LECTURE 20 Last time: • Gaussian channels with feedback • Upper bound to benefit of capacity Lecture outline • Multiple access channels • Coding theorem • Capacity region for Gaussian channels Reading: Section 14.1-14.3. Multiple access channels Several users share the same medium What is the right metric? Joint informa­ tion? User 1 has rate R1 and user 2 has rate R2 How do we relate them to mutual informa­ tion? Model: Yi = X1i + X2i + Ni Liao and Ahlswede (independently, 1972) R1 ≤ I(X1; Y |X2) R2 ≤ I(X2; Y |X1) R1 + R2 ≤ I((X1, X2); Y ) Coding theorem mi: message sent by user i � i: decoded message for user i m P e1 (P e2): probability that the decoded codeword for user 1 (2) is different from that sent by user 1 (2) while the decoded codeword for user 2 (1) is the same as the one sent by user 2 (1) (such errors will be de­ noted as errors of type 1 (2)) P e1,2: probability that the decoded codewords for both users 1 and 2 are different from those sent by those users (such an error will be denoted as error of type 3) We begin by bounding the probability of er­ ror with an exponential argument and then we explore the behavior of that argument Coding theorem We first consider errors of type 1 - results for errors of type 2 can be derived analo­ gously We denote P e1,m1,m2 the probability that an error of type 1 occurs conditioned on messages m1 and m2 being sent Using the overbar to denote expectation P e1,m1,m2 � � � = y x1 x2 fX 1 (x1) fX 2 (x2) fY |X 1,X 2 (y|x1, x2) P �� � �1 = � 2 = m2|y, x1, x2 � m1) ∩ m (m ��� dx2dx1dy. Using the union bound, we obtain P �� � 1 �= m1) ∩ m � 2 = m2|y, x1, x2 (m ⎧ ⎨ � ⎩ � P m= � m1 ∀0 ≤ ρ ≤ 1. �� � ��� ≤ � 1 = m) ∩ m � 2 = m2|y, x1, x2 (m ⎫ρ ���⎬ ⎭ Coding theorem Using arguments similar to those for the single user strong coding theorem, we can establish ∀ρ ∈ [0, 1], fX 1 (x1) and fX 2 (x2) probabil­ ity density functions for X1 and X2, respec­ tively, we have P e1,m1,m2 ≤ � exp −N � −ρR1 + E01 � ρ, fX 1 (x1) , fX 2 (x2) ��� where we have defined, � � 1 E0 ρ, fX 1 (x1) , fX 2 (x2) �� � ln �� x y x2 =− 1 N fX 2 (x2) � fX 1 (x) fY |X,X 2 y |x, x2 � 1 1+ρ ⎫ ⎬ �1+ρ dx dx2dy ⎭ Coding theorem It now suffices to determine the behavior of the exponent to determine whether the upper bound to error probability becomes vanishingly small The following lemma parallels the one for the one-user case If I (X 1; Y |X 2) > 0, then for all 1 ≥ ρ ≥ 0 we have I (X 1; Y |X 2) ≥ E01 � ∂ 2N E01 ∂N E01 � � ρ, fX 1 (x1) , fX 2 (x2) ∂ρ (1) � ρ, fX 1 (x1) , fX 2 (x2) ≥ 0 � ρ, fX 1 (x1) , fX 2 (x2) ∂ρ2 >0 (2) � ≤0 (3) � �⎪ ⎪ ⎪ ∂E01 ρ, fX 1 (x1) , fX 2 (x2) ⎪ ⎪ I (X 1; Y |X 2) ⎪ ⎪ = . ⎪ ⎪ ⎪ ∂ρ N ⎪ ⎪ρ=0 (4) Coding theorem Let q1,N , q2,N be a pair of probability den­ sity functions for the codewords of length N of users 1 and 2 E01 � � 1,N 2,N ρ, q ,q In order for − ρR1 to be strictly positive for some ρ in [0, 1], it is necessary and sufficient that � � ⎧ ⎫⎪ ⎪ 1 1,N 2,N ⎨ ∂E ρ, q ⎬⎪ ⎪ , q ⎪ 0 ⎪ − R1 ⎪ > 0. ⎪ ⎪ ⎩ ⎭ ⎪ ∂ρ ⎪ ⎪ρ=0 We have that: For all fX 1 (x1) , fX 2 (x2) probability density functions for X1, X2, we have I(X 1 ;Y |X 2 ) N > R1 ≥ 0 ⇒ ∃ρ ∈ [0, 1] s.t. E1 0 � � ρ, fX 1 (x1) , fX 2 (x2) − R1ρ > 0 We can establish analogous results for er­ rors of type 2 and 3 Coding theorem Let us define Emin = � � � � � 1 1,N 2,N min maxρ E0 ρ, q ,q − R1 ρ , � � � � 2 maxρ E0 ρ, fX 1 (x1) , fX 2 (x2) − R2ρ , � � � �� 3 1,N 2,N maxρ E0 ρ, q ,q − (R1 + R2) ρ where E02 and E03 is defined analogously to E01. We may state the following theorem: For allfX 1 (x1) , fX 2 (x2) probability density functions for X1, X1, for any messages m1 and m2 of users 1 and 2, we have P em1,m2 ≤ 3e−N Emin and I �� � � X1 , X 2 ; Y N � I X1; Y |X2 > R1 ≥ 0 and N � � I X2; Y |X1 > R2 ≥ 0 and N � > R1 + R2 ≥ 0 ⇒ Emin > 0 Capacity region Cover-Wyner region for two users Capacity region Consider AWGN Multiple-access channel, 2 user i has energy σX 1 Pentagon: dominant face corresponds to � 1 ln 2 1+ 2 +σ 2 σX X2 1 2 σN � Interference cancellation at the corners: � � 1 ln 2 � 1+ � 1 ln 2 1+ 2 σX 1 � ,1 2 ln 1 + 2 +σ 2 σX N 2 2 σX � 1 2 σN � � ,1 2 ln 1 + 2 σX 2 2 σN 2 σX �� �� 2 2 2 σX +σN 1 without interference cancellation: � � 1 ln 2 1+ 2 σX � 1 2 +σ 2 σX N 2 � ,1 2 ln 1 + Recall DS-CDMA example 2 σX 2 2 2 σX +σN 1 �� Capacity region � FDMA: � W1 2 ln 1 + 2 σX 1 2 W1 σN � � , W22 ln 1 + 2 σX 2 2 W2 σN for equal energies, equal W s desirable TDMA: let α be the fraction of time that user 1 transmits � � α ln 2 1+ 2 σX 1 2 ασN � � α ln 1 + , 1− 2 2 σX �� 2 2 (1−α)σN for equal energies, α = 0.5 desirable How do we achieve points on the dominant face, that yields maximum sum rate? First way: time share between the corners Other way: rate splitting �� Capacity region Make one user (say user 1) into two virtual users (virtual user 1 and virtual user 3) and split energy between these two virtual users Virtual user 1 rate: � 1 ln 2 1+ 2 ασX 1 2 2 +σ 2 σN +(1−α)σX X2 1 User 2 rate: � 1 ln 2 1+ � 2 σX 2 2 2 σN +(1−α)σX 1 Virtual user 3 rate: � 1 ln 2 1+ 2 (1−α)σX 2 σN � 1 � Capacity region We have ⎛ ⎞ 2 ασX 1 1 ⎝ ⎠ ln 1 + 2 2 2 2 σN + (1 − α)σX + σX 1⎞ 2 ⎛ 2 σX 1 ⎝ 2 ⎠ + ln 1 + 2 2 2 σN + (1 − α)σX ⎛ 2 ⎞ (1 − α)σX 1 ⎝ 1⎠ + ln 1 + 2 2 σN ⎛ = = ⎞ 2 σX 1 ⎝ 1 ⎝ 2 ⎠ ⎠ + ln 1 + ln 1 + 2 2 2 2 2 σX + σN σN 2 ⎛ ⎞ 2 2 σX + σX 1 ⎝ 2 ⎠ ln 1 + 1 2 2 σN 2 σX 1 ⎞ 1 ⎛ Capacity region If we have � R2 = 1 2 ln 1 + 2 σX 2 2 2 σN +(1−α)σX 1 � , then R1 is defined as ⎛ = + ⎞ 2 + σX 1 ⎝ 2 ⎠ ln 1 + − R2 2 2 σN ⎛ ⎞ 2 ασX 1 ⎝ 1 ⎠ ln 1 + 2 2 2 2 σN + (1 − α)σX + σX 1 2 ⎛ ⎞ 2 (1 − α)σX 1 ⎝ 1⎠ ln 1 + 2 2 σN 2 σX 1 One variable provides all the necessary de­ grees of freedom Capacity region In general, for µ users, the capacity region is � i∈S Ri ≤ I((Xi )i∈S ; Y |(Xi)i�∈S ), ∀S ⊂ {1, . . . , µ} We have 2µ − 1 pseudo-users are sufficient to achieve any point on the multiple-access dominant face MIT OpenCourseWare http://ocw.mit.edu 6.441 Information Theory Spring 2010 For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.