CAPACITY OF GAUSSIAN CHANNELS WITH NOISE UNCERTAINTY Stojan Z. Denic1 , Charalambos D. Charalambous 1,3, Seddik M. Djouadi2 1 School of Information Technology and Engineering, University of Ottawa, Ottawa, Canada 2 Electrical and Computer Engineering Department, University Tennessee, Knoxville, USA sdenic@site.uottawa.ca , chadcha@site.uottawa.ca , djouadi@ece.utk.edu Abstract In this paper the problem of defining, and computing the capacity of a communication channel when the statistic of an additive noise is not fully known, is addressed. The communication channel is specified as a continuous time channel with the known transfer function, where the transmitted signal is constrained in power, and an additive Gaussian noise channel is assumed. The power spectral density of the noise although unknown belongs to a known set defined through the uncertainty of the filter the shapes the power spectral density of the noise. The channel capacity is defined as the max-min of mutual information rate between the transmitted, and received signals, where the infimum is taken over the set of all possible power spectral densities of the noise, and supremum is taken over all power spectral densities of transmitted signal with constrained power. It is shown that the so defined channel capacity is equal to the operational capacity that represents the supremum of all attainable rates over a given channel. Keywords: Channel capacity; Uncertain noise. 1. INTRODUCTION In the classical information, and communication theory, it is assumed very often that the communication channel is fully known to a transmitter and receiver. That means that both transmitter, and receiver are perfectly aware of all channel parameters such as the parameters of the channel frequency or impulse response, and the statistic of the noise. Although this may be true for some communication channels when it is possible to measure a channel with high accuracy, there are many situations when the channel is not perfectly known to the transmitter, and receiver, which affects the performance of a communication system. Some examples of communication systems with channel uncertainty include wireless communication 3 systems, communication networks, communication systems in the presence of jamming. For instance, in wireless communication, the channel parameters such as attenuation, delay, phase, and Doppler spread constantly change with time that gives rise to uncertainty. In order to enable reliable, and efficient communication, the receiver has to estimate channel parameters. Also, the receiver, which operates in communication network, has to cope with the interference from other users that transmit signals on the same channel, and whose signals could have characteristics unknown to the receiver. In the case of adversary jamming, the parameters of the jamming signal are usually unknown to the transmitter, and receiver, making the communication channel uncertain. The above discussion just partially explains the importance of channel uncertainty in communications. An interested reader is referred to the papers [1], [2], [3] that give excellent overview of the topic, and represent good source of other important references. From above discussion, it can be concluded that there are two major sources of a channel uncertainty. One is the channel response uncertainty, and the other is the lack of knowledge of noise or interference characteristics affecting the transmitted signal. This paper is concerned with the information theoretic limits for the latter case, for a continuous time channel with additive Gaussian noise when the power spectral density of the noise is just partially known to the receiver. It is assumed that the transmitted signal is power limited, and frequency response of the channel is perfectly known. The channel capacity in the presence of a noise uncertainty will be defined, and explicit formula for the channel capacity will be derived. The problem of defining, and computing the capacity of the channel with a noise uncertainty is alleviated by using the appropriate uncertainty model. In this paper, a basic model borrowed from the robust control theory is used [4]. In particular, the additive uncertainty model of frequency response is employed to model the uncertainty of the power spectral density of the noise, giving the explicit formula for the Also with the Department of Electrical and Computer Engineering, University of Cyprus, Cyprus, and Adjunct Professor with the Department of Electrical and Computer Engineering, McGill University, Montreal, P.Q., Canada. This work was supported by the Natural and Science and Engineering Research Council of Canada under an operating grant. channel capacity. The obtained formula describes how the channel capacity decreases when the uncertainty of the power spectral density of the noise increases. The other important result stemming from the channel capacity is the water-filling formula that shows the effect of the noise uncertainty on the optimal transmission power. At the end it is shown that there exists a code that enables the reliable transmission over the channel with uncertain noise if the code rate is less then the channel capacity, and that the channel capacity as defined in the paper, is equal to the operational capacity. 