Difference Threshold Test for M -ary Orthogonal FSK Signaling in Rayleigh Fading Channels Young Gil Kim Sang Wu Kim Dept. of Electrical Engineering Dept. of Electrical Engineering University of Seoul Korea Advanced Institute of Science and Technology ygkim@ieee.org swkim@san.kaist.ac.kr Abstract-We present the difference threshold test (DTT) for M -ary orthogonal frequency shift keying (FSK) signaling in Rayleigh flat fading channels. The proposed DTT declares an erasure whenever the difference between the largest and the second largest energy detector outputs does not exceed a fixed threshold. We show that the DTT is an approximation of the Bayesian erasure test, but provides almost the same performance as the Bayesian erasure test. For (7, 3) and (31, 15) Reed-Solomon (RS)-coded systems, the DTT provides power gains of 0.8 dB and 0.4 dB over the ratio threshold test (RTT), respectively. We prove that the limiting performance of the DTT approaches that of the RTT for very large M . I. I NTRODUCTION The use of side information permits identification and erasure of symbols that have been impaired by channel effects such as fading, jamming, background noise, etc. Since more erasures can be corrected than errors, it is advantageous to determine the reliability of the received symbols and erase unreliable symbols prior to the decoding process. There are a number of methods for generating side information, and their performances have been analyzed in [1]-[7]. Particularly in [7], the difference threshold test (DTT) for binary orthogonal signaling was proposed. In this paper, we present the DTT for M -ary orthogonal frequency shift keying (FSK) signaling in Rayleigh fading channels. The DTT declares an erasure whenever the difference between the largest and the second largest energy detector outputs does not exceed a fixed threshold, γ. Hard decision corresponds to a special case of the DTT with γ set to zero. We show that the DTT is an approximation of the Bayesian erasure test, but provides almost the same performance as the Bayesian erasure test. A similar reliability test, called ratio threshold test (RTT), generates side information about the reliability of the received symbol by comparing the ratio of the largest to the second largest decision statistic [1],[8]. We show that the DTT provides power gains of 0.8 dB and 0.4 dB over the RTT for (7, 3) and (31, 15) Reed-Solomon (RS)-coded systems, respectively, in Rayleigh flat fading channels. We prove that the limiting performance of the DTT approaches that of the RTT for very large M . This paper consists of six sections. In Section II, we describe the system model. In Section III, we show that the DTT is an approximation of the Bayesian erasure test. In Section IV, we derive the probabilities of symbol error, symbol erasure, and decoding error probability. Also, the limiting performance of the DTT for very large M is analyzed. In Section V, numerical results are presented. Finally, conclusions are provided in Section VI. II. S YSTEM M ODEL Let si (t) = Acos(2πfi t), i = 1, 2, ..., M, 0≤t≤T (1) be the transmitted signal, where A is the signal amplitude and fi is the ith tone frequency. We assume that the tone frequencies {fi } are chosen such that the signals {si (t), i = 1, 2, ..., M } are noncoherently orthogonal, i.e., |fi+1 − fi | = 1/T . The received signal r(t) at the input of the energy detector given that s1 (t) is transmitted is r(t) = gAcos(2πf1 t + θ) + n(t), 0≤t≤T (2) where g is Rayleigh distributed, θ is uniformly distributed over [0, 2π), and n(t) is the white Gaussian noise of one-sided spectral density N0 . We assume that g and θ are independent, and are also independent of n(t). Then, the noncoherent mth energy detector output Zm , given that s1 (t) is transmitted, is Zm 2 2 T = r(t) cos(2πfm t)dt T 0 2 2 T + r(t) sin(2πfm t)dt (3) T 0 2 A2 T g cos θ + n c,1 2 2 (4) = A2 T + , m=1 2 g sin θ + ns,1 m = 1, (nc,m )2 + (ns,m )2 , where nc,m = ns,m = 2 T 2 T T 0 0 n(t) cos(2πfm t)dt, m = 1, 2, ..., M n(t) sin(2πfm t)dt, m = 1, 2, ..., M T are independent Gaussian random variables each with mean zero and variance N0 /2. Since g is Rayleigh distributed and θ is uniformly distributed over [0, 2π], g cos θ is a Gaussian random variable with mean zero and variance E[g 2 ]/2. Thus, 0-7803-7802-4/03/$17.00 © 2003 IEEE 2743 the conditional probability density function (pdf) PZm (z|1) of Zm , given that s1 (t) is transmitted, is 1 −z/2σ2 1, e m=1 2σ12 (5) PZm (z|1) = 1 −z/2σ02 e , m = 1 2σ 2 0 where 2σi2 = Ēs + N0 , i = 1 i=0 N0 , III. D IFFERENCE T HRESHOLD T EST In this section, we derive the DTT for M -ary orthogonal signaling from the Bayesian erasure test. With the Bayesian erasure test, the received signal is erased if the a posteriori probability of correct decision is less than a given threshold φ [2]: (7) Pr(s = sd |Z) ≤ φ where s ∈ {s1 , s2 , ..., sM } is the transmitted signal when represented in an orthogonal expansion form, Z = (Z1 , Z2 , ..., ZM ), and max m=1,2,...,M (8) Zm . We define log-likelihood ratio (LLR) Λm as Λm = ln Pr(s = sd |Z) . Pr(s = sm |Z) (9) Then e−Λm Pr(s = sd |Z) M e−Λm Pr(s = sd |Z) = = m=1 Pr(s = sm |Z) M M e−Λm = Assuming that each symbol is transmitted with equal probability, the LLR Λm in (9) can be expressed as Pr(Z|s = sd ) . Pr(Z|s = sm ) (12) Since the conditional joint pdf of Z is P (Z|s = sm ) = M −1 Z 1 1 1 m · 2 exp − 2 − 2 2 2σ1 2σ0 2σ1 2σ0 M i=1,i=m the LLR Λm in (12) is given by M 1 Zm 1 Zd 1 Λm = + Zi − 2 − 2 2 2 2 σ1 σ0 σ1 σ0 i=1,i=m 1 1 1 = − 2 (Zd − Zm ). 2 σ02 σ1 − 12 Pr(s = sd |Z) ≈ 1/ 1 + e where d = arg 1 σ2 0 − 1 σ2 1 max m=1,2,...,M,m=d (Zd −Zd ) Zm Zd − Zd ≤ γ , . (16) (17) (18) where Zd − Zd is the difference between the largest and the second largest energy detector outputs. IV. P ERFORMANCE A NALYSIS The probability of not correctly decoding PE for (N, K) RS-coded system with errors-and-erasures decoding is N −K−e N −K 2 n [per (γ)]e [pe (γ)]t [pc (γ)]n−e−t e, t e=0 t=0 (19) where per (γ) is the probability of symbol erasure, pe (γ) is the probability of symbol error, pc (γ) is the probability of correct decision, and N = M − 1. It follows from (18) that the DTT generates an erasure if PE = 1− max Zk ≤ Zm < max Zk + γ k=m k=m Zm ≥ max Zk + γ k=m (20) (21) for some m = 1, when s1 (t) is transmitted. Notice that conventional hard decision corresponds to the special case of γ = 0. Because of the symmetry of the signal set, the probability of symbol error is independent of the transmitted signal if each signal is transmitted equally likely. Thus, for our analysis we assume that s1 (t) is transmitted without loss of generality. Then, the probability of symbol error pe (γ) is (13) Zi , pe (γ) = Pr(∪M m=2 Zm ≥ max Zk + γ|1) k=m ∞ − z−γ 2 = (M − 1) (1 − e 2σ1 ) Zi i=1,i=d (14) 2744 (22) γ − z−γ 2 M is the argument of the second largest decision statistic. Since (16) is a monotonically increasing function of Zd − Zd , the DTT is to erase the symbols if (11) 1. m=1 Λm = ln for some parameter γ ≥ 0 and each m ∈ {1, 2, ..., M }, and produces an error if m=1 Pr(s = sd |Z) Thus, (10) Pr(s = sm |Z) (15) m=1 (6) where Ēs = E[g 2 ]A2 T /2 is the average received code symbol energy. If an RS code of rate r is used, then the average received information bit energy Ēb is Ēs /(r log2 M ). d = arg From (11) and (14), we get M − 12 σ12 − σ12 (Zd −Zm ) 0 1 e . Pr(s = sd |Z) = 1/ 1 − 2σz2 e 0 dz 2σ02 M −2 M −2 = (M − 1)e−γ/N0 k ·(1 − e · 2σ 0 )M −2 (23) k=0 k (−1) . (k + 1){(k + 1)(Ēs /N0 + 1) + 1} (24) Note that if we let γ = 0 in (24), then we get the wellknown probability of symbol error formula with hard decision in Rayleigh fading channels [9]. The probability of correct decision pc (γ) is Pr(Z1 ≥ max Zm + γ|1) m=1 ∞ − z−γ 1 − z2 2 (1 − e 2σ0 )M −1 2 e 2σ1 dz = 2σ1 γ M −1 M − 1 γ − E¯ +N = e s 0 k pc (γ) = (25) (26) k=0 (−1)k · 1 + k(Ēs /N0 + 1) The right-hand side of (33) is the maximum possible code rate for error-free communication. Following the same argument in [8], the optimum code rate ropt that minimizes (Ēb /N0 )min can be found from (34) as −1 e , γ>0 ropt = (36) 0.46, γ = 0. (27) For values of r given in (36), the (Ēb /N0 )min is e ln 2(= 2.75 dB), DTT (Ēb /N0 )min = 4.79(= 6.80 dB), hard decision. (37) This is equivalent to the minimum required Ēb /N0 for errorfree communication with the RTT [8]. whereas the probability of symbol erasure per (γ) is per (γ) = 1 − pc (γ) − pe (γ). In Appendix A, we derive the probability of symbol error pe (γ) and the probability of correct decision pc (γ) in detail. Limit Analysis In this subsection, we show that the performance of the DTT is the same as that of the RTT as M goes to infinity. The following propositions are proved in Appendix B. Proposition 1: lim pc (γ) = 2−(γ +1)N0 /r E¯b M →∞ where γ = V. N UMERICAL R ESULTS (28) (29) γ N0 ln M . Figure 1 is a plot of the probability of not correctly decoding PE versus Ēb /N0 for a (7, 3) RS code with M = 8 and (31, 15) RS code with M = 32 in a Rayleigh fading channel. The probability of not correctly decoding with the Bayesian erasure test is evaluated by computer simulation. The thresholds of the DTT and the RTT are optimally chosen for each Ēb /N0 . We find that the DTT provides power gains of 0.8 dB and 0.4 dB over the RTT for (7, 3) and (31, 15) RS codes, respectively, when the probability of not correctly decoding PE is 10−3 . Also, the DTT provides almost the same performance as the Bayesian erasure test. The power gain provided by the DTT over the RTT diminishes with increasing M (or N ), and in the limit of large M both the DTT and the RTT provides the same performance in accordance with the limit analysis. Proposition 2: lim pe (γ) = M →∞ VI. C ONCLUSION 0, γ>0 N − rE¯0 1−2 b , γ = 0. (30) Notice that γ = 0 and r = 1 in (30) yields the same result given in [10]. The asymptotic probability of symbol erasure can be obtained by applying Propositions 1 and 2 in (28), which results in ¯ 1 − 2−(γ +1)N0 /rEb , γ > 0 lim pe (γ) = (31) 0 γ = 0. M →∞ Thus, for large M , all errors becomes erasures if γ > 0. Following the analysis in [8], error free communication is possible with the DTT and errors-and-erasures decoding provided r 1 − per (γ) − 2pe (γ) −(γ +1)N /rE¯ 0 b 2 , γ>0 = −N0 /r E¯b − 1 γ = 0. 2·2 < or equivalently Ēb N0 > γ +1 r log2 (1/r) 1 r{1−log2 (r+1)} = (Ēb /N0 )min . γ>0 γ = 0. We presented the DTT for M -ary orthogonal FSK signaling in Rayleigh flat fading channels. The proposed DTT declares an erasure whenever the difference between the largest and the second largest energy detector outputs does not exceed a fixed threshold. We have shown that the DTT is an approximation of the Bayesian erasure test, but provides almost the same performance as the Bayesian erasure test. For (7, 3) and (31, 15) RS-coded systems, the DTT provides power gains of 0.8 dB and 0.4 dB over the RTT, respectively. The power gain provided by the DTT over the RTT diminishes with increasing M (or N ), and in the limit of large M both the DTT and the RTT provides the same performance. (32) (33) (34) (35) R EFERENCES [1] A. J. Viterbi, “A robust ratio-threshold technique to mitigate tone and partial band jamming in coded MFSK systems,” IEEE MILCOM, pp. 22.4.1-22.4.5, Oct. 1982. [2] C. W. Baum and M. B. Pursley, “Bayesian methods for erasure insertion in frequency-hop communication systems with partialband interference,” IEEE Trans. Commun., vol. 40, No. 7, pp. 1231-1238, Jul. 1992. [3] C. M. Keller and M. B. Pursley, “Diversity combining for channels with fading and partial-band interference,” IEEE J. Select. Areas Commun., vol. SAC-5, No. 2, pp. 248-260, Feb. 1987. 2745 [4] L. Yang and L. Hanzo, “Performance analysis of coded M ary orthogonal signaling errors-and-erasures decoding over frequency-selective fading channels,” IEEE J. Select. Areas Commun., vol. 19, No. 2, pp. 211-221, Feb. 2001. [5] L. Yang, K. Yen and L. Hanzo, “A Reed-Solomon coded DSCDMA system using noncoherent M -ary orthogonal modulation over multipath fading channels,” IEEE J. Select. Areas Commun., vol. 18, No. 11, pp. 2240-2251, Nov. 2000. [6] L. Yang and L. Hanzo, “Low-complexity erasure insertion in frequency-hopping spread-spectrum communications subject to fading and partial-band interference,” IEEE Globecom, Nov. 2529, 2001, San Antonio, Texas, USA, pp. 796-800. [7] I. Su and J. Wu, “Difference threshold test for asynchronous BFSK frequency-hopped multiple access systems over Rician channels,” Electronic Letters, vol. 35, No. 18, pp. 1512-1513, Sept. 1999. [8] S. W. Kim and W. Stark, “Performance limits of Reed-Solomon coded CDMA with orthogonal signaling in a Rayleigh-fading channel,” IEEE Trans. Commun., vol. 46, No. 9, pp. 1125-1134, Sep. 1998. [9] J. G. Proakis, Digital Communications. pp. 834, 4th Ed., McGraw-Hill, 2001. [10] M. V. Hegde and W. Stark, “Asymptotic performance of M -ary orthogonal signals in worst case partial-band and Rayleigh fading,” IEEE Trans. Commun., vol. 36, pp. 989-992, Aug. 1988. A PPENDIX B We prove the two propositions in Section 5. The derivation follows the same mathematical techniques in [8]. Proposition 1: lim pc (γ) = 2−(γ (38) = (M − 1) · Pr(Z2 ≥ max Zk + γ|1) (39) k=2 ∞ 1 − z2 = (M − 1) Pr(max Zk ≤ z − γ|1) 2 e 2σ0 dz k=2 2σ0 0 ∞ − z−γ 2 = (M − 1) (1 − e 2σ1 ) where ·(1 − e 2σ 0 )M −2 − 1 e 2σ02 = (M − 1)e−γ/N0 z 2σ 2 0 M −2 = → k=0 k (49) ln L = (M − 1) ln(1 − M −au ) 1 1 = (M − 1)(−M −au − M −2au − M −3au − · · ·) 2 3 (50) ≈ −M (1−au) , for large M implying 0, 1, lim L = M →∞ u < 1/a u > 1/a. (51) Therefore, we get lim pc (γ) ∞ = M →∞ 1/a − 2 1 − u −γ/2σ12 e 2 ·e du 2 (γ +1)N0 r E¯b (52) (53) . Proposition 2: (41) lim pe (γ) = 0, γ>0 N − rE¯0 1−2 M →∞ , γ = 0. b (54) Proof: Letting u = (z − γ)/σ12 in (23) gives The probability of correct decision pc (γ) for the DTT is (42) pc (γ) = Pr(Z1 ≥ max Zm + γ|1) m=1 ∞ 1 − z2 = Pr(max Zm ≤ z − γ|1) 2 e 2σ1 dz (43) m=1 2σ1 0 ∞ z − z−γ − 1 2 2 = (1 − e 2σ0 )M −1 2 e 2σ1 dz (44) 2σ1 γ M −1 γ M −1 − = e E¯s +N0 k lim pe (γ) = ∞ (M − 1)M −bu (1 − e−u/2 ) lim b M →∞ M →∞ 0 ·(1 − M −bu M −2 ) − ln M · e γ 2σ 2 0 du (55) where k=0 k (−1) . 1 + k(Ēs /N0 + 1) (48) (40) dz (−1) · . (k + 1){(k + 1)(Ēs /N0 + 1) + 1} · 1 2 ln 2 as M → ∞. If we let L = (1 − M −au )M −1 , then = M −2 k rĒb 1 + N0 log2 M rĒb 1 N0 2 ln 2 a γ − z−γ 2 (46) We derive (24) and (27). The probability of symbol error pe (γ) for the DTT is k=m +1)N0 /r E¯b γ where γ = N0 ln M. Proof: Letting u = (z − γ)/σ12 in (26) gives ∞ 1 u − γ2 lim pc (γ) = lim (1 − M −au ) e− 2 e 2σ1 du (47) M →∞ M →∞ 0 2 A PPENDIX A pe (γ) = Pr(∪M m=2 Zm ≥ max Zk + γ|1) M →∞ (45) 2746 b = → σ12 ln M rĒb . 2N0 ln 2 2σ02 (56) (57) as M → ∞. If we let L = (M − 1)M −bu (1 − M −bu )M −2 ln M · M −γ − where M −γ = e γ 2σ 2 0 (58) , then for large M ln L ≈ (1 − bu − γ ) ln M + ln(ln M ) + (M − 2) ln(1 − M −bu ) = ln(ln M ) + (1 − bu − γ ) ln M + (M − 2) 1 1 ·(−M −bu − M −2bu − M −3bu − · · ·). 2 3 (59) (60) Since −(M − 2)M −bu term in (60) predominates over all others, if 1 − bu > 0 (61) lim ln L = −∞, M →∞ implying lim L = 0, (62) if u < 1/b. M →∞ Thus, we can write lim pe (γ) = ∞ (M − 1)M −bu−γ lim b M →∞ M →∞ 1/b −u/2 ·(1 − e (63) ) ln M du. Note that the term (1 − M −bu )M −2 in (55) has in effect been eliminated from the integrand, but has been compensated for by increasing the lower limit from zero to 1/b. The compensation is exact as M approaches infinity. Performing integration in (63) gives lim pe (γ) M →∞ = = − N0 lim M −γ (1 − 2 rE¯b ) M →∞ 0, γ>0 N − rE¯0 1−2 b , γ = 0. (64) (65) Fig. 1. The probabilities of not correctly decoding for the hard decision, the RTT, the DTT, and the Bayesian erasure test for (7, 3) and (31, 15) RS-codes 2747