Book Title: Orthogonal Frequency Division Multiple Access. Editors

Size: px
Start display at page:

Download "Book Title: Orthogonal Frequency Division Multiple Access. Editors"

Transcription

1 Book Ttle: Orthogonal Frequency Dvson Multple Access Edtors August 18, 2009

2

3 Contents 1 Schedulng and Resource Allocaton n OFDMA Introducton Related Work on OFDMA resource allocaton OFDMA Schedulng and Resource Allocaton Gradent-based Wreless Schedulng and Resource Allocaton Problem Formulaton General OFDMA rate regons Optmal Algorthms Prmal optmal soluton OFDMA Feasblty Power allocaton gven subchannel allocaton Low Complexty Suboptmal Algorthms CA n SOA1: Progressve Subchannel Allocaton Based on Metrc Sortng CA n SOA2: tone Number Assgnment & tone User Matchng

4 CONTENTS Power Allocaton (PA) phase Complexty and performance of Suboptmal Algorthms for the Uplnk Scenaro Conclusons and Open Problems Acknowledgement

5 Chapter 1 Schedulng and Resource Allocaton n OFDMA Wreless Systems Janwe Huang, Vjay Subramanan, Randall Berry, and Rajeev Agrawal Dynamc schedulng and resource allocaton are key components of emergng broadband wreless standards based on Orthogonal Frequency Dvson Multple Access (OFDMA). However, schedulng and resource allocaton n an OFDMA system s complcated due to the dscrete nature of channel assgnments and the heterogenety of the users channel condtons, applcaton requrements, and constrants. In ths chapter, we provde a framework for jont schedulng and resource allocaton for OFDMA communcatons systems that operate n an nfrastructure/cellular mode, such as IEEE (WMax) and 3GPP LTE. Ths framework, whch ncludes both uplnk and downlnk resource allocaton problems as specal cases, assumes a (centralzed) scheduler per access pont/base staton that determnes the assgnment of OFDMA tones to users as well as the allocaton of power across these tones, based on the avalable channel qualty feedback. Physcal layer resources are allocated n each tme slot to maxmze the projecton of the users rates onto the gradent of a total system utlty functon that models the applcaton-layer Qualty of Servce (QoS). Although the optmzaton problem at every schedulng nstance s a mxed nteger and nonlnear op- 1

6 2 CHAPTER 1. SCHEDULING AND RESOURCE ALLOCATION IN OFDMA tmzaton problem, we show that ts optmal soluton can often be acheved by solvng a related convex optmzaton problem usng the Lagrangan dual. In general, the resultng optmal algorthms have hgh complexty, but they provde ntutons that enable the desgn of low complexty heurstc algorthms that acheve close to optmal performance n smulatons. All algorthms take nto account many ssues and constrants encountered n practcal OFDMA systems. 1.1 Introducton Schedulng and resource allocaton are essental components of wreless data systems. Here by schedulng we refer the problem of determnng whch users wll be actve n a gven tme-slot; resource allocaton refers to the problem of allocatng physcal-layer resources such as bandwdth and power among these actve users. In modern wreless data systems, frequent channel qualty feedback s avalable enablng both the scheduled users and the allocaton of physcal layer resources to be dynamcally adapted based on the users channel condtons and qualty of servce (QoS) requrements. Ths has led to a great deal of nterest both n practce and n the research communty on varous channel aware schedulng and resource allocaton algorthms. Many of these algorthms can be vewed as gradent-based algorthms, whch select the transmsson rate vector that maxmzes the projecton onto the gradent of the system s total utlty [1 4,8,9,25,28,29]. One example s the proportonally far rule [3,4] frst proposed for CDMA 1xEVDO based on a logarthmc utlty functon of each user s throughput. A larger class of throughput-based utltes s consdered n [2] where effcency and farness are allowed to be traded-off. The Max Weght polcy (e.g. [6 8]) can also be vewed as a gradent-based polcy, where the utlty s now a functon of a user s queue-sze or delay. Compared to TDMA and CDMA technologes, OFDMA dvdes the wreless resource nto non-overlappng frequency-tme chunks and offers more flexblty for resource allocaton. It has many advantages such as robustness aganst ntersymbol nterference and multpath fadng as well as and lower complexty of recever equalzaton. Owng to these OFDMA has

7 1.2. RELATED WORK ON OFDMA RESOURCE ALLOCATION 3 been adopted the core technology for most recent broadband wreless data systems, such as IEEE (WMAX), IEEE a/g (Wreless LANs), and LTE for 3GPP. Ths chapter dscusses gradent-based schedulng and resource allocaton n OFDMA systems. Ths bulds on prevous work specfc to the sngle cell downlnk [28] and uplnk [25] settng (e.g., Fg. 1.1). The key contrbuton of the book chapter s provdng a general framework that ncludes each of these as specal cases and also apples to multple cell/sector downlnk transmssons (e.g., Fg. 1.2). In partcular, several mportant practcal constrants are ncluded n ths framework, namely, 1) nteger constrants on the tone allocaton,.e., a tone can be allocated to at most one user; 2) constrants on the maxmum SNR (.e., rate) per tone, whch models a lmtaton on the avalable modulaton and codng schemes; 3) self-nose on tones due to channel estmaton errors (e.g., [11]) or phase nose [24]; and 4) user-specfc mnmum and maxmum rate constrants. We not only provde the optmal algorthm for solvng the optmzaton problem correspondng to the generalzed model, but also provde low complexty heurstc algorthms that acheve close to optmal performance. Most prevous work on OFDMA systems focused on solvng the resource allocaton problem wthout jontly consderng the problem of user schedulng. We wll brefly survey ths work n the next secton. Then we descrbe our general formulaton together wth the optmal and heurstc algorthms to solve the problem. Fnally, we wll summarze the chapter and outlne some future research drectons. 1.2 Related Work on OFDMA resource allocaton A number of formulatons for sngle cell downlnk OFDMA resource allocaton have been studed (e.g., [12 21]). In [13, 14], the goal s to mnmze the total transmt power gven target bt-rates for each user. In [14], the target bt-rates are determned by a far queueng algorthm, whch does not take nto account the users channel condtons. A number of papers ncludng [15 18, 20, 21] have studed varous sum-rate maxmzaton problems, gven a total power constrant. In [16 18] there s also a mnmum bt-rate per user that must be met. [21] consders both mnmum and maxmum rate targets for each user and also takes

8 4 CHAPTER 1. SCHEDULING AND RESOURCE ALLOCATION IN OFDMA Base Staton Base Staton User 1 User 1 User K User K User 2 User 2 Sngle Cell Downlnk Communcatons Sngle Cell Uplnk Communcatons Fgure 1.1: Example of a sngle cell downlnk (left) and uplnk scenero (rght). User 1 User 1 User 2 Base Staton 2 User 2 Base Staton 1 User K2 User K1 Base Staton 3 User K3 User 1 User 2 Fgure 1.2: Example of a multple cell/sector downlnk scenero (dfferent base statons could represent dfferent sectors of the same base staton shown by the crcle).

9 1.2. RELATED WORK ON OFDMA RESOURCE ALLOCATION 5 nto account several constrants that arse n Moble WMax. In [20], certan delay senstve users are modeled as havng fxed target bt-rate (.e. ther maxmum and mnmum rates are the same), whle other best effort users have no bt-rate constrants. Thus the scheduler attempts to maxmze the sum-rate of the best effort users whle meetng the rate-targets of the delay senstve ones. In [12,19], weghted sum-rate maxmzaton s consdered. Ths s a specal case of the resource allocaton problem we study here for a gven tme-slot but does not account for constrants on the SNR per carrer, rate constrants, or self-nose. In [12], a suboptmal algorthm wth constant power per tone was shown n smulatons to have lttle performance loss. Other heurstcs that use a constant power per tone are gven n [15 17]; we wll brefly dscuss a related approach n Secton 1.4. In [19], a dual-based algorthm smlar to ours s consdered, and smulatons are gven whch show that the dualty gap of ths problem quckly goes to zero as the number of tones ncreases. In [22], the nformaton theoretc capacty regon of a sngle cell downlnk broadcast channel wth frequency-selectve fadng usng a TDM scheme s gven; the feasble rate regon we consder, wthout any maxmum SNR and rate constrants, can be vewed as a specal case of ths regon. None of these papers consder self-nose, rate constrants or per user SNR constrants. Moreover, most of these papers optmze a statc objectve functon, whle we are nterested n a dynamc settng where the objectve changes over tme accordng to a gradent-based algorthm. It s not a pror clear f a good heurstc for a statc problem appled to each tme-step wll be a good heurstc for the dynamc case, snce the optmalty result n [1 3, 6 8, 29] s predcated on solvng the weghted-rate optmzaton problem exactly n each tme-slot. Smulaton results n [28] show that ths does hold for the heurstcs presented n Secton 1.4. Resource allocaton for a sngle cell OFDMA uplnk has been presented n [32 39]. In [32], a resource allocaton problem was formulated n the framework of Nash Barganng, and an teratve algorthm was proposed wth relatvely hgh complexty. The authors of [33] proposed a heurstc algorthm that tres to mnmze each user s transmsson power whle satsfyng the ndvdual rate constrants. In [34], the author consdered the sum-rate maxmzaton problem, whch s a specal case of the problem consdered here wth equal weghts. The algorthm derved n [34] assumes Raylegh fadng on each subchannel; we do not make such an assumpton here. In [35], an uplnk problem wth multple antennas at the base staton was consdered; ths enables spatal multplexng of subchannels among

10 6 CHAPTER 1. SCHEDULING AND RESOURCE ALLOCATION IN OFDMA multple users. Here, we focus on sngle antenna systems where at most one user can be assgned per sub-channel. The work n [36 39] s closer to our model. The authors n [36] also consdered a weghted rate maxmzaton problem n the uplnk case, but assumed statc weghts. They proposed two algorthms, whch are smlar to one of the algorthms descrbed n ths chapter. We propose several other algorthms that outperform those n [36] wth smlar or slghtly hgher complexty. Paper [37] generalzed the results n [36] by consderng utlty maxmzaton n one tme-slot, where the utlty s a functon of the nstantaneous rate n each tme-slot. Another work that focused on per tme-slot farness s [39]. Fnally, [38] proposed a heurstc algorthm based on Lagrangan relaxaton, whch has hgh complexty due to a subgradent search of the dual varables. Resource allocaton and nterference management of mult-cell downlnk OFDMA systems were presented n [42 49]. A key focus of these works s on nterference management among multple cells. Our general formulaton ncludes the case where resource coordnaton leads to no nterference among dfferent cells/sectors/stes. In our model, ths s acheved by dynamcally parttonng the subchannels across the dfferent cells/sectors/stes. In addton to beng easer to mplement, the nterference free operaton assumed n our model allows us to optmze over a large class of achevable rate regons for ths problem. If the nterference strength s of the order of the sgnal strength, as would be typcal n the broadband wreless settng, then ths parttonng approach could also be the better opton n an nformaton theoretc sense [31]. 1 1 We note that our dscussons do not drectly apply to the case of frequency reuse, where dfferent non-adjacent cells may use the same frequency bands. In practce, frequency reuse s typcally consdered together wth fxed frequency allocatons, whle here we consder dynamc frequency allocatons across dfferent cells.

