Make your paper interactiveCreate

Text from the interactive explanation

On the MIMO Channel Capacity of Multi-Dimensional Signal Sets

S. X. Ng and L. Hanzo

Original paper

Ng & Hanzo · IEEE Transactions on Vehicular Technology · 2006

Your phone and the mast it talks to can each carry several antennas. Every spare antenna faces a choice: repeat the message so it survives the flicker of a city’s radio echoes — or carry a new message and go faster. In 2006 Soon Xin Ng and Lajos Hanzo worked out how many bits each choice can carry when radios speak with real, finite alphabets — including new multi-dimensional ones.

Capacity in bits per symbol (16-QAM ceiling: 4).

Each glowing cluster is one of 16 “letters” of the 16-QAM alphabet, as the receiver hears them. When the path fades, the letters shrink into the noise and blur together — misread samples glow coral. The strip beneath each lane shows the last 8 seconds of path strength (deep fades in coral); the bar shows bits per symbol right now, its white tick the running average.

Simulated Rayleigh fading · capacity from the paper’s Eq. 11

Chapter 1

An alphabet of points

A digital radio doesn’t send words. It sends symbols, one after another, each picked from a small menu of points on a plane. Each point is one combination of a wave’s strength (amplitude) and its timing (phase), and the whole menu is called a constellation . 16-QAM has 16 points on a 4 × 4 grid, so every symbol picks one of 16: that is log2 16 = 4 bits .

Noise blurs every received point into a little cloud. When clouds overlap, the receiver sometimes picks the wrong neighbour. Capacity is the most bits per symbol that can get through reliably, assuming ideal error-correcting codes.

Shannon’s unrestricted bound, log2(1 + SNR)

chosen alphabet, still air

Turn the noise right down. Will 16-QAM ever carry more than 4 bits per symbol?

Make a guess to see the answer.

In 1948 Claude Shannon found the limit for an idealised signal whose points are spread continuously, like a bell curve: log2(1 + SNR) bits per symbol. The paper calls this the unrestricted bound . Real radios use finite alphabets — PSK and QAM — and their capacity, the one this paper computes, can never exceed log2 M bits for M points. Try 256-QAM: more letters, higher ceiling, but the points sit closer together and noise misreads them more easily.

Chapter 2

The air fights back

In a city, your signal never travels one clean path. It bounces off buildings, buses and windows and arrives as a crowd of echoes, each delayed a little differently. The receiver hears their sum. As you move, the echoes slide in and out of step: sometimes they add up, sometimes they nearly cancel. That flicker is Rayleigh fading , and the moments of near-cancellation are deep fades .

Fading costs signal even on average. Compare 16-QAM on a still link — only noise, which engineers call AWGN — with the same alphabet over Rayleigh fading:

still air (noise only)

Rayleigh fading

—

The closer you push toward the ceiling, the more the fades hurt: the receiver must ride out moments when the letters collapse into the noise. This is the gap that spare antennas can attack.

Chapter 3

Say it twice: diversity

The first use for spare antennas: send the same message along several independent paths. A deep fade on one path is unlikely to coincide with a deep fade on the others.

The neatest way to do it is a space-time block code (STBC). In Alamouti’s two-antenna version, two antennas send the same pair of symbols over two time slots — the second time swapped and mirrored. That pattern is orthogonal: the receiver untangles the two symbols with simple arithmetic and is left with one clean channel whose strength is the sum of all the path strengths (the paper’s Eq. 4). For 16-QAM it only ever compares 16 candidates. The total transmit power stays fixed, so each of Nt antennas gets 1/Nt of it.

With Nt transmit and Nr receive antennas there are Nt × Nr paths. The paper’s Eq. 11 gives the capacity of this “repeat” link for any alphabet, including multi-dimensional ones; the authors evaluated it by averaging over many random fades.

chosen antennas

a single path, for comparison

still air

Rayleigh fading, repeat (STBC)

the gap

Gap at 3 bits: —

With 4 receive antennas, will extra transmit antennas still help as much?

Make a guess to see the answer.

Chapter 4

Say two things at once: multiplexing

The other use for spare antennas: let each send a different symbol at the same moment. This is spatial multiplexing, the idea behind Bell Labs’ BLAST. Two antennas with 16-QAM carry 4 + 4 = 8 bits per symbol period — double.

The catch is at the receiver. Each receive antenna hears both symbols added together, each scaled and rotated by its own fading path. The 16 × 16 = 256 combinations land as 256 points that slide around as the paths change. To decode, the receiver weighs every combination — maximum-likelihood (ML) detection — and when two combinations land close together, noise can swap them.

