Interactions of Information Theory and Estimation in Single- and Multi-user Communications

Similar documents
Decoupling of CDMA Multiuser Detection via the Replica Method

THIS paper is centered around two basic quantities in information

Information, Estimation, and Lookahead in the Gaussian channel

Single-letter Characterization of Signal Estimation from Linear Measurements

Coding over Interference Channels: An Information-Estimation View

much more on minimax (order bounds) cf. lecture by Iain Johnstone

Random Matrices and Wireless Communications

information estimation feedback

Shannon meets Wiener II: On MMSE estimation in successive decoding schemes

SOME fundamental relationships between input output mutual

On the Shamai-Laroia Approximation for the Information Rate of the ISI Channel

Principles of Communications

ELEC546 Review of Information Theory

Estimation in Gaussian Noise: Properties of the Minimum Mean-Square Error

Appendix B Information theory from first principles

Relations between Information and Estimation in the presence of Feedback

ANALYSIS OF A PARTIAL DECORRELATOR IN A MULTI-CELL DS/CDMA SYSTEM

Design of MMSE Multiuser Detectors using Random Matrix Techniques

Optimal Data Detection in Large MIMO

Interleave Division Multiple Access. Li Ping, Department of Electronic Engineering City University of Hong Kong

The Robustness of Dirty Paper Coding and The Binary Dirty Multiple Access Channel with Common Interference

These outputs can be written in a more convenient form: with y(i) = Hc m (i) n(i) y(i) = (y(i); ; y K (i)) T ; c m (i) = (c m (i); ; c m K(i)) T and n

UCSD ECE250 Handout #27 Prof. Young-Han Kim Friday, June 8, Practice Final Examination (Winter 2017)

Performance Analysis of Spread Spectrum CDMA systems

Minimum Energy Per Bit for Secret Key Acquisition Over Multipath Wireless Channels

Competition and Cooperation in Multiuser Communication Environments

Mismatched Estimation in Large Linear Systems

Lecture 9: Diversity-Multiplexing Tradeoff Theoretical Foundations of Wireless Communications 1. Overview. Ragnar Thobaben CommTh/EES/KTH

Lecture 8: MIMO Architectures (II) Theoretical Foundations of Wireless Communications 1. Overview. Ragnar Thobaben CommTh/EES/KTH

Mismatched Estimation in Large Linear Systems

Linear and Nonlinear Iterative Multiuser Detection

Samah A. M. Ghanem, Member, IEEE, Abstract

The Effect upon Channel Capacity in Wireless Communications of Perfect and Imperfect Knowledge of the Channel

One Lesson of Information Theory

x log x, which is strictly convex, and use Jensen s Inequality:

Direct-Sequence Spread-Spectrum

ee378a spring 2013 April 1st intro lecture statistical signal processing/ inference, estimation, and information processing Monday, April 1, 13

BASICS OF DETECTION AND ESTIMATION THEORY

Joint FEC Encoder and Linear Precoder Design for MIMO Systems with Antenna Correlation

Lecture 9: Diversity-Multiplexing Tradeoff Theoretical Foundations of Wireless Communications 1

Minimum Feedback Rates for Multi-Carrier Transmission With Correlated Frequency Selective Fading

Capacity Pre-log of Noncoherent SIMO Channels via Hironaka s Theorem

LECTURE 3. Last time:

Output MAI Distributions of Linear MMSE Multiuser Receivers in DS-CDMA Systems

Information Theory for Wireless Communications, Part II:

Estimation of the Capacity of Multipath Infrared Channels

Chapter 4: Continuous channel and its capacity

Revision of Lecture 4

RADIO SYSTEMS ETIN15. Lecture no: Equalization. Ove Edfors, Department of Electrical and Information Technology

MMSE Dimension. snr. 1 We use the following asymptotic notation: f(x) = O (g(x)) if and only

Memory in Classical Information Theory: A Brief History

UCSD ECE153 Handout #40 Prof. Young-Han Kim Thursday, May 29, Homework Set #8 Due: Thursday, June 5, 2011

Sparse Superposition Codes for the Gaussian Channel

On the Optimum Asymptotic Multiuser Efficiency of Randomly Spread CDMA

What is the Value of Joint Processing of Pilots and Data in Block-Fading Channels?

