On the Low-SNR Capacity of Phase-Shift Keying with Hard-Decision Detection ustafa Cenk Gursoy Department of Electrical Engineering University of Nebraska-Lincoln, Lincoln, NE 68588 Email: gursoy@engr.unl.edu arxiv:0705.04v [cs.it] ay 007 Abstract The low-snr capacity of -ary PSK transmission over both the additive white Gaussian noise AWGN) fading channels is analyzed when hard-decision detection is employed at the receiver. Closed-form expressions for the first second derivatives of the capacity at zero SNR are obtained. The spectralefficiency/bit-energy tradeoff in the low-snr regime is analyzed by finding the wideb slope the bit energy required at zero spectral efficiency. Practical design guidelines are drawn from the information-theoretic analysis. The fading channel analysis is conducted for both coherent noncoherent cases, the performance penalty in the low-power regime for not knowing the channel is identified. I. INTRODUCTION Phase modulation is a widely used technique for information transmission, the performance of coded phase modulation has been of interest in the research community since the 960s. One of the early works was conducted in [4] the capacity error exponents of a continuous-phase modulated system, in which the transmitted phase can assume any value in [, ), is studied. ore recent studies include [], [], [5], [6], [7], [8]. Kaplan Shamai studied in [5] the achievable information rates of differential phase-shift keying DPSK) while reference [6] investigated the capacity of - ary PSK over an additive white Gaussian noise AWGN) channel with unknown phase that stays constant for a block of symbols. Pierce in [] considered hard-decision detection of PSK signals trasmitted over the AWGN channel compared the performances of -, - 4-phase modulations. Pierce also provided in [] an expression for the bit energy required by -ary PSK at zero spectral efficiency. The authors in [] analyzed the spectral efficiency of coded PSK DPSK with soft- hard-decision detection. Reference [7] analyzed the energy efficiency of PSK when it is combined with onoff keying for transmission over noncoherent Rician fading channels. Recent work by Zhang Laneman [8] investigated the achievable rates of PSK over noncoherent Rayleigh fading channels with memory. The low-snr capacity of PSK with soft detection is wellunderstood. For instance, Verdú [] has shown that quaternary PSK QPSK) transmission over the AWGN or coherent fading channels is optimally efficient in the low-snr regime, achieving both the minimum bit energy of.59 db optimal wideb slope which is defined as the slope of the spectral This work was supported in part by the NSF CAREER Grant CCF- 054684. efficiency curve at zero spectral efficiency. Although soft detection gives the best performance, hard-decision detection decoding is preferred when reduction in the computational burden is required [0]. Such a requirement, for instance, may be enforced in sensor networks [9]. oreover, at very high transmission rates such as in fiber optic communications, obtaining multiple-bit resolution from A/D converters may not be possible []. Finally, it is of interest to underst the fundamental limits of hard-decision detection so that the performance gains of soft detection can be identified weighed with its increased complexity requirements. otivated by these considerations the fact that the performance difference of hard soft detections are more emphasized at low power levels, we study in this paper the low-snr capacity of - ary PSK over both the AWGN fading channels when a hard-decision detection is employed at the receiver end. II. CHANNEL ODEL We consider the following channel model r k = h k s xk +n k k =,,... ) x k is the discrete input, s xk is the transmitted signal when the input is x k, r k is the received signal during the the k th symbol