0.9 Error analysis of digital communications

Recall the figure below, from the module Discrete-Time Implementation of Digital Communication . When the channel is trivial and noiseless and thepulses satisfy the Nyquist criterion (i.e., $g\left[k\right]*q\left[k\right]=\delta \left[k\right]$ ), the digital comm system will work perfectly, yielding $y\left[n\right]=a\left[n\right]$ .

In practice, however,

• the pulses $g\left[k\right]$ and $q\left[k\right]$ will be truncated to finite length,
• the channel will not be trivial (i.e., $\tilde{h}\left[k\right]\ne \delta \left[k\right]$ ), and
• the channel will not be noiseless (i.e., $\tilde{w}\left[k\right]\ne 0$ ),

leading to $y\left[n\right]\ne a\left[n\right]$ , in which case we must infer the value of $a\left[n\right]$ from the received samples ${\left\{y\left[m\right]\right\}}_{m=-\infty }^{\infty }$ . For now, we consider using only the single sample $y\left[n\right]$ to infer $a\left[n\right]$ .

Key question: What are the mechanisms by which errors are made?

To better understand error behavior, we can plot the “eye diagram” or the “constellation diagram” and calculate the symbol error rate (SER).

Eye diagrams

Usually used when $a\left[n\right]\in \mathbb{R}$ , the eye diagram is a plot which superimposes T -second segments of $Re\left\{y\right(t\left)\right\}$ over the time intervals $t\in [nT-\frac{T}{2},nT+\frac{T}{2})$ for many integers n .

In MATLAB, the eye diagram can be made by superimposing P -sample segments of $Re\left\{y_{\uparrow }\left[k\right]\right\}$ corresponding to intervals $k\in \{nP-\frac{P}{2},\cdots ,nP+\frac{P-1}{2}\}$ for many n . (Usually $P\ge 8$ .)

If $a\left[n\right]\in \mathbb{C}$ , eye diagrams can be plotted for both the “I” and “Q” channels using $Re\left\{y\right(t\left)\right\}$ and $Im\left\{y\right(t\left)\right\}$ , respectively.MATLAB for digital mod/demod with eye diagram:

"% design SRRC P = 16; % oversampling factoralpha = 0.5; % SRRC rolloff param D = 2; % truncation to [-DT,DT]g = srrc(D,alpha,P); % SRRC pulse Ng = length(g);% generate symbols N = 100; % # symbolsM = 4; % alphabet size sig2a = 1; % symbol variancea = pam(N,M,sig2a); % symbol sequence % pulse-amplitude modulatea_up = zeros(1,N*P); a_up(1:P:end) = a; % upsampled symbolsm = conv(a_up,g); % PAM % matched-filter demodulatey_up = conv(m,g); % use SRRC again % remove causal filtering delayleft = -P/2; % left edge of eye diagram k = left+[1:P*N]; % desired time indices dly = (Ng-1)/2+(Ng-1)/2;% delay due to pulsesy_up = y_up(k+dly); % remove delay y = y_up([1-left:P:P*N]);% downsample % plot received signalfigure(1) plot(left/P+[0:P*N-1]/P,y_up,’r’,[0:N-1],y,’.’);axis(‘tight’) title([‘SRRC (\alpha=‘,num2str(alpha),...‘) truncated to \pm’,num2str(D),’T’])xlabel([‘symbol index’])% plot eye diagram figure(2)Y_up = reshape(y_up,P,N); % extract segments plot(left/P+[0:P-1]/P,Y_up) % superimpose them title([‘SRRC (\alpha=‘,num2str(alpha),...‘) truncated to \pm’,num2str(D),’T’])xlabel([‘relative symbol index’])

Constellation diagrams

The constellation diagram is a plot of $Im\left\{y\right[n\left]\right\}$ vs. $Re\left\{y\right[n\left]\right\}$ for many integers n . When the comm system is working well, the points cluster around thesymbol alphabet values:

Recall that $y\left[n\right]\in \mathbb{C}$ due to the complex-baseband channel model, regardless of whether $a\left[n\right]\in \mathbb{R}$ or $a\left[n\right]\in \mathbb{C}$ .

Sometimes it is instructive to superimpose a plot of $Im\left\{y_{\uparrow }\left[n\right]\right\}$ vs. $Re\left\{y_{\uparrow }\left[n\right]\right\}$ , which approximates the trajectory of $y\left(t\right)$ in the complex plane:

MATLAB for digital mod/demod with constellation diagram:

% design SRRC P = 8; % oversampling factoralpha = 0.5; % SRRC rolloff param D = 2; % truncation to [-DT,DT]g = srrc(D,alpha,P); % SRRC pulse Ng = length(g);% generate symbols N = 100; % # symbolsM = 4; % (sqrt) alphabet size sig2a = 1; % symbol variancea = qam(N,M,sig2a); % symbol sequence % pulse-amplitude modulatea_up = zeros(1,N*P); a_up(1:P:end) = a; % upsampled symbolsm = conv(a_up,g); % PAM % matched-filter demodulatey_up = conv(m,g); % use SRRC again % remove causal filtering delayk = [1:P*N]; % desired time indicesdly = (Ng-1)/2+(Ng-1)/2;% delay due to pulses y_up = y_up(k+dly); % remove delayy = y_up(1:P:end); % downsample % plot received signalfigure(1) plot3([0:P*N-1]/P,real(y_up),imag(y_up),... ‘r’,[0:N-1],real(y),imag(y),’.’); xlabel(‘symbol index’);ylabel(‘I’); zlabel(‘Q’); view(20,30); % to see full trajectory%also try view(0,90), view(0,0), view(90,0) % plot constellation diagramfigure(2) plot(real(y_up),imag(y_up),’y’,real(y),imag(y),’.’);xlabel(‘I’); ylabel(‘Q’); title([‘SRRC (\alpha=‘,num2str(alpha),...‘) truncated to \pm’,num2str(D),’T’])axis(‘equal’);

