.** Everything else is refinement of that one trick.
2. **M points carry log₂M bits.** Doubling throughput requires squaring the constellation.
3. **Bandwidth comes from symbol rate and roll-off. Modulation order costs signal quality, not spectrum.**
4. **Every two extra bits per symbol costs about 6 dB.** This never stops being true, and it explains most of what you will see in the field.
5. **The “SNR” your modem reports is RxMER, not CNR.** Learn the difference and you will save yourself and your colleagues a great deal of wasted time.
6. **A clean spectrum display does not mean a healthy channel.** Linear distortion is invisible there and fatal to high-order QAM.
And the habit that matters more than any of the facts: when someone hands you a number, ask what produced it. Specification requirement, theoretical bound, lab measurement, or field estimate — those are four different things, and telling them apart is most of what separates a competent engineer from one people trust.
Now go and look at a constellation. It will teach you things this document cannot.
About the author
Greg Phillips is the founder and principal engineer of Think IT Data Solutions LLC. He has been building and operating physical infrastructure since 1984 — industrial control systems, the U.S. Navy Advanced Electronics Program, satellite communications at Comsat World Systems and Intelsat, and terrestrial transport and video distribution networks since. He founded Think IT Data Solutions in 2013.
Think IT Data Solutions builds and operates the layers underneath AI workloads — power and facilities, network transport, and GPU compute — for organizations that can’t run their most sensitive work on shared infrastructure.
Building something like this and want a second set of eyes on it? greg@thinkitdata.com · 678-643-4065