Computing Probability of Symbol Error

Transcription

1 Computing Probability of Symbol Error I When decision boundaries intersect at right angles, then it is possible to compute the error probability exactly in closed form. I The result will be in terms of the Q-function. I This happens whenever the signal points form a rectangular grid in signal space. I Examples: QPSK and 16-QAM I When decision regions are not rectangular, then closed form expressions are not available. I Computation requires integrals over the Q-function. I We will derive good bounds on the error rate for these cases. I For exact results, numerical integration is required. 2017, B.-P. Paris ECE 630: Statistical Communication Theory 207

2 Illustration: 2-dimensional Rectangle I Assume that the n-th signal was transmitted and that the representation for this signal is~s n =(s n,0, s n,1 ) 0. I Assume that the decision region G n is a rectangle G n = {~r =(r 0, r 1 ) 0 :s n,0 a 1 < r 0 < s n,0 + a 2 and s n,1 b 1 < r 1 < s n,1 + b 2 }. I I Note: we have assumed that the sides of the rectangle are parallel to the axes in signal space. Since rotation and translation of signal space do not affect distances this can be done without affecting the error probability. I Question: What is the conditional error probability, assuming that s n (t) was sent. 2017, B.-P. Paris ECE 630: Statistical Communication Theory 208

3 Illustration: 2-dimensional Rectangle I In terms of the random variables R k = hr t, F k i, with k = 0, 1, an error occurs if z error event 1 } { (R 0 apple s n,0 a 1 or R 0 s n,0 + a 2 ) or (R 1 apple s n,1 b 1 or R {z 1 s n,1 + b 2 ). } error event 2 I Note that the two error events are not mutually exclusive. I Therefore, it is better to consider correct decisions instead, i.e., ~ R 2 Gn : s n,0 a 1 < R 0 < s n,0 + a 2 and s n,1 b 1 < R 1 < s n,1 + b , B.-P. Paris ECE 630: Statistical Communication Theory 209

4 Illustration: 2-dimensional Rectangle I We know that R 0 and R 1 are I independent - because F k are orthogonal I with means s n,0 and s n,1, respectively I variance 2. I Hence, the probability of a correct decision is Pr{c s n } = Pr{ a 1 < < a 2 } Pr{ b 1 < N 1 < b 2 } Z a2 = p R0 s n (r 0 ) dr 0 a 1 a1 =(1 Q p N0 /2 b1 (1 Q p N0 /2 Z b2 p R1 s n (r 1 ) dr 1 b 1 a2 Q p ) N0 /2 b2 Q p ). N0 /2 2017, B.-P. Paris ECE 630: Statistical Communication Theory 210

5 Exercise: QPSK I Find the error rate for the signal set s n (t) = p 2E s /T cos(2pf c t + n p/2 + p/4), for n = 0,..., 3. I Answer: (Recall h P = d min 2 E b = 4 for QPSK) s! s E s Pr{e} = 2Q Q 2 = 2Q = 2Q s s 2E b h P E b 2!! Q 2 Q 2 E s! s! 2E b s h P E b 2!. 2017, B.-P. Paris ECE 630: Statistical Communication Theory 211

6 Exercise: 16-QAM = 5 3 (Recall h P = d min 2 E b for 16-QAM) I Find the error rate for the signal set (a I, a Q 2{ 3, 1, 1, 3}) s n (t) = p 2E 0 /Ta I cos(2pf c t)+ p 2E 0 /Ta Q sin(2pf c t) I Answer: (h P = d min 2 E b = 4) s! 2E Pr{e} = 0 3Q s! 4E = b 3Q 5 s! h = P E b 3Q Q2 9 4 Q2 9 4 Q2 s! 2E 0 s! 4E b 5 s h P E b 2!. 2017, B.-P. Paris ECE 630: Statistical Communication Theory 212

7 N-dimensional Hypercube I Find the error rate for the signal set with 2 N signals of the form (b k,n 2{ 1, 1}): r 2Es s n (t) = NT b k,n cos(2pnt/t ), for 0 apple t apple T I Answer: N Â k=1 Pr{e} = 1 1 Q = 1 1 Q = 1 1 Q s 2E s!! N N s!! N 2E b s!! N h P E b N Q 2 s! h P E b , B.-P. Paris ECE 630: Statistical Communication Theory 213