Lecture 18: Gaussian Channel

Multiple-Input Multiple-Output Systems

Revision of Lecture 5

Information Theory. Lecture 5 Entropy rate and Markov sources STEFAN HÖST

Shannon Meets Carnot: Mutual Information Via Thermodynamics

Joint Channel Estimation and Co-Channel Interference Mitigation in Wireless Networks Using Belief Propagation

On the Low-SNR Capacity of Phase-Shift Keying with Hard-Decision Detection

Chapter 9 Fundamental Limits in Information Theory

Communication Theory II

Lecture 22: Final Review

(Classical) Information Theory III: Noisy channel coding

Signal Estimation in Gaussian Noise: A Statistical Physics Perspective

ENGR352 Problem Set 02

Lecture 1. Introduction

Low-High SNR Transition in Multiuser MIMO

Capacity of Block Rayleigh Fading Channels Without CSI

Power Control in Multi-Carrier CDMA Systems

Lattice Reduction Aided Precoding for Multiuser MIMO using Seysen s Algorithm

On the Applications of the Minimum Mean p th Error to Information Theoretic Quantities

Achieving the Full MIMO Diversity-Multiplexing Frontier with Rotation-Based Space-Time Codes

Capacity of AWGN channels

Shannon and Poisson. sergio verdú

Multiuser Detection of Sparsely Spread CDMA

ECE 564/645 - Digital Communications, Spring 2018 Homework #2 Due: March 19 (In Lecture)

This examination consists of 11 pages. Please check that you have a complete copy. Time: 2.5 hrs INSTRUCTIONS

Multiuser Receivers, Random Matrices and Free Probability

IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 49, NO. 10, OCTOBER

Performance Regions in Compressed Sensing from Noisy Measurements

Dirty Paper Coding vs. TDMA for MIMO Broadcast Channels

The Noncoherent Rician Fading Channel Part II : Spectral Efficiency in the Low-Power Regime

5958 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 56, NO. 12, DECEMBER 2010


National University of Singapore. Ralf R. Müller

Lecture 4 Noisy Channel Coding

Outline of the Lecture. Background and Motivation. Basics of Information Theory: 1. Introduction. Markku Juntti. Course Overview

Approximately achieving the feedback interference channel capacity with point-to-point codes

Lecture 15: Thu Feb 28, 2019

This examination consists of 10 pages. Please check that you have a complete copy. Time: 2.5 hrs INSTRUCTIONS

Capacity Penalty due to Ideal Zero-Forcing Decision-Feedback Equalization

Generalized Writing on Dirty Paper

Functional Properties of MMSE

A General Formula for Compound Channel Capacity

Lecture 7: Wireless Channels and Diversity Advanced Digital Communications (EQ2410) 1

A Simple Memoryless Proof of the Capacity of the Exponential Server Timing Channel

MULTICARRIER code-division multiple access (MC-

Transcription:

Interactions of Information Theory and Estimation in Single- and Multi-user Communications Dongning Guo Department of Electrical Engineering Princeton University March 8, 2004 p 1 Dongning Guo

Communications S Source X Encoder Noisy Y channel S Decoder Dest p 2 Dongning Guo

Communications S Source X Encoder Noisy Y channel S Decoder Dest Tension: Noise causes errors (X Y ) Reliable transmission at a good rate (S S ) p 2 Dongning Guo

Communications S Source X Encoder Noisy Y channel S Decoder Dest Tension: Noise causes errors (X Y ) Reliable transmission at a good rate (S S ) Shannon (1948): Given a noisy channel, arbitrarily reliable transmission is possible up to a certain rate p 2 Dongning Guo

Communications S Source X Encoder Noisy Y channel S Decoder Dest Tension: Noise causes errors (X Y ) Reliable transmission at a good rate (S S ) Shannon (1948): Given a noisy channel, arbitrarily reliable transmission is possible up to a certain rate Estimation vs Information Analog vs digital p 2 Dongning Guo

Communications S Source X Encoder Noisy Y channel S Decoder Dest Tension: Noise causes errors (X Y ) Reliable transmission at a good rate (S S ) Shannon (1948): Given a noisy channel, arbitrarily reliable transmission is possible up to a certain rate Estimation vs Information Analog vs digital This talk: Interactions, (new) surprises, and applications p 2 Dongning Guo

Example: Mars Rover Received power as low as 10 18 watt Rate as high as 10 Kbits/sec p 3 Dongning Guo

Wireless Network (X 1, Y 1 ) (X k, Y k ) p 4 Dongning Guo

Wireless Network (X 1, Y 1 ) (X k, Y k ) Tension: Throughput Interferences (complicated) Resources are scarce and shared p 4 Dongning Guo

Wireless Network (X 1, Y 1 ) (X k, Y k ) Tension: Throughput Interferences (complicated) Resources are scarce and shared Interferences are information Think smart! p 4 Dongning Guo

Wireless Network (X 1, Y 1 ) (X k, Y k ) Tension: Throughput Interferences (complicated) Resources are scarce and shared Interferences are information Think smart! Wanted: Simple laws of the seemingly disordered system p 4 Dongning Guo

