You are on page 1of 14

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO.

12, DECEMBER 2008

6003

ARQ by Subcarrier Assignment for OFDM-Based Systems
Chin Keong Ho, Hongming Yang, Ashish Pandharipande, and Jan W. M. Bergmans, Senior Member, IEEE
Abstract—We consider two automatic repeat request (ARQ) schemes based on subcarrier assignment in orthogonal frequency-division multiplexing (OFDM)-based systems: single ARQ subcarrier assignment (single ARQ-SA) and multiple ARQ-SA. In single ARQ-SA, data transmitted on a subcarrier in a failed transmission are repeated on a single assigned subcarrier in the ARQ transmission. In multiple ARQ-SA, the data are repeated on multiple assigned subcarriers in the ARQ transmission. At the receiver, maximum ratio combining is performed on subcarriers that carry the same data. Our goal is to optimize certain system utility functions (such as to minimize bit error rates or to maximize sum capacity) through the choice of the subcarrier assignment. We show that a large class of reasonable system utility functions that we wish to maximize are characterized as Schur-concave. For this class of utility functions, we obtain the optimum subcarrier assignment for single ARQ-SA and propose a suboptimum (heuristic) subcarrier assignment scheme for multiple ARQ-SA. Further, to lower the overhead of signaling the subcarrier assignment information, we consider subcarrier grouping methods. Numerical results indicate that substantial throughput improvement can be achieved by appropriate assignments, especially with the use of incremental redundancy at high signal-to-noise ratios. Index Terms—Automatic repeat request (ARQ), majorization, orthogonal frequency-division multiplexing (OFDM), subcarrier assignment.

I. INTRODUCTION

O

RTHOGONAL frequency-division multiplexing (OFDM) is an effective solution for delivering high data rates over wireless channels with frequency-selective fading. A number of wireless standards such as IEEE 802.11a [1], WiMax [2], and long-term 3G evolution [3] have adopted OFDM-based solutions for physical-layer transmission. An OFDM-based system combats multipath fading with the use of

Manuscript received April 03, 2007; revised July 01, 2008. First published August 29, 2008; current version published November 19, 2008. The associate editor coordinating the review of this manuscript and approving it for publication was Dr. Gerald Matz. C. K. Ho was with the Department of Electrical Engineering, Eindhoven University of Technology, Eindhoven, The Netherlands, and also with Philips Research Laboratories Eindhoven, The Netherlands. He is now with the Institute for Infocomm Research, A*STAR, Singapore (e-mail: hock@i2r.a-star.edu.sg). H. Yang and J. W. M. Bergmans are with the Department of Electrical Engineering, Eindhoven University of Technology, Eindhoven, The Netherlands. They are also with Philips Research Laboratories Eindhoven, 5656 AE Eindhoven, The Netherlands (e-mail: h.m.yang@tue.nl; j.w.m.bergmans@tue.nl). A. Pandharipande is with Philips Research Laboratories Eindhoven, 5656 AE Eindhoven, The Netherlands (e-mail: ashish.p@philips.com). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TSP.2008.2005093

a cyclic prefix that, in conjunction with the Fourier transform, converts the frequency-selective fading channel into a set of parallel subcarriers experiencing flat fading. It is common to use an automatic repeat request (ARQ) mechanism [4], [5] in OFDM systems when a packet transmission fails. In this mechanism, the transmitter retransmits the data when it fails to receive an acknowledgment (ACK) or receives an explicit negative ACK. We shall study ARQ schemes involving subcarrier assignment in OFDM-based systems. The system under consideration is a general OFDM system where a linear unitary pretransform may be applied before the application of the inverse discrete Fourier transform at the transmitter. Such pretransformed OFDM (PT-OFDM) systems have been known to offer various advantages such as improved block error rates [6] and reduced transmitter complexity [7]. A PT-OFDM system can also be shown to be equivalent to a system with parallel subcarriers in the frequency domain. In such systems, under the ARQ mechanism, the data in the failed transmission have to be retransmitted over the parallel subcarriers in the event of a packet failure. For clarity, we shall call the failed transmission the original transmission and the associated subcarriers the original subcarriers. The retransmission will be called ARQ transmission and the associated subcarriers will be termed ARQ subcarriers. In this paper, we consider two ARQ schemes: single ARQ subcarrier assignment (ARQ-SA) scheme and multiple ARQ-SA. In single ARQ-SA, data on an original subcarrier are repeated on a single ARQ subcarrier, which may be different from the original subcarrier. We say that the ARQ subcarrier is assigned to the original subcarrier. In multiple ARQ-SA, however, zero, one, or more ARQ subcarriers may be assigned to an original subcarrier. At the receiver, maximum ratio combining (MRC) is performed on the original subcarrier, and all ARQ subcarriers that carry the same data. Subsequently, a single stage of equalization and decoding is carried out. Our goal is to optimize a certain system metric by choosing the assignment, under the assumption that full channel state information (CSI) is available at both the transmitter and the receiver. We first phrase this optimization problem as one of maximizing a utility function and show that many utility functions of practical interest that we wish to maximize are Schur-concave. Examples of such utility functions are the sum capacity and the probability of correct reception. Under single ARQ-SA, we show that for Schur-concave utility functions, the optimum assignment is to assign the ARQ subcarrier with the strongest signal-to-noise ratio (SNR) to the original subcarrier with the weakest SNR, and so on. That is, if the original subcarriers are ordered with their respective SNRs in decreasing order, then