2. NOISE UNCERTAINTY MODEL W(f ) The model of communication system is depicted in Fig. 1. The input signal x = {x (t );−∞ < t < +∞}, received {y(t );−∞ < t < +∞} , and noise n = {n (t );−∞ < t < +∞} are wide-sense stationary processes with power spectral densities S x ( f ), S y ( f ) , and S n ( f ) . All three power spectral densities are known. y = The noise n is an additive Gaussian random process. The frequency response of the channel H ( f ) is a fixed known transfer function. The uncertainty in the noise power spectral density is modeled through the additive uncertainty model of the filter W ( f ) that shapes the power spectral density of the noise S n ( f ) . The overall power spectral density of the noise is S n ( f )W ( f ) 2 . The additive uncertainty model W ( f ) = Wnom( f ) + W1 ( f )∆( f ), where is defined by Wnom ( f ) represents the nominal transfer function that can be chosen such that it reflects one’s limited knowledge or belief regarding the power spectral density of the noise. The second term represents a perturbation where W1 ( f ) is a fixed known transfer function, and ∆( f ) is unknown transfer function with ∆( f ) ∞ ≤ 1 . The norm . is called the infinity norm, and it is defined ∞ as H ( f ) ∞ := sup H ( f ) . The set of all transfer functions f defined by W1 ( f ) . Thus, the amplitude of uncertainty varies with frequency and it is determined by the fixed transfer function W1 ( f ) . The lager W1 ( f ) , the larger + Fig. 1 Communication system signal is the nominal transfer function Wnom ( f ) , and radius is 3. CHANNEL CAPACITY y H(f ) be proven that this space is a Banach space. All transfer functions mentioned until now belong to this normed linear space H ∞ . It should be noted that the uncertainty in the frequency response of the filter W ( f ) can be seen as a ball in a frequency domain , where the center of the ball W ( f ) − Wnom ( f ) ≤ W1 ( f ) the uncertainty. The transfer function W1 ( f ) can be determined from the measured data. Based on this uncertainty model the channel capacity will be defined, and computed in the following section. n x that have a finite . norm is denoted as H ∞ , and it can ∞ Define the following two sets ∞ A1 = S x ( f ); ∫ S x ( f )df ≤ P −∞ A2 ={W ∈H∞; W( f ) =Wnom( f ) + ∆( f )W1( f ), Wnom∈H∞, ∆ ∈H∞, W1 ∈H∞, ∆( f )W1( f ) ∞ ≤ γ , γ > 0} A2 controls the radius of uncertainty. The larger the radius of uncertainty (e.g., γ ) the larger Clearly, the set the uncertainty set A2 . Definition 1. The capacity of an additive Gaussian continuous time channel with noise uncertainty, is defined by S x ( f )H ( f ) 2 1 df (1) C n = sup inf ∫ log 1 + S ( f )W ( f ) 2 Sx ∈ A1 W ∈ A2 2 n The interval of integration will become clear from the discussion below. Although, in (1) the capacity is determined by the infimum over the set of noises A2 , which can be conservative, it provides the limit of reliable communication, when the noise is unknown and belongs to an uncertainty set. The better the noise knowledge, the smaller the uncertainty set, which then implies a less conservative value for the channel capacity. Clearly, the channel capacity definition is a variant of the Shannon capacity for additive Gaussian continuous time channels, subject to an input power and frequency constraints [5]. Theorem 1. Consider an additive uncertainty description H(f ) 2 ( ) ( ( ) S n f Wnom f − W1 ( f ) ) bounded, and integrable, and Wnom ( f ) ≠ W1 ( f ) . for W ( f ) , and assume 2 is i) The robust information capacity of an additive Gaussian continuous time channel with additive uncertainty shown in Fig. 1, and defined by (1), is given parametrically by 2 ν * H( f ) 1 df (2) C n = ∫ log 2 S ( f )( W ( f ) + W ( f ) ) 2 nom 1 n where the Lagrange multiplier ν* is found via 2 ν * − S n ( f )(Wnom( f ) + W1 ( f ) ) df = P ∫ 2 H( f ) (3) noise S n ( f )W ( f ) 2 − S n ( f )(Wnom ( f ) + W1 ( f ) )2 > 0, ν * > 0 of the noise is W ( f (4) solution of the equation (3). ii) The infimum over the channel uncertainty in (1) is achieved at ∆( f ) = exp (− j arg (W1 ( f )) + j arg(W ( f ))), ∆( f ) ∞ = 1 (5) and the resulting mutual information rate after the minimization is given by 2 S x ( f )H ( f ) df inf ∫ log1 + 2 ∆ ∞ ≤1 S ( f )W ( f ) + W ( f )∆( f ) n nom 1 2 (6) S ( f )+ Sn ( f )(W nom ( f ) + W1 ( f ) ) 2 H (f )2 W1 ( f ) = 0 , the standard formula for channel capacity is obtained [5], which corresponds to the case when the power spectral density of the noise is perfectly known. If the noise is not known the amplitude of uncertainty W1 ( f ) is different than zero, and the channel capacity decreases. If it is assumed that both the transmitter, and receiver have the partial knowledge of the channel then the modified water-filling equation is given by (7) describing how the uncertainty affects the optimal transmitted power. Formula (7) suggests how the transmitted power decreases with uncertainty of the overall noise power spectral density. In this section, it is shown that under certain conditions the coding theorem, and its converse hold for the set of communication channels