11 1.3. OFDMA SCHEDULING AND RESOURCE ALLOCATION OFDMA Schedulng and Resource Allocaton Gradent-based Wreless Schedulng and Resource Allocaton Problem Formulaton Let us consder a network wth a total of K users. In each tme-slot t, the schedulng and resource allocaton decson can be vewed as selectng a rate vector r t = (r 1,t,..., r K,t ) from the current feasble rate regon R(e t ) R K +. If a user s not scheduled hs rate s smply zero. Here e t ndcates the tme-varyng channel state nformaton of all users avalable at the scheduler at tme t. The decson on the rate vector s made accordng to the gradent-based schedulng framework n [1 3, 29] that s bascally a stochastc verson of the condtonal gradent/frank-wolfe algorthm [26]. Namely, an r t R(e t ) s selected that has the maxmum projecton onto the gradent of the system s total utlty functon U(W t ) := K U (W,t ), (1.1) =1 where U ( ) s an ncreasng concave utlty functon that measures user s satsfacton for dfferent values of throughput, and W,t s user s average throughput up to tme t. In other words, the schedulng and resource allocaton decson s the soluton to max U(W t) T r t = r t R(e t) max r t R(e t) K U (W,t )r,t, (1.2) where U ( ) s the dervatve of U ( ). As a concrete example, t s useful to consder the class of commonly used so-elastc utlty functons gven n [2, 5], =1 c U (W,t ) = (W α,t) α, α 1, α 0, c log(w,t ), α = 0, (1.3)

12 8 CHAPTER 1. SCHEDULING AND RESOURCE ALLOCATION IN OFDMA where α 1 s a farness parameter and c s a QoS weght. dervatves, (1.2) becomes max r t R(e t) In ths case, after takng c (W,t ) α 1 r,t. (1.4) Wth equal class weghts (c = c for all ), settng α = 1 results n a schedulng rule that maxmzes the total throughput durng each slot. For α = 0, ths results n the proportonally far rule, and as α ncreases wthout bound, we get closer to a max-mn far soluton. Thus, ths famly of utlty functons yelds a flexble class of polces: the α parameter allows for the choce of an approprate farness objectve whle the c parameter allows one to dstngush relatve prortes wthn each farness class. However, more generally, we consder the problem of max r t R(e t) w,t r,t, (1.5) where w,t 0 s a tme-varyng weght assgned to the th user at tme t. In the case of (1.4), we let w,t = c (W,t ) α 1. In (1.4) these weghts are gven by the gradents of throughputbased utltes; however, other methods for generatng the weghts (possbly dependng upon queue-lengths and/or delays [6 8]) are also possble. We note that (1.5) must be re-solved at each schedulng nstance because of changes n both the channel state and the weghts (e.g., the gradents of the utltes). Whle the former changes are due to the tme-varyng nature of wreless channels, the latter changes are due to new arrvals and past servce decsons General OFDMA rate regons The soluton to (1.5) depends on the channel state dependent rate regon R(e), where we suppress the dependence on tme for smplcty. We consder a model approprate for general OFDMA systems ncludng sngle cell downlnk and uplnk as well as multple cell/sector/ste downlnk wth frequency sharng; related sngle cell downlnk and uplnk models have been consdered n [12,22,25,28]. In ths model, R(e) s parameterzed by the allocaton of tones to users and the allocaton of power across tones. In a tradtonal OFDMA system at most

13 1.3. OFDMA SCHEDULING AND RESOURCE ALLOCATION 9 one user may be assgned to any tone. Intally, as n [13, 14], we make the smplfyng assumpton that multple users can share one tone usng some orthogonalzaton technque (e.g. TDM). 2 In practce, f a schedulng nterval contans multple OFDMA symbols, we can mplement such sharng by gvng a fracton of the symbols to each user; of course, each user wll be constraned to use an nteger number of symbols. Also, wth a large number of tones, adjacent tones wll have nearly dentcal gans, n whch case ths tme-sharng can also be approxmated by frequency sharng. The two approxmatons becomes tght as the number of symbols or tones ncreases, respectvely. We dscuss the case where only one user can use a tone n Secton 1.4. Let N = {1,..., N} denote the set of tones 3 and K = {1, 2,..., K} the set of users. For each j N and user K, let e j be the receved sgnal-to-nose rato (SNR) per unt transmt power. We denote the transmt power allocated to user on tone j by p j, and the fracton of that tone allocated to user by x j. As tones are shared resources, the total allocaton for each tone j must satsfy x j 1. For a gven allocaton, wth perfect channel estmaton, user s feasble rate on tone j s ( r j = x j B log 1 + p ) je j, x j whch corresponds to the Shannon capacty of a Gaussan nose channel wth bandwdth x j B and receved SNR p j e j /x j. 4 Ths SNR arses from vewng p j as the energy per tme-slot user uses on tone j; the correspondng transmsson power becomes p j /x j when only a fracton x j of the tone bandwdth s allocated. Smlarly ths can also be explaned by tme-sharng as follows: a channel of bandwdth B s used only a fracton x j of the tme wth average power p j whch leads to the power durng channel usage to be p j /x j. Wthout loss of generalty we set B = 1 n the followng. 2 We focus on systems that do not use superposton codng and successve nterference cancellaton wthn a tone, as such technques are generally consdered too complex for practcal systems. 3 In practce, tones may be grouped nto subchannels and allocated at the granularty of subchannels. As dscussed n [28], our model can be appled to such settngs as well by approprately redefnng the sub-channel gans {e j } and nterpretng N as the set of sub-channels. 4 To better model the achevable rates n a practcal system we can re-normalze e j by γe j, where γ [0, 1] represents the system s gap from capacty.

14 10 CHAPTER 1. SCHEDULING AND RESOURCE ALLOCATION IN OFDMA Self-nose In a realstc OFDMA system, mperfect carrer synchronzaton and channel estmaton may result n self-nose (e.g. [11,24]). We follow a smlar approach as n [11] to model self-nose. Let the receved sgnal on the jth tone of user be gven by y j = h j s j + n j, where h j, s j and n j are the (complex) channel gan, transmtted sgnal and addtve nose, respectvely, wth n j CN (0, σ 2 ). 5 Assume that h j = h j + h j,δ, where h j s recever s estmate of h j and h j,δ CN (0, δj). 2 After matched-flterng, the receved sgnal wll be z j = h jy j resultng n an effectve SNR of Eff-SNR = where p j = E( s j 2 ), β j = h j 4 p j σ 2 j h j 2 + δ 2 j p j h j 2 = δ2 j and e h j 2 j = h j 2 σj 2 p je j 1 + β j p j e j, (1.6). 6 Here, β j p j e j s the self-nose term. As n the case wthout self-nose (β j = 0), the effectve SNR s stll ncreasng n p j. However, t now has a maxmum of 1/β j. In general, β j may depend on the channel qualty e j. For example, ths happens when self-nose arses prmarly from estmaton errors. The exact dependence wll depend on the detals of channel estmaton. As an example, usng the model n [23, Secton IV] t can be shown that when the plot power s ether constant or nversely proportonal to channel qualty subject to maxmum and mnmum power constrants (modelng power control), β s nversely proportonal to the channel condton for large e. On the other hand β j = β s a constant when self-nose s due to phase nose as n [24]. For smplcty of presentaton, we assume constant β j = β n the remander of the paper (except n Fg. 1.4 where we we allow β(e) 1/e to llustrate the mpact of self-nose on the optmal power allocaton). The analyss s almost dentcal f users have dfferent β j s. We assume that e j s known by the scheduler for all and j as s β. For example, n a frequency dvson duplex (FDD) downlnk system, ths knowledge can be acqured 5 We use the notaton x CN (0, b) to denote that x s a 0 mean, complex, crcularly-symmetrc Gaussan random varable wth varance b := E( x 2 ). 6 Ths s slghtly dfferent from the Eff-SNR n [11] n whch the sgnal power s nstead gven by h j 4 p j ; the followng analyss works for such a model as well by a smple change of varables. For the problem at hand, (1.6) seems more reasonable n that the resource allocaton wll depend only on h j and not on h j. We also note that (1.6) s shown n [23] to gve an achevable lower bound on the capacty of ths channel.

15 1.3. OFDMA SCHEDULING AND RESOURCE ALLOCATION 11 by havng the base staton transmt plot sgnals, from whch the users can estmate ther channel gans and feedback to the base staton. In a tme dvson duplex (TDD) system, these gans can also be acqured by havng the users transmt uplnk plots; for the downlnk case, the base staton can then explot recprocty to measure the channel gans. In both cases, ths feedback nformaton would need to be provded wthn the channel s coherence tme. Wth self-nose, user s feasble rate on tone j becomes ( r j = x j log 1 + ) ( p j e j pj e j =: x j f x j + βp j e j x j where agan x j models tme-sharng of a tone and the functon f( ) s gven by ), (1.7) ( ) 1 f(s) = log 1 +, β 0. (1.8) β + 1/s More generally, we assume that a user s rate on channel j s gven by ( pj e j r j = x j f x j ), (1.9) for some functon f : R + R + that s non-decreasng, twce contnuously dfferentable and concave wth f(0) = 0, (wthout loss of generalty) 7 f (0) := df ds (0) = lm s 0 f(s) sup s>0 f(s) s s = df = 1, and lm t + (t) = 0. We also assume by ds contnuty8 that xf(p/x) s 0 at x = 0 for every p 0. From the assumptons on the functon f( ) t follows that xf(p/x) s jontly concave n x, p; ths can be easly proved by showng that the Hessan s negatve semdefnte [26, 27]. It s easy to verfy that f gven by (1.8) satsfes the above propertes. We should, however, pont out that usng the theory of subgradents [26, 27], our mathematcal results easly extend to a general f( ) that s only non-decreasng and concave. For nstance, t can be easly proved from frst prncples that xf(p/x) s jontly 7 Usng the dea that Shannon capacty log(1 + s) s a natural upper bound for f(s), t follows that 0 < df (0) 1. Therefore, ds f f (0) 1, then we can solve the problem usng a scaled verson of functon,.e., f(s) = f(s)/ df (0), after scalng the rate ds constrants by the same amount; the power and subchannel allocatons wll be the same n the two cases. The Shannon capacty df upper bound also yelds that 0 lm t + ds (t) lm s + f(s) s mply that df f(t) (t) for all t > 0. ds t lm s + log(1+s) s = 0, as concavty of f( ) and f(0) = 0 8 Usng the Shannon capacty functon, log(1 + s), upper bound, we have for p > 0, that lm x 0 xf(p/x) = p lm t + f(t) t p lm t + log(1+t) t = 0. For p = 0, we drectly get the property from f(0) = 0.

