Dynamic Power Allocation in MIMO Fading Systems Without Channel Distribution Information

Similar documents
Consider a system of 2 simultaneous first order linear equations

Summary: Solving a Homogeneous System of Two Linear First Order Equations in Two Unknowns

State Observer Design

The Variance-Covariance Matrix

innovations shocks white noise

Advanced Queueing Theory. M/G/1 Queueing Systems

Boosting and Ensemble Methods

9. Simple Rules for Monetary Policy

Frequency Response. Response of an LTI System to Eigenfunction

t=0 t>0: + vr - i dvc Continuation

Ergodic Capacity of a SIMO System Over Nakagami-q Fading Channel

Supplementary Figure 1. Experiment and simulation with finite qudit. anharmonicity. (a), Experimental data taken after a 60 ns three-tone pulse.

ELEN E4830 Digital Image Processing

SIMEON BALL AND AART BLOKHUIS

Lecture 4 : Backpropagation Algorithm. Prof. Seul Jung ( Intelligent Systems and Emotional Engineering Laboratory) Chungnam National University

Problem 1: Consider the following stationary data generation process for a random variable y t. e t ~ N(0,1) i.i.d.

CIVL 8/ D Boundary Value Problems - Triangular Elements (T6) 1/8

Analysis of decentralized potential field based multi-agent navigation via primal-dual Lyapunov theory

arxiv: v1 [math.ap] 16 Apr 2016

FAULT TOLERANT SYSTEMS

Wave Superposition Principle

Chapter 9 Transient Response

Guaranteed Cost Control for a Class of Uncertain Delay Systems with Actuator Failures Based on Switching Method

Safety and Reliability of Embedded Systems. (Sicherheit und Zuverlässigkeit eingebetteter Systeme) Stochastic Reliability Analysis

Applying Software Reliability Techniques to Low Retail Demand Estimation

Chapter 13 Laplace Transform Analysis

Safety and Reliability of Embedded Systems. (Sicherheit und Zuverlässigkeit eingebetteter Systeme) Stochastic Reliability Analysis

Control Systems (Lecture note #6)

Theoretical Seismology

Lucas Test is based on Euler s theorem which states that if n is any integer and a is coprime to n, then a φ(n) 1modn.

Lecture 1: Numerical Integration The Trapezoidal and Simpson s Rule

Transient Analysis of Two-dimensional State M/G/1 Queueing Model with Multiple Vacations and Bernoulli Schedule

CPSC 211 Data Structures & Implementations (c) Texas A&M University [ 259] B-Trees

CONTINUOUS TIME DYNAMIC PROGRAMMING

Engineering Circuit Analysis 8th Edition Chapter Nine Exercise Solutions

Implementation of the Extended Conjugate Gradient Method for the Two- Dimensional Energized Wave Equation

A Note on Estimability in Linear Models

Homework: Introduction to Motion

Gaussian Random Process and Its Application for Detecting the Ionospheric Disturbances Using GPS

Copyright 2000, Kevin Wayne 1

On the Derivatives of Bessel and Modified Bessel Functions with Respect to the Order and the Argument

Dr. Junchao Xia Center of Biophysics and Computational Biology. Fall /21/2016 1/23

Bethe-Salpeter Equation Green s Function and the Bethe-Salpeter Equation for Effective Interaction in the Ladder Approximation

EE243 Advanced Electromagnetic Theory Lec # 10: Poynting s Theorem, Time- Harmonic EM Fields

Chap 2: Reliability and Availability Models

10/7/14. Mixture Models. Comp 135 Introduction to Machine Learning and Data Mining. Maximum likelihood estimation. Mixture of Normals in 1D

CSE 245: Computer Aided Circuit Simulation and Verification

A universal saturation controller design for mobile robots

Grand Canonical Ensemble

Dynamic Team Decision Theory. EECS 558 Project Shrutivandana Sharma and David Shuman December 10, 2005

Soft k-means Clustering. Comp 135 Machine Learning Computer Science Tufts University. Mixture Models. Mixture of Normals in 1D

Dynamic Controllability with Overlapping Targets: Or Why Target Independence May Not be Good for You

Reduced The Complexity of Soft Input Soft Output Maximum A posteriori Decoder of Linear Block Code By Using Parallel Trellises Structure

Geo-LANMAR Routing: Asymptotic Analysis of a Scalable Routing Scheme with Group Motion Support

Least squares and motion. Nuno Vasconcelos ECE Department, UCSD

Prediction of channel information in multi-user OFDM systems

Final Exam : Solutions

Convergence of Quintic Spline Interpolation

Comparative Study of Finite Element and Haar Wavelet Correlation Method for the Numerical Solution of Parabolic Type Partial Differential Equations

Physics 160 Lecture 3. R. Johnson April 6, 2015

8. Queueing systems. Contents. Simple teletraffic model. Pure queueing system

RELATIONSHIPS BETWEEN SPECTRAL PEAK FREQUENCIES OF A CAUSAL AR(P) PROCESS AND ARGUMENTS OF ROOTS OF THE ASSOCIATED AR POLYNOMIAL.

Midterm exam 2, April 7, 2009 (solutions)

Review - Probabilistic Classification

Yutaka Suzuki Faculty of Economics, Hosei University. Abstract

Chapter 7 Stead St y- ate Errors

1. Inverse Matrix 4[(3 7) (02)] 1[(0 7) (3 2)] Recall that the inverse of A is equal to:

Mechanics Physics 151

Published in: Proceedings of the Twenty Second Nordic Seminar on Computational Mechanics

Charging of capacitor through inductor and resistor

Valuation and Analysis of Basket Credit Linked Notes with Issuer Default Risk

Robust decentralized control with scalar output of multivariable structurally uncertain plants with state delay 1

Microscopic Flow Characteristics Time Headway - Distribution

Solution in semi infinite diffusion couples (error function analysis)

Conventional Hot-Wire Anemometer

AR(1) Process. The first-order autoregressive process, AR(1) is. where e t is WN(0, σ 2 )

CHAPTER 33: PARTICLE PHYSICS

Gauge Theories. Elementary Particle Physics Strong Interaction Fenomenology. Diego Bettoni Academic year

Lecture 12: Introduction to nonlinear optics II.

Lectures 9-11: Fourier Transforms

CHAPTER CHAPTER14. Expectations: The Basic Tools. Prepared by: Fernando Quijano and Yvonn Quijano

NAME: ANSWER KEY DATE: PERIOD. DIRECTIONS: MULTIPLE CHOICE. Choose the letter of the correct answer.