Example: Sensor Network p 5 Dongning Guo

Mutual Information & MMSE Probabilistic point of view: X P X X P Y X Y X and Y are random variables/vectors or processes p 6 Dongning Guo

Mutual Information & MMSE Probabilistic point of view: X P X X P Y X Y X and Y are random variables/vectors or processes Minimum mean-squared error (MMSE): mmse(x; Y ) = min f E X f(y ) 2, Achieved by X(Y ) = E {X Y } p 6 Dongning Guo

Mutual Information & MMSE Probabilistic point of view: X P X X P Y X Y X and Y are random variables/vectors or processes Minimum mean-squared error (MMSE): mmse(x; Y ) = min f E X f(y ) 2, Achieved by X(Y ) = E {X Y } Mutual information: I(X; Y ) = E { log p } XY (X, Y ) p X (X)p Y (Y ) p 6 Dongning Guo

Multi-user Perspective Examples: cellular telephony, sensor networks, DSL X 1 X K Channel P Y X Y Multiuser detector X 1 X K p 7 Dongning Guo

Multi-user Perspective Examples: cellular telephony, sensor networks, DSL X 1 X K Channel P Y X Y Multiuser detector X 1 X K Mean-square error: E X k X k 2 p 7 Dongning Guo

Multi-user Perspective Examples: cellular telephony, sensor networks, DSL X 1 X K Channel P Y X Y Multiuser detector X 1 X K Mean-square error: E X k X k 2 Mutual information: I (X k ; X k ) p 7 Dongning Guo

Outline Part I: Canonical Gaussian channel: Y = α X + W A fundamental I-mmse relationship By-products and applications p 8 Dongning Guo

Outline Part I: Canonical Gaussian channel: Y = α X + W A fundamental I-mmse relationship By-products and applications Part II: Multiuser (vector) channel: Y = S N K X + W Large-system analysis via statistical physics Result: Multiuser channel can be decoupled p 8 Dongning Guo

Part I: Y = α X + W p 9 Dongning Guo

An Observation Scalar Gaussian channel: Y = snr X + W, W N (0, 1) p 10 Dongning Guo

An Observation Scalar Gaussian channel: Y = snr X + W, W N (0, 1) Gaussian input: X N (0, 1) I(X; Y ) = 1 2 mmse(snr) = E log(1 + snr), [ snr X 1 + snr Y ] 2 = 1 1 + snr p 10 Dongning Guo

An Observation Scalar Gaussian channel: Y = snr X + W, W N (0, 1) Gaussian input: X N (0, 1) I(X; Y ) = 1 2 mmse(snr) = E log(1 + snr), [ snr X 1 + snr Y ] 2 = 1 1 + snr Immediately, d dsnr I(X; Y ) = 1 mmse(snr) log e 2 ( ) p 10 Dongning Guo

Another Observation Scalar Gaussian channel: Y = snr X + W, W N (0, 1) p 11 Dongning Guo

Another Observation Scalar Gaussian channel: Y = snr X + W, W N (0, 1) Binary input: X = ±1 equally likely I(snr) = snr e y2 2 2π log cosh(snr snr y) dy, mmse(snr) = 1 e y2 2 2π tanh(snr snr y) dy p 11 Dongning Guo

Another Observation Scalar Gaussian channel: Y = snr X + W, W N (0, 1) Binary input: X = ±1 equally likely I(snr) = snr e y2 2 2π log cosh(snr snr y) dy, mmse(snr) = 1 e y2 2 2π tanh(snr snr y) dy Again, d dsnr I(X; Y ) = 1 2 mmse(snr) ( ) p 11 Dongning Guo

d dsnr I(snr) = 1 2 mmse(snr) ( ) 12 1 08 06 04 02 mmse snr I snr 2 4 6 8 10 snr Gaussian input p 12 Dongning Guo

d dsnr I(snr) = 1 2 mmse(snr) ( ) 12 1 08 06 04 02 mmse snr I snr 2 4 6 8 10 snr Gaussian input, binary input p 12 Dongning Guo

I-mmse Theorem Theorem 1 (Guo, Shamai & Verdú) Y = snr X + W P X with EX 2 <, d dsnr I(snr) = 1 2 mmse(snr) ( ) p 13 Dongning Guo

I-mmse Theorem Theorem 1 (Guo, Shamai & Verdú) Y = snr X + W P X with EX 2 <, ( ) also proved for: Vector channel: Continuous-time: Discrete-time: d dsnr I(snr) = 1 2 mmse(snr) Y = snr HX + W Y t = snr X t + W t, t [0, T ] Y n = snr X n + W n, n = 1, 2, ( ) p 13 Dongning Guo