duration. h k is the channel gain. h k is a fixed constant in unfaded AWGN channels, while in flat fading channels, h k denotes the fading coefficient. n k } is a sequence of independent identically distributed i.i.d.) zero-mean circularly symmetric Gaussian rom variables denoting the additive background noise. The variance of n k is E n k } =. We assume that the system has an average power constraint of E s xk } E k. At the transmitter, -ary PSK modulation is employed for transmission. Hence, the discrete input, x k, takes values from 0,,..., }, if x k = m, then the transmitted signal in the k th symbol duration is s xk = s m = Ee jθm ) θ m = m m = 0,,...,, is one of the phases available in the constellation. At the receiver, the detector makes hard decisions for every received symbol. Therefore, each received signal r k is mapped to one of the points in the constellation set Ee jm/,m = 0,,..., } before the decoding step. We assume that maximum likelihood decision rule is used at the detector. Note that with hard-decision detection, the channel can be
now regarded as a symmetric discrete channel with inputs outputs. III. PSK OVER AWGN CHANNELS We first consider the unfaded AWGN channel assume that h =. In this case, the conditional probability density function of the channel output given the channel input is f r x r x = m) = f r sm r s m ) ) = e r sm m = 0,...,. 4) It is well-known that the maximum likelihood detector selects the constellation point closest to the received signal r. We denote the signal at the output of the detector by y assume that y 0,,..., }. Note that y = l for l = 0,,..., means that the detected signal is Ee jl/. The decision region for y = l is the two-dimensional region D l = r = r e jθ : l ) θ < l+) }. 5) With quantization at the receiver, the resulting channel is a symmetric, discrete, memoryless channel with input x 0,,..., } output y 0,,..., }. The transition probabilities are given by P l,m = Py = l x = m) 6) l ) = P θ < l+) ) x = m 7) = l+) l ) f θ sm θ s m )dθ 8) f θ sm θ s m ) is the conditional probability density function of the phase of the received signal given that the input is x = m, hence the transmitted signal is s m. It is wellknown that the capacity of this symmetric channel is achieved by equiprobable inputs the resulting capacity expression [] is C SNR) = log Hy x = 0) 9) = log + P l,0 logp l,0 0) SNR = E, H ) is the entropy function, P l,0 = Py = l x = 0). In order to evaluate the capacity of general -ary PSK transmission with a hard-decision detector, the transition probabilities P l,0 } should be computed. Starting from 4), we can easily find that f θ s0 θ s 0 ) = e SNR + Qx) = x SNR θ cosθe SNRsin Q ) SNRcos θ) ) e t / dt. ) Since the channel is memoryless, we henceforth, without loss of generality, drop the time index k in the equations for the sake of simplification. Since f θ s0 is rather complicated, closed-form expressions for the capacity is available only for the special cases of = 4: C SNR) = log hq SNR SNR)) C 4 SNR) = C hx) = xlogx x)log x). For the other cases, the channel capacity can only be found by numerical computation. On the other h, the behavior of the capacity in the low- SNR regime can be accurately predicted through the secondorder Taylor series expansion of the capacity, which involves Ċ 0) C 0), the first second derivatives of the channel capacity in nats/symbol) with respect to SNR at SNR = 0. In the following result, we provide closed-form expressions for these derivatives. Theorem : The first second derivatives of C SNR) respectively, ψ) = 6 Ċ 0) = = 4 sin, ) 8 ) = C 0) = = 4 ) = 4 ψ) 5 )sin 4) + 4)sin 4 sin ) sin. 5) Proof : The main approach is to obtain Ċ0) C 0) by first finding the derivatives of the transition probabilities P l,0 }. This can be accomplished by finding the first second derivatives of f θ s0 with respect to SNR. However, the presence of SNR in second part of ) complicates this approach because df θ s 0 =. In order to circumvent SNR=0 dsnr this problem, we define the new variable a = SNR consider f θ s0 θ s 0 ) = e a + a cosθe a sin θ Q ) a cos θ). 