V.Abramov - FURTHER ANALYSIS OF CONFIDENCE INTERVALS FOR LARGE CLIENT/SERVER COMPUTER NETWORKS

Almost unbiased exponential estimator for the finite population mean

Node Placement and Mobility Control in Mobile Wireless Sensor Networks

MATCHED FILTER BOUND OPTIMIZATION FOR MULTIUSER DOWNLINK TRANSMIT BEAMFORMING

The Fourier Transform

Real-Time Scheduling for Event-Triggered and Time-Triggered Flows in Industrial Wireless Sensor-Actuator Networks

Chapter 10. The singular integral Introducing S(n) and J(n)

Continous system: differential equations

DEPARTMENT OF ELECTRICAL &ELECTRONICS ENGINEERING SIGNALS AND SYSTEMS. Assoc. Prof. Dr. Burak Kelleci. Spring 2018

Probabilistic Reasoning; Graphical models

4.1 The Uniform Distribution Def n: A c.r.v. X has a continuous uniform distribution on [a, b] when its pdf is = 1 a x b

Robustness Experiments with Two Variance Components

ADAPTIVE PRE-EMPTIVE CONTROL OF VACUUM DEWATERING IN PAPER MANUFACTURING 1. Petar Bjegovic 3 Perry Y. Li 2

The Matrix Exponential

Relative controllability of nonlinear systems with delays in control

Poisson process Markov process

IMPROVED RATIO AND PRODUCT TYPE ESTIMATORS OF FINITE POPULATION MEAN IN SIMPLE RANDOM SAMPLING

The Matrix Exponential

Transcription:

PROC. IEEE INFOCOM 06 Dynamc Powr Allocaon n MIMO Fadng Sysms Whou Channl Dsrbuon Informaon Hao Yu and Mchal J. Nly Unvrsy of Souhrn Calforna Absrac Ths papr consdrs dynamc powr allocaon n MIMO fadng sysms wh unknown channl sa dsrbuons. Frs, h dal cas of prfc nsananous channl sa nformaon a h ransmr (CSIT) s rad. Usng h drf-plus-pnaly mhod, a dynamc powr allocaon polcy s dvlopd and shown o approach opmaly, rgardlss of h channl sa dsrbuon and whou rqurng knowldg of hs dsrbuon. Nx, h cas of dlayd and quanzd channl sa nformaon s consdrd. Opmal uly s fundamnally dffrn n hs cas, and a dffrn onln algorhm s dvlopd ha s basd on convx projcons. Th proposd algorhm for hs dlayd-csit cas s shown o hav an O( ) opmaly gap, whr s h quanzaon rror of CSIT. I. INTRODUCTION Durng h pas dcad, h mulpl-npu mulpl-oupu (MIMO) chnqu has bn rcognzd as on of h mos mporan chnqus for ncrasng h capabls of wrlss communcaon sysms. In h wrlss fadng channl, whr h channl changs ovr m, h problm of powr allocaon s o drmn h ransm covaranc of h ransmr o maxmz h rgodc capacy subjc o boh long rm and shor rm powr consrans. I s ofn rasonabl o assum ha nsananous channl sa nformaon (CSI) s avalabl a h rcvr hrough ranng. Mos works on powr allocaon n MIMO fadng sysms also assum ha sascal nformaon abou h channl sa, rfrrd o as channl dsrbuon nformaon (CDI), s avalabl a h ransmr. Undr h assumpon of prfc nsananous channl sa nformaon a h rcvr (CSIR) and prfc channl dsrbuon nformaon a h ransmr (CDIT), pror work on powr allocaon n MIMO fadng sysms can b cagorzd no wo cass: Prfc nsananous channl sa nformaon a h ransmr (dal-csit): In h dal cas of prfc CSIT, opmal powr allocaon s known o b a warfllng soluon []. Compuaon of war-lvls nvolvs a on-dmnsonal ngral quaon for fadng channls wh..d. Raylgh nrs or a mul-dmnsonal ngral quaon for gnral fadng channls []. No CSIT: If CSIT s unavalabl, h opmal powr allocaon s n gnral sll opn. If h channl marx has..d. Raylgh nrs, hn h opmal powr allocaon s known o b h dny ransm covaranc scald o Hao Yu and Mchal J. Nly ar wh h Dparmn of Elcrcal Engnrng, Unvrsy of Souhrn Calforna, Los Angls, USA. Ths work s suppord n par by h NSF gran CCF-074755. sasfy h powr consran []. Th opmal powr allocaon n MIMO fadng channls wh corrlad Raylgh nrs s oband n [3], [4]. Th powr allocaon n MIMO fadng channls s furhr consdrd n [5] undr a mor gnral channl corrlaon modl. Ths pror work rls on accura CDIT and/or on rsrcv channl dsrbuon assumpons. I can b dffcul o accuraly sma h CDI, spcally whn hr ar complcad corrlaons. Soluons ha bas dcsons on prfc CDIT can b subopmal du o msmachs. Ths papr dsgns algorhms ha do no rqur pror knowldg of h channl dsrbuon, y prform arbrarly clos o h opmal valu ha can b achvd by havng hs knowldg. Furhr, h convrgnc m s compud and shown o b sgnfcanly smallr han h m rqurd o accuraly sma h channl dsrbuon. Th dal-csit assumpon s rasonabl n m-dvson duplx (TDD) sysms wh symmrc TDD wrlss channls. Howvr, n frquncy-dvson duplx (FDD) scnaros and ohr scnaros whou channl symmry, h CSI mus b smad a h rcvr, quanzd, and rpord back o h ransmr wh a m dlay. Ths papr frs consdrs h dal-csit cas and dvlops a soluon ha dos no rqur CDIT. Nx, h cas of dlayd and quanzd CSIT s consdrd and a fundamnally dffrn algorhm s dvlopd for ha cas. Th lar algorhm agan dos no us CDIT, bu achvs a uly whn O( ) of h bs uly ha can b achvd vn wh prfc CDIT, whr s h quanzaon rror. Ths shows ha dlayd bu accura CSIT (wh 0) s almos as good as havng prfc CDIT. A. Rlad work and our conrbuons In h dal CSIT cas, h proposd dynamc powr allocaon polcy s an adapon of h gnral drf-plus-pnaly algorhm for sochasc nwork opmzaon [6], [7]. In hs MIMO conx, h currn papr shows h algorhm provds srong sampl pah and convrgnc m guarans. Th dynamc of h drf-plus-pnaly algorhm s smlar o ha of h sochasc dual subgradn algorhm, alhough h opmaly analyss and prformanc bounds ar dffrn. Th sochasc dual subgradn algorhm has bn appld n opmzaon of h wrlss fadng channl whou CDI,.g., downlnk powr schdulng n sngl annna cllular sysms [8], powr allocaon n sngl annna broadcas OFDM channls [9], schdulng and rsourc allocaon n