Popular symbol alphabets

QAM : “quadrature amplitude modulation”
PAM : “pulse amplitude modulation”
PSK : "phase shift keying"

Note that: "QPSK" = 4-QAM = 4-PSK and "BPSK" = 2-PAM = 2-PSK

When the alphabet entries are spaced by Δ and picked with equal probability, the symbol variance ${\sigma }_a^2={E\left\{\right|a\left[n\right]|}^2\}$ obeys:

Decision regions

A reasonable way to infer the transmitted symbol $a\left[n\right]$ from the received sample $y\left[n\right]$ is to decide that $a\left[n\right]$ was the alphabet element nearest to $y\left[n\right]$ .

Nearest-element decision making is equivalent to using decision regions whose boundaries are equidistant from the two nearest alphabet elements:

When $a\left[n\right]=a$ , the symbol error rate (SER) equals the probability that $y\left[n\right]$ lies outside the decision region corresponding to alphabet member a . Writing $y\left[n\right]=a\left[n\right]+e\left[n\right]$ , we represent the cumulative effect of noise and ISI by the error $e\left[n\right]$ . Usually we model $e\left[n\right]$ as a Gaussian random variable with mean 0 and variance σ _e ² .

Symbol error rate (ser) for M -pam

Let's first consider an M -PAM alphabet, where $a\left[n\right]\in \mathbb{R}$ . Since the decision regions show that $Im\left\{y\right[n\left]\right\}$ is not useful, we'll consider only the real parts of $y\left[n\right]$ and $e\left[n\right]$ .

When $a\left[n\right]=a$ , we have $y\left[n\right]=a+e\left[n\right]$ , implying that $y\left[n\right]$ is Gaussian with mean a and variance σ _e ² , abbreviated as “ $\mathcal{N}(a,{\sigma }_e^2)$ ”. This is illustrated below for the case of 4-PAM:

Formally, we say that $p_{y\left[n\right]\left|a\right[n]}\left(y\right|a)$ , the probability density function (pdf) of $y\left[n\right]$ conditioned on $a\left[n\right]=a$ , obeys

Basically, $p_{y\left[n\right]\left|a\right[n]}\left(y\right|a)$ tells us how likely it is that $y\left[n\right]=y$ given that $a\left[n\right]=a$ .

Consider first the case where a is an “interior” (not an “edge”) element of the symbol alphabet.Given that $a\left[n\right]=a$ , we make an error when $y\left[n\right]<a-\frac{\Delta }{2}$ or when $y\left[n\right]>a+\frac{\Delta }{2}$ . To find the probability of the latter error event, i.e.,

we integrate $p_{y\left[n\right]\left|a\right[n]}\left(y\right|a)$ over $y\in (a+\frac{\Delta }{2},\infty )$ :

The integral represents the shaded area below:

This integral is often solved via

using the “Q function”:

While the Q function is not represented in MATLAB, it can be calculated using the “complementary error function” $erfc$ :

In any case, the latter error event occurs with probability

By symmetry, the former error event probability is also

Since these two events are disjoint, the probability of making a decision error on an interior symbol equals their sum:

For edge symbols, we experience half the decision error probability, since there is only one decision boundary to cross.

Finally, we average over the conditional error probabilities:

Using ${\sigma }_a^2=\frac{{\Delta }^2}{12}(M^2-1)$ , we can finally write

Symbol error rate (ser) for M ² -qam

With QAM, we have complex-valued $y\left[n\right]$ , $a\left[n\right]$ , $e\left[n\right]$ . We'll assume that $Re\left\{e\right[n\left]\right\}$ and $Im\left\{e\right[n\left]\right\}$ are uncorrelated and equal variance.To calculate SER, we can re-use the PAM approach with a few modifications:

1. integration is done on the complex plane,
2. σ _e ² -variance $e\left[n\right]$ $\Rightarrow$ $\frac{{\sigma }_e^2}{2}$ -variance $Re\left\{e\right[n\left]\right\}$ & $Im\left\{e\right[n\left]\right\}$ ,
3. M ² -QAM has 4 corner points, $4(M-2)$ edge points, and $M^2-4M+4$ interior points,
4. calculate $Pr\left\{\text{error}\right|a\left[n\right]=a\}$ via $1-Pr\left\{\text{correct}\right|a\left[n\right]=a\}$ , since the regions of integration are simpler:

   

After a bit of algebra, we find

Bit error rate (ber) and gray coding

With an M -ary alphabet, there are ${log}_{2}M$ bits per symbol, so 1 symbol error could cause up to ${log}_{2}M$ bit errors.

Gray coding is a clever way of mapping bits to symbols so that neighboring symbols differ by only a single bit.Since the vast majority of errors occur when $y\left[n\right]$ falls into a neighboring decision region, Gray coding yields BER $\approx$ SER.