16 12 CHAPTER 1. SCHEDULING AND RESOURCE ALLOCATION IN OFDMA concave n (x, p) f f( ) s merely concave. We conscously choose the smpler settng of twce contnuously dfferentable functons to keep the level of dscusson smple, but to ad a more nterested reader, we wll strve to pont out the loosest condtons needed for each of our results. Before proceedng we should pont out that, operatonally, f( ) s a functon of the receved sgnal-to-nose rato, and thus, abstracts the usage of all possble sngle-user decoders, ncludng the optmal decoder that yelds Shannon capacty. General power constrant - sngle cell downlnk, uplnk and mult-cell downlnk wth frequency sharng Let {K m } M m=1 be non-empty subsets of the set of users K that form a coverng,.e., M m=1k m = K. We assume that there s a vector of non-negatve power budgets {P m } M m=1 assocated wth these subsets, so that K m j p j P m for each m. Ths condton ensures that there s no user who s unconstraned n ts power usage. Ths provdes a common formulaton of the sngle cell downlnk and uplnk schedulng problems as descrbed n [28] and [25], respectvely. For the sngle cell downlnk problem M = 1 and K 1 = K, and for the sngle cell uplnk problem M = K and K = {} for K. More generally, f {K m } M m=1 s a partton,.e., mutually dsjont, then we can vew the transmtters for users K m as colocated wth a sngle power amplfer. For example, such a model may arse n the downlnk case where M := {1, 2,..., M} represents sectors or stes across whch we need to allocate common frequency/channel resources, but whch have ndependent power budgets. A key assumpton, however, s that we can make the transmssons from the dfferent sectors/stes non-nterferng by tme-sharng or by some other sutable orthogonalzaton technque.

17 1.3. OFDMA SCHEDULING AND RESOURCE ALLOCATION 13 Capacty Regon - max SNR and mn/max rate constrants Under these assumptons, the rate regon can be wrtten as R(e) = { r : r = j K m j x j f ( ) pj e j x j p j P m, m, and R mn r R max,, } x j 1, j, (x, p) X, (1.10) where X := { (x, p) 0 : x j 1, p j x js j e j }, j. (1.11) Here and n the followng, a boldfaced symbol wll ndcate the vector of the correspondng scalar quanttes, e.g. x := (x j ) and p := (p j ). Also, any nequalty such as x 0 should be nterpreted componentwse. The lnear constrant on (x j, p j ) n (1.11) usng s j models a constrant on the maxmum rate per subchannel due to a lmtaton on the avalable modulaton and codng schemes; f user can send at a maxmum rate of r j on tone j, then s j = f 1 ( r j ). We have also assumed that each user K has maxmum and mnmum rate constrants R max and R mn, respectvely. In order to have a soluton we assume that the vector of mnmum rates {R mn } K s feasble. For the vector of maxmum rates, t s more convenent to assume that {R max } K s nfeasble. Otherwse the optmzaton problem assocated wth feasblty (see Secton 1.3.5) wll yeld an optmal soluton. Typcally we wll set R mn = 0 and R max to be the (tme-varyng) buffer occupancy. However, wth tght mnmum throughput demands one can magne usng a non-zero R mn to guarantee ths.

18 14 CHAPTER 1. SCHEDULING AND RESOURCE ALLOCATION IN OFDMA Optmal Algorthms From (1.5) and (1.10), the optmal schedulng and resource allocaton problem can be stated as: max V (x, p) := (x,p) X ( pj e j subject to: x j f j x j w x j f ) ( ) pj e j x j f j x j j ( ) pj e j x j R mn K (η ) R max K (γ ) x j 1 j N (µ j ) p j P m m = 1, 2,..., M (λ m ) K m j (P2) where set X s gven n (1.11). As a rule, varables at the rght of constrants wll ndcate the dual varables that we wll use to relax those constrants whle constructng the dual problem later. One mportant pont to note s that as descrbed above, the optmzaton problem (P2) s not convex and so we can not appeal to standard results such as Slater s condtons to guarantee that s has zero dualty gap [26, 27]. In partcular, note that the maxmum rate constrants have a concave functon on the left sde. To show that we stll have no dualty gap, we wll consder a related convex problem n hgher dmensons that has the same prmal

19 1.3. OFDMA SCHEDULING AND RESOURCE ALLOCATION 15 soluton and the same dual. The new optmzaton problem (P1) s gven by max w r subject to: r ( ) pj e j x j f, K (α ) x j j x j 1, j N (µ j ) p j P m, m = 1, 2,..., M (λ m ) K m R mn j r R max, K (x, p) X. (P1) Ths problem s easly seen to be convex due to the jont concavty of xf(p/x) as a functon of (x, p) and also wll satsfy Slater s condton. 9 Hence, t wll have zero dualty gap [26,27]. The problem (P1) can be practcally motvated as follows: the physcal (PHY) layer gves the scheduler (at the MAC layer) a maxmum rate that t can serve per user based upon power and subchannel allocatons, and the scheduler then drans from the queue an amount that obeys the mnmum and maxmum rate constrants (mposed by the network layer) and the maxmum rate constrant from the PHY layer output. If the scheduler chooses not to use the complete allocaton gven by the PHY layer, then the fnal packet sent by the MAC layer s assumed to be constructed usng an approprate number of padded bts. However, we wll now show that at the optmal, there s no of loss optmalty n assumng that the scheduler never sends less than what the PHY layer allocates,.e., the frst constrant n Problem (P1) s always made tght at an optmal soluton. Ths pont of vew s exemplfed n schematc shown n Fgure 1.3. Assume that there s an optmzer of (P1) at whch for some user, r < j x jf( p je j x j ). We wll now construct another feasble soluton that wll satsfy the above relatonshp wth equalty. Let γ [0, 1] and set p j := γp j. Note that by convexty, both the power and subchannel constrants are satsfed for every value of γ. Now j x jf(γ p je j x j ) s a non- 9 More precsely, Slater s condton wll be satsfed provded that the mnmum rate (R mn ) are strctly n the nteror of rate-regon R(e). If R mn = 0 for all ths wll trvally be true.

20 16 CHAPTER 1. SCHEDULING AND RESOURCE ALLOCATION IN OFDMA Network Layer QoS Constrants Buffer Sze Transmsson Rate MAC Layer Scheduler Channel Qualty Measurements Maxmum Transmsson Rate Transmsson Rate Physcal Layer Fgure 1.3: Schematc of a scheduler that has cross-layer vsblty.

21 1.3. OFDMA SCHEDULING AND RESOURCE ALLOCATION 17 decreasng and contnuous functon of γ takng values 0 at γ = 0 and j x jf( p je j x j ) at γ = 1. Therefore, there exsts a γ (0, 1) such that r = j x jf(γ p j e j x j ) as desred. Ths procedure can be followed for every user for whom r < j x jf( p je j x j ), so that at the end we satsfy r = j x jf( p je j x j ) for a feasble (x, p). Therefore both the optmal value and an optmzer of problem (P1) concdes wth those for problem (P2). The loosest condton needed for the above to hold s f( ) beng non-decreasng and concave wth f(0) = 0. Henceforth, we wll only work wth Problem (P1). Before proceedng to solve the problem by dual methods, we frst defne some key notaton. For two numbers, x, y R we set x y := mn(x, y), x y := max(x, y) and (x) + = [x] + := x 0. Dual of Problem We now proceed to derve a closed-form expresson for the dual functon for problem (P1). The Lagrangan obtaned by relaxng the marked constrants of (P1) usng the correspondng dual varables s gven by L(r, x, p, α, µ, λ) = (w α )r + j µ j x j j m µ j + M ( ) pj e j α x j f x j λ m P m + m=1,j λ m p j. (1.12) The correspondng dual functon s then gven by maxmzng ths Lagrangan over r, x and p. Frst optmzng over rate r [R mn we get L(x, p, α, µ, λ) = +,j (w α ) + R max K m j, R max ] and notng that the Lagrangan s lnear n r α x j f( p je j ) x j j (α w ) + R mn + M µ j + λ m P m j m=1 µ j x j λ m p j. m K m j

22 18 CHAPTER 1. SCHEDULING AND RESOURCE ALLOCATION IN OFDMA The optmzng r s gven by the followng {R max } f α < w ; K, r {R mn } f α > w ; and [R mn, R max ] f α = w (1.13) Note that the last term of equaton (1.12) can be rewrtten as λ m p j = m j,j K m p j λ m = p jˆλ (1.14) m: K m,j where ˆλ := m: K m λ m. Now maxmzng the Lagrangan over power p requres us to maxmze ( ) pj e j α x j [f ˆλ ] p j e j x j α e j x j (1.15) over p j for each, j. From the assumptons on the functon f, t s easy to check that the maxmzng p j wll be of the form p je j x j = g ( ˆλ α e j ) s j, (1.16) for some functon g : R + [0, ] wth g(x) = 0 for x f (0). Specfcally f df/ds s monotoncally decreasng, we may show that g( ) = ( df ds) 1 ( ),.e., the nverse of the dervatve of f( ). Otherwse, snce df/ds s stll a non-ncreasng functon we can set g(x) = nf{t : df/ds(t) = x}. Usng the non-ncreasng property of df/ds we can see that g(x) y = g ( x df ds (y)). Note that we have assumed df/ds(0) = 1 and lm t + df/ds(t) = 0 but we do not assume that lm s + f(s) = + (e.g., see the self-nose example). In case f( ) s not dfferentable, then we would defne the functon g( ) usng the subgradents of f( ). In all cases, the key concluson from (1.16) s that the optmal value of p j s always a lnear functon of x j.

23 1.3. OFDMA SCHEDULING AND RESOURCE ALLOCATION β=0 β=0.01 Optmal power p j * β=10/e β= Channel condton e j (db) Fgure 1.4: Optmal power p j as a functon of the channel condton e j. Here x j = 1, α = 1, s j = +, and ˆλ = 15. ( [28] c 2009 IEEE) Note that when f = log(1 + 1 ), wth β 0, as gven by (1.8), then β+1/s g(x) = q((1/x 1) + ), where z, f β = 0, q(z) = ( ) ( ) 2β β(β+1) z 1, f β > 0. 2β(β+1) (2β+1) 2 Fgure 1.4 shows p j n (1.16) as a functon of e j for the specfc choce of f from (1.8) wth three dfferent values of β = 0, 0.01, 0.1. When β = 0, (1.16) becomes a water-fllng type of soluton n whch p j s non-decreasng n e j. For a fxed β > 0, ths s not necessarly true,.e., due to self-nose, less power may be allocated to better subchannels. We also consder the case where β = 10/e to model the case where self-nose s due to channel estmaton error.

24 20 CHAPTER 1. SCHEDULING AND RESOURCE ALLOCATION IN OFDMA Insertng the expresson for p j nto the Lagrangan yelds L(x, α, µ, λ) = (w α ) + R max (α w ) + R mn + M µ j + λ m P m j m=1 + ( ( ) ) ˆλ x j [α f g s j ˆλ ( ( ) ) ] ˆλ g s j µ j, (1.17) α,j e j e j α e j whch s now a lnear functon of {x j }. Thus, optmzng over x j yelds the dual functon for (P1), L(α, λ, µ) = +,j = + j (w α ) + R max [ α f ( g α e j ( (w α ) + R max ( ( ) ) ˆλ s j (α w ) + R mn ˆλ e j (α w ) + R mn ( ) ] ˆλ [µ j α, α e j µ j + + µ j + λ m P m j m ( ( ) ) ] ˆλ g s j µ j α e j ) + λ m P m m + µ j ), (1.18) + where ( ) ( ) µ j (a, b) := a f (g(b) ) s j b g(b) s j. Note that any choce such that {1}, f µ j (α, x j [0, 1], f µ j (α, {0}, f µ j (α, ˆλ α e j ) > µ j, ˆλ α e j ) = µ j, ˆλ α e j ) < µ j (1.19) wll optmze the Lagrangan n (1.17).