PROC. IEEE INFOCOM 06 random accss channls [0], powr allocaon n mul-carrr MIMO nworks []. In h dlayd and quanzd CSIT cas, h suaon s smlar o h scnaro of onln convx opmzaon [] xcp ha w ar unabl o obsrv ru hsory rward funcons du o channl quanzaon. Th proposd dynamc powr allocaon polcy can b vwd as an onln algorhm wh naccura hsory nformaon. Th currn papr analyzs h prformanc loss du o CSI quanzaon rror and provds srong sampl pah and convrgnc m guarans of hs algorhm. Accordng o h auhors knowldg, onln convx opmzaon wh naccura hsory nformaon has no bn sudd bfor. Th analyss n hs MIMO conx can b xndd o mor gnral onln convx opmzaon wh naccura hsory nformaon. Onln opmzaon has bn appld n powr allocaon n wrlss fadng channls whou CDIT and wh dlayd and accura CSIT,.g., subopmal onln powr allocaon n sngl annna sngl usr channls [3], subopmal onln powr allocaon n sngl annna mulpl usr channls [4]. A clos rlad rcn work s [5], whr onln powr allocaon n MIMO sysms s consdrd. Th onln algorhm n [5] s dffrn from our algorhm and follows a marx xponnal larnng schm rqurng h compuaon of marx xponnals a ach slo. In conras, our onln algorhm only nvolvs a smpl projcon a ach slo and a closd-form soluon of hs projcon s drvd n hs papr. Work [5] also consdrs h ffc of mprfc CSIT by assumng CSIT s unbasd,.., xpcd CSIT rror condonal on obsrvd prvous CSIT s zro. Ths assumpon of mprfc CSIT s suabl o modl h CSIT masurmn rror or fdback rror bu can no capur h CSI quanzaon rror. In conras, h currn papr only rqurs ha CSIT rror s boundd. II. SIGNAL MODEL AND PROBLEM FORMULATIONS A. Sgnal modl Consdr a pon-o-pon MIMO fadng channl ha opras n slod m wh normalzd m slos {,,...}. Thr ar N T annnas a h ransmr and N R annnas a h rcvr. Th channl can b modld as y() =H()x()+z() whr {,,...} s h m ndx, z() C N R s h addv nos vcor, x() C N T s h ransmd sgnal vcor, H() C N R N T s h channl marx, and y() C N R s h rcvd sgnal vcor. Assum ha nos vcors z() ar ndpndn and dncally dsrbud (..d.) ovr m slos and ar normalzd crcularly symmrc complx Gaussan random vcors wh E[z()z H ()] = I NR, whr I NR dnos an N R N R dny marx. Assum ha channl marcs H() ar..d. across m and hav a fxd bu arbrary probably dsrbuon, possbly on wh corrlaons bwn nrs of h marx. Assum hr xss If h sz of h dny marx s clar, w ofn smply wr I. a consan B>0 such ha khk F appl B wh probably on, whr k k F dnos h Frobnus norm of marcs. Assum ha h rcvr can rack channl marcs H() xacly hrough ranng. In symmrc TDD scnaros, s rasonabl o assum h ransmr has prfc CSIT. In mor gnral scnaros, h channl marx H() s masurd a h rcvr a ach slo, a quanzd vrson H() s crad as a funcon of H(), and hs quanzd vrson s fd back o h ransmr wh on slo of dlay. W assum ha h quanzaon rror s boundd,.., hr xss >0 such ha kh() H()k F appl for all. Du o h on slo dlay, a slo h ransmr only knows H( ). Snc channls ar..d. ovr slos, hs dlayd nformaon s ndpndn of h currn (and unknown) H(). Rmarkably, urns ou ha h oudad nformaon s sll usful. B. Opmal powr allocaon wh prfc CDIT If h channl marx s fxd a H and h ransm covaranc s fxd a Q, h MIMO capacy s gvn by []: log d(i + HQH H ) whr suprscrp H dnos Hrman ranspos and d( ) dnos h drmnan opraor of marcs. If H s random hn h avrag capacy, formally calld h rgodc capacy [6], s gvn by E H log d(i + HQH H ). W consdr wo yps of powr consrans a h ransmr: An avrag powr consran E H [r(q)] appl P and an nsananous powr consran r(q) appl P, whr r( ) dnos h rac opraor of marcs. Th dal-csit problm s o choos Q as a (possbly random) funcon of h obsrvd H o maxmz h rgodc capacy subjc o powr consrans: max Q(H) E H log d(i + HQ(H)H H ) () s.. E H [r(q(h))] appl P, () Q(H) Q, 8H, (3) whr Q s a s ha nforcs h nsananous powr consran: Q = Q S N T + : r(q) appl P (4) whr S N T + dnos h N T N T posv smdfn marx spac. To avod rvals s assumd ha P P. In ()-(3) w us noaon Q(H) o mphasz ha Q can dpnd on H,.., adapv o channl ralzaons. If h ransmr has no CSIT, h opmal powr allocaon problm s dffrn, gvn as follows. max Q E H log d(i + HQH H ) (5) s.. E H [r(q)] appl P, (6) Q Q, (7) whr s Q s dfnd n (4). Agan assum P P. Snc h nsananous CSIT s unavalabl, h ransm covaranc A boundd Frobnus norm always holds n h physcal world bcaus h channl anuas h sgnal. Parcular modls such as Raylgh and Rcan fadng vola hs assumpon n ordr o hav smplr dsrbuon funcons.