8 Comparison Symbol Error Probability BPSK (η P =4) QPSK (η P =4) 16-QAM (η P =1.6) 3D-Hypercube (η=4) E b / (db) I Better power efficiency h P leads to better error performance (at high SNR). 2017, B.-P. Paris ECE 630: Statistical Communication Theory 214

9 What if Decision Regions are not Rectangular? I Example: For 8-PSK, the probability of a correct decision is given by the following integral over the decision region for s 0 (t) Pr{c} = Z 1 p 2pN0 /2 exp( (x 0 Z x tan(p/8) p Es ) 2 2N o /2 1 p x tan(p/8) 2pN0 /2 exp( y 2 2 /2 ) dy {z } =1 2Q( x p tan(p/8) ) N0 /2 dx I This integral cannot be computed in closed form. 2017, B.-P. Paris ECE 630: Statistical Communication Theory 215

10 Union Bound I When decision boundaries do not intersect at right angles, then the error probability cannot be computed in closed form. I An upper bound on the conditional probability of error (assuming that s n was sent) is provided by: Pr{e s n }apple  Pr{k~ R ~s k k < k~ R ~s n k ~s n } k6=n k ~s = k ~s n k  Q k6=n 2 p. /2 I Note that this bound is computed from pairwise error probabilities between s n and all other signals. 2017, B.-P. Paris ECE 630: Statistical Communication Theory 216

11 Union Bound I Then, the average probability of error can be bounded by k ~s Pr{e} = k ~s n k  p n  Q p. n k6=n 2N0 I This bound is called the union bound; it approximates the union of all possible error events by the sum of the pairwise error probabilities. 2017, B.-P. Paris ECE 630: Statistical Communication Theory 217

12 Example: QPSK I For the QPSK signal set s n (t) = p 2E s /T cos(2pf c t + n p/2 + p/4), for n = 0,..., 3 the union bound is s! s! Pr{e} apple2q E s + Q 2E s. I Recall that the exact probability of error is s! s! E s Pr{e} = 2Q Q 2 E s. 2017, B.-P. Paris ECE 630: Statistical Communication Theory 218

13 "Intelligent" Union Bound I The union bound is easily tightened by recognizing that only immediate neighbors of s n must be included in the bound on the conditional error probability. I Define the the neighbor set N ML (s n ) of s n as the set of signals s k that share a decision boundary with signal s n. I Then, the conditional error probability is bounded by Pr{e s n }apple  Pr{k~ R ~s k k < k~ R ~s n k ~s n } k2n ML (s n ) k ~s = k ~s n k  Q k2n ML (s n ) 2 p. /2 2017, B.-P. Paris ECE 630: Statistical Communication Theory 219

14 "Intelligent" Union Bound I Then, the average probability of error can be bounded by k ~s Pr{e} k ~s n k appleâ p n  Q p. n k2n ML (s n ) 2N0 I We refer to this bound as the intelligent union bound. I It still relies on pairwise error probabilities. I It excludes many terms in the union bound; thus, it is tighter. 2017, B.-P. Paris ECE 630: Statistical Communication Theory 220

15 Example: QPSK I For the QPSK signal set s n (t) = p 2E s /T cos(2pf c t + n p/2 + p/4), for n = 0,..., 3 the intelligent union bound includes only the immediate neighbors of each signal: s! E s Pr{e} apple2q. I Recall that the exact probability of error is s! s! E s Pr{e} = 2Q Q 2 E s. 2017, B.-P. Paris ECE 630: Statistical Communication Theory 221

16 Example: 16-QAM I For the 16-QAM signal set, there are I 4 signals s i that share a decision boundary with 4 neighbors; bound on conditional error probability: Pr{e s i } = 4Q(q 2E0 ). I 8 signals s c that share a decision boundary with 3 neighbors; bound on conditional error probability: Pr{e s c } = 3Q(q 2E0 ). I 4 signals s o that share a decision boundary with 2 neighbors; bound on conditional error probability: Pr{e s o } = 2Q(q 2E0 N ). 0 I The resulting intelligent union bound is Pr{e} apple3q s 2E 0!. 2017, B.-P. Paris ECE 630: Statistical Communication Theory 222