25 1.3. OFDMA SCHEDULING AND RESOURCE ALLOCATION 21 Optmzng the Dual Functon over µ From the dualty theory of convex optmzaton [26,27] the optmal soluton to problem P1 s gven by mnmzng the dual functon n (1.18) over all (α, λ, µ) 0. We do ths coordnatewse startng wth the µ varables. The followng lemma characterzes ths optmzaton. Lemma 1 For all α, λ 0, L(α, λ) := mn L(α, λ, µ) µ 0 = ( (w α ) + R max (α w ) + R mn ) + λ m P m + m j where for every tone j, the mnmzng value of µ j s acheved by µ j(α, λ), (1.20) µ j(α, λ) := max µ j ( ) ˆλ α,. (1.21) α e j The proof of Lemma 1 follows from a smlar argument as n [9]. Note that (1.21) requres searchng for the maxmum value of the metrcs µ j across all users for each tone j. Snce L(α, λ) s the mnmum of a convex functon over a convex set, t s a convex functon of (α, λ). Optmzng the Dual Functon over (α, λ) Now we are ready to optmze the remanng varables n the dual functons, namely, (α, λ). In the sngle cell downlnk case wth no rate constrants (and thus no α varables), ths reduces to a one dmensonal problem n λ and hence, t can be mnmzed usng an terated one dmensonal search (e.g., the Golden Secton method [26]). Snce there s no dualty gap, at λ = arg mn λ 0 L(λ), L(λ ) gves the optmal objectve value of problem (P1). Smlarly, n the absence of rate constrants, the multple stes/sectors problem wth a partton of the users {K m } M m=1 also leads to a one dmensonal problem wthn each partton.

26 22 CHAPTER 1. SCHEDULING AND RESOURCE ALLOCATION IN OFDMA In general, however, one would need to use subgradent methods [26, 27] to numercally solve for the optmal (α, λ). The followng lemma characterzes the set of subgradents of L(α, λ) wth respect to (α, λ). Lemma 2 About any (α 0, λ 0 ) 0, L(α, λ) d(α 0 )(α α 0 ) + m d(λ 0 m)(λ m λ 0 m), (1.22) wth d(λ m ) = P m ( ) p j = P m x j ˆλ g s j (1.23) e j α e j K m K m d(α m ) = ( ( ) ) x ˆλ jf g s j r (1.24) α j e j where x js satsfy ( x j 1 and µ j (α, λ) 1 x j ) = 0; j, and satsfy the equaton (1.19) wth µ j = µ j(α, λ) as gven n equaton (1.21), and r satsfy equaton (1.13). Thus the subgradents d(λ m ) and d(α ) are parameterzed by (r, x ) and are lnear n these varables. Moreover, the permssble values of r le n a hypercube and those of x n a smplex. Observe that the dual functon at any pont (α, λ) s obtaned by takng the maxmum of the Lagrangan over (r, p, x ) satsfyng x j 1, j N, (x, p) X. In case (r, p, x ) s unque, then the resultng Lagrangan s a gradent to the dual functon at (α, λ). In case there are multple optmzers, the resultng Lagrangans are each a subgradent, and every subgradent can be obtaned by a convex combnaton of these subgradents so that the set of subgradents s convex. The lemma follows easly by substtutng for the optmal (r, p, x ). Havng characterzed the set of subgradents, a method smlar to that used n [25] for

27 1.3. OFDMA SCHEDULING AND RESOURCE ALLOCATION 23 the sngle cell uplnk problem can be used to solve for the optmal dual varables (α, λ ) numercally. In each step of ths method we change the dual varables along the drecton gven by a subgradent subject to non-negatvty of the dual varables. The convergence of ths procedure (for a proper step-sze choce) s once agan guaranteed by the convexty of L(α, λ) (see [26, Exer ], [25]). Optmzng the dual functon over α Snce the dmenson of α equals the number of users and the dmenson of µ equals the number of tones, t may be computatonally better to optmze over α nstead of µ f the number of users s greater, and then use numercal methods to solve the problem. Next we detal the means to optmze over α before µ. The dual functon contans many terms that have defntons wth ( ) +, and therefore we would need to dentfy exactly when these terms are non-zero. For ths we need to solve a non-lnear equaton whch s guaranteed to have a unque soluton. We frst dscuss ths and then apply t to optmzng the dual functon over α. Gven y, z 0, defne by v(y, z) the unque soluton wth 1 x < + to ( ( 1 xf g x df )) ( 1 (z) g ds x df ) ds (z) = y, ( ( ) ) ( ) 1 where t s easy to show that xf g df (z) 1 g df (z) s a monotoncally ncreasng x ds x ds functon takng value 0 at x = 1 and ncreasng wthout bound as x +. If y ( ) f(z)/(df/ds(z)) z 0, then v(y, z) = (y + z)/f(z) where t s easy to verfy that v(y, z) z/f(z) 1/(df/ds(z)) 1/(df/ds(0)) = 1 from the concavty of f( ) and from f(0) = 0. Otherwse we need to solve for the unque 1 x 1/(df/ds(z)) such that ( ( 1 ) ) ( 1 xf g g = y. x x)

28 24 CHAPTER 1. SCHEDULING AND RESOURCE ALLOCATION IN OFDMA ( µj e For our results we wll be nterested n v j, s ˆλ j ), usng whch we also defne ν j := ˆλ v ( µj e j ˆλ, s j ) s j ˆλ e j and ζ j := µ j + e j f(s j ), where ν j = ζ j f µ je j ˆλ f(s j) df(s j ) ds s j. Frst note that we can rewrte the functon n (1.18) as follows L(α, µ, λ) = j µ j + m λ m P m + L, where L = (w α ) + R max + [ ˆλ α e j f e j j ˆλ (α w ) + R mn ( ( ) ) ˆλ g s j α e j ( ( ) ) ] ˆλ g s j µ je j. α e j ˆλ + Now usng the quanttes defned earler n ths secton, one can wrte L as follows L = j ˆλ e j [ 1 {0 α ζ j } 1 {ζj <α ν j } + (w α ) + R max ( α e j f(s j ) s j µ ) je j + ˆλ ˆλ ( ( α e ( j ˆλ )) )] ( ) ˆλ f g g µ je j ˆλ α e j α e j ˆλ (α w ) + R mn. Mnmzng L over α 0 can now be accomplshed by a smple one dmensonal search; we defne the optmal vector of α s to be α (λ, µ). Thereafter one would need to use a subgradent method [25, 26] to numercally mnmze over (µ, λ). A subgradent of L wth respect to λ m s gven by P m K m p j where p j s taken from (1.16) where one substtutes x j from (1.19). A subgradent of L wth respect to µ j s gven by 1 x j where we substtute for x j from (1.19). Note, however, that t s mportant that we also meet the

29 1.3. OFDMA SCHEDULING AND RESOURCE ALLOCATION 25 followng constrants for all, namely, R mn x jf ( p ) j x j f α < w, then j f α > w, then j x jf x jf R max ; ( p ) j x j ( p ) j x j = R max ; and = R mn. The proof of ths follows by retracng the steps of the proof of Lemma 2 wth the roles of α and µ beng swtched Prmal optmal soluton For the general OFDMA problem we presented two methods to solve for V : n the frst method we showed how to characterze the dual varables µ(α, λ) and then we proposed numercally solvng for the optmal (α, λ ) usng subgradent methods, whle n the second method followed the same strategy after swtchng the roles of µ and α. However, we stll need to solve for the values of the correspondng optmal prmal varables. Concentratng on the frst method, we know by dualty theory [26] that gven (α, λ ) we need to fnd one vector from the set of (r, x, p ) that also satsfes prmal feasblty and complementary slackness. These constrants can easly be seen to translate to the followng: d(λ m) 0, d(λ m)λ m = 0, m; (1.25) d(α ) 0, d(α )α = 0,. (1.26) From the lnearty of d(λ m), d(α ) n (r, x ) t follows that the prmal optmal (r, x, p) are the soluton of a lnear program n (r, x ). For the sngle cell downlnk case wth no rate constrants, as we have prevously noted searchng for the dual optmal s a one dmensonal numercal search n λ. In that case, the search for prmal optmal soluton turns out to have addtonal structure as shown n [28].

30 26 CHAPTER 1. SCHEDULING AND RESOURCE ALLOCATION IN OFDMA OFDMA Feasblty Next we turn to the correspondng feasblty problem, whch can be stated as: V = mn σ (1.27) subject to: R x j f( p je j ), (α ) x j j x j 1 j (µ j ) p j σ m (λ m ) P m K m j (x, p) X. The vector of rates (R ) s feasble f V 1,.e., all the power constrants wll also be satsfed by a vector (x, p ). As mentoned earler, we need to check that (R ) = (R mn ) s ndeed feasble; otherwse problems (P1) and (P2) are both nfeasble as well. Moreover, f (R ) = (R max ) s also feasble, then r = (R max ) s the optmzer for problems (P1) and (P2). In whch case, the optmal soluton to the problem above wth (R ) = (R max ) wll also yeld an optmal soluton to the schedulng problem. Observe that problem (1.27) s convex and satsfes Slater s condtons. Fnally, we also note that other alternate formulatons of the feasblty problem are possble where one could ether apply the σ constrant also on the subchannel utlzaton or swtch the roles of subchannel and power utlzaton. All of these wll yeld the same concluson about feasblty although the actual solutons, n terms of (x, p ), would possbly be dfferent. The Lagrangan consderng the marked constrants s L(σ, x, p, α, µ, λ) = σ ( + j 1 m λ m ) µ j x j j µ j + α R j ( ( ) ) pj e j α x j f + p j λ x j where λ := m: K m λ m Pm. As before, mnmzng over p j yelds p j e j x j = g ( λ α e j ) s j.

