2012 IEEE International Symposium on Information Theory Proceedings On the Capacity of Additive White Alpha-Stable Noise Channels Jihad Fahs, Ibrahim Abou-Faycal Dept. of Elec. and Comp. Engineering, American University of Beirut Beirut 1107 2020, Lebanon jjf03@aub.edu.lb, iaf@alum.mit.edu Abstract— Many communication channels are reasonably modeled to be impaired by additive noise. Recent studies suggest that many of these channels are affected by additive noise that is best explained by alpha-stable statistics. We study in this work such channel models and we characterize the capacity-achieving input distribution for those channels under fractional order moment constraints. We prove that the optimal input is necessarily discrete with a compact support for all such channels. Interestingly, if the second moment is viewed as a measure of power, even when the channel input is allowed to have infinite second moment, the optimal one is found to have finite power. I. I NTRODUCTION In modeling the noise effect in a communication channel, it is common to assume that the noise is additive and Gaussian distributed due to the Central Limit Theorem which in layman terms states that, if the noise is due to multiple independent sources, then their cumulative effect is asymptotically Gaussian distributed. Recent studies however suggest that in some circumstances, the additive noise is better explained by non-Gaussian statistics as is the case when modeling the multiuser interference in a network [1], [2]. This was also observed in the context of modeling the Radio-Frequency interference in embedded wireless laptop transceivers [3]. In both setups, the noise was found to be better modeled as a Symmetric Alpha-Stable (SαS ) random variable. This noise model poses multiple challenges to system designers and we intend to address one of them in this work: • Despite the fact that the Probability Density Function (PDF) of a SαS random variable was proven to exist and exhibit rather “nice” properties, no closed-form expression is known except in two special cases: the Cauchy and the Gaussian ones. • Such noise distributions have infinite variance, which implies that the received signal has potentially infinite power and any analysis that is based on a Hilbert space approach is not valid anymore. The “classical” signal processing techniques used in receiver design are hence not necessarily applicable, let alone optimal. • The channel capacity of a basic linear channel where the output is simply a noisy version of the input is not known and optimal signaling schemes are not known either. A few attempts were made along this direction and as far This work was supported by AUB’s University Research Board. 978-1-4673-2579-0/12/$31.00 ©2012 IEEE 299 as the authors know, only numerical evaluation of some achievable rates have been conducted [4], [5]. In this work, we attempt to partially answer these questions. Earlier information theoretic studies of linear channels under general additive non-Gaussian noise have been conducted as in [6], [7]. However, typically these studies either require a finite noise variance which is not the case in this model, or impose a peak input power condition, a constraint that we will not impose in this work. It is worth noting as well that, it is not clear whether the SαS noise model fits within the general categories described in [7]. In this paper, we study the additive linear channel model, where a SαS noise variable is added to the input that is generally constrained to satisfy a fractional r-th order moment constraint: E [|X|r ] ≤ a, for any real value of r > 1. Clearly, the