Information vs Estimation Information theory I(X; Y ): (coded) reliable rate detection theory likelihood ratio estimation theory mmse(x; Y ): (uncoded) accuracy p 14 Dongning Guo

Information vs Estimation Information theory I(X; Y ): (coded) reliable rate detection theory likelihood ratio estimation theory mmse(x; Y ): (uncoded) accuracy History: Wiener, Shannon, Kolmogorov Price, Kailath, 1950-60s Duncan, 1970: Mutual information causal MMSE in continuous-time filtering p 14 Dongning Guo

Proof of d dsnr I(snr) = 1 2 mmse(snr) ( ) Equivalent to: I(snr + δ) I(snr) = δ 2 mmse(snr) + o(δ) p 15 Dongning Guo

Proof of d dsnr I(snr) = 1 2 mmse(snr) ( ) Equivalent to: I(snr + δ) I(snr) = δ 2 mmse(snr) + o(δ) Incremental channel: X σ 1 W 1 σ 2 W 2 Y 1 snr + δ Y 2 snr p 15 Dongning Guo

Proof of d dsnr I(snr) = 1 2 mmse(snr) ( ) Equivalent to: I(snr + δ) I(snr) = δ 2 mmse(snr) + o(δ) Incremental channel: Markov property: X σ 1 W 1 σ 2 W 2 Y 1 snr + δ Y 2 snr I(X; Y 1 ) I(X; Y 2 ) = I(X; Y 1, Y 2 ) I(X; Y 2 ) = I(X; Y 1 Y 2 ) p 15 Dongning Guo

Proof of d dsnr I(snr) = 1 2 mmse(snr) ( ) Equivalent to: I(snr + δ) I(snr) = δ 2 mmse(snr) + o(δ) Incremental channel: Markov property: X σ 1 W 1 σ 2 W 2 Y 1 snr + δ Y 2 snr I(X; Y 1 ) I(X; Y 2 ) = I(X; Y 1, Y 2 ) I(X; Y 2 ) = I(X; Y 1 Y 2 ) Given Y 2, the channel X Y 1 is Gaussian: (snr + δ) Y 1 = snr Y 2 + δ X + N (0, δ) p 15 Dongning Guo

Proof of d dsnr I(snr) = 1 mmse(snr) (Cont) 2 Lemma 1 (Verdú 90, Lapidoth-Shamai 02, Guo et al 04) Y = δ Z + U, U N (0, 1) As δ 0, I(Y ; Z) = δ 2 E (Z EZ)2 + o(δ) p 16 Dongning Guo

Proof of d dsnr I(snr) = 1 mmse(snr) (Cont) 2 Lemma 1 (Verdú 90, Lapidoth-Shamai 02, Guo et al 04) Y = δ Z + U, U N (0, 1) As δ 0, I(Y ; Z) = δ 2 E (Z EZ)2 + o(δ) Apply Lemma 1 to X Y 1 conditioned on Y 2 : I(X; Y 1 Y 2 ) = δ 2 E (X E {X Y 2}) 2 + o(δ) p 16 Dongning Guo

Proof of d dsnr I(snr) = 1 mmse(snr) (Cont) 2 Lemma 1 (Verdú 90, Lapidoth-Shamai 02, Guo et al 04) Y = δ Z + U, U N (0, 1) As δ 0, I(Y ; Z) = δ 2 E (Z EZ)2 + o(δ) Apply Lemma 1 to X Y 1 conditioned on Y 2 : I(X; Y 1 Y 2 ) = δ 2 E (X E {X Y 2}) 2 + o(δ) Increase due to SNR increment: I(snr + δ) I(snr) = I(X; Y 1 Y 2 ) = δ mmse(snr) + o(δ) 2 p 16 Dongning Guo

Proof of d dsnr I(snr) = 1 mmse(snr) (Cont) 2 Lemma 1 (Verdú 90, Lapidoth-Shamai 02, Guo et al 04) Y = δ Z + U, U N (0, 1) As δ 0, I(Y ; Z) = δ 2 E (Z EZ)2 + o(δ) Apply Lemma 1 to X Y 1 conditioned on Y 2 : I(X; Y 1 Y 2 ) = δ 2 E (X E {X Y 2}) 2 + o(δ) Increase due to SNR increment: I(snr + δ) I(snr) = I(X; Y 1 Y 2 ) = δ mmse(snr) + o(δ) 2 Key property: infinite divisibility of Gaussian p 16 Dongning Guo