1053-587X/$25.00 © 2008 IEEE

assuming th order finite-impulse response channel a sample-spaced . The ARQ-SA schemes we propose here are aimed at exploiting CSI effectively. II. In Section VIII. we find that determining the optimum assignment is an NP-hard problem. so as to realize the maximum gain offered through an appropriate assignment. 1. VOL. This clearly offers an improvement over a simple scheme where no assignment is done. Fig. even in quasi-static channels. represented as . for some channel realizations data on a weak original subcarrier may again be retransmitted on a weak ARQ subcarrier. the received subcarrier vector in the frequency docan be expressed as main (1) is independent identically distributed where (i. For simplicity. . cyclic assignments are used. however. After performing a parallel-to-serial (P/S) conversion (its inverse operation is denoted as S/P). 12. Under multiple ARQ-SA. After the ARQ round has ended. After FFT. we insert a cyclic prefix with duration not shorter than the maximum channel delay spread so as to avoid inter-OFDM symbol interference. In Section II. we show a general OFDM system including a possible pretransform. NO. The sudenote complex conjugate. In [18]. where is an transformation matrix. a diversity effect is realized by retransmitting data on an ARQ subcarrier with a channel that is independent from that of the original subcarrier. OFDM Systems the assigned ARQ subcarriers should be ordered with their respective SNRs in increasing order. SYSTEM DESCRIPTION A. Here. 1. In these works. the problem of how to exactly perform the assignment was not addressed. To obtain our results. in cyclic assignments. At the receiver.6004 IEEE TRANSACTIONS ON SIGNAL PROCESSING. . simulation results are presented to test the efficacy of our algorithms. Notation: We use bold lower case letters to denote column vectors and bold upper case letters to denote matrices. Hence. is the th channel response coefficient in the frequency domain. we employ the theory of majorization [10]. However. while is a diagonal matrix. we consider either an OFDM is an identity matrix or the pretransformed system where is unitary with constant-amplitude OFDM system. We address this shortcoming by using the available CSI to develop better assignment strategies. The optimality of these algorithms is investigated in Section V. resulting in poor performance. perscripts . transpose. The block of modulation symbols is then passed through an inverse discrete Fourier transform. where ARQ subcarriers are cyclic-shifted versions of the original subcarriers. while the all-ones noted vector is denoted as . before the next original transmission begins. Fixed assignments that do not exploit CSI have been considered in the literature [15]–[17]. Algorithms for subcarrier assignment that solve these problems and their complexity are presented in Section IV. The vector is linearly transformed into subcarriers in the frequency domain as . If the cyclic shift is larger than the coherence bandwidth. where subcarrier grouping techniques for reducing this overhead are presented. The overhead of signaling the subcarrier assignment is analyzed in Section VI. DECEMBER 2008 Section VII presents throughput analysis for the subcarrier assignment schemes. . Although [18] considered the problem of choosing the group of ARQ subcarriers for retransmission (specifically those with SNRs above a certain threshold). To this end. Transmission Scheme We define an ARQ round to consist of an original transmission and the subsequent ARQ transmissions. usually implemented using the inverse fast Fourier transform (IFFT). or if a maximum number of ARQ transmissions has been reached. we propose a heuristic scheme. see [11] and [12] for a recent review and also [13] and [14] for applications of majorization theory. CSI was also employed to develop an ARQ scheme based on Chase combining. model with coefficients B. The th element of matrix is denoted by and the th element of vector is de. This paper is organized as follows. where entries [19]. Conclusions are drawn in Section IX. In Fig.) circularly symmetric complex additive white Gaussian noise (AWGN) with zero mean and unit variance. The identity matrix is denoted as . the PT-OFDM symbol is transmitted. The ARQ round ends if all the data sent so far have been recovered. respectively. An OFDM system with M subcarriers. Similar results have been obtained recently in the context of relay-assisted communications [8]. the samples of the received signal corresponding to the cyclic prefix are removed. We consider a system with symbols. Finally.i. each of unit power. then data are transmitted on subcarriers that experience independent channel fades. [9].d. we describe the general OFDM system with ARQ-SA and consider a number of utility functions that characterize system performance. and and Hermitian. all . 56. The assignment problems for single ARQ-SA and multiple ARQ-SA are formulated in Section III. These choices are prevalent in current wireless systems [1]–[3]. may either carry data or reWe shall see that each symbol dundancy that is used to improve the probability of detecting previously failed transmissions.