The paper’s Eq. 20 gives the capacity of this “say more” link (ML-detected MIMO) for any alphabet; Eq. 11 gives the “repeat” link. Same two transmit antennas, same 16-QAM:

say more: ML-detected MIMO (Eq. 20)

repeat: orthogonal STBC (Eq. 11)

—

With one receive antenna, the two strategies give the same capacity at low SNR; the extra bits of “say more” only appear once the signal is strong. With two receive antennas the gap widens as SNR grows. Saying more reaches higher capacity — at the cost of far more detection effort (256 candidates instead of 16) and a stronger signal.

The 4-bit race

You need 4 bits per symbol, essentially error-free, with 2 transmit and 2 receive antennas. Three ways to get there. Which needs the least energy per bit ?

Make a guess to see the answer.

Chapter 5

A bigger alphabet in time: extra dimensions

Now the paper’s own ingredient. So far every symbol was a point on a flat, two-dimensional plane (a cosine and a sine component). Ng and Hanzo study multi-dimensional alphabets called L-orthogonal PSK and QAM.

Split each symbol period into V shorter time slots. Light up exactly one slot with an ordinary L-point symbol and keep the others silent. Which slot is lit carries log2 V extra bits. The alphabet grows to M = V × L symbols spread over D = 2V dimensions (each slot brings its own cosine and sine). The price: each slot is V times shorter, so the signal needs V times more bandwidth. The paper’s Fig. 1 shows 8-PSK with V = 2: 16 symbols, 4 bits, 4 dimensions. The authors also extended the idea from PSK to L-orthogonal QAM.

—

—

—

With several antennas saying different things, each antenna picks its own slot and point, so an ML receiver faces (V·L)Nt combinations (the paper’s Fig. 2 and Eq. 17). Does the bigger alphabet pay off?

V = 1 (2 dimensions), fading

V = 2 (4 dimensions), fading

same, still air

—

Per symbol, the four-dimensional alphabet wins: it starts with an extra bit and holds more capacity through the fades. But it uses twice the bandwidth. Count bits per second per hertz against the energy spent on each bit and V = 2 falls below V = 1 — although at very low energy both curves run into the same unrestricted bound, Shannon’s limit of −1.59 dB. More bits per symbol, fewer bits per hertz.

Chapter 6

Design challenge: eight bits, five recipes

Every design below has exactly M = 256 possible transmissions, so each tops out at the same 8 bits per symbol. All use 2 receive antennas over Rayleigh fading. They differ in how they spend their resources — alphabet size L, time slots V, transmit antennas Nt — and in whether the antennas repeat or say more. (After the paper’s Figs. 7–8.)

1 Predict. Pick the recipe you think carries the most bits per symbol at 10 dB.

2 Run the experiment. Your browser evaluates the paper’s Eq. 11 and Eq. 20 by Monte Carlo averaging — thousands of random fades and noise draws per point — and the estimates settle as samples pile up.

samples per point: 0

3 The verdict

Pick a recipe, then run the experiment to rank them.

Build your own link

Choose an alphabet, slots, antennas and a strategy. Your browser computes the capacity curve from the paper’s formulas on the spot (Rayleigh fading). Pin up to four curves to compare.

Chapter 7

Back to the question

Repeating narrows the fading gap

Transmit diversity pulls the capacity of a fading link toward the still-air capacity (Fig. 3).

…but matters less with many receive antennas

When the receiver already has plenty of diversity, the gap is small and extra transmit diversity adds little (Fig. 4).

Saying more carries more, at a price

ML-detected multiplexing reaches higher capacity than orthogonal STBC, but needs more SNR and much more detection effort (Figs. 5–6).

Extra dimensions trade bandwidth for bits

Multi-dimensional alphabets raise bits per symbol but lower bits per hertz (Figs. 3–4, 7–8).

In other words: once the receiver listens with enough antennas, spend the spare transmit antennas on saying more, and get robustness from coding over time instead. For the most bits per hertz, the authors found the best design to be two-dimensional ML-detected MIMO with the smallest alphabet and the most transmit antennas (more hardware). For the most bits per symbol at a given ceiling: ML MIMO with the most dimensions, a small alphabet and many antennas (more bandwidth).

Go back to the opening scene with four receive antennas: the lanes become almost equally steady, and repeating buys little.

That is what these formulas give designers: a way to compare alphabets, dimensions and antenna strategies on paper, before building any hardware.