average power constrained channel is a special case of our setup. More importantly, if r is chosen to be less than 2, then the input is technically allowed to have infinite power similarly to the noise and the received signal: a primary motivation behind this particular choice of the input constraint. Since the received signal is of infinite power, and if a transceiver’s hardware is capable of coping with such a received signal, then it should be able to cope with a similar transmitted one. The rest of the paper is organized as follows: We first describe the channel model in Section II and derive a lower bound on the output PDF in Section III. Our main results are presented in Section IV, and Section V concludes the paper. Two Appendices I and II establish useful technical lemmas. II. C HANNEL MODEL We consider the real linear discrete-time memoryless channel: Y = X + N, (1) where Y is the channel output and X is its input which is subject to the fractional r-th order moment constraint: E [|X|r ] ≤ a for some real-valued r > 1 and a > 0. The additive noise variable N , which is independent of the input, is assumed to behave statistically as an alpha-stable Random Variable (RV) whose characteristic function is defined by: α Φ(ω) = eiγω−|cω| , where α ∈ [1, 2), c > 0 and γ ∈ R. We note first that, from an information theoretic perspective, assuming the values γ = 0 and c = 1, does not affect the generality of the problem since it relates to the equivalent channel where the output is scaled (Y − γ)/c. Henceforth we assume that the noise, the PDF of which is pN (·), is a standard SαS RV whose characteristic α function is e−|ω| . For additive channels with absolutely continuous noise, the output PDF exists [8]. We denote by pY (y) = p(y; F )† the output density induced by an input probability distribution function F . Note that, for α ∈ [1, 2), Z 1 1 1 1 −|ω|α < √ < 1, Γ e dω = pN (n) ≤ 2π πα α π the bound in the two ranges follow similar steps. We present them hereafter for y ≤ −y0 : Z pN y − x dF (x) p(y; F ) ≥ 1 r Zx:|x|≤(2a) ≥ (1 − ǫ) g(y − x) dF (x) (4) 1 for all n ∈ R and therefore, since pN (·) is continuous [9, p.11 Th.1.9], less than 1 and has R as support, so is pY (·). where we used the lower bound in (2) to obtain (4), and (5) is justified by the fact that g(y − x) is a decreasing function of x on the given interval. The last inequality is an application of Markov’s inequality and the r-th moment constraint. III. P RELIMINARIES In order to study the capacity of channel (1) and characterize its achieving input, we derive in this section a lower bound on the output PDF. It is known that alpha-stable distributions possess heavy tails that decay asymptotically as power laws. More specifically [9, Th.1.12, p.14], let 0 < ǫ < 12 , there exists n1 > 0 such that ∀ |n| ≥ n1 , (1 − ǫ) g(n) < pN (n) < (1 + ǫ) g(n), (2) πα −(α+1) α = |n|4kα+1 . where g(n) = α π sin( 2 )Γ(α)|n| Additionally, alpha-stable distribution functions are known to be analytic on R whenever α ∈ [1, 2) [10, p.183 Sec.36]. Whenever α > 1 the radius of convergence is ∞ and is no less than one if α = 1. Hence, if Sδ = {z ∈ C : |ℑ(z)| < δ}, pN (·) is analytically extendable over Sδ for δ small enough. Moreover, we prove in Appendix I the following novel bound: there exists κ > 0 and n2 > 0 such that κ , |pN (z)| ≤ |ℜ(z)|α+1 ∀ z ∈ Sδ : |ℜ(z)| ≥ n2 , (3) which we use to establish Lemma 2 in Appendix II. In the remainder of this paper, we fix ǫ and a small enough δ, and we define n0 = max{n1 , n2 } where n1 and n2 are as above. Lemma 1. For any input probability distribution function, The output PDF pY (y) of channel (1) is