Why d dsnr I(snr) = 1 2 mmse(snr)? snr 1 snr 2 snr 3 0 X Y 1 Y 2 Y 3 p 17 Dongning Guo

Why d dsnr I(snr) = 1 2 mmse(snr)? snr 1 snr 2 snr 3 0 X Y 1 Y 2 Y 3 Using mutual information chain rule: I(snr 1 ) = I(X; Y 1 ) = I(X; Y 1 Y 2 ) + I(X; Y 2 ) = n=1 I(X; Y n Y n+1 ) 1 (snr n snr n+1 ) mmse(snr n ) 2 1 2 n=1 snr1 0 mmse(γ) dγ p 17 Dongning Guo

Continous-time Channel Channel model: R t = dy t = snr X t + N t, snr X t dt + db t or, equivalently, {B t } Brownian motion (Wiener process) p 18 Dongning Guo

Continous-time Channel Channel model: R t = snr X t + N t, or, equivalently, dy t = snr X t dt + db t {B t } Brownian motion (Wiener process) Consider t [0, T ] Information rate: I(snr) = 1 T I(XT 0 ; Y T 0 ) Causal and non-causal MMSEs: cmmse(t, snr) = E ( X t E { }) 2, X t Y0 t mmse(t, T, snr) = E ( X t E { }) 2 X t Y T 0 p 18 Dongning Guo

Triangle Relationship Theorem 2 If T 0 EX2 t dt < (finite-power input), d dsnr I(snr) = 1 2 T 0 mmse(t, T, snr) dt T Proof: Radon-Nikodym derivatives Stochastic calculus p 19 Dongning Guo

Triangle Relationship Theorem 2 If T 0 EX2 t dt < (finite-power input), d dsnr I(snr) = 1 2 T 0 mmse(t, T, snr) dt T Proof: Radon-Nikodym derivatives Stochastic calculus Theorem 3 (Duncan 1970) For finite-power input, I(snr) = snr 2 T 0 cmmse(t, snr) dt T p 19 Dongning Guo

Triangle Relationship Theorem 2 If T 0 EX2 t dt < (finite-power input), d dsnr I(snr) = 1 2 T 0 mmse(t, T, snr) dt T Proof: Radon-Nikodym derivatives Stochastic calculus Theorem 3 (Duncan 1970) For finite-power input, I(snr) = snr 2 T 0 cmmse(t, snr) dt T Theorem 4 a T 0 cmmse(t, snr) dt T = T 0 snr 0 mmse(t, T, γ) dγ snr dt T p 19 Dongning Guo

Stationary Inputs Theorem 5 For stationary finite-power input, d dsnr I(snr) = 1 2 mmse(snr) ( ) p 20 Dongning Guo

Stationary Inputs Theorem 5 For stationary finite-power input, d dsnr I(snr) = 1 2 mmse(snr) ( ) Theorem 6 For stationary finite-power input, cmmse(snr) = 1 snr snr 0 mmse(γ) dγ ( ) p 20 Dongning Guo

Stationary Inputs Theorem 5 For stationary finite-power input, d dsnr I(snr) = 1 2 mmse(snr) ( ) Theorem 6 For stationary finite-power input, cmmse(snr) = 1 snr snr 0 mmse(γ) dγ ( ) Can be checked: Gaussian input with spectrum S X (ω) I(snr) [Shannon 49], cmmse(snr) [Yovits-Jackson 55] Random telegraph waveform (2-state Markov) input MMSEs due to Wonham 65 and Yao 85 p 20 Dongning Guo

cmmse(snr) = 1 snr snr 0 mmse(γ) dγ ( ) 1 An example: Random telegraph waveform input 08 06 04 02 cmmse snr mmse snr 5 10 15 20 25 30 snr p 21 Dongning Guo

Extensions Vector channel Y = snr H X + W : d dsnr I(X; Y ) = 1 2 E H X H E {X Y } 2 p 22 Dongning Guo

Extensions Vector channel Y = snr H X + W : d dsnr I(X; Y ) = 1 2 E H X H E {X Y } 2 Discrete-time model: Y n = snr X n + W n, n = 1, 2, Via piecewise constant continuous-time input p 22 Dongning Guo

Extensions Vector channel Y = snr H X + W : d dsnr I(X; Y ) = 1 2 E H X H E {X Y } 2 Discrete-time model: Y n = snr X n + W n, n = 1, 2, Via piecewise constant continuous-time input More general models: dy t = snr h t (X t ) dt + db t, t R MMSEs errors in estimating channel input h t (X t ) p 22 Dongning Guo