6) The following can be easily verified. f θ s0 θ s 0 ) =, df θ s0 da = cosθ, d f θ s0 da = cosθ, df θ s 0 da = cosθsin θ, df θ s 4 0 da 4 = 6cos θ 8cos4 θ. Using the above derivatives, we can find the first through fourth derivatives of P l,0 with respect to a at a = 0. Using the derivatives ofp l,0 performing several algebraic operations, )
we arrive to the following Taylor expansion for C a) at a = 0: C a) = φ )a +φ )a +φ )a 4 +oa 4 ) 7) = φ )SNR+φ )SNR / +φ )SNR +osnr ) 8) 8) follows due to the fact that a = SNR. In the above expansion, φ ) = sin cos i, 9) i= φ ) = sin sin ) 6 sin cos i, i= 0) φ ) = +) sin + 6 6 sin i= + 4 )sin ) sin sin + 4 sin sin i= cos 4i i= cos 4 i cos i. ) We immediately conclude from 8) that Ċ 0) = φ ). Note that the expansion includes the term SNR / which implies that C 0) = ± for all. However, it can be easily seen that φ ) = 0 for all, at =, φ ) = 0.78. Therefore, while C 0) =, C 0) = φ ) for. Further algebraic steps simplification yields ) 4). Remark: We should note that the first derivative expression ) has previously been given in [] through the bit energy expressions. In addition, Verdú in [] has provided the second derivative expression for the special case of = 4. Hence, the main novelty in Theorem is the second derivative expression for general. First second derivative expressions are given together for completeness. The following corollary provides the asymptotic behavior as. Corollary : In the limit as, the first second derivatives of the capacity at zero SNR converge to lim Ċ 0) = 4 lim C 0) = 8+8. ) 6 In the low-power regime, the tradeoff between bit energy spectral efficiency is a key measure of performance. The normalized energy per bit can be obtained from = SNR CSNR) ) CSNR) is the channel capacity in bits/symbol. The maximum achievable spectral efficiency in bits/s/hz is given by ) Eb C = CSNR) bits/s/hz 4) if we, without loss of generality, assume that one symbol occupies a s Hz time-frequency slot. Two important notions regarding the spectral-efficiency/bit-energy tradeoff in the low power regime are the bit-energy required at zero spectral efficiency, = log e C=0 Ċ0), 5) the wideb slope, S 0 = Ċ0)) C0), 6) which gives the slope of the spectral efficiency curve E C / ) at zero spectral efficiency []. Therefore, b C=0 S 0 constitute a linear approximation to the spectral efficiency curve in the low-snr regime. Since these quantities depend only Ċ0) C0), the bit energy at zero spectral efficiency wideb slope achieved by-ary PSK signals with a hard-decision detector can be readily obtained by using the formulas ) 4). Corollary : The bit energy at zero spectral efficiency wideb slope achieved by -ary PSK signaling are given by = log e = log C=0 e 7) S 0 = 4 sin = 0 = 6 = 4 4 8 sin4 ψ) 5 8) respectively. As it will be evident in numerical results, generally the C=0 is the minimum bit energy required for reliable transmission when. However, for =, the minimum bit energy is achieved at a nonzero spectral efficiency. Corollary : For -PSK modulation, the minimum bit energy is achieved at a nonzero spectral efficiency. This corollary follows immediately from the fact that C 0) = which implies that the slope at zero SNR of SNR/C SNR) is. This lets us conclude that the bit energy required at zero spectral efficiency cannot be the minimum one. The fact that -PSK achieves its minimum bit energy at a nonzero spectral efficiency is also pointed out in [] through numerical results. Here, this result is shown analytically through the second derivative expression. Figure plots the spectral efficiency curves as a function of the bit energy for hard-detected PSK with different constellation sizes. As observed in this figure, the informationtheoretic analysis conducted in this paper provides several practical design guidelines. We note that although -PSK 4-PSK achieve the same minimum bit energy of 0.69 db at zero spectral efficiency, 4-PSK is more efficient at low but nonzero spectral efficiency values due to its wideb slope being twice that of -PSK. -PSK is better than -PSK for