31 1.3. OFDMA SCHEDULING AND RESOURCE ALLOCATION 27 Substtutng ths n the Lagrangan, we get L(σ, x, α, µ, λ) = α R ( µ j + σ 1 ) λ m j m [ x j α f(g( ˆλ ) s j ) ˆλ (g( ˆλ ] ) s j ) µ j. α,j e j e j α e j Mnmzng over 0 x j 1 yelds L(σ, α, µ, λ) = L j µ j + σ ( 1 m λ m ) where L = α R j [ α f(g( ˆλ α e j ) s j ) ˆλ e j (g( ˆλ α e j ) s j ) µ j ] +. Next we mnmze L over all values of σ. Snce there are no constrants on σ, t follows that the resultng L s fnte only when m λ m = 1; for all other values we would get L =. Hereafter we wll assume that m λ m = 1. Thus L(σ, x, α, µ, λ) = L j µ j. Note that as before, as a functon of α the problem s now separable. Therefore we only need to maxmze L over α 0. Smlarly we can wrte L as L(σ, x, α, µ, λ) = j ˆL j + α R, where we have ( ˆL j = µ j + [ α f(g( ˆλ α e j ) s j ) ˆλ e j (g( ˆλ α e j ) s j ) µ j ] + )

32 28 CHAPTER 1. SCHEDULING AND RESOURCE ALLOCATION IN OFDMA As a functon of µ j the problem s now separable, and we only need to maxmze ˆL j over µ 0. Thus, we could optmze frst over ether µ or α, once agan based upon whether the number of users or subchannels s smaller. In ether case, the methodology and the functons that appear are very smlar to the correspondng problem n the schedulng problem (P1), and due to space constrants we do not elaborate on ths. Care must be take, however, whle evaluatng subgradents wth respect to λ. Here we propose usng a projected gradent method [26, 27] based upon the constrant m λ m = 1 to numercally solve for the optmal λ Power allocaton gven subchannel allocaton In many of the suboptmal schedulng algorthms that we wll dscuss, a central feature wll be a computatonally smpler (but stll close to optmal) method to provde a subchannel allocaton. Once the subchannel allocaton has been made, all that wll reman s the power allocaton problem, subject to the varous constrants that we dscussed earler. Here we dscuss how ths can be solved n an optmal manner. A smlar queston can also be asked about the feasblty problem, hence we also dscuss ths here. In all cases, we assume that we are gven a feasble subchannel allocaton. Snce we are gven a feasble subchannel allocaton x, the Lagrangan of the new schedulng problem (power allocaton only) can be easly derved by settng µ = 0. For ths we once agan use the formulaton based upon Problem (P1). The optmal power allocaton s then gven by p j = x j s e j ( g ( ) ˆλ sj α e j ). The Lagrangan that results from substtutng ths formula L(x, α, λ) = λ m P m + (w α ) + R max (α w ) + R mn m + ( ( ) ) ˆλ α x j f g s j ˆλ ( x ( ) j ˆλ g s j ). e j j α e j e j α

33 1.4. LOW COMPLEXITY SUBOPTIMAL ALGORITHMS 29 Now t s easy to argue that f R mn = 0 and R max = + and f the K m s form a partton, then wthn each partton the λ m s can be solved for as n Secton In any case, n ths settng solvng for the optmal α 0 s easer, but uses some of the functons descrbed at the end of Secton However, after ths step we would stll need to solve for λ numercally; f the parttons assumpton holds, then t would only need a sngle dmensonal search wthn each partton. A fnte-tme algorthm for achevng the optmal λ has been gven n [25, 28] under the assumpton that f( ) represents the Shannon capacty as n (1.8) wth β = 0. Feasblty check Under the assumpton that a feasble subchannel allocaton has already been provded, even the feasblty check problem becomes a lot easer. As before we can assume m λ m = 1, and that the optmal power allocaton s gven by p j = x j ths we get e j ( g ( ) λ e j α sj ), and substtutng L(x, α, λ) = α ˆR j ( ( ) ) λ x j [α f g s j λ ( ( ) ) ] λ g s j. e j α e j e j α Agan solvng for the optmal α s smpler. Once agan the λ vector would need to be computed numercally, subject to t beng a probablty vector,.e., m λ m = 1 and λ m 0 for each m. 1.4 Low Complexty Suboptmal Algorthms wth Integer Channel Allocaton There are two shortcomngs wth usng the optmal algorthm outlned n the prevous secton for schedulng and resource allocaton: () the complexty of the algorthm n general s not computatonally feasble for even moderate szed systems; () the soluton found may requre a tme-sharng channel allocaton, whle practcal mplementatons typcally requre a sngle

34 30 CHAPTER 1. SCHEDULING AND RESOURCE ALLOCATION IN OFDMA user per sub-channel. One way to address the second pont s to frst fnd the optmal prmal soluton as n the prevous secton and then project ths onto a nearby nteger soluton. Such an approach s presented n [28] for the case of a sngle cell downlnk system (M = 1) wthout any rate constrants. In that settng, after mnmzng the dual functon over µ, one optmzes the functon L(λ), whch only depends on a sngle varable. Ths functon wll have scalar subgradents whch can then be used to develop rules for mplementng such an nteger projecton. Moreover, n ths case snce L(λ) s a one-dmensonal functon the search for the optmal dual values s greatly smplfed. However, n the general settng, ths type of approach does not appear to be promsng. 10 In ths secton we dscuss a famly of sub-optmal algorthms (SOA s) for the general settng that try to reduce the complexty of the optmal algorthm, whle sacrfcng lttle n performance. These algorthms seek to explot the problem structure revealed by the optmal algorthm. Furthermore, all of these sub-optmal algorthms enforce an nteger tone allocaton durng each schedulng nterval. In the followng we consder the general model from Secton wth the restrcton that {K m } forms a partton of the user groups (.e. each user s n only one of these sets) and that R mn of these assumptons wll be true. = 0 for all. In a typcal settng both In the optmal algorthm, gven the optmal λ and α, the optmal tone allocaton up to any tes s determned by sortng the users on each tone accordng to the metrc µ j (α, (cf. (1.19)). ˆλ α e j ) Gven an optmal tone allocaton, the optmal power allocaton s gven by (1.16). In each SOA, we use the same two phases wth some modfcatons to reduce the complexty of computng (λ, α) and the optmal tone allocaton. Specfcally, we begn wth a subchannel Allocaton (CA) phase n whch we assgn each tone to at most one user. We consder two dfferent SOAs that mplement the CA phase dfferently. In SOA1, nstead of usng the metrc gven by the optmal λ and α, we consder metrcs based on a constant power allocaton over all tones assgned to a partton. In SOA2, we fnd the tone allocaton, once agan through a dual based approach, but here we frst determne the number of tones assgned to each user and then match specfc tones and users. In all cases we assgn the tones to dstnct parttons whch wll, n turn, yeld an nterference-free operaton. After the 10 See [25] for a more detaled dscusson of ths n the context of the uplnk scenaro.

35 1.4. LOW COMPLEXITY SUBOPTIMAL ALGORITHMS 31 tone allocaton s done n both SOAs, we execute the Power Allocaton (PA) phase n whch each user s power s allocated across the assgned tones usng the optmal power allocaton n (1.16) CA n SOA1: Progressve Subchannel Allocaton Based on Metrc Sortng In ths famly of SOAs, tones are assgned sequentally n one pass based on a per user metrc for each tone,.e., we terate N tmes, where each teraton corresponds to the assgnment of one tone. Let N (n) denote the set of tones assgned to user after the nth teraton. Let g (n) denote user s metrc durng the nth teraton and let l (n) be the tone ndex that user would lke to be assgned f he/she s assgned the nth tone. The resultng CA algorthm s gven n Algorthm 1. Note that all the user metrcs are updated after each tone s assgned. Algorthm 1 CA Phase for SOA1 1: Intalzaton: set n = 0 and N (n) = for each user. 2: whle n < N do 3: 4: n + 1. Update tone ndex l (n) for each user. 5: Update metrc g (n) for each user. 6: Fnd (n) = arg max g (n) (break tes arbtrarly). 7: f g (n)(n) 0 then 8: Assgn the nth tone to user (n): { N (n 1) {l N (n) = (n)}, f = n; N (n 1), otherwse. 9: else 10: Do not assgn the nth tone. 11: end f 12: end whle We consder several varatons of Algorthm 1 whch correspond to dfferent choces for steps 4 and 5. The choces for step 4 are: (4A): Sort the tones based on the best channel condton among all users. Ths nvolves two steps. Frst, for each tone j, fnd the best channel condton among all users and denote t by µ j := max e j. Second, fnd a tone permutaton {α j } j N such that µ α1 µ α2 µ αn, and set l (n) = α n for each user at the nth teraton. Each max operaton

ECE559VV Project Report

ECE559VV Project Report ECE559VV Project Report (Supplementary Notes Loc Xuan Bu I. MAX SUM-RATE SCHEDULING: THE UPLINK CASE We have seen (n the presentaton that, for downlnk (broadcast channels, the strategy maxmzng the sum-rate

More information

princeton univ. F 17 cos 521: Advanced Algorithm Design Lecture 7: LP Duality Lecturer: Matt Weinberg

princeton univ. F 17 cos 521: Advanced Algorithm Design Lecture 7: LP Duality Lecturer: Matt Weinberg prnceton unv. F 17 cos 521: Advanced Algorthm Desgn Lecture 7: LP Dualty Lecturer: Matt Wenberg Scrbe: LP Dualty s an extremely useful tool for analyzng structural propertes of lnear programs. Whle there

More information

Module 3 LOSSY IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur

Module 3 LOSSY IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur Module 3 LOSSY IMAGE COMPRESSION SYSTEMS Verson ECE IIT, Kharagpur Lesson 6 Theory of Quantzaton Verson ECE IIT, Kharagpur Instructonal Objectves At the end of ths lesson, the students should be able to:

More information

Joint Scheduling and Resource Allocation in CDMA Systems

Joint Scheduling and Resource Allocation in CDMA Systems TO APPEAR IEEE TRANSACTIONS ON INFORMATION THEORY Jont Schedulng and Resource Allocaton n CDMA Systems Vjay G. Subramanan, Randall A. Berry, and Rajeev Agrawal Abstract In ths paper, the schedulng and

More information

Joint Scheduling and Resource Allocation in CDMA Systems

Joint Scheduling and Resource Allocation in CDMA Systems TECHNICAL REPORT - JUNE 2009 1 Jont Schedulng and Resource Allocaton n CDMA Systems Vjay G. Subramanan, Randall A. Berry, and Rajeev Agrawal Expanded Techncal Report: A shorter verson of ths paper wll

More information

Assortment Optimization under MNL

Assortment Optimization under MNL Assortment Optmzaton under MNL Haotan Song Aprl 30, 2017 1 Introducton The assortment optmzaton problem ams to fnd the revenue-maxmzng assortment of products to offer when the prces of products are fxed.

More information

Joint Scheduling and Resource Allocation in CDMA Systems

Joint Scheduling and Resource Allocation in CDMA Systems ACCEPTED TO IEEE TRANSACTIONS ON INFORMATION THEORY 1 Jont Schedulng and Resource Allocaton n CDMA Systems Vjay G. Subramanan, Randall A. Berry, and Rajeev Agrawal Abstract We consder schedulng and resource

More information

Joint Scheduling and Resource Allocation in CDMA Systems

Joint Scheduling and Resource Allocation in CDMA Systems SUBMITTED TO IEEE TRANSACTIONS ON INFORMATION THEORY 1 Jont Schedulng and Resource Allocaton n CDMA Systems Vjay Subramanan, Randall A. Berry, and Rajeev Agrawal Abstract We consder schedulng and resource

More information

Module 9. Lecture 6. Duality in Assignment Problems

Module 9. Lecture 6. Duality in Assignment Problems Module 9 1 Lecture 6 Dualty n Assgnment Problems In ths lecture we attempt to answer few other mportant questons posed n earler lecture for (AP) and see how some of them can be explaned through the concept

More information

College of Computer & Information Science Fall 2009 Northeastern University 20 October 2009

College of Computer & Information Science Fall 2009 Northeastern University 20 October 2009 College of Computer & Informaton Scence Fall 2009 Northeastern Unversty 20 October 2009 CS7880: Algorthmc Power Tools Scrbe: Jan Wen and Laura Poplawsk Lecture Outlne: Prmal-dual schema Network Desgn:

More information

MMA and GCMMA two methods for nonlinear optimization

MMA and GCMMA two methods for nonlinear optimization MMA and GCMMA two methods for nonlnear optmzaton Krster Svanberg Optmzaton and Systems Theory, KTH, Stockholm, Sweden. krlle@math.kth.se Ths note descrbes the algorthms used n the author s 2007 mplementatons

More information

Resource Allocation with a Budget Constraint for Computing Independent Tasks in the Cloud

Resource Allocation with a Budget Constraint for Computing Independent Tasks in the Cloud Resource Allocaton wth a Budget Constrant for Computng Independent Tasks n the Cloud Wemng Sh and Bo Hong School of Electrcal and Computer Engneerng Georga Insttute of Technology, USA 2nd IEEE Internatonal

More information

Structure and Drive Paul A. Jensen Copyright July 20, 2003

Structure and Drive Paul A. Jensen Copyright July 20, 2003 Structure and Drve Paul A. Jensen Copyrght July 20, 2003 A system s made up of several operatons wth flow passng between them. The structure of the system descrbes the flow paths from nputs to outputs.