lowerbounded by kα h(2a) r1 −yi(α+1) y ≤ −y0 h(y) = kα y ≥ y0 , i(α+1) 1 h (2a) r +y 1 where y0 = n0 + (2a) r . Proof: First note that the statement of this lemma does not hold for all inputs in the Gaussian case. The derivations of †p Y (y) = p(y; F ) = R p(y|x) dF (x) = R x:|x|≤(2a) r 1 1 1 g y − (2a) r ≥ Pr |X| ≤ (2a) r 2 h i−(α+1) 1 ≥ kα (2a) r − y , IV. M AIN R ESULT In the following we state and prove the main theorem of this paper Theorem. The capacity achieving input of channel (1), whenever r > 1 and α ∈ [1, 2) is compactly supported, discrete with no accumulation point except possibly at zero. Proof: With the sufficient topological and functional requirements satisfied, we proceed as in [8] and we write the Karush-Kuhn-Tucker (KKT) conditions as being necessary and sufficient conditions for the unique optimal input to satisfy. These conditions state that an input random variable X ∗ with Cumulative Distribution Function F ∗ achieves the capacity C of an r-th moment constrained channel if and only if there exists ν ≥ 0 such that, Z r ν(|x| − a) + C + H + pN (y − x) ln p(y; F ∗ ) dy ≥ 0, (6) for all x ∈ R, with equality if x is a point of increase of F ∗ . H is the entropy of the noise and is finite by virtue of equation (2) along with the continuity of pN (·). Assume that the points of increase of F ∗ have a positive accumulation point and let Z s(z) = ν(z r − a) + C + H + pN (y − z) ln p(y; F ∗ ) dy. The function s(z) is analytic on Sδ \R− by virtue of Lemma 2 proven in Appendix II and by using for z r the principal branch of the logarithm [11, p.85 Prop.1.6.4]. Since s(z) has accumulating zeros on the positive real axis, by the identity theorem [12], s(·) is identically null on Sδ \ R− . Therefore, Z ν(xr − a) + C + H = − pN (y) ln p(y − x; F ∗ ) dy, (7) for all x ∈ R+ . Examining the integral, let Z ∞ pN (y) ln p(y − x; F ∗ ) dy t(x) = − −∞ = pN (y − x) dF (x). 300 (5) I1 + I2 + I3 where the interval of integration is divided into three subintervals: (−∞, x − y0 ), [x − y0 , x + y0 ] and (x + y0 , ∞). Using Lemma 1, Z x−y0 pN (y) ln p(y − x; F ∗ ) dy I1 = − −∞ ≤ − (ln kα ) [1 − Pr(N > x − y0 )] + Z x−y0 h i 1 pN (y) ln (2a) 2r + x − y dy. (α + 1) −∞ 1 Using the fact that ln x ≤ x α+1 , Z x−y0 i h 1 pN (y) ln (2a) 2r + x − y dy −∞ Z x−y0 1 h i α+1 1 pN (y) (2a) 2r + x − y ≤ dy −∞ 1 α+1 1 ≤ EN (2a) 2r + x − N h i 1 1 1 ≤ (2a) 2r(α+1) + |x| α+1 + EN |N | α+1 . (8) Pn Pn β We justify (8) by the hfact thati| i=1 Xi |β ≤ i=1 |Xi | 1 for β ≤ 1. Since EN |N | α+1 is finite [10, p.179 Sec.35 Th.3], An inspection of equation (8) along with the fact that Pr(N > x − y0 ) ∼ 4kαα (x − y0 )−α as x → ∞ [9, Th.1.12, p.14] shows that I1 is o(|x|r ). Similarly, it can be shown that I3 = o(|x|r ). As for I2 , Z x+y0 pN (y) ln p(y − x; F ∗ ) dy I2 = − x−y0 ≤ FN (x + y0 ) − FN (x − y0 ) max ln |y|≤y0 1 , pY (y) where the maximum exists and is positive since pY (y) is continuous and 0 < pY (y) < 1. I2 is also o(|x|r ) and hence t(x) = o(xr ) for which equation (7) is impossible unless ν = 0 which is non-sensible since the constraint is biding, and can be formally ruled out [13]. This leads to a contradiction and rules out the assumption of having a positive accumulation point. Similarly, we rule out a negative accumulation point and in conclusion, the set of points of increase of F ∗ has no accumulation point except possibly at 0. Compact support of the optimal input Having ruled out the possibility that an optimal distribution has a set of points of increase with a non-zero accumulation point, we can immediately conclude that this set is countable on R by virtue of the fact that R is Lindelöf, and X ∗ is discrete. Denoting by {xi }i≥0 the points of increase of F ∗ (x) and pi = Pr(X = xi ) > 0, the output probability density can be lowerbounded by pY (y) > pi p(y|xi ), ∀y ∈ R; ∀i ≥ 0. Examining the last term in the Left-Hand Side (LHS) of the KKT condition (6), it can be proven to be o(|x|r ) as in the preceding section. Hence, the LHS of (6), LHS(x) = ν|x|r + o(|x|r ) diverges to infinity as |x| goes to infinity for any ν > 0. 