Other Channels & Applications Incremental channel device works if noise has independent increments (Lévy processes) Eg, Gaussian channel MMSE: X 2 t X 2 t ; Poisson channel φ(λ t ) φ( λ t ), where φ(x) = x log x p 23 Dongning Guo

Other Channels & Applications Incremental channel device works if noise has independent increments (Lévy processes) Eg, Gaussian channel MMSE: X 2 t X 2 t ; Poisson channel φ(λ t ) φ( λ t ), where φ(x) = x log x Bounds on MMSE bounds on mutual information Linear estimation upper bound Intersymbol interference channel p 23 Dongning Guo

Other Channels & Applications Incremental channel device works if noise has independent increments (Lévy processes) Eg, Gaussian channel MMSE: X 2 t X 2 t ; Poisson channel φ(λ t ) φ( λ t ), where φ(x) = x log x Bounds on MMSE bounds on mutual information Linear estimation upper bound Intersymbol interference channel MMSE estimator is often a key component in capacity achieving receivers p 23 Dongning Guo

Other Channels & Applications Incremental channel device works if noise has independent increments (Lévy processes) Eg, Gaussian channel MMSE: X 2 t X 2 t ; Poisson channel φ(λ t ) φ( λ t ), where φ(x) = x log x Bounds on MMSE bounds on mutual information Linear estimation upper bound Intersymbol interference channel MMSE estimator is often a key component in capacity achieving receivers More to be discovered p 23 Dongning Guo

Part II: Y = SX + W p 24 Dongning Guo

Multiuser Channel Channel model: K Y = s k snrk X k + W where W N (0, I) k=1 = S N K X + W, p 25 Dongning Guo

Multiuser Channel Channel model: Y = where W N (0, I) K s k snrk X k + W k=1 Assumptions: Iid input: X k P X Random signatures: s k Fading: snr k P snr = S N K X + W, p 25 Dongning Guo

Examples Code-division multiple access (CDMA): X 1 X 2 snr1 s 1 snr2 s 2 W N (0, I) K Y = s k snrk X k + W k=1 X K snrk s K Y = S N K X + W Raondom signatures in Qualcomm CDMA, 3G wireless Also: multiple-antenna, multi-carrier systems p 26 Dongning Guo

Joint Decoding Joint decoding Encoder Encoder Encoder X 1 snr1 s 1 X 2 snr2 s 2 X K snrk s K N (0, I) Y Joint decoding p 27 Dongning Guo

Joint Decoding Joint decoding Encoder Encoder Encoder X 1 snr1 s 1 X 2 snr2 s 2 X K snrk s K N (0, I) Y Joint decoding Spectral efficiency: C joint = 1 N I(X; Y ) p 27 Dongning Guo

Separate Decoding Multiuser detection + single-user decoding: Encoder Encoder Encoder X 1 snr1 s 1 X 2 snr2 s 2 X K snrk s K N (0, I) Y Multiuser detector Decoder X 1 Decoder X 2 Decoder X K p 28 Dongning Guo

Separate Decoding Multiuser detection + single-user decoding: Encoder Encoder Encoder X 1 snr1 s 1 Xk snrk s k X K snrk s K N (0, I) Multiuser Y detector Decoder X 1 X k Decoder Decoder X K Channel for user k: Capacity of user k: P Xk X k I(X k ; X k ) p 28 Dongning Guo

Multiuser Detection Detection function: X k = f k (Y, S) p 29 Dongning Guo

Multiuser Detection Detection function: X k = f k (Y, S) Optimal: X k = E {X k Y, S} Wrt p Xk Y,S (induced from p X and p Y X,S ) p 29 Dongning Guo

Multiuser Detection Detection function: X k = f k (Y, S) Optimal: X k = E {X k Y, S} Wrt p Xk Y,S (induced from p X and p Y X,S ) Suboptimal: X k q = E q {X k Y, S} Wrt q Xk Y,S (induced from q X and q Y X,S ) p 29 Dongning Guo

Multiuser Detection Detection function: X k = f k (Y, S) Optimal: X k = E {X k Y, S} Wrt p Xk Y,S (induced from p X and p Y X,S ) Suboptimal: X k q = E q {X k Y, S} Wrt q Xk Y,S (induced from q X and q Y X,S ) Postulated input q X Postulated channel: Y = S X + σ W p 29 Dongning Guo

Special Cases If q X N (0, 1), then X q = [ S S + σ 2 I ] 1 S Y σ : matched filter; σ = 1: linear MMSE; σ 0: decorrelator p 30 Dongning Guo

Special Cases If q X N (0, 1), then X q = [ S S + σ 2 I ] 1 S Y σ : matched filter; σ = 1: linear MMSE; σ 0: decorrelator If q X = p X, then σ 0: jointly optimal; σ = 1: individually optimal p 30 Dongning Guo