PROC. IEEE INFOCOM 06 canno adap o H. By h convxy of hs problm and Jnsn s nqualy, a randomzd Q can b shown o b uslss. I suffcs o consdr a consan Q. Snc P P, hs mpls h problm s quvaln o a problm ha rmovs h consran (6) and ha changs h consran (7) o: Q Q = {Q S N T + : r(q) appl P } Th problms ()-(3) and (5)-(7) ar fundamnally dffrn and hav dffrn opmal objcv funcon valus. Opmaly for hs problms s dfnd by h channl dsrbuon nformaon (CDI). In hs papr, h problms ar solvd va dynamc algorhms ha do no rqur CDI. Th algorhms ar dffrn for h wo cass, and us dffrn chnqus. C. Lnar algbra and marx drvavs Rcall ha f A C m n and B C n m hn r(ab) = r(ba). Ths subscon prsns addonal usful facs abou Frobnus norms and complx marcs. Proofs ar gvn n [7] for complnss. Fac. For any A, B C m n and C C n k w hav: ) kak F = ka H k F = ka T k = k Ak F. ) ka + Bk F applkak F + kbk F. 3) kack F applkak F kck F. 4) r(a H B) applkak F kbk F. Fac. For any A S n + w hav kak F appl r(a). Fac 3 ([8]). Th funcon f : S n +! R dfnd by f(q) = log d(i + HQH H ) s concav and s gradn s gvn by r Q f(q) =H H (I + HQH H ) H, 8Q S n +. Th nx fac s h complx marx vrson of h frs ordr condon for concav funcons of ral numbr varabls,.., f(y) appl f(x)+f 0 (x)(y x), 8x, y domf f f s concav. Fac 4. L funcon f(q) :S n +! R b a concav funcon and hav gradn r Q f(q) S n a pon Q. Thn, f( b Q) appl f(q)+r [r Q f(q)] H ( b Q Q), 8 b Q S n +. III. IDEAL CSIT CASE Consdr h cas of prfc nsananous CSIT, calld h dal-csit cas. Th problm o solv s ()-(3). A h bgnnng of ach slo {,,...} h channl H() s known and a covaranc marx Q() can b chosn basd on hs nformaon. Ths s don whou usng CDI va h drfplus-pnaly chnqu of [7]. For ach slo {,,...} dfn h rward R(): R() = log d(i + H()Q()H() H ) (8) Th avrag powr consran () s nforcd va a vrual quu Z() wh Z(0) = 0 and wh upda: Z( + ) = max[z()+r(q()) P ] In h drf-plus-pnaly algorhm, vry slo a marx Q() Q s slcd o maxmz VR() Z()r(Q()), whr V s a posv wgh. Ths rsuls n Algorhm blow. Algorhm Dynamc powr allocaon wh dal CSIT L V > 0 b a consan paramr and Z(0) = 0. A ach m {,,...}, obsrv H() and Z(). Thn do h followng: Choos ransm covaranc Q() Qo maxmz: V log d(i + H()Q()H() H ) Z()r(Q()) Upda Z( + ) = max[z()+r(q()) P ]. Dfn R op as h opmal avrag uly n (). Th valu R op dpnds on h (unknown) dsrbuon for H(). Fx > 0 and dfn V =(P + P ) /( ). A horm n [7] nsurs ha, rgardlss of h dsrbuon of H(): X E[R( )] R op, 8 >0 (9) lm! X E[r(Q( ))] appl P (0) Ths holds for arbrarly small valus of > and so h algorhm coms arbrarly clos o opmaly. Noc ha Algorhm dos no us channl dsrbuon nformaon (.., no CDI). Th nx subscons show how o solv h covaranc slcon problm for choosng Q() n Algorhm, and shows ha h spcal srucur of hs MIMO problm producs a sampl pah guaran ha s sgnfcanly srongr han (0) and ha dmonsras convrgnc m ha s ypcally much fasr han h m ha would b rqurd o accuraly sma h CDI nformaon. A. Transm covaranc updas n Algorhm Ths subscon shows h Q() slcon n Algorhm can b asly solvd and has an (almos) closd-form soluon. Th convx program nvolvd n h ransm covaranc upda of Algorhm s n h form max log d(i + HQH H Z ) r(q) () Q V s.. r(q) appl P () Q S N T + (3) Ths convx program s smlar o h convnonal problm of ransm covaranc dsgn wh a drmnsc channl H, xcp ha objcv () has an addonal pnaly rm (Z/V )r(q). I s wll known ha, whou hs pnaly rm, h soluon s o dagonalz h channl marx and alloca powr ovr gn-mods accordng o a war-fllng chnqu []. Th nx horm shows ha h opmal soluon o problm ()-(3) has a smlar srucur. Thorm. Consdr h SVD H H H = U H U, whr U s a unary marx and s a dagonal marx wh non-ngav nrs,..., N T. Thn h opmal soluon o ()-(3) s

PROC. IEEE INFOCOM 06 gvn by Q = U H U, whr s a dagonal marx wh nrs,..., N T gvn by: = max µ, 8 {,...,NT }, whr µ s chosn such ha P N T = appl P, µ 0 and µ P N T = P =0. Th xac µ can b drmnd wh complxy O(N T log N T ), dscrbd n Algorhm. Proof: S Appndx A. Algorhm Algorhm o solv problm ()-(3) ) Chck f P N T = max{ Z/V } appl P holds. If ys, l µ = 0 and = max{ Z/V }, 8 {,,...,N T } and rmna h algorhm; ls, connu o h nx sp. ) Sor all, {,,...,N T } n a dcrasng ordr such ha () () (N T ). Dfn S 0 =0. 3) For =o N T L S = S +. L µ = () S +P (Z/V ). If µ > 0 and () appl hn rmna h loop; ls, connu (+) o h nx raon n h loop. 4) L = max, 8 {,,...,NT } and rmna h algorhm. Th complxy of Algorhm s domnad by h sorng of all n sp (). Rcall ha h war-fllng soluon of powr allocaon n mulpl paralll channls can also b found by an xac algorhm (s Scon 6 n [9]), whch s smlar o Algorhm. Th man dffrnc s ha Algorhm has a frs sp o vrfy f µ =0. Ths s bcaus unlk h powr allocaon n mulpl paralll channls, whr h opmal soluon always uss full powr, h opmal soluon o problm ()-(3) may no us full powr for larg Z du o h pnaly rm (Z/V )r(q) n objcv (). B. Drmnsc bounds Rcall ha kh()k F appl B for all, for som consan B. Lmma. In Algorhm, f Z() VB, hn Q() =0. Proof: Suppos h SVD of H H ()H() s gvn by H H ()H() =U H U, whr dagonal marx has nonngav dagonal nrs,..., N T. No ha r(h H ()H()) (b) appl kh()k F appl B whr (a) follows from r(h H ()H()) = P N T = ; and (b) follows from Fac. By Thorm, Q() =U H U, whr s a dagonal marx wh nrs,..., N T gvn by = max µ +Z()/V, 8 {,,...,NT }, whr µ 0. Snc appl B, 8 {,,...,N T }, w know ha f Z() VB, hn µ+z()/v appl 0 for all µ 0 and hnc = 8 {,,...,N T }. (a) appl Lmma. L Z() b yldd by Algorhm. For all slos {,,...}, w hav Z() appl VB +(P P ). Proof: By Lmma, Z() can no ncras f Z() VB. If Z() appl VB, hn Z(+) s a mos VB +(P P ) by h upda quaon of Z(+) and h nsananous powr consran. C. Prformanc of Algorhm (dal-csit) Thorm. Fx >0 and dfn V =(P + P ) /( ). Undr Algorhm w hav for all >0: X X E[R( )] R op r(q( )) appl P + B (P + P ) + (P P ) In parcular, h sampl pah m avrag powr s whn of s rqurd consran P whnvr (/ ). Proof: Th frs nqualy s h sam as (9). I rmans o prov h scond nqualy. For all slos h Algorhm upda for Z( ) sasfs: Z( + ) = max[z( )+r(q( )) P ] Z( )+r(q( )) P Rarrangng rms gvs: r(q( )) P appl Z( + ) Z( ). Fx >0. Summng ovr {..., } and dvdng by gvs: X Z() Z(0) r(q( )) P appl appl (VB +(P P )) whr h las nqualy holds bcaus Z(0) = 0 and Z() appl VB +(P P ) by Lmma. Thorm provds a sampl pah guaran on avrag powr, whch s much srongr han h guran n (0). I also shows ha convrgnc m o rach an -approxma soluon s O(/ ). Typcally, hs s dramacally mor ffcn han h convrgnc m rqurd o oban vn a coars sma of h jon dsrbuon for h nrs of H(). Indd, f ach channl nry h j wr quanzd no / dsnc lvls, hr would b (/ ) N T N R dffrn possbl (quanzd) marx ralzaons. Wang for (/ ) N T N R slos would a bs allow ach ralzaon o appar onc, whch s sll no nough for accura smaon of h probabls assocad wh ach ralzaon. Forunaly, Thorm shows ha such smaon s no ndd. IV. DELAYED AND QUANTIZED CSIT CASE Consdr h cas of dlayd and quanzd CSIT. A h bgnnng of ach slo {,,...}, channl H() s unknown and only quanzd channls of prvous slos H( ), {,..., } ar known.