More information

Feature Selection: Part 1

Feature Selection: Part 1 CSE 546: Machne Learnng Lecture 5 Feature Selecton: Part 1 Instructor: Sham Kakade 1 Regresson n the hgh dmensonal settng How do we learn when the number of features d s greater than the sample sze n?

More information

Lecture Notes on Linear Regression

Lecture Notes on Linear Regression Lecture Notes on Lnear Regresson Feng L fl@sdueducn Shandong Unversty, Chna Lnear Regresson Problem In regresson problem, we am at predct a contnuous target value gven an nput feature vector We assume

More information

Some modelling aspects for the Matlab implementation of MMA

Some modelling aspects for the Matlab implementation of MMA Some modellng aspects for the Matlab mplementaton of MMA Krster Svanberg krlle@math.kth.se Optmzaton and Systems Theory Department of Mathematcs KTH, SE 10044 Stockholm September 2004 1. Consdered optmzaton

More information

Solutions HW #2. minimize. Ax = b. Give the dual problem, and make the implicit equality constraints explicit. Solution.

Solutions HW #2. minimize. Ax = b. Give the dual problem, and make the implicit equality constraints explicit. Solution. Solutons HW #2 Dual of general LP. Fnd the dual functon of the LP mnmze subject to c T x Gx h Ax = b. Gve the dual problem, and make the mplct equalty constrants explct. Soluton. 1. The Lagrangan s L(x,

More information

Lecture 10 Support Vector Machines II

Lecture 10 Support Vector Machines II Lecture 10 Support Vector Machnes II 22 February 2016 Taylor B. Arnold Yale Statstcs STAT 365/665 1/28 Notes: Problem 3 s posted and due ths upcomng Frday There was an early bug n the fake-test data; fxed

More information

2E Pattern Recognition Solutions to Introduction to Pattern Recognition, Chapter 2: Bayesian pattern classification

2E Pattern Recognition Solutions to Introduction to Pattern Recognition, Chapter 2: Bayesian pattern classification E395 - Pattern Recognton Solutons to Introducton to Pattern Recognton, Chapter : Bayesan pattern classfcaton Preface Ths document s a soluton manual for selected exercses from Introducton to Pattern Recognton

More information

COS 521: Advanced Algorithms Game Theory and Linear Programming

COS 521: Advanced Algorithms Game Theory and Linear Programming COS 521: Advanced Algorthms Game Theory and Lnear Programmng Moses Charkar February 27, 2013 In these notes, we ntroduce some basc concepts n game theory and lnear programmng (LP). We show a connecton

More information

Real-Time Systems. Multiprocessor scheduling. Multiprocessor scheduling. Multiprocessor scheduling

Real-Time Systems. Multiprocessor scheduling. Multiprocessor scheduling. Multiprocessor scheduling Real-Tme Systems Multprocessor schedulng Specfcaton Implementaton Verfcaton Multprocessor schedulng -- -- Global schedulng How are tasks assgned to processors? Statc assgnment The processor(s) used for

More information

Problem Set 9 Solutions

Problem Set 9 Solutions Desgn and Analyss of Algorthms May 4, 2015 Massachusetts Insttute of Technology 6.046J/18.410J Profs. Erk Demane, Srn Devadas, and Nancy Lynch Problem Set 9 Solutons Problem Set 9 Solutons Ths problem

More information

CHAPTER 5 NUMERICAL EVALUATION OF DYNAMIC RESPONSE

CHAPTER 5 NUMERICAL EVALUATION OF DYNAMIC RESPONSE CHAPTER 5 NUMERICAL EVALUATION OF DYNAMIC RESPONSE Analytcal soluton s usually not possble when exctaton vares arbtrarly wth tme or f the system s nonlnear. Such problems can be solved by numercal tmesteppng

More information

3.1 Expectation of Functions of Several Random Variables. )' be a k-dimensional discrete or continuous random vector, with joint PMF p (, E X E X1 E X

3.1 Expectation of Functions of Several Random Variables. )' be a k-dimensional discrete or continuous random vector, with joint PMF p (, E X E X1 E X Statstcs 1: Probablty Theory II 37 3 EPECTATION OF SEVERAL RANDOM VARIABLES As n Probablty Theory I, the nterest n most stuatons les not on the actual dstrbuton of a random vector, but rather on a number

More information

Linear Approximation with Regularization and Moving Least Squares

Linear Approximation with Regularization and Moving Least Squares Lnear Approxmaton wth Regularzaton and Movng Least Squares Igor Grešovn May 007 Revson 4.6 (Revson : March 004). 5 4 3 0.5 3 3.5 4 Contents: Lnear Fttng...4. Weghted Least Squares n Functon Approxmaton...

More information

Winter 2008 CS567 Stochastic Linear/Integer Programming Guest Lecturer: Xu, Huan

Winter 2008 CS567 Stochastic Linear/Integer Programming Guest Lecturer: Xu, Huan Wnter 2008 CS567 Stochastc Lnear/Integer Programmng Guest Lecturer: Xu, Huan Class 2: More Modelng Examples 1 Capacty Expanson Capacty expanson models optmal choces of the tmng and levels of nvestments

More information

Simultaneous Optimization of Berth Allocation, Quay Crane Assignment and Quay Crane Scheduling Problems in Container Terminals

Simultaneous Optimization of Berth Allocation, Quay Crane Assignment and Quay Crane Scheduling Problems in Container Terminals Smultaneous Optmzaton of Berth Allocaton, Quay Crane Assgnment and Quay Crane Schedulng Problems n Contaner Termnals Necat Aras, Yavuz Türkoğulları, Z. Caner Taşkın, Kuban Altınel Abstract In ths work,

More information

An Interactive Optimisation Tool for Allocation Problems

An Interactive Optimisation Tool for Allocation Problems An Interactve Optmsaton ool for Allocaton Problems Fredr Bonäs, Joam Westerlund and apo Westerlund Process Desgn Laboratory, Faculty of echnology, Åbo Aadem Unversty, uru 20500, Fnland hs paper presents

More information

VQ widely used in coding speech, image, and video

VQ widely used in coding speech, image, and video at Scalar quantzers are specal cases of vector quantzers (VQ): they are constraned to look at one sample at a tme (memoryless) VQ does not have such constrant better RD perfomance expected Source codng

More information

Solutions to exam in SF1811 Optimization, Jan 14, 2015

Solutions to exam in SF1811 Optimization, Jan 14, 2015 Solutons to exam n SF8 Optmzaton, Jan 4, 25 3 3 O------O -4 \ / \ / The network: \/ where all lnks go from left to rght. /\ / \ / \ 6 O------O -5 2 4.(a) Let x = ( x 3, x 4, x 23, x 24 ) T, where the varable

More information

Generalized Linear Methods

Generalized Linear Methods Generalzed Lnear Methods 1 Introducton In the Ensemble Methods the general dea s that usng a combnaton of several weak learner one could make a better learner. More formally, assume that we have a set

More information

The Minimum Universal Cost Flow in an Infeasible Flow Network

The Minimum Universal Cost Flow in an Infeasible Flow Network Journal of Scences, Islamc Republc of Iran 17(2): 175-180 (2006) Unversty of Tehran, ISSN 1016-1104 http://jscencesutacr The Mnmum Unversal Cost Flow n an Infeasble Flow Network H Saleh Fathabad * M Bagheran

More information

The Geometry of Logit and Probit

The Geometry of Logit and Probit The Geometry of Logt and Probt Ths short note s meant as a supplement to Chapters and 3 of Spatal Models of Parlamentary Votng and the notaton and reference to fgures n the text below s to those two chapters.

More information

A 2D Bounded Linear Program (H,c) 2D Linear Programming

A 2D Bounded Linear Program (H,c) 2D Linear Programming A 2D Bounded Lnear Program (H,c) h 3 v h 8 h 5 c h 4 h h 6 h 7 h 2 2D Lnear Programmng C s a polygonal regon, the ntersecton of n halfplanes. (H, c) s nfeasble, as C s empty. Feasble regon C s unbounded

More information

Difference Equations

Difference Equations Dfference Equatons c Jan Vrbk 1 Bascs Suppose a sequence of numbers, say a 0,a 1,a,a 3,... s defned by a certan general relatonshp between, say, three consecutve values of the sequence, e.g. a + +3a +1

More information

U.C. Berkeley CS294: Spectral Methods and Expanders Handout 8 Luca Trevisan February 17, 2016

U.C. Berkeley CS294: Spectral Methods and Expanders Handout 8 Luca Trevisan February 17, 2016 U.C. Berkeley CS94: Spectral Methods and Expanders Handout 8 Luca Trevsan February 7, 06 Lecture 8: Spectral Algorthms Wrap-up In whch we talk about even more generalzatons of Cheeger s nequaltes, and

More information

Kernel Methods and SVMs Extension

Kernel Methods and SVMs Extension Kernel Methods and SVMs Extenson The purpose of ths document s to revew materal covered n Machne Learnng 1 Supervsed Learnng regardng support vector machnes (SVMs). Ths document also provdes a general

More information

CSci 6974 and ECSE 6966 Math. Tech. for Vision, Graphics and Robotics Lecture 21, April 17, 2006 Estimating A Plane Homography

CSci 6974 and ECSE 6966 Math. Tech. for Vision, Graphics and Robotics Lecture 21, April 17, 2006 Estimating A Plane Homography CSc 6974 and ECSE 6966 Math. Tech. for Vson, Graphcs and Robotcs Lecture 21, Aprl 17, 2006 Estmatng A Plane Homography Overvew We contnue wth a dscusson of the major ssues, usng estmaton of plane projectve

More information

Lecture 4: November 17, Part 1 Single Buffer Management

Lecture 4: November 17, Part 1 Single Buffer Management Lecturer: Ad Rosén Algorthms for the anagement of Networs Fall 2003-2004 Lecture 4: November 7, 2003 Scrbe: Guy Grebla Part Sngle Buffer anagement In the prevous lecture we taled about the Combned Input

More information

Lecture 12: Discrete Laplacian

Lecture 12: Discrete Laplacian Lecture 12: Dscrete Laplacan Scrbe: Tanye Lu Our goal s to come up wth a dscrete verson of Laplacan operator for trangulated surfaces, so that we can use t n practce to solve related problems We are mostly

More information

Chapter 5. Solution of System of Linear Equations. Module No. 6. Solution of Inconsistent and Ill Conditioned Systems

Chapter 5. Solution of System of Linear Equations. Module No. 6. Solution of Inconsistent and Ill Conditioned Systems Numercal Analyss by Dr. Anta Pal Assstant Professor Department of Mathematcs Natonal Insttute of Technology Durgapur Durgapur-713209 emal: anta.bue@gmal.com 1 . Chapter 5 Soluton of System of Lnear Equatons