Special Cases If q X N (0, 1), then X q = [ S S + σ 2 I ] 1 S Y σ : matched filter; σ = 1: linear MMSE; σ 0: decorrelator If q X = p X, then σ 0: jointly optimal; σ = 1: individually optimal In principle, multiuser detector X q = E q {X Y, S}, parameterized by q X and σ p 30 Dongning Guo

Problem Multiuser detection + single-user codes Encoder Encoder Encoder X 1 snr1 s 1 Xk snrk s k X K snrk s K N (0, I) Multiuser Y detector Decoder X 1 q Decoder X k q Decoder X K q What is P Xk q X k? I(X k ; X k q )? I(X; Y )? p 31 Dongning Guo

Problem Multiuser detection + single-user codes Encoder Encoder Encoder X 1 snr1 s 1 Xk snrk s k X K snrk s K N (0, I) Multiuser Y detector Decoder X 1 q Decoder X k q Decoder X K q What is P Xk q X k? I(X k ; X k q )? I(X; Y )? Large system: user number (K), dimensionality (N), lim K K N β p 31 Dongning Guo

New Result: Decoupling Multiuser channel: Encoder Encoder Encoder X 1 snr1 s 1 Xk snrk s k X K snrk s K N (0, I) Multiuser Y detector Decoder X 1 q Decoder X k q Decoder X K q p 32 Dongning Guo

New Result: Decoupling Multiuser channel: Encoder Encoder Encoder X 1 snr1 s 1 Xk snrk s k X K snrk s K N (0, I) Multiuser Y detector Decoder X 1 q Decoder X k q Decoder X K q Equivalent single-user scalar channel: N ( 0, η 1) Encoder X k snrk Y Decision function Decoder X k q p 32 Dongning Guo

Decision Function Implicit decision function: N ( 0, η 1) X p X snr + Z Decision function X q p 33 Dongning Guo

Decision Function Implicit decision function: N ( 0, η 1) X p X snr + Z Decision function X q Best estimate of X given Z assuming a postulated channel: N ( 0, ξ 1) X q X snr + Z Ie, X q = E q {X Z, snr; ξ} p 33 Dongning Guo

Retrochannel The equivalent channel & retrochannel: N ( 0, η 1) X p X snr + Z Decision function E q {X Z, snr; ξ} X q Retrochannel q X Z,snr;ξ X p 34 Dongning Guo

Retrochannel The equivalent channel & retrochannel: N ( 0, η 1) X p X snr + Z Decision function E q {X Z, snr; ξ} X q Retrochannel q X Z,snr;ξ [ ] 2; The MSE: E(snr; η, ξ) = E X X q [ ] 2 the variance: V(snr; η, ξ) = E X X q X p 34 Dongning Guo

Main Result: Decoupling Theorem 7 Equivalent channel for a user with snr k = snr: N ( 0, η 1) X p X snr + Z Decision function X q E q {X Z, snr; ξ} Retrochannel X q X Z,snr;ξ ie, Gaussian followed by a decision function, where η 1 =1 + β E {snr E(snr; η, ξ)}, ξ 1 =σ 2 + β E {snr V(snr; η, ξ)} ( ) E(snr; η, ξ) = E[X X q ] 2 ; V(snr; η, ξ) = E[X X q ] 2 p 35 Dongning Guo

Spectral Efficiencies Corollary 1 The capacity of user k is I(η snr k ) N ( 0, η 1) X p X snrk + Z p 36 Dongning Guo

Spectral Efficiencies Corollary 1 The capacity of user k is I(η snr k ) N ( 0, η 1) X p X snrk + Z Corollary 2 The overall spectral efficiency: C sep = β E {I(η snr)} p 36 Dongning Guo

Spectral Efficiencies Corollary 1 The capacity of user k is I(η snr k ) N ( 0, η 1) X p X snrk + Z Corollary 2 The overall spectral efficiency: C sep = β E {I(η snr)} Theorem 8 The spectral efficiency under joint decoding is C joint = C sep + (η 1 log η)/2 Generalization of Shamai-Verdú, Tanaka, Müller-Gerstaker p 36 Dongning Guo

Special Case Postulate q X N (0, 1), σ 2 = 1 Then linear MMSE p 37 Dongning Guo

Special Case Postulate q X N (0, 1), σ 2 = 1 Then linear MMSE MMSE: mmse(snr; η) = 1 1 + η snr p 37 Dongning Guo