PROC. IEEE INFOCOM 06 Ths s smlar o h scnaro of onln opmzaon whr h dcson makr slcs x() X a ach slo o maxmz an unknown rward funcon f (x) basd on h nformaon of prvous rward funcons f (x( )), {,..., }. Th goal s o mnmz avrag rgr max P xx f (x) P f (x( )). Th bs known avrag rgr of onln opmzaon wh Lpschz connuous and convx rward funcons s O( p ) n []. Ths s dffrn from convnonal onln opmzaon bcaus a ach slo, h rwards of prvous slos,.., R( ) = log d(i + H( )Q( )H H ( )), {,..., }, ar sll unknown du o h fac ha h rpord channls H( ) ar h quanzd vrsons. Nvrhlss, an onln algorhm whou usng CDIT s dvlopd n Algorhm 3. Algorhm 3 Dynamc Powr Allocaon wh Dlayd and Quanzd CSIT L >0 b a consan paramr and Q(0) Qb arbrary. A ach m {,,...}, obsrv H( ) and do h followng: L D( ) = H H ( )(I NR + H( )Q( ) H H ( )) H( ). Choos ransm covaranc Q() =P Q Q( ) + D( ), whr P Q [ ] s h projcon ono convx s Q = {Q S N T + : r(q) appl P }. Dfn Q Q as an opmal soluon o problm (5)-(7), whch dpnds on h (unknown) dsrbuon for H(). Dfn R op () = log d(i + H()Q H H ()) as h uly a slo aand by Q. If h channl s no quanzd,.., H( ) = H( ), 8 {,,...}, hn D( ) s h gradn of R( ) a pon Q( ). Fx >0 and ak =. Th rsuls n [] nsur ha, rgardlss of h dsrbuon of H(): X R( ) X R op ( ) P N R B 4, 8 >0 (4) X r(q( )) appl P,8 >0 (5) Th nx subscons analyz h prformanc of Algorhm 3 wh quanzd channls and shows ha h prformanc dgrads lnarly wh rspc o h quanzaon rror. If = hn (4) and (5) ar rcovrd. A. Transm Covaranc Updas n Algorhm 3 Ths subscon shows h Q() slcon n Algorhm 3 can b asly solvd and has an (almos) closd-form soluon. Th projcon opraor nvolvd n Algorhm 3 by dfnon s mn kq Xk F (6) s.. r(q) appl P (7) Q S N T + (8) whr X = Q( ) + D( ) s an Hrman marx a ach m. Whou consran r(q) appl P, h projcon of Hrman marx X ono h posv smdfn con S n + s smply akng h gnvalu xpanson of X and droppng rms assocad wh ngav gnvalus (s Scon 8... n [0]). Work [] consdrd h projcon ono h nrscon of h posv smdfn con S n + and an affn subspac gvn by {Q : r(a Q)=b, {,,...,p}, r(b j Q) appl d j,j {,,...,m}} and dvlopd h dual-basd rav numrcal algorhm o calcula h projcon. Problm (6)- (8) s a spcal cas, whr h affn subspac s gvn by r(q) appl P, of h projcon consdrd n []. Insad of solvng problm (6)-(8) usng numrcal algorhms, hs subscon shows ha problm (6)-(8) has an (almos) closd-form soluon. Thorm 3. Consdr SVD X = U H U, whr U s a unary marx and s a dagonal marx wh nrs,..., N T. Thn h opmal soluon o problm (6)-(8) s gvn by Q = U H U, whr s a dagonal marx wh nrs,..., N T gvn by, = max[ µ ], 8 {,,...,N T }, whr µ s chosn such ha P N T = appl P, µ 0 and µ P N T = P =0. Th xac µ can b drmnd wh complxy O(N T log N T ), dscrbd n Algorhm 4. Proof: Th proof s skchd as follows. Frs, problm (6)-(8) s rducd o a smplr convx program wh a ral vcor varabl by characrzng h srucur of s opmal soluon. Thn, an (almos) closd-form soluon o h smplr convx program s oband by sudyng s KKT condons. S Appndx B for dals. Algorhm 4 Algorhm o solv problm (6)-(8) ) Chck f P N T = max[ ] appl P holds. If ys, l µ =0 and = max[ ], 8 {,,...,N T } and rmna h algorhm; ls, connu o h nx sp. ) Sor all, {,,...,N T } n a dcrasng ordr such ha () () (N T ). Dfn S 0 =0. 3) For =o N T L S = S +. L µ = S P. If µ () µ > 0 and (+) µ appl hn rmna h loop; ls, connu o h nx raon n h loop. 4) L = max[ µ ], 8 {,,...,N T } and rmna h algorhm.