More information

Lagrange Multipliers Kernel Trick

Lagrange Multipliers Kernel Trick Lagrange Multplers Kernel Trck Ncholas Ruozz Unversty of Texas at Dallas Based roughly on the sldes of Davd Sontag General Optmzaton A mathematcal detour, we ll come back to SVMs soon! subject to: f x

More information

CSC 411 / CSC D11 / CSC C11

CSC 411 / CSC D11 / CSC C11 18 Boostng s a general strategy for learnng classfers by combnng smpler ones. The dea of boostng s to take a weak classfer that s, any classfer that wll do at least slghtly better than chance and use t

More information

Foundations of Arithmetic

Foundations of Arithmetic Foundatons of Arthmetc Notaton We shall denote the sum and product of numbers n the usual notaton as a 2 + a 2 + a 3 + + a = a, a 1 a 2 a 3 a = a The notaton a b means a dvdes b,.e. ac = b where c s an

More information

Maximizing the number of nonnegative subsets

Maximizing the number of nonnegative subsets Maxmzng the number of nonnegatve subsets Noga Alon Hao Huang December 1, 213 Abstract Gven a set of n real numbers, f the sum of elements of every subset of sze larger than k s negatve, what s the maxmum

More information

Stanford University CS359G: Graph Partitioning and Expanders Handout 4 Luca Trevisan January 13, 2011

Stanford University CS359G: Graph Partitioning and Expanders Handout 4 Luca Trevisan January 13, 2011 Stanford Unversty CS359G: Graph Parttonng and Expanders Handout 4 Luca Trevsan January 3, 0 Lecture 4 In whch we prove the dffcult drecton of Cheeger s nequalty. As n the past lectures, consder an undrected

More information

Lecture 14: Bandits with Budget Constraints

Lecture 14: Bandits with Budget Constraints IEOR 8100-001: Learnng and Optmzaton for Sequental Decson Makng 03/07/16 Lecture 14: andts wth udget Constrants Instructor: Shpra Agrawal Scrbed by: Zhpeng Lu 1 Problem defnton In the regular Mult-armed

More information

An Admission Control Algorithm in Cloud Computing Systems

An Admission Control Algorithm in Cloud Computing Systems An Admsson Control Algorthm n Cloud Computng Systems Authors: Frank Yeong-Sung Ln Department of Informaton Management Natonal Tawan Unversty Tape, Tawan, R.O.C. ysln@m.ntu.edu.tw Yngje Lan Management Scence

More information

Temperature. Chapter Heat Engine

Temperature. Chapter Heat Engine Chapter 3 Temperature In prevous chapters of these notes we ntroduced the Prncple of Maxmum ntropy as a technque for estmatng probablty dstrbutons consstent wth constrants. In Chapter 9 we dscussed the

More information

x = x 1 + :::+ x K and the nput covarance matrces are of the form ± = E[x x y ]. 3.2 Dualty Next, we ntroduce the concept of dualty wth the followng t

x = x 1 + :::+ x K and the nput covarance matrces are of the form ± = E[x x y ]. 3.2 Dualty Next, we ntroduce the concept of dualty wth the followng t Sum Power Iteratve Water-fllng for Mult-Antenna Gaussan Broadcast Channels N. Jndal, S. Jafar, S. Vshwanath and A. Goldsmth Dept. of Electrcal Engg. Stanford Unversty, CA, 94305 emal: njndal,syed,srram,andrea@wsl.stanford.edu

More information

Singular Value Decomposition: Theory and Applications

Singular Value Decomposition: Theory and Applications Sngular Value Decomposton: Theory and Applcatons Danel Khashab Sprng 2015 Last Update: March 2, 2015 1 Introducton A = UDV where columns of U and V are orthonormal and matrx D s dagonal wth postve real

More information

EEE 241: Linear Systems

EEE 241: Linear Systems EEE : Lnear Systems Summary #: Backpropagaton BACKPROPAGATION The perceptron rule as well as the Wdrow Hoff learnng were desgned to tran sngle layer networks. They suffer from the same dsadvantage: they

More information

Lecture 12: Classification

Lecture 12: Classification Lecture : Classfcaton g Dscrmnant functons g The optmal Bayes classfer g Quadratc classfers g Eucldean and Mahalanobs metrcs g K Nearest Neghbor Classfers Intellgent Sensor Systems Rcardo Guterrez-Osuna

More information

More metrics on cartesian products

More metrics on cartesian products More metrcs on cartesan products If (X, d ) are metrc spaces for 1 n, then n Secton II4 of the lecture notes we defned three metrcs on X whose underlyng topologes are the product topology The purpose of

More information

Additional Codes using Finite Difference Method. 1 HJB Equation for Consumption-Saving Problem Without Uncertainty

Additional Codes using Finite Difference Method. 1 HJB Equation for Consumption-Saving Problem Without Uncertainty Addtonal Codes usng Fnte Dfference Method Benamn Moll 1 HJB Equaton for Consumpton-Savng Problem Wthout Uncertanty Before consderng the case wth stochastc ncome n http://www.prnceton.edu/~moll/ HACTproect/HACT_Numercal_Appendx.pdf,

More information

Supplement: Proofs and Technical Details for The Solution Path of the Generalized Lasso

Supplement: Proofs and Technical Details for The Solution Path of the Generalized Lasso Supplement: Proofs and Techncal Detals for The Soluton Path of the Generalzed Lasso Ryan J. Tbshran Jonathan Taylor In ths document we gve supplementary detals to the paper The Soluton Path of the Generalzed

More information

Joint Scheduling and Power-Allocation for Interference Management in Wireless Networks

Joint Scheduling and Power-Allocation for Interference Management in Wireless Networks Jont Schedulng and Power-Allocaton for Interference Management n Wreless Networks Xn Lu *, Edwn K. P. Chong, and Ness B. Shroff * * School of Electrcal and Computer Engneerng Purdue Unversty West Lafayette,

More information

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix Lectures - Week 4 Matrx norms, Condtonng, Vector Spaces, Lnear Independence, Spannng sets and Bass, Null space and Range of a Matrx Matrx Norms Now we turn to assocatng a number to each matrx. We could

More information

Distributed Non-Autonomous Power Control through Distributed Convex Optimization

Distributed Non-Autonomous Power Control through Distributed Convex Optimization Dstrbuted Non-Autonomous Power Control through Dstrbuted Convex Optmzaton S. Sundhar Ram and V. V. Veeravall ECE Department and Coordnated Scence Lab Unversty of Illnos at Urbana-Champagn Emal: {ssrnv5,vvv}@llnos.edu

More information

NUMERICAL DIFFERENTIATION

NUMERICAL DIFFERENTIATION NUMERICAL DIFFERENTIATION 1 Introducton Dfferentaton s a method to compute the rate at whch a dependent output y changes wth respect to the change n the ndependent nput x. Ths rate of change s called the

More information

EEL 6266 Power System Operation and Control. Chapter 3 Economic Dispatch Using Dynamic Programming

EEL 6266 Power System Operation and Control. Chapter 3 Economic Dispatch Using Dynamic Programming EEL 6266 Power System Operaton and Control Chapter 3 Economc Dspatch Usng Dynamc Programmng Pecewse Lnear Cost Functons Common practce many utltes prefer to represent ther generator cost functons as sngle-

More information

Complete subgraphs in multipartite graphs

Complete subgraphs in multipartite graphs Complete subgraphs n multpartte graphs FLORIAN PFENDER Unverstät Rostock, Insttut für Mathematk D-18057 Rostock, Germany Floran.Pfender@un-rostock.de Abstract Turán s Theorem states that every graph G

More information

Errors for Linear Systems

Errors for Linear Systems Errors for Lnear Systems When we solve a lnear system Ax b we often do not know A and b exactly, but have only approxmatons  and ˆb avalable. Then the best thng we can do s to solve ˆx ˆb exactly whch

More information

For now, let us focus on a specific model of neurons. These are simplified from reality but can achieve remarkable results.

For now, let us focus on a specific model of neurons. These are simplified from reality but can achieve remarkable results. Neural Networks : Dervaton compled by Alvn Wan from Professor Jtendra Malk s lecture Ths type of computaton s called deep learnng and s the most popular method for many problems, such as computer vson

More information

Perfect Competition and the Nash Bargaining Solution

Perfect Competition and the Nash Bargaining Solution Perfect Competton and the Nash Barganng Soluton Renhard John Department of Economcs Unversty of Bonn Adenauerallee 24-42 53113 Bonn, Germany emal: rohn@un-bonn.de May 2005 Abstract For a lnear exchange

More information

LINEAR REGRESSION ANALYSIS. MODULE IX Lecture Multicollinearity

LINEAR REGRESSION ANALYSIS. MODULE IX Lecture Multicollinearity LINEAR REGRESSION ANALYSIS MODULE IX Lecture - 30 Multcollnearty Dr. Shalabh Department of Mathematcs and Statstcs Indan Insttute of Technology Kanpur 2 Remedes for multcollnearty Varous technques have

More information

= z 20 z n. (k 20) + 4 z k = 4

= z 20 z n. (k 20) + 4 z k = 4 Problem Set #7 solutons 7.2.. (a Fnd the coeffcent of z k n (z + z 5 + z 6 + z 7 + 5, k 20. We use the known seres expanson ( n+l ( z l l z n below: (z + z 5 + z 6 + z 7 + 5 (z 5 ( + z + z 2 + z + 5 5

More information

Lecture 20: Lift and Project, SDP Duality. Today we will study the Lift and Project method. Then we will prove the SDP duality theorem.

Lecture 20: Lift and Project, SDP Duality. Today we will study the Lift and Project method. Then we will prove the SDP duality theorem. prnceton u. sp 02 cos 598B: algorthms and complexty Lecture 20: Lft and Project, SDP Dualty Lecturer: Sanjeev Arora Scrbe:Yury Makarychev Today we wll study the Lft and Project method. Then we wll prove

More information

Econ107 Applied Econometrics Topic 3: Classical Model (Studenmund, Chapter 4)

Econ107 Applied Econometrics Topic 3: Classical Model (Studenmund, Chapter 4) I. Classcal Assumptons Econ7 Appled Econometrcs Topc 3: Classcal Model (Studenmund, Chapter 4) We have defned OLS and studed some algebrac propertes of OLS. In ths topc we wll study statstcal propertes

More information

U.C. Berkeley CS294: Beyond Worst-Case Analysis Luca Trevisan September 5, 2017

U.C. Berkeley CS294: Beyond Worst-Case Analysis Luca Trevisan September 5, 2017 U.C. Berkeley CS94: Beyond Worst-Case Analyss Handout 4s Luca Trevsan September 5, 07 Summary of Lecture 4 In whch we ntroduce semdefnte programmng and apply t to Max Cut. Semdefnte Programmng Recall that

More information

Week 5: Neural Networks

Week 5: Neural Networks Week 5: Neural Networks Instructor: Sergey Levne Neural Networks Summary In the prevous lecture, we saw how we can construct neural networks by extendng logstc regresson. Neural networks consst of multple

More information

Lossy Compression. Compromise accuracy of reconstruction for increased compression.