Special Case Postulate q X N (0, 1), σ 2 = 1 Then linear MMSE MMSE: mmse(snr; η) = 1 1 + η snr Tse-Hanly 99 is a special case of ( ): { } η 1 1 = 1 + β E snr 1 + η snr p 37 Dongning Guo

Special Case Postulate q X N (0, 1), σ 2 = 1 Then linear MMSE MMSE: mmse(snr; η) = 1 1 + η snr Tse-Hanly 99 is a special case of ( ): { } η 1 1 = 1 + β E snr 1 + η snr Single-user capacity C(snr; η) = 1 2 log (1 + ηsnr) p 37 Dongning Guo

Joint vs Separate Theorem 9 P X, C joint (β) = β 0 1 β C sep(β ) dβ p 38 Dongning Guo

Joint vs Separate Theorem 9 P X, Proof: Use C joint (β) = β d dβ C joint = 1 β C sep + dη dβ 0 1 β C sep(β ) dβ [ { } ] d E dη I(ηsnr) + 1 η 1 d dsnr I(snr) = 1 mmse(snr) ( ), and fixed-point eqn ( ) 2 p 38 Dongning Guo

Joint vs Separate Theorem 9 P X, β 1 C joint (β) = 0 β C sep(β ) dβ d Proof: dβ C joint = 1 β C sep + dη [ { } ] d E dβ dη I(ηsnr) + 1 η 1 d Use dsnr I(snr) = 1 mmse(snr) ( ), and fixed-point eqn ( ) 2 Mutual information chain rule Successive cancellation 1 N I(X; Y S) = 1 K I(X k ; Y S, X k+1,, X K ) N = 1 N k=1 K I k=1 ( η ( k N ) ) snr k p 38 Dongning Guo

Statistical Physics From microscopic interactions to macroscopic properties p 39 Dongning Guo

Statistical Physics From microscopic interactions to macroscopic properties Spin glass: p 39 Dongning Guo

Statistical Physics From microscopic interactions to macroscopic properties Spin glass: Posterior distribution of transmitted symbols = configuration distribution under external field p X Y,S (x y, S) exp [ H y,s (x)] p 39 Dongning Guo

Statistical Physics From microscopic interactions to macroscopic properties Spin glass: Posterior distribution of transmitted symbols = configuration distribution under external field p X Y,S (x y, S) exp [ H y,s (x)] Multiuser system spin glass Spins bits External field received signal & channel state p 39 Dongning Guo

Replica Trick Mutual information 1 K I(X; Y S) = 1 K E { log p Y S (Y S) S } const = F(S) const p 40 Dongning Guo

Replica Trick Mutual information 1 K I(X; Y S) = 1 K E { log p Y S (Y S) S } const = F(S) const Asymptotic equipartition property Free energy F(S) F = lim K (1/K) E { log p Y S (Y S) } p 40 Dongning Guo

Replica Trick Mutual information 1 K I(X; Y S) = 1 K E { log p Y S (Y S) S } const = F(S) const Asymptotic equipartition property Free energy F(S) F = lim K (1/K) E { log p Y S (Y S) } Replica trick: lim u 0 u log E E {Θ u log Θ} {Θu } = lim u 0 E {Θ u } = E {log Θ} p 40 Dongning Guo

Replica Trick Mutual information 1 K I(X; Y S) = 1 K E { log p Y S (Y S) S } const = F(S) const Asymptotic equipartition property Free energy F(S) F = lim K (1/K) E { log p Y S (Y S) } Replica trick: lim u 0 u log E E {Θ u log Θ} {Θu } = lim u 0 E {Θ u } = E {log Θ} Introducing replicas X ak, a = 1,, u, p u Y S (y S) E { u a=1 p Y X,S(y X a, S) S } p 40 Dongning Guo

Conclusion I-mmse relationship: Under arbitrary input distribution, d dsnr I(snr) = 1 2 mmse(snr) ( ) Incremental channel proof Relationship between causal & non-causal MMSEs p 41 Dongning Guo

Conclusion I-mmse relationship: Under arbitrary input distribution, d dsnr I(snr) = 1 2 mmse(snr) ( ) Incremental channel proof Relationship between causal & non-causal MMSEs Multiuser channel: Multiuser detector optimal for a postulated system Decoupling: Equivalent Gaussian single-user channel Multiuser efficiency, spectral efficiencies p 41 Dongning Guo

Future Work Application and practical implications of d dsnr I(snr) = 1 2 mmse(snr) ( ) Eg, ISI channels, coding Further extensions MMSE structures in capacity-achieving receivers Eg, ISI channel, CDMA channel, lattice codes Applications of the decoupling result Eg, power control, signaling optimization Applications of statistical physics methodologies Eg, large sensor networks p 42 Dongning Guo