Redundancy for the original data symbols is sent generally by using (a) incremental redundancy symbols or. simulation results in Section VIII show that substantial performance can already be achieved. Here. In general. C. The set of indexes of the ARQ subcarriers assigned to orig. Clearly. Recall that and are the sets of subcarrier indexes used by the original DSs and incremental and . we refer to these RSs RSs. If the original DSs or ARQ DSs are still not recovered after the first ARQ transmission is sent. the ARQ transmission is triggered. as a special case. 2(a). Clearly. subcarriers used for a common purpose are grouped for clarity. To initiate another ARQ round.. The set carries ARQ DSs that are used carries ARQ RSs that are used as to send more data. extensions to multiple OFDM symbols are straightforward and are not treated in this paper. we have is commonly considered in the literature and is covered in our formulation. 2. With this restriction. RSs composed of with size are sent as redundancy for the original DSs. in the original transare used to mission. ). Incremental RSs for Original Data Symbols We consider how to assign incremental RSs (and full RSs as a special case) to original DSs. each symbol in (1) is used for transmission either as a data symbol (DS) or as a redundancy symbol (RS). e. Suppose that where at least one bit in is not received correctly and ARQ is trigcarried by original subcarrier is then gered. The set . original DSs composed of send data. past transmissions are discarded from memory. Transmission structure for the original and the first ARQ transmission. an independent original transmission is sent.: ARQ BY SUBCARRIER ASSIGNMENT FOR OFDM-BASED SYSTEMS 6005 Fig.g. (b) full redundancy symbols. This scenario redundancy is used. We call th ARQ subcarrier. as depicted in Fig. We assume that each transmission uses one OFDM symbol with subcarriers in the set . In general. with full redundancy as a speeral (with arbitrary . transmissions are denoted as the th original subcarrier and the respectively. by a straightforward generalization of Fig. In the ARQ transmission. 2(a). redundancy for the ARQ DSs. [15]–[17] and [20]. henceforth we allow at most one ARQ transmission to be sent. the remaining ARQ transmission are then split into two disjoint sets of size and of size . but they need not be neighboring subcarriers. The ARQ schemes cial case (with in the literature typically employ only full redundancy. respectively. SNRs of the original and ARQ subcarrier. If specifically as full RSs and Fig. The received signal (1) can be expressed on a per-subcarrier basis as (2) is the th row of the transform .HO et al. We refer to these RSs generally as incremental and thus . let us first focus on Fig. These ARQ subinal subcarrier is denoted as carriers are received as (3) . a second ARQ transmission consisting of more RSs and DSs can be sent. The signal repeated in the assigned ARQ subcarriers in the ARQ transmission. Fig. 2(b) as a spesubcarriers in the cial case. although the redundancy is referred to as incremental redundancy. hence . 2(a) becomes Fig. assuming full for . respectively. 2(a). Specifically. When at least one bit error octhe size of the set is curs in the original transmission. The channel RSs. 2 shows the transmission structure of an ARQ round consisting of the original transmission and the first ARQ transmission. all and subcarriers should be used for transmission. For exposition. where coefficients in the frequency domain in the original and ARQ and . For clarity of presentation. Hence. ARQ subcarrier by and are also the Since the noise variance is set to one. respectively. For a time-invariant channel. by feeding back a negative ACK (NACK) to the transmitter. our use of incremental redundancy is more gen). where is the set of subcarriers used. We denote the power of the original subcarrier and and . for example. better performance would be achieved with more ARQ transmissions. In this paper. the transmissions within any ARQ round are independent of those in other ARQ rounds.