PROC. IEEE INFOCOM 06 B. Propry of D( ) Dfn D( ) = H H ( )(I NR + H( )Q( )H H ( )) H( ), whch s h gradn of R( ) a pon Q( ) and s unknown o h ransmr du o h unavalably of H( ). Th nx lmma rlas D( ) and D( ). Lmma 3. For all slos {,,...}, w hav ) kd( )k F appl p N R B. ) kd( ) D( )kf appl ( ), whr ( ) = p NR B + p N R (B + )+(B + ) N R P (B + ) sasfyng ( )! 0 as!.., ( ) O( ). 3) k D( )k F appl ( )+ p N R B Proof: S full vrson [7] for dals. C. Prformanc of Algorhm 3 Thorm 4. Fx >0 and dfn =. Undr Algorhm 3, w hav for all >0: X R( ) X R op ( ) X r(q( )) appl P ( ) P P ( ( )+ p N R B ) whr ( ) s h consan dfnd n Lmma 3. In parcular, h sampl pah m avrag uly s whn + ( ) P of h opmal m avrag uly for problm (5)-(7) whnvr (/ ). Proof: Th scond nqualy rvally follows from h fac ha Q() Q, 8 {,...}. I rmans o prov h frs nqualy. Ths proof xnds h rgr analyss of convnonal onln convx opmzaon [] by consdrng nxac gradn D( ). For all slos {,,...}, h ransm covaranc upda n Algorhm 3 sasfs: kq( ) Q k F =kp Q Q( ) + D( ) (a) applkq( ) + D( ) Q k F Q k F =kq( ) Q k F + r D H ( )(Q( ) Q ) + k D( )k F =kq( ) Q k F + r D H ( )(Q( ) Q ) + r ( D( ) D( )) H (Q( ) Q ) + k D( )k F, whr (a) follows from h non-xpansv propry of projcons ono convx ss. Dfn () =kq( + ) Q k F kq() Q k F. Rarrangng rms n h las quaon and dvdng by facor mpls r D H ( )(Q( ) Q ) ( ) k D( )k F r ( D( ) D( )) H (Q( ) Q ) (9) Dfn f(q) = log d(i+h( )QH H ( )). By Fac 3, f( ) s concav ovr Q. No ha D( ) = r Q f(q( )) by Fac 3 and Q Q. By Fac 4, w hav f(q( )) f(q ) r(d H ( )(Q( ) Q )). (0) No ha f(q( )) = R( ) and f(q )=R op ( ). Combnng (9) and (0) ylds R( ) R op ( ) ( ) k D( )k F (a) (b) r ( D( ) D( )) H (Q( ) Q ) ( ) k D( )k F k D( ) D( )k F kq( ) Q k F ( ) ( ( )+p N R B ) ( ) P whr (a) follows from Fac and (b) follows from Lmma 3 and h fac ha kq( ) Q k F applkq( )k F +kq k F appl r(q( )) + r(q ) appl P, whch s mpld by Fac, Fac and fac Q( ), Q Q. Rplacng wh ylds for all {,...}, R( ) R op ( ) ( ) p ( ( )+ NR B ) ( ) P. Fx >0. Summng ovr {,..., }, dvdng by facor and nong ha P ( ) s a lscop sum gvs P R( ) P Rop ( ) (kq() Q k F kq(0) Q k F ) ( ( )+p N R B ) ( ) P P ( ( )+p N R B ) ( ) P, whr h las nqualy follows bcaus kq(0) Q k F appl kq(0)k F + kq k F appl r(q(0)) + r(q ) appl P and kq() Q k F 0. Thorm 4 provs a sampl pah guaran on h avrag uly. I shows ha h convrgnc m o rach an + ( ) P approxma soluon s O(/ ). No ha f = hn quaons (4) and (5) ar rcovrd by Thorm 4. Thorm 4 also solas h ffcs of dlay and quanzaon. Th obsrvaon s ha h ffc of CSIT dlay vanshs as Algorhm 3 runs for a suffcnly long m. In som sns, dlayd bu accura CSIT s almos as good as prfc CDIT. In conras, h ffc of CSIT quanzaon dos no vansh as Algorhm 3 runs for a suffcnly long m. Th prformanc dgradaon du o quanzaon scals lnarly wh rspc o h quanzaon rror snc ( ) O( ). Inuvly, hs s rasonabl snc h powr allocaon basd on quanzd CSIT s acually opmzng anohr dffrn MIMO sysm.

PROC. IEEE INFOCOM 06 D. Exnsons ) T -Slo Dlayd and Quanzd CSIT: Thus far, w hav assumd ha CSIT s dlayd by on slo. In fac, f CSIT s dlayd by T slos, w can modfy h upda of ransm covarancs n Algorhm 3 as Q() =P Q [Q( T )+ D( T )]. AT -slo vrson of Thorm 4 can b smlarly provn. ) Algorhm 3 wh Tm Varyng : Algorhm 3 can b xndd o hav m varyng sp sz () = p a m. Th full vrson [7] provs ha such an algorhm P P Rop ( ) ylds R( ) Pp p ( ( )+ p NR B ) ( ) P for all >0. Ths shows h convrgnc m o an + ( ) P approxma soluon s agan O(/ ). Howvr, an advanag of m varyng sp szs s h prformanc auomacally gs mprovd as h algorhm runs and hr s no nd o rsar h algorhm wh a dffrn consan sp sz f a br prformanc s dmandd. V. SIMULATIONS Consdr a MIMO sysm whr boh h ransmr and h rcvr hav wo annnas. Th powr consrans ar P = 5and P = 0. Th channl has appl wo ralzaons wh j0.84 qual probably 0.5,.., H =0.5 j.58 j.83 j.97 and appl j.3 H = j.69 j0.07 j.86. If h channl s quanzd, hy ar quanzd as H and H, rspcvly. Th algorhms n hs papr can b asly appld o xampls wh nfn possbl oucoms for h channl marx. Ths smpl xampl of wo possbls s consdrd bcaus an offln opmal soluon basd on prfc CDIT can only b compud whn h numbr of sampls s small. 3 Fgur compars h prformanc of Algorhm wh prfc CSIT and h opmal soluon o problm ()-(3). In h smulaon, w ak V = 000. Fgur compars h prformanc of Algorhm 3 wh on slo dlayd and quanzd CSIT and h opmal soluon o problm (5)-(7). To sudy h ffc of quanzaon rror, w consdr 3 dffrn quanzaon lvls. Cas : H appl appl = j0.8 0.5 j.5 j.8 j.9 and H j.3 = j.6 j0 j.8 ; Cas appl appl : H j =0.5 j.5 j j and H j.5 = j.5 j0 j ; appl appl Cas 3: H = 0.5 and H =. In h smulaon, w ak Q(0) = 0 and = 0 3. I can b obsrvd ha prformanc bcoms wors as CSIT quanzaon gs coarsr, whl h avrag powr consrans ar srcly sasfd vn wh quanzd CSIT. 3 Ths s known as h curs of dmnsonaly for sochasc opmzaon du o h larg sampl sz. Tha s, vn wh prfc CDIT, problm ()-(3) and problm (5)-(7) can b numrcally hard o solv whn h sampl sz of H s larg. In conras, h dynamc algorhms proposd n hs papr can dal wh problms vn wh an nfn numbr of sampls and h prformanc guarans ar ndpndn of h sampl sz. Uly Avrag Powr Uly 3.5 3 Prformanc of Algorhm.5 0 000 000 3000 4000 5000 6000 7000 8000 Slos: Avrag Powr 0 9 8 7 6 Algorhm Algorhm opmal soluon o problm () (3) opmal soluon o problm () (3) 5 0 000 000 3000 4000 5000 6000 7000 8000 Slos:.5.5 0.5 Fg.. Prformanc of Algorhm. Prformanc of Algorhm 3 0 0 0.5.5 Slos:.5 3 3.5 4 x 0 5 5 4 3 quanzaon cas opmal soluon o problm (5) (7) quanzaon cas quanzaon cas quanzaon cas 3 quanzaon cas quanzaon cas 3 0 0 0.5.5 Slos:.5 3 3.5 4 x 0 5 Fg.. Prformanc of Algorhm 3. VI. CONCLUSION Ths papr consdrs dynamc powr allocaon n MIMO fadng sysms whou CDIT. In h cas of dal CSIT, h proposd dynamc powr polcy can approach opmaly. In h cas of dlayd and quanzd CSIT, h proposd dynamc powr allocaon polcy can achv O( ) sub-opmaly, whr s h quanzaon rror. APPENDIX A PROOF OF THEOREM Th proof mhod s an xnson of Scon 3. n [], whch gvs h srucur of h opmal ransm covaranc n drmnsc MIMO channls. No ha log d(i + HQH H ) (a) = log d(i + QH H H) (b) = log d(i + QU H U) (c) = log d(i + / UQU H / ), whr (a) and (c) follows from h lmnary dny