17 Example: 16-QAM I The resulting intelligent union bound is s! 2E Pr{e} 0 apple3q. I Recall that the exact probability of error is s! s 2E Pr{e} = 3Q 0 9 2E 0 4 Q2!. 2017, B.-P. Paris ECE 630: Statistical Communication Theory 223

18 Nearest Neighbor Approximation I At high SNR, the error probability is dominated by terms that involve the shortest distance d min between any pair of nodes. I The corresponding error probability is proportional to Q(q dmin 2 ). I For each signal s n, we count the number N n of neighbors at distance d min. I Then, the error probability at high SNR can be approximated as s s Pr{e} 1 d N n Q( min )= N min Q( M 2 M 1 Â n=1 d min 2 ). 2017, B.-P. Paris ECE 630: Statistical Communication Theory 224

19 Example: 16-QAM I In 16-QAM, the distance between adjacent signals is d min = 2 p E s. I There are: I I I 4 signals with 4 nearest neighbors 8 signals with 3 nearest neighbors 4 signals with 2 nearest neighbors I The average number of neighbors is N min = 3. I The error probability is approximately, s! 2E Pr{e} 0 3Q. I same as the intelligent union bound. 2017, B.-P. Paris ECE 630: Statistical Communication Theory 225

20 Example: 8-PSK I For 8-PSK, eachq signal has 2 nearest neighbors at p distance d min = (2 2)Es. I Hence, both the intelligent union bound and the nearest neighbor approximation yield 0s p 1 0s Pr{e} (2 2)Es A = 3(2 2 2 p 2)Eb 1 A I Since, E b = 3E s. 2017, B.-P. Paris ECE 630: Statistical Communication Theory 226

21 Comparison Symbol Error Probability (with bound) BPSK (η P =4) QPSK (η P =4) 16-QAM (η P =1.6) 8-PSK (η=1.76) E b / (db) Solid: exact P e, dashed: approximation. For 8PSK, only approximation is shown. I The intelligent union bound is very tight for all cases considered here. I It also coincides with the nearest neighbor approximation 2017, B.-P. Paris ECE 630: Statistical Communication Theory 227

22 General Approximation for Probability of Symbol Error I From the above examples, we can conclude that a good, general approximation for the probability of error is given by s! dmin h Pr{e} N min Q p = P E b N min Q. 2N0 2 I Probability of error depends on I signal-to-noise ratio (SNR) E b / and I geometry of the signal constellation via the average number of neighbors N min and the power efficiency h P. 2017, B.-P. Paris ECE 630: Statistical Communication Theory 228

23 Bit Errors I So far, we have focused on symbol errors; however, ultimately we are concerned about bit errors. I There are many ways to map groups of log 2 (M) bits to the M signals in a constellation. I Example QPSK: Which mapping is better? QPSK Phase Mapping 1 Mapping 2 p/ p/ p/ p/ , B.-P. Paris ECE 630: Statistical Communication Theory 229

24 Bit Errors I Example QPSK: QPSK Phase Mapping 1 Mapping 2 p/ p/ p/ p/ I Note, that for Mapping 2 nearest neighbors differ in exactly one bit position. I That implies, that the most common symbol errors will induce only one bit error. I That is not true for Mapping , B.-P. Paris ECE 630: Statistical Communication Theory 230

25 Gray Coding I A mapping of log 2 (M) bits to M signals is called Gray Coding if I The bit patterns that are assigned to nearest neighbors in the constelation I differ in exactly one bit position. I With Gray coding, the most likely symbol errors induce exactly one bit error. I Note that there are log 2 (M) bits for each symbol. I Hence, with Gray coding the bit error probability is well approximated by Pr{bit error} N s! min log 2 (M) Q h P E b dmin. Q p. 2 2N0 2017, B.-P. Paris ECE 630: Statistical Communication Theory 231