with uncertain noise defined by A2 . It means that there exists a code, whose code rate R is less than the channel capacity Cn given by formula (2), for which the error probability is arbitrary small over the set of noises A2 . This result is obtained in [6], by combining two approaches found in [5], and [7]. First define the frequency response of the equivalent communication channel by F ( f ) = (S x ( f ) H ( f ) 2 / Sn ( f )W ( f ) 2 )1/ 2 and denote its inverse Fourier transform by f (t ) . Further define two sets A3 , and B as follows A3 = {F ( f );W ( f ) ∈ A2 } , B = { f (t ); F ( f ) ∈ A3 , f (t ) satisfies i), ii), iii)} where Moreover, the supremum of (6) over A1 yields the waterfilling equation * x ) 2 . If 4. CODING AND CONVERSE TO CODING THEOREMS in which the integrations in (2), and (3) are over the frequency interval over which 2 2 ν * > S n ( f )( Wnom ( f ) + W1 ( f ) ) / H ( f ) , and ν* is the Sx ( f ) H ( f ) df = ∫ log1 + 2 S ( f )( W ( f ) + W ( f ) ) n nom 1 affects the capacity. To understand this point better, assume that the noise n is a white Gaussian noise with S n ( f ) = 1 W / Hz over all frequencies such that the overall power spectral density subject to the condition ν * H (f )2 = ν * (7) Proof. Proof will be omitted due to the space constraint. The formula for the channel capacity (2) shows how the uncertainty in overall power spectral density of the i) ii) f (t ) has finite duration δ , f (t ) is square integrable ( f (t ) ∈ L2 ) −A iii) ∫ −∞ F( f +∞ ) 2 df + ∫ F ( f ) 2 df → 0, when A → +∞ A The set of all f (t ) that satisfy these conditions is conditionally compact set in L2 (see [7]), and this enables the proof of coding theorem, and its converse. Note that the condition i) can be relaxed (see Lemma 4 [8]). Now, the definition of the code for the set of channels B is given as well as the definition of the attainable rate R, and operational capacity C. The channel code (M , ε , T ) for the set of communication channels B is defined as the set of M distinct time-functions {x1 (t ),K, xM (t )}, in the interval − T / 2 ≤ t ≤ T / 2 , and the set of M disjoint sets D1 , …, DM, of the space of output signal y such that 1 T /2 x (t )dt ≤ P T − T∫/ 2 k model mitigates the computation of the channel capacity, and provides very intuitive result that describes how the channel capacity decreases when the size of the uncertainty set increases. Also, the modified water-filling equation is derived showing how the optimal transmitted power changes with the noise uncertainty. At the end, it is shown that the channel capacity as introduced in the paper is equal to the operational capacity, i.e., the channel coding theorem, and its converse hold. Refereces for each k, and such that the error probability for each codeword is Pr y(t ) ∈ Dkc | xk (t ) transmitted ≤ ε , [1] Medard, M., “Channel uncertainty in communications,” IEEE Information Theory Society Newsletters, vol. 53, no. 2, p. 1, pp. 10-12, June, 2003. {( [2] Biglieri, E., Proakis, J., Shamai, S., “Fading channels: information-theoretic and communications aspects,” IEEE Transactions on Information Theory, vol. 44, no. 6, pp. 2619-2692, October, 1998. ( ) k = 1,..., M , for all f (t ) ∈ B . For a positive number R is said to be an attainable rate for the set of channels B if there is a sequence of codes eTn R , ε n , Tn , such that when lim Tn → ∞ , lim ε n → 0 , )} n →∞ n→ ∞ uniformly over set B, where Tn is a codeword time duration. The operational channel capacity C is defined as a supremum of attainable rate R. Theorem 2. The operational capacity C for the set of communication channels with the noise uncertainty B is given by formula (2), and is equal to Cn . Proof. The proof is omitted, and is given in [6]. 5. CONCLUSION This paper concerns the problem of the channel capacity of continuous time additive Gaussian channels when the power spectral density of the Gaussian noise is not completely known. The capacity is defined as the max-min of a mutual information rate between the transmitted, and received signals, where the maximum is taken over all power spectral densities of the transmitted signal with the constrained power, and minimum is taken over all power spectral densities of the noise signal that belong to uncertainty set. The uncertainty set is defined by using the additive uncertainty model of the filter that shapes the power spectral density of the noise. This [3] Lapidoth, A., Narayan, P., “Reliable communication under channel uncertainty,” IEEE Transactions on Information Theory, vol. 44, no. 6, pp. 2148-2177, October, 1998. [4] Doyle, J.C., Francis, B.A., Tannenbaum, A.R., Feedback control theory, New York: McMillan Publishing Company, 1992. [5] Gallager, G.R., Information theory and reliable communication. New York: Wiley, 1968. [6] Denic, S.Z., Charalambous, C.D., Djouadi, S.M., “Robust capacity for additive colored Gaussian uncertain channels,” preprint. [7] Root, W.L., Varaiya, P.P., “Capacity of classes of Gaussian channels,” SIAM J. Appl. Math., vol. 16, no. 6, pp. 1350-1353, November, 1968. [8] Forys, L.J., Varaiya, P.P., “The ε-capacity of classes of unknown channels,” Information and control, vol. 44, pp. 376-406, 1969.