We see that both the BER and BLER increases. we define the utility functions with respect to the original DSs. It is common to use the zero-forcing (ZF) or minimum mean-square-error (MMSE) equalizer. assuming that is i. we only need to replace . The assignsigned to original DSs via the assignment ment of ARQ RSs to ARQ DSs is carried out in essentially the same way. the effective SNR subcarrier remains the same after ZF or MMSE equalization. DECEMBER 2008 where . We detect based on . and a to ARQ DSs by replacing is used. see Fig.6006 IEEE TRANSACTIONS ON SIGNAL PROCESSING. When entries. We seek to maximize these vector with elements utility functions by appropriately choosing the subcarrier assignment. such as the maximum likelihood detector (MLD). as a block. 12. D. the signal before the slicers . For ZF equalization. Utility Functions We describe several utility functions commonly used to reflect system performance. but with the “retransmission” always triggered. is the minimum of the effective SNR over all subcarriers. iterative detectors. it follows that the effective SNR of in the frequency domain is given by summing the SNRs of the original subcarrier and the assigned ARQ subcarriers. Both are appropriate measures to maximize since error probabilities typically decrease as SNR or is unitary with constant-amplitude SINR increases. The block error rate (BLER) is defined as the probability that at least one bit error occurs in a block. to reflect the new utility funcwith and with . This is appropriate if each block uses a separate error detection code. We use quadrature phase-shift keying (QPSK) modulation throughout this paper. be the equalized signal before slicing. VOL. It can be expressed as is a linear equalizer. The problem remains essentially the same if we instead maximize the utility with respect to the ARQ DSs. where subcarrier) is thus given by . The uncoded symbol error performance is often dominated by weak subcarriers experiencing deep fades. E. given. since the problem of assigning ARQ RSs to ARQ DSs is similar. since these equalizations involves only a scalar multiplication. Henceforth. Two important measures of performance are the BER and BLER. where is denoted as . To give an intuitive explanation of our subsequent remize . we saw that incremental RSs are as. the effective SNR should be made as flat as possible across the subcarriers. and a block is discarded when any bit error occurs. Details are provided in Appendix A. In particular. Various detectors can be used. by (6a) (6b) where diag . we will make frequent use of and in (4) after MRC of the mutual information between is performed. The BLER is given by 1 1 . since there are 2 bits in an OFDM symbol with QPSK modulation. indeed decrease monotonically as for 3) OFDM: For OFDM systems. or linear-equalizer based detectors. Gaussian distributed. so that the subcarrier assignment problem remains essentially unchanged. respectively. An appropriate measure to minimize is the expected BER. In (7b). We refer to all the original DSs and their redundancy symbols. 1) OFDM and PT-OFDM: For OFDM and PT-OFDM systems. 2) PT-OFDM: For PT-OFDM systems. the noise after equalization is Gaussian distributed. (2)–(5) can be easily generalized to any arbitrary number of ARQ transmissions. we obtain (see. it sufnew assignment in place of fices to consider the problem of assigning incremental RSs to original DSs as considered in Section II-C. which gives (5) Besides the assignment carried out via . tions. we employ MRC on all the to give received signals that carry (4) Since the noise in all received signals is independent. [19]) the SNR and SINR for subcarrier as (7c) (7d) respectively. We note that the SNR or SINR is independent of . NO. To detect . at the receiver. we may use either the ZF or MMSE equalizer (6). for example. To illuminate this problem. as a function of the effective-SNR from (5). we note that the and effective SNR depends also on the subcarrier sets used for (re)transmissions. For linear-equalizer based detectors. to simplify implementations. Redundancy for ARQ Data Symbols In the previous section. 1. respectively. This is because. is the sum sults. The SNR of Let after ZF equalization and the signal-to-interference noise ratio (SINR) after MMSE equalization is denoted as and . 56. we consider these utility functions (7a) (7b) In (7a). or all the ARQ DSs and their redundancy symbols. one way to do this is to maxi. We indicates the number of bits that can be reliably note that transmitted with a Gaussian codebook if ideal channel coding is carried out. To reduce the effects of fading.d. (2)–(5) apply for the case of assigning ARQ RSs with and with . Although we consider at most one ARQ transmission.i. and the BER (on any .

we impose the condition that each original subcarrier is either assigned to one ARQ subcarrier or not assigned at all. say. if otherwise. henceforth we add subcarriers with zero SNRs to the set of ARQ subcarriers . we pad with zeros so that its length becomes . the choice of these subcarrier sets can be predetermined to optimize system performance for a given average SNR. the imposed condition is equivalent to setting the row sums of to one. We assume that the CSI in the form of the SNRs of the original and ARQ subcarriers is known to the transmitter. carrier is assigned to original subcarrier . more advanced techniques. The ARQ-SA is completely we define matrix with binary entries . and so becomes an square matrix. with SNR Under single ARQ-SA. such as MLD or interference cancellation. Since is a binary matrix. By including virtual subcarriers.. we defer problem formulation holds for arbitrary the choice of the subcarrier sets to Section VIII. to keep the receiver processing simple. as much as possible. respectively. . Together with constraint (9). In practice. which is used in long-term 3G evolution systems [3]. while the subcarrier assignment is optimized during run time whenever small-scale fading leads to changes in the channel. described by an Hence. Hence. This is equivalent to maximizing the probability that all the bits in the block are successfully detected. In (8). summed over all . We see that each ARQ subcarrier. Hence. i.HO et al. the optimization problem becomes the following. is assigned to one original subcarrier. single ARQ-SA and multiple ARQ-SA. would . To improve system performance and to provide a