Lossy Compression. Compromise accuracy of reconstruction for increased compression. Lossy Compresson Compromse accuracy of reconstructon for ncreased compresson. The reconstructon s usually vsbly ndstngushable from the orgnal mage. Typcally, one can get up to 0:1 compresson wth almost

More information

On the Multicriteria Integer Network Flow Problem

On the Multicriteria Integer Network Flow Problem BULGARIAN ACADEMY OF SCIENCES CYBERNETICS AND INFORMATION TECHNOLOGIES Volume 5, No 2 Sofa 2005 On the Multcrtera Integer Network Flow Problem Vassl Vasslev, Marana Nkolova, Maryana Vassleva Insttute of

More information

Chapter - 2. Distribution System Power Flow Analysis

Chapter - 2. Distribution System Power Flow Analysis Chapter - 2 Dstrbuton System Power Flow Analyss CHAPTER - 2 Radal Dstrbuton System Load Flow 2.1 Introducton Load flow s an mportant tool [66] for analyzng electrcal power system network performance. Load

More information

Pricing and Resource Allocation Game Theoretic Models

Pricing and Resource Allocation Game Theoretic Models Prcng and Resource Allocaton Game Theoretc Models Zhy Huang Changbn Lu Q Zhang Computer and Informaton Scence December 8, 2009 Z. Huang, C. Lu, and Q. Zhang (CIS) Game Theoretc Models December 8, 2009

More information

The Second Anti-Mathima on Game Theory

The Second Anti-Mathima on Game Theory The Second Ant-Mathma on Game Theory Ath. Kehagas December 1 2006 1 Introducton In ths note we wll examne the noton of game equlbrum for three types of games 1. 2-player 2-acton zero-sum games 2. 2-player

More information

Chapter Newton s Method

Chapter Newton s Method Chapter 9. Newton s Method After readng ths chapter, you should be able to:. Understand how Newton s method s dfferent from the Golden Secton Search method. Understand how Newton s method works 3. Solve

More information

THE CHINESE REMAINDER THEOREM. We should thank the Chinese for their wonderful remainder theorem. Glenn Stevens

THE CHINESE REMAINDER THEOREM. We should thank the Chinese for their wonderful remainder theorem. Glenn Stevens THE CHINESE REMAINDER THEOREM KEITH CONRAD We should thank the Chnese for ther wonderful remander theorem. Glenn Stevens 1. Introducton The Chnese remander theorem says we can unquely solve any par of

More information

Error Probability for M Signals

Error Probability for M Signals Chapter 3 rror Probablty for M Sgnals In ths chapter we dscuss the error probablty n decdng whch of M sgnals was transmtted over an arbtrary channel. We assume the sgnals are represented by a set of orthonormal

More information

Computing Correlated Equilibria in Multi-Player Games

Computing Correlated Equilibria in Multi-Player Games Computng Correlated Equlbra n Mult-Player Games Chrstos H. Papadmtrou Presented by Zhanxang Huang December 7th, 2005 1 The Author Dr. Chrstos H. Papadmtrou CS professor at UC Berkley (taught at Harvard,

More information

Parametric fractional imputation for missing data analysis. Jae Kwang Kim Survey Working Group Seminar March 29, 2010

Parametric fractional imputation for missing data analysis. Jae Kwang Kim Survey Working Group Seminar March 29, 2010 Parametrc fractonal mputaton for mssng data analyss Jae Kwang Km Survey Workng Group Semnar March 29, 2010 1 Outlne Introducton Proposed method Fractonal mputaton Approxmaton Varance estmaton Multple mputaton

More information

Annexes. EC.1. Cycle-base move illustration. EC.2. Problem Instances

Annexes. EC.1. Cycle-base move illustration. EC.2. Problem Instances ec Annexes Ths Annex frst llustrates a cycle-based move n the dynamc-block generaton tabu search. It then dsplays the characterstcs of the nstance sets, followed by detaled results of the parametercalbraton

More information

Single-Facility Scheduling over Long Time Horizons by Logic-based Benders Decomposition

Single-Facility Scheduling over Long Time Horizons by Logic-based Benders Decomposition Sngle-Faclty Schedulng over Long Tme Horzons by Logc-based Benders Decomposton Elvn Coban and J. N. Hooker Tepper School of Busness, Carnege Mellon Unversty ecoban@andrew.cmu.edu, john@hooker.tepper.cmu.edu

More information

6.854J / J Advanced Algorithms Fall 2008

6.854J / J Advanced Algorithms Fall 2008 MIT OpenCourseWare http://ocw.mt.edu 6.854J / 18.415J Advanced Algorthms Fall 2008 For nformaton about ctng these materals or our Terms of Use, vst: http://ocw.mt.edu/terms. 18.415/6.854 Advanced Algorthms

More information

PHYS 705: Classical Mechanics. Calculus of Variations II

PHYS 705: Classical Mechanics. Calculus of Variations II 1 PHYS 705: Classcal Mechancs Calculus of Varatons II 2 Calculus of Varatons: Generalzaton (no constrant yet) Suppose now that F depends on several dependent varables : We need to fnd such that has a statonary

More information

Power Allocation/Beamforming for DF MIMO Two-Way Relaying: Relay and Network Optimization

Power Allocation/Beamforming for DF MIMO Two-Way Relaying: Relay and Network Optimization Power Allocaton/Beamformng for DF MIMO Two-Way Relayng: Relay and Network Optmzaton Je Gao, Janshu Zhang, Sergy A. Vorobyov, Ha Jang, and Martn Haardt Dept. of Electrcal & Computer Engneerng, Unversty

More information

Composite Hypotheses testing

Composite Hypotheses testing Composte ypotheses testng In many hypothess testng problems there are many possble dstrbutons that can occur under each of the hypotheses. The output of the source s a set of parameters (ponts n a parameter

More information

NP-Completeness : Proofs

NP-Completeness : Proofs NP-Completeness : Proofs Proof Methods A method to show a decson problem Π NP-complete s as follows. (1) Show Π NP. (2) Choose an NP-complete problem Π. (3) Show Π Π. A method to show an optmzaton problem

More information

LOW BIAS INTEGRATED PATH ESTIMATORS. James M. Calvin

LOW BIAS INTEGRATED PATH ESTIMATORS. James M. Calvin Proceedngs of the 007 Wnter Smulaton Conference S G Henderson, B Bller, M-H Hseh, J Shortle, J D Tew, and R R Barton, eds LOW BIAS INTEGRATED PATH ESTIMATORS James M Calvn Department of Computer Scence

More information

APPROXIMATE PRICES OF BASKET AND ASIAN OPTIONS DUPONT OLIVIER. Premia 14

APPROXIMATE PRICES OF BASKET AND ASIAN OPTIONS DUPONT OLIVIER. Premia 14 APPROXIMAE PRICES OF BASKE AND ASIAN OPIONS DUPON OLIVIER Prema 14 Contents Introducton 1 1. Framewor 1 1.1. Baset optons 1.. Asan optons. Computng the prce 3. Lower bound 3.1. Closed formula for the prce

More information

Time-Varying Systems and Computations Lecture 6

Time-Varying Systems and Computations Lecture 6 Tme-Varyng Systems and Computatons Lecture 6 Klaus Depold 14. Januar 2014 The Kalman Flter The Kalman estmaton flter attempts to estmate the actual state of an unknown dscrete dynamcal system, gven nosy

More information

Case A. P k = Ni ( 2L i k 1 ) + (# big cells) 10d 2 P k.

Case A. P k = Ni ( 2L i k 1 ) + (# big cells) 10d 2 P k. THE CELLULAR METHOD In ths lecture, we ntroduce the cellular method as an approach to ncdence geometry theorems lke the Szemeréd-Trotter theorem. The method was ntroduced n the paper Combnatoral complexty

More information

1 Convex Optimization

1 Convex Optimization Convex Optmzaton We wll consder convex optmzaton problems. Namely, mnmzaton problems where the objectve s convex (we assume no constrants for now). Such problems often arse n machne learnng. For example,

More information

Games of Threats. Elon Kohlberg Abraham Neyman. Working Paper

Games of Threats. Elon Kohlberg Abraham Neyman. Working Paper Games of Threats Elon Kohlberg Abraham Neyman Workng Paper 18-023 Games of Threats Elon Kohlberg Harvard Busness School Abraham Neyman The Hebrew Unversty of Jerusalem Workng Paper 18-023 Copyrght 2017

More information

COMPARISON OF SOME RELIABILITY CHARACTERISTICS BETWEEN REDUNDANT SYSTEMS REQUIRING SUPPORTING UNITS FOR THEIR OPERATIONS

COMPARISON OF SOME RELIABILITY CHARACTERISTICS BETWEEN REDUNDANT SYSTEMS REQUIRING SUPPORTING UNITS FOR THEIR OPERATIONS Avalable onlne at http://sck.org J. Math. Comput. Sc. 3 (3), No., 6-3 ISSN: 97-537 COMPARISON OF SOME RELIABILITY CHARACTERISTICS BETWEEN REDUNDANT SYSTEMS REQUIRING SUPPORTING UNITS FOR THEIR OPERATIONS

More information

Yong Joon Ryang. 1. Introduction Consider the multicommodity transportation problem with convex quadratic cost function. 1 2 (x x0 ) T Q(x x 0 )

Yong Joon Ryang. 1. Introduction Consider the multicommodity transportation problem with convex quadratic cost function. 1 2 (x x0 ) T Q(x x 0 ) Kangweon-Kyungk Math. Jour. 4 1996), No. 1, pp. 7 16 AN ITERATIVE ROW-ACTION METHOD FOR MULTICOMMODITY TRANSPORTATION PROBLEMS Yong Joon Ryang Abstract. The optmzaton problems wth quadratc constrants often

More information

4DVAR, according to the name, is a four-dimensional variational method.

4DVAR, according to the name, is a four-dimensional variational method. 4D-Varatonal Data Assmlaton (4D-Var) 4DVAR, accordng to the name, s a four-dmensonal varatonal method. 4D-Var s actually a drect generalzaton of 3D-Var to handle observatons that are dstrbuted n tme. The

More information

Two-Way and Multiple-Access Energy Harvesting Systems with Energy Cooperation

Two-Way and Multiple-Access Energy Harvesting Systems with Energy Cooperation Two-Way and Multple-Access Energy Harvestng Systems wth Energy Cooperaton Berk Gurakan, Omur Ozel, Jng Yang 2, and Sennur Ulukus Department of Electrcal and Computer Engneerng, Unversty of Maryland, College

More information

Psychology 282 Lecture #24 Outline Regression Diagnostics: Outliers

Psychology 282 Lecture #24 Outline Regression Diagnostics: Outliers Psychology 282 Lecture #24 Outlne Regresson Dagnostcs: Outlers In an earler lecture we studed the statstcal assumptons underlyng the regresson model, ncludng the followng ponts: Formal statement of assumptons.

More information

Appendix for Causal Interaction in Factorial Experiments: Application to Conjoint Analysis

Appendix for Causal Interaction in Factorial Experiments: Application to Conjoint Analysis A Appendx for Causal Interacton n Factoral Experments: Applcaton to Conjont Analyss Mathematcal Appendx: Proofs of Theorems A. Lemmas Below, we descrbe all the lemmas, whch are used to prove the man theorems

More information