PROC. IEEE INFOCOM 06 log d(i n + AB) = log d(i n + BA), 8A, B C n n ; and (b) follows from h fac ha H H H = U H U. Dfn Q = UQU H, whch s smdfn posv f and only f Q s. No ha r( Q)=r(UQU H )=r(q) by h fac ha r(ab) =r(ba), 8A C m n, B C n m. Thus, problm ()-(3) s quvaln o max Q log d(i + / Q / ) Z V r( Q) () s.. r( Q) appl P () Q S N T + (3) Fac 5 (Hadamard s Inqualy, Thorm 7.8. n []). For all A S n +, d(a) appl Q n = A wh qualy f A s dagonal. Th nx clam can b provn usng Hadamard s nqualy. Clam. Problm ()-(3) has a dagonal opmal soluon. Proof: Suppos problm ()-(3) has a non-dagonal opmal soluon gvn by marx Q. Consdr a dagonal marx Q b whos nrs ar dncal o h dagonal nrs of Q. No ha r( Q) b = r( Q). To show Q b s a soluon no wors han Q, suffcs o show ha log d(i + / Q b / ) log d(i + / Q / ). Ths s ru bcas d(i + / Q b / ) = Q N T = ( + Q b ) = Q N T = ( + Q ) d(i + / Q / ), whr h las nqualy follows from Hadamard s nqualy. Thus, Q b s a soluon no wors han Q and hnc opmal. By Clam, w can consdr Q = = dag(,,..., NT ) and problm ()-(3) s quvaln o max s.. XN T log( + ) = N T Z V XN T (4) = X appl P (5) = 8 {,,...,N T } (6) No ha problm (4)-(6) sasfs Slar s condon. So h opmal soluon o problm (4)-(6) s characrzd by KKT condons [0]. Th rmanng par s smlar o h drvaon of h war-fllng soluon of powr allocaon n paralll channls,.g., h proof of Exampl 5. n [0]. Inroducng Lagrang mulplrs µ R + for nqualy consran P N T = appl P and = [,..., NT ] T R N T + for nqualy consrans {,,...,N T }. L = [,..., N T ] T and (µ, ) b any prmal and dual opmal pons wh zro dualy gap. By KKT condons, w hav + + µ = 8 {,,...,N T }; P N T = appl P ; µ 0; µ P N T = P = 0; 8 {,,...,N T }; 8 {,,...,N T }; = 8 {,,...,N T }. Elmnang, 8 {,,...,N T } n all quaons ylds µ +, 8 {,,...,N T }; P N T = appl P ; µ 0; µ P N T = P = 0; 8 {,,...,N T };(µ + ) = 8 {,,...,N T }. For all {,,...,N T }, w consdr µ < and µ sparaly: ) If µ <, hn µ + holds only whn > whch by (µ +Z/V + ) mpls ha µ + =.., =. ) If µ, hn > 0 s mpossbl, bcaus > 0 mpls ha µ + > whch oghr wh > 0 conradc h slacknss condon (µ + ) =0. Thus, f µ, w mus hav =0. Summarzng boh cass, w hav = max, 8 {,,...,NT }, whr µ s chosn such ha P n = appl P, µ 0 and µ P N T = P =0. To fnd such µ, w frs chck f µ = 0. If µ = 0 s ru, h slacknss condon µ P N T = = 0 s guarand o hold and w nd o furhr rqur P N T P = = NT = max appl P. Thus µ =0f and only f P N T = max Z/V appl P. Thus, Algorhm chcks f P N T = max Z/V appl P holds a h frs sp and f hs s ru, hn w conclud µ =0and w ar don! Ohrws, w know µ > 0. By h slacknss condon µ P N T = P = w mus hav P N T = = P NT = max = P. To fnd µ > 0 such ha P NT = max = P, w could apply a bscon sarch by nong ha all ar dcrasng wh rspc o µ. Anohr algorhm of fndng µ s nsprd by h obsrvaon ha f j k, 8j, k {,,...,N T }, hn j k. Thus, w frs sor all n a dcrasng ordr, say s h prmuaon such ha () () (N T ); and hn squnally chck f {,,...,N T } s h ndx such ha () µ 0 and (+) µ appl 0. To chck hs, w frs assum s ndd such an ndx and solv h quaon = P o oban µ ; (No ha n P j= (j) Algorhm, o avod rcalculang h paral sum P j= (j) for ach, w nroduc h paramr S = P j= and (j) upda S ncrmnally. By dong hs, h complxy of ach raon n h loop s only O().) hn vrfy h assumpon by chckng f 0 and () appl 0. (+) Ths algorhm s dscrbd n Algorhm. APPENDIX B PROOF OF THEOREM 3 Clam. If b s an opmal soluon o h followng convx program: mn k k F (7) s.. r( ) appl P (8) S N T + (9) hn b Q = U H b U s an opmal soluon o problm (6)-(8).