C = log2 M − (1/M) Σm E[ log2 Σn exp(Ψm,n) | xm ]

For two-dimensional signals, Ψm,n = Σi ( −| χi(xm − xn) + Ωi |² + | Ωi |² ) / (χi N0), where χi = Σj |hi,j|² is the summed strength of the paths into receive antenna i (Eq. 4) and Ωi its noise. For D-dimensional signals the sum also runs over every dimension (Eq. 10).

C = log2 M − (1/M) Σm E[ log2 Σn exp(Ψm,n) | Xm ], Ψm,n = Σv ( −‖H[v](xm[v] − xn[v]) + n[v]‖² + ‖n[v]‖² ) / N0

with M = (V·L)Nt candidate transmissions (Eq. 17).

Ψm,n measures how much closer the received signal sits to the true symbol xm than to a rival xn. If every rival is clearly farther away, the sum inside the logarithm is just 1 and the capacity reaches its ceiling, log2 M. When rivals crowd close, the logarithm eats into that ceiling. The expectation E[·] averages over random fades and noise. No neat closed form exists, so the authors estimated it by Monte Carlo averaging : draw many random fades and noise samples and average the result. The curves on this page were recomputed the same way.

Uncorrelated Rayleigh fading: every transmit–receive path fades independently.

The transmitter has no knowledge of the channel.

All symbols are equally likely.

Total energy is split equally over the Nt transmit antennas (amplitude scaled by 1/√Nt).

SNR (Eq. 12) = Es / (D·N0/2): average symbol energy over the noise energy in its D dimensions.

Bandwidth efficiency (Eq. 23): η = C / (D/2) = C / V bits/s/Hz, plotted against energy per bit Eb/N0 = SNR / η.

The unrestricted bounds (Eq. 13 for STBC, Eq. 22 for general MIMO) assume idealised Gaussian signals instead of a finite alphabet.

The moving scenes use a sum-of-sinusoids (Clarke/Jakes) model of Rayleigh fading, slowed so that fades last long enough to watch; at walking speed real fades come and go in milliseconds. The opening scene’s “now” value is the still-air 16-QAM capacity at the instantaneous strength of the combined paths; its running average converges on the Eq. 11 value printed beside it.

Sources & credits

S. X. Ng and L. Hanzo, “On the MIMO Channel Capacity of Multi-Dimensional Signal Sets,” IEEE Transactions on Vehicular Technology, vol. 55, no. 2, pp. 528–536, March 2006. DOI: 10.1109/TVT.2005.863357 · author copy (PDF)

Visuals are original. Curves are a Monte Carlo re-computation of the paper’s capacity formulas (Eq. 11, 13, 20); highlighted dB figures are quoted from the paper. The hero’s fading is simulated at an illustrative speed.

Background ideas named along the way: C. E. Shannon (1948), the capacity of a noisy channel; S. M. Alamouti (1998), the two-antenna space-time block code; V. Tarokh, N. Seshadri and A. R. Calderbank (1998), space-time trellis codes; G. J. Foschini (1996), layered space-time architecture (BLAST).

Headings are set in Instrument Serif (SIL Open Font License).

Topics in this iPaper

Engineering
the capacity of the DCMC is more pertinent in the design of channel coded modulation schemes

Section I (Introduction), page 528

Mathematics
the Monte Carlo averaging method is the most efficient approximation technique of computing the expectation terms

Section IV, page 533

multiple input multiple output
the capacity of Multi-Input Multi-Output (MIMO) systems

Abstract, page 528

channel capacity
we derive channel capacity formulae applicable to MIMO systems employing multi-dimensional signal sets

Section I (Introduction), page 529

space time block coding
Space-Time Block Coding (STBC) [6], [9] were designed for achieving diversity

Section I (Introduction), page 528

spatial multiplexing
transmits independent information in parallel over the MIMO channel for the sake of achieving multiplexing gain

Section I (Introduction), page 528

quadrature amplitude modulation
we extended the concept of L-orthogonal PSK signalling to L-orthogonal QAM signalling

Section II (Multidimensional Signal Set), page 529

rayleigh fading
the achievable capacity of the Rayleigh-fading channel increases

Section V (Numerical Results), page 533

monte carlo method
the Monte Carlo averaging method is the most efficient approximation technique of computing the expectation terms

Section IV, page 533

On the MIMO Channel Capacity of Multi-Dimensional Signal Sets — Interactive explanation | iPaper.ai