301 However, LHS(xi ) = 0 for all points of increase xi . If these were arbitrarily large, then ν will be zero which is not possible. V. C ONCLUSION We studied the channel capacity and the capacity-achieving input distributions for channels affected by SαS additive noise. We proved that the optimal input is discrete and compactly supported for any fractional order input moment constraint larger than one. We note that, even when imposing a fractional lower order moment constraint on the input and allowing it to be potentially of infinite variance, since the optimal input is compactly supported, its variance is nevertheless finite. This indicates that over these types of channels, surprisingly, it does not increase the rates to transmit with infinite power. We acknowledge that measuring the power of a signal via the second moment might not be adequate in this context. We also note that the methodology developed in this paper may be readily used for Pareto distributions. Finally, the results may be readily generalized for non-linear channels of the form Y = sgn(X)|X|s + N , where N is a SαS noise random variable, and where the input is subject to E [|X|t ] ≤ a, t > s. An appropriate change of variable as in [14] yields readily a similar result. A PPENDIX I We study in this appendix the rate of decay of standard SαS distributions with α ∈ [1, 2) on the horizontal strip S = {z ∈ C : |ℑ(z)| < 1}. We prove that |pN (z)| = 1 , for z ∈ S as |ℜ(z)| → ∞‡ . O |ℜ(z)| α+1 1 For the case α = 1, pN (z) = π1 1+z 2 which clearly satisfies our assertion. When α > 1, pN (·) can be formally extended on C as: Z ∞ Z ∞ α α 1 1 e−izt−|t| dt = e−ixt+yt−|t| dt pN (z) = 2π −∞ 2π −∞ Z ∞ 1 X y n ∞ n −ixt−|t|α = t e dt, (9) 2π n=0 n! −∞ where we wrote z = x + iy, and used Lebesgue’s Dominated Convergence Theorem (DCT) for the interchange in (9). This is justified since N ∞ X yn X α α |y|n n −|t|α n −ixt−|t| t e |t| e = e|y||t|−|t| , ≤ n! n! n=0 n=0 which is integrable for α > R1. α ∞ We study next Jn (x) = ˆ −∞ tn eixt−|t| dt†† . It is known that limx→∞ xα+1 J0 (x) = 2 Γ(α + 1) sin πα 2 [15, p.134]. ‡ By definition, given three positive real-valued functions f (.), g(.) and h(.) we write f (z) = O (g(z)) as h(z) → ∞ if and only if there exist two positive scalars, c > 0 and ho such that f (z) ≤ cg(z), ∀ h(z) ≥ ho . †† Note that J (−x) = (−1)n J (x) and the n-th derivative of p (x) is, n n N (n) pN (x) = in (−i)n Jn (−x) = Jn (x), 2π 2π n ∈ N. For n ≥ 1 and x > 0, we have Z ∞ Z ∞ α α xtn eixt−t dt + (−1)n xtn e−ixt−t dt xJn (x) = 0 0 Z Z ∞ ∞ α n−1 ixt−|t|α tn+α−1 eixt−t dt t e dt − iα = in 0 −∞ Z ∞ n+α−1 −ixt−tα n−1 t e dt (10) +(−1) 0 iα (11) In + (−1)n−1 In xn+α where equation (10) is obtained by integration by parts and regrouping, and where In is the complex conjugate of In : Z ∞ Z ∞ γ αγ α eiv −ζv dv, tn+α−1 eixt−t dt = γ In = xn+α = i n Jn−1 (x) − 0 0 1 where γ = n+α , ζ = x−α and the change of variable is v = (xt)n+α . Since as x → ∞, ζ → 0+ , Z ∞ γ αγ lim In = γ lim eiv −ζv dv x→∞ ζ→0 0 Z ∞ γ iγθ αγ iαγθ +iθ = γ lim lim eiv e −ζv e dv (12) ζ→0 θ→0 0 Z ∞ γ iγθ eiv e +iθ dv (13) = γ lim θ→0 0 Z γ = γ lim lim eiz dz, θ→0 R→∞,ρ→0 L1 iθ where z = ve for some small positive θ and L1 = {z ∈ C : z = veiθ , 0 < ρ ≤ v ≤ R}. Equation (12) is justified by ζ αγ DCT since the integrand’s norm is less than e− 2 v which is integrable. Similarly, (13) is justified since the order between the two limits is interchangeable and the integrand in (12) is γ upperboundedR by e−v sin γθ which is integrable. To evaluate γ the limit of L1 eiz dz, we use contour integration over C shown in Figure 1. where C1 and C2 are arcs of radius R, 6 C2 Fig. 1. L2 ) L1 1 C1 - The contour C. and ρ respectively between angles θ and ϕ = ˆ mod 2π. Note that since we are interested in the limit as θ goes to zero, we can always choose it small enough in order to have the contour counter-clockwise. Finally, L2 is a line connecting the extremities of the arcs. γ Now since f (z) = eiz is analytic on and inside C (by choosing an appropriate branch cut in the plane), by Cauchy’s theorem [11, p.111 Sec.2.2], I Z Z Z Z 0= f (z) dz = f (z) + f (z) + f (z) + f (z). L1 C1 R→∞ θ R→∞ θ =0 γ where the interchange is valid because Re−R Rsin(γφ) is def (z) = 0. creasing as 0 < γθ ≤ γφ ≤ π2 . Hence, limR→∞ C1 R Similarly, it can be shown that limρ→0 C2 f (z) = 0. It π remains to evaluate the integral on L2 where z = tei 2γ , Z ∞ Z π γ iπ ei 2γ eit e 2 dt f (z)dz = − lim R→∞,ρ→0 L 0 2 Z ∞ γ π π 1 1 i 2γ i 2γ −t = −e . Γ dt = −e e γ γ 0 π In conclusion limx→∞ In = ei 2 (n+α) Γ(n + α), and by (11), we can write for n ≥ 1 lim xn+α+1 Jn (x) − i nxn+α Jn−1 (x) = Wn x→∞ π π = ˆ − iαΓ(n + α) ei 2 (n+α) + (−1)n−1 e−i 2 (n+α) (14) Since lim xα+1 J0 (x) exists and is equal to U0 = 2 Γ(α + ˆ lim xn+α+1 Jn (x) exists for all 1) sin( πα 2 ), then Un = x→∞ non-negative n and is equal to, h i Un = inUn−1 + Wn = Pn−1 ik n n! i U0 + k=0 (n−k)! Wn−k . Furthermore, for n ≥ 0, # " n−1 X |Wn−k | |Un | ≤ n! |U0 | + (n − k)! k=0 # " n−1 X (n + 1 − k) = 2 n! n2 + 3n + 2 . ≤ 4 n! 1 + k=0 The last inequality is justified using the fact that for α ∈ (1, 2), Γ(α + l) is increasing for l ∈ N∗ . Now using (9), lim xα+1 |pN (z)| Z ∞ X 1 y n ∞ n −ixt−|t|α α+1 = lim x t e dt 2π x→∞ n! −∞ n=0 ∞ 1 X yn lim xα+1 Jn (−x) = 2π n=0 n! x→∞ x→∞ π 2γ C On C1 , we have: Z Z ϕ iφ iRγ eiγφ lim f (z)dz = lim iRe e dφ R→∞ R→∞ C1 θ Z ϕ Z ϕ γ γ −R sin(γφ) lim Re−R sin(γφ) dφ Re dφ = ≤ lim L2 C2 302 ≤ ≤ (15) ∞ ∞ 1 X |y|n 1 X |y|n lim xn+α+1 |Jn (x)| = |Un | 2π n=0 n! x→∞ 2π n=0 n! ∞ 1X n 2 |y| n + 3n + 2 , π n=0 which is finite because |y| < 1, and where we used the fact that f (x) = |x| is continuous. The interchange in (15) is valid because the end result is finite. In conclusion, lim xα+1 |pN (z)| < ∞ which concludes our proof. x→∞ A PPENDIX II In this appendix we prove the analyticity of the integral function on a strip around the real axis. Lemma 2. The function i(·; F ) : Sδ → C defined by: Z ∞ pN (y − z) ln p(y; F ) dy, z → i(z; F ) = − (16) −∞ is analytic over Sδ , for all F such that R |x|r dF (x) ≤ a. Proof: To prove this lemma, we will make use of Morera’s theorem: a) We start first by proving the continuity of i(·; F ) on Sδ . In fact, let z0 ∈ Sδ , ρ > 0. and z ∈ Sδ such that |z − z0 | ≤ ρ, Z pN (y − z) ln p(y; F ) dy lim i(z; F ) = − lim z→z z→z0 Z 0 (17) =− lim pN (y − z) ln p(y; F ) dy z→z0 Z = − pN (y − z0 ) ln p(y; F )dy = i(z0 ; F ). (18) Equation (18) is justified by pN (y − z) being a continuous function of z on Sδ , and (17) by DCT. Indeed, in what follows we find an integrable function r(y) such that, pN (y − z) ln p(y; F ) = −pN (y − z) ln p(y; F ) ≤ r(y), for all z ∈ Sδ such that |z − z0 | ≤ ρ and for all y ∈ R. Using (3), we upperbound first |pN (y − z)|. For z ∈ Sδ such that |z − z0 | ≤ ρ and y ≤ −(n0 + |ℜ(z0 )| + ρ), then y ≤ −n0 + ℜ(z0 ) − ρ and κ κ |pN (y − z)| ≤ ≤ max |y − ℜ(z)|α+1 ζ∈Sδ :|ζ−z0 |≤ρ |y − ℜ(ζ)|α+1 κ = , |y − ℜ(z0 ) + ρ|α+1 Similarly, for y ≥ (n0 + |ℜ(z0 )| + ρ) ≥ (n0 + ℜ(z0 ) + ρ), κ pN (y − z) ≤ . |y − ℜ(z0 ) − ρ|α+1 Since p(y; F ) < 1, we upperbound − ln p(y; F ) using Lemma 1 to obtain # " 1 r − y (2a) (α + 1)κ y ≤ −ψ ln 1 α+1 |y − ℜ(z0 ) + ρ| kα(α+1) −M ln M1 r(y) = # |y| < ψ " 1 r (α + 1)κ (2a) + y y ≥ ψ. 1 |y − ℜ(z0 ) − ρ|α+1 ln kα(α+1) where o n 1 ψ = max n0 + |ℜ(z0 )| + ρ, n0 + (2a) r M = max max |y|≤ψ ζ∈Sδ :|ζ−z0 |≤ρ |pN (y − ζ)| M1 = min py (y). |y|≤ψ 303 Note that M is finite since the maximization of pN (·) is taken over a compact set that is included in Sδ where it is analytic, and 0 < M1 < 1 since py (·) is positive and continuous. The fact that r(y) is integrable concludes the proof of continuity of i(z; F ). b) To continue the proof of analyticity, we need to integrate i(·; F ) on the boundary ∂∆ of a compact triangle ∆ ⊂ Sδ . We denote by |∆| its perimeter, η0 = minz∈∂∆ ℜ(z), η1 = maxz∈∂∆ ℜ(z) and φ = max{n0 + max{|η0 |, |η1 |}, n0 + 1 (2a) r }. By similar arguments as above and using the fact that R pN (·) is analytic on Sδ one can prove that ∂∆ i(z; F )dz = 0, which in addition to the continuity of i(·; F ) insure its analyticity on Sδ by applying Morera’s theorem. ACKNOWLEDGMENTS The authors would like to thank AUB Professor Farouk Abi-Khuzam for helpful discussions on upper-bounding the analytic extension of the alpha-stable density function. R EFERENCES [1] M. Win, P. Pinto, and L. Shepp, “A mathematical theory of network interference and its applications,” Proceedings of the IEEE, vol. 97, no. 2, pp. 205 –230, feb. 2009. [2] B. Hughes, “Alpha-stable models of multiuser interference,” in IEEE ISIT, Sorrento, Italy, 2000, p. 383. [3] M. Nassar, K. Gulati, A. Sujeeth, N. Aghasadeghi, B. Evans, and K. Tinsley, “Mitigating near-field interference in laptop embedded wireless transceivers,” in IEEE ICASSP, 2008, pp. 1405 –1408. [4] J. Wang, E. Kuruoglu, and T. Zhou, “Alpha-stable channel capacity,” Communications Letters, IEEE, vol. 15, no. 10, pp. 1107 –1109, october 2011. [5] H. Ben Mâad, A. Goupil, L. Clavier and G. Gelle, “Asymptotic performance of LDPC codes in impulsive non-Gaussian channel,” in IEEE SPAWC, 2010. [6] A. Das, “Capacity-Achieving Distributions for Non-Gaussian Additive Noise Channels,” in IEEE ISIT, Sorrento, Italy, 2000, p. 432. [7] A. Tchamkerten, “On the Discreteness of Capacity-Achieving Distributions,” IEEE Trans. Inf. Theory, vol. 50, no. 11, pp. 2773–2778, November 2004. [8] J. G. Smith, “The information capacity of peak and average power constrained scalar Gaussian channels,” Inform. Contr., vol. 18, pp. 203– 219, 1971. [9] J. P. Nolan, Stable Distributions - Models for Heavy Tailed Data. Boston: Birkhauser, 2012, in progress, Chapter 1 online at academic2.american.edu/∼jpnolan. [10] B. V. Gnedenko and A. N. Kolmogorov, Limit Distributions for Sums of Independent Random Variables. Reading Massachusetts: AddisonWesley Publishing Company, 1968. [11] J. E. Marsden and M. J. Hoffman, Basic Complex Analysis, 3rd ed. W. H. Freeman and Company, 1999. [12] H. Silverman, Complex Variables. Houghton Mifflin Company, 1975. [13] I. Abou-Faycal, M. D. Trott, and S. Shamai, “The capacity of discretetime memoryless Rayleigh-fading channels,” IEEE Trans. Inf. Theory, vol. 47, no. 4, pp. 1290–1301, May 2001. [14] J. Fahs and I. Abou-Faycal, “On the Detrimental Effect of Assuming a Linear Model for Non-Linear AWGN Channels,” in IEEE ISIT, Saint Petersburg, Russia, 2011, pp. 1693–1697. [15] G. Pólya and G. Szegö, Problems and Theorems in Analysis I. SpringerVerlag, 1978.