PROC. IEEE INFOCOM 06 Proof: Ths clam can b provn by conradcon. L b b an opmal soluon o convx program (7)-(9) and dfn Q b = U H U. b Assum ha hr xss Q S N T + such ha Q 6= Q b and s a soluon o problm (6)-(8) ha s srcly br han Q. b Consdr = UQU H and rach a conradcon by showng s srcly br han b as follows: No ha r( )=r(u QU H )=r( Q) appl P, whr h las nqualy follows from h assumpon ha Q s soluon o problm (6)-(8). Also no ha S N T + snc Q S N T +. Thus, s fasbl o problm (7)-(9). No ha k (a) k F = ku H U U H (b) Uk F = kq (c) Xk F < kq b (d) Xk F = kuqu b H UXU H () k F = k b k F, whr (a) and (d) follow from h fac Frobnus norm s unary nvaran 4 ; (b) follows from h fac ha = UQU H and X = U H U; (c) follows from h fac ha Q s srcly br han Q; b and () follows from h fac ha Q b = U H U b and X = U H U. Thus, s srcly br han. b A conradcon! Clam 3. Th opmal soluon o problm (7)-(9) mus b a dagonal marx. Proof: Ths clam can b provn by conradcon. Assum ha problm (7)-(9) has an opmal soluon ha s no dagonal. Snc s posv smdfn, all h dagonal nrs of ar non-ngav. Dfn b as a dagonal marx whos h -h dagonal nry s qual o h -h dagonal nry of for all {,,...,N T }. No ha r( b )=r( ) appl P and b S n +. Thus, b s fasbl o problm (7)-(9). No ha k b k F < k k F snc s dagonal. Thus, b s a soluon srcly br han. A conradcon! So h opmal soluon o problm (7)-(9) mus b a dagonal marx. By h abov wo clams, suffcs o assum ha h opmal soluon o problm (6)-(8) has h srucur ˆQ = U H U, whr s a marx wh non-ngav nrs. To solv problm (6)-(8), suffcs o consdr h followng convx program. mn s.. XN T ( ) (30) = XN T appl P (3) = 8 {,,...,N T } (3) No ha problm (30)-(3) sasfs Slar s condon. So h opmal soluon o problm (30)-(3) s characrzd by KKT condons [0]. Th rmanng par s smlar o h proof of Thorm and can b found n h full vrson [7]. REFERENCES [] S. K. Jayawra and H. V. Poor, Capacy of mulpl-annna sysms wh boh rcvr and ransmr channl sa nformaon, IEEE Transacons on Informaon Thory, vol. 49, no. pp. 697 709, 003. [3] S. A. Jafar, S. Vshwanah, and A. Goldsmh, Channl capacy and bamformng for mulpl ransm and rcv annnas wh covaranc fdback, n Procdngs of IEEE Inrnaonal Confrnc on Communcaons (ICC), 00. [4] E. Jorswck and H. Boch, Channl capacy and capacy-rang of bamformng n MIMO wrlss sysms undr corrlad fadng wh covaranc fdback, IEEE Transacons on Wrlss Communcaons, vol. 3, no. 5, pp. 543 553, 004. [5] V. V. Vravall, Y. Lang, and A. M. Sayd, Corrlad MIMO wrlss channls: capacy, opmal sgnalng, and asympocs, IEEE Transacons on Informaon Thory, vol. 5, no. 6, pp. 058 07, 005. [6] M. J. Nly, Dynamc powr allocaon and roung for sall and wrlss nworks wh m varyng channls, Ph.D. dssraon, Massachuss Insu of Tchnology, 003. [7], Sochasc nwork opmzaon wh applcaon o communcaon and quung sysms. Morgan & Claypool Publshrs, 0 vol. 3, no.. [8] J.-W. L, R. R. Mazumdar, and N. B. Shroff, Opporunsc powr schdulng for dynamc mul-srvr wrlss sysms, IEEE Transacons on Wrlss Communcaons, vol. 5, no. 6, pp. 506 55, 006. [9] A. Rbro, Ergodc sochasc opmzaon algorhms for wrlss communcaon and nworkng, IEEE Transacons on Sgnal Procssng, vol. 58, no., pp. 6369 6386, 00. [0] Y. Hu and A. Rbro, Adapv dsrbud algorhms for opmal random accss channls, IEEE Transacons on Wrlss Communcaons, vol. no. 8, pp. 703 75, 0. [] J. Lu, Y. T. Hou, Y. Sh, and H. D. Shral, On prformanc opmzaon for mul-carrr mmo ad hoc nworks, n Procdngs of h 0h ACM nrnaonal symposum on Mobl ad hoc nworkng and compung (MobHoc), 009, pp. 43 54. [] M. Znkvch, Onln convx programmng and gnralzd nfnsmal gradn ascn, n Procdngs of h 0h Inrnaonal Confrnc on Machn Larnng (ICML), 003. [3] N. Buchbndr, L. Lwn-Eyan, I. Mnach, J. S. Naor, and A. Orda, Dynamc powr allocaon undr arbrary varyng channls-an onln approach, n Procdngs of IEEE Annual Jon Confrnc of h IEEE Compur and Communcaons Socs (INFOCOM), 009. [4], Dynamc powr allocaon undr arbrary varyng channls h mul-usr cas, n Procdngs of IEEE Annual Jon Confrnc of h IEEE Compur and Communcaons Socs (INFOCOM), 00. [5] I. Sakogannaks, P. Mrkopoulos, and C. Toua, Adapv powr allocaon and conrol n m-varyng mul-carrr MIMO nworks, arxv prprn arxv:503.055, 05. [6] A. Goldsmh, Wrlss communcaons. Cambrdg Unvrsy Prss, 005. [7] H. Yu and M. J. Nly, Dynamc powr allocaon n MIMO fadng sysms whou channl dsrbuon nformaon, arxv prprn arxv:5.0849, 05. [8] A. Fn, S. Hanly, and R. Mahar, Drvavs of muual nformaon n gaussan vcor channls wh applcaons, n Procdngs of IEEE Inrnaonal Symposum on Informaon Thory (ISIT), 007, pp. 96 300. [9] D. P. Palomar and M. A. Lagunas, Jon ransm-rcv spac-m qualzaon n spaally corrlad MIMO channls: A bamformng approach, IEEE Journal on Slcd Aras n Communcaons, vol., no. 5, pp. 730 743, 003. [0] S. Boyd and L. Vandnbrgh, Convx Opmzaon. Cambrdg Unvrsy Prss, 004. [] S. Boyd and L. Xao, Las-squars covaranc marx adjusmn, SIAM Journal on Marx Analyss and Applcaons, vol. 7, no., pp. 53 546, 005. [] R. A. Horn and C. R. Johnson, Marx Analyss. Cambrdg Unvrsy Prss, 985. [] I. E. Tlaar, Capacy of mul-annna gaussan channls, Europan ransacons on lcommuncaons, vol. no. 6, pp. 585 596, 999. 4 Tha s kauk F = kak F for all A C n n and all unary marx U.