Spectral efficiency bits/s/hz) 0.9 0.8 0.7 0.6 0.5 0.4 0. 0. = 04 = = 6 = 0 = 8 = 4 = = P l,0,h = f θ s0,hθ s 0,h) = l+) l ) e h f θ s0,hθ s 0,h)dθ 0) + SNR h SNR cosθe h SNRsin θ Q ) h SNRcos θ) ) with SNR = E/. Through a similar analysis as in Section III, we have the following result on the derivatives of the capacity. 0. 0 0.5 0 0.5.5 E /N db) b 0 Fig.. Spectral efficiency C / ) vs. bit energy / for -ary PSK with a hard-decision detection in the AWGN channel. spectral efficiency values greater than 0.08 bits/s/hz below which -PSK performes worse than both 4-PSK. - PSK achieves its minimum bit energy of 0.88 db at 0.4 bits/s/hz. Operation below this level of spectral efficiency should be avoided as it only increases the energy requirements. We further observe that increasing the constellation size to 8 provides much improvement over 4-PSK. 8-PSK achieves a minimum bit energy of 0.8 db. Further increase in provides diminishing returns. For instance, there is little to be gained by increasing the constellation size more than as -PSK achieves a minimum bit energy of 0.58 db the minimum bit energy as is 0.54 db. Note that 0.54 db still presents a loss of approximately.05 db with respect to the fundamental limit of.59 db achieved by soft detection. We find that the wideb slopes of = 8,0,6,, 04 are.44,.5,.64,.69,.7. The similarity of the wideb slope values is also apparent in the figure. As, the wideb slope is.77. Finally, note that the wideb slope of -PSK, as predicted, is 0. IV. PSK OVER FADING CHANNELS A. Coherent Fading Channels In this section, we consider fading channels assume that the fading coefficients h k } are known at the receiver but not at the transmitter. The only requirements on the fading coefficients are that their variations are ergodic they have finite second moments. Due to the presence of receiver channel side information CSI), scaled nearest point detection is employed, the analysis follows along lines similar to those in the previous section. Hence, the treatment will be brief. Note that in this case, the average capacity is C SNR) = log + E h P l,0,h logp l,0,h } 9) Theorem : The first second derivatives of C SNR) Ċ 0) = E h } = 4 sin E h }, ) 8 ) E h 4 } = C 0) = = 4 ) E h 4 ) } = 4 ψ)e h 4 } 5 respectively, ψ) is given in 5). Note that the first derivative second derivatives of the capacity at zero SNR are essentially equal to the scaled versions of those obtained in the AWGN channel. The scale factors are E h } E h 4 } for the first second derivatives, respectively. In the fading case, we can define the received bit energy as Eb r = E h }SNR C SNR) 4) as E h }SNR is the received signal-to-noise ratio. It immediately follows from Theorem that E r b / C=0 in the coherent fading channel is the same as that in the AWGN channel. On the other h, the wideb slope is scaled by E h }) /E h 4 }. B. Noncoherent Fading Channels In this section, we assume that neither the receiver nor the transmitter knows the fading coefficients h k }. We further assume that h k } are i.i.d. proper complex Gaussian rom variables with mean Eh k } = d 0 variance E h k d } = γ. Now, the conditional probability density function of the channel output given the input is r dsm f r sm r s m ) = γ s m + ) e γ sm +. 5) Recall that s m = Ee jθm } are the PSK signals hence s m = E for all m = 0,...,. Due to this constant magnitude property, it can be easily shown that the maximum likelihood detector selects s k as the transmitted signal if 4 Rers k ) > Rers i ) i k 6) d 0 is required because phase cannot be used to transmit information in a noncoherent Rayleigh fading channel d = 0. 4 6) is obtained when we assume, without loss of generality, that d = d. 4
s k is the complex conjugate of s k, Re denotes the operation that selects the real part. Therefore, the decision regions are the same as in the AWGN channel case. In this case, the channel capacity is C SNR) = log + P l,0 logp l,0 7) P l,0 = l+) l ) f θ s0 θ s 0 )dθ 8) f θ s0 θ s 0 ) = d SNR e γ SNR+ 9) d + d SNR SNR γ SNR+) cosθe γ SNR+ sin θ )) Q d SNR γ SNR+ cos θ. 40) The following results provide the first second derivatives of the capacity at zero SNR, the bit energy wideb slope in the low-snr regime. Theorem : The first second derivatives of C SNR) C 0) = Ċ 0) = d = d 4 sin, 4) 8 ) d 4 4 d γ = = 4 ) d 4 4 d γ = 4 ψ) d 4 d γ sin 5 4) respectively, ψ) is given in 5). Corollary 4: In the limit as, the first second derivatives of the capacity at zero SNR converge to lim Ċ 0) = d 4. 4) lim C 0) = 8 +8) d 4 d γ. 44) 6 In the noncoherent fading case, the received bit energy is E r b = d +γ )SNR. 45) C SNR) Corollary 5: The received bit energy at zero spectral efficiency wideb slope achieved by -ary PSK signaling are given by + = K) loge = + C=0 K) loge 46) 4 sin = + K 0 = S 0 = 6 = 4 + K 5 4 8 sin4 ψ)+ K sin 47) Spectral Efficiency bits/s/hz) 0.5 0. 0.5 0. 0.05 = 04 = 0 = 8 = 6 = = 4 = = 0.5.5 4 4.5 5 / db) Fig.. Spectral efficiency C / ) vs. bit energy / in the noncoherent Rician fading channel with Rician factor K = d γ =. respectively, ψ) is given in 5), K = d γ is the Rician factor. Remark: If we let d = γ = 0, or equivalently let K, the results provided above coincide with those given for the AWGN channel. Fig. plots the spectral efficiency curves as a function of the bit energy for -ary PSK transmission over the noncoherent Rician fading channel with K =. Note that conclusions similar to those given for Fig. also apply for Fig.. The main difference between the figures is the substantial increase in the bit energy values as a penalty of not knowing the channel. For instance, 4-PSK now achieves a minimum bit energy of.79 db while 8-PSK attains.69 db. As, the minimum bit energy goes to.467 db. REFERENCES [] J. R. Pierce, Comparison of three-phase modulation with two-phase four-phase modulation, IEEE Trans. Commun, vol. 8, pp. 098-099, July 980. [] G. Kramer, A. Ashikhmin, A. J. van Wijngaarden, X. Wei, Spectral efficiency of coded phase-shift keying for fiber-optic communication, IEEE/OSA J. Lightwave Technol., vol., pp. 48-445, Oct. 00. [] S. Verdú, Spectral efficiency in the wideb regime, IEEE Trans. Inform. Theory, vol. 48, pp. 9-4, June 00. [4] A. D. Wyner, Bounds on communication with polyphase coding, Bell Syst. Tech. J., vol. XLV, pp. 5-559, Apr. 966. [5] G. Kaplan S. Shamai Shitz), On the achievable information rates of DPSK, IEE Proceesings, vol. 9, pp. -8, June 99. [6]. Peleg Shlomo Shamai Shitz), On the capacity of the blockwise incoherent PSK channel, IEEE Trans. Commun, vol. 46, pp. 60-609, ay 998. [7]. C. Gursoy, H. V. Poor, S. Verdú, The noncoherent Rician fading channel Part II : Spectral efficiency in the low power regime, IEEE Trans. Wireless Commun., vol. 4, no. 5, pp. 07-, Sept. 005. [8] W. Zhang J. N. Laneman, How good is phase-shift keying for peaklimited Rayleigh fading channels in the low-snr regime?, to appear in IEEE Trans. Inform. Theory. [9] X. Luo G. B. Giannakis, Energy-constrained optimal quantization for wireless sensor networks, IEEE SECON, pp. 7-78, 4-7 Oct. 004. [0] J. G. Proakis, Digital Communications. New York: cgraw-hill, 995. [] T.. Cover J. A. Thomas, Elements of Information Theory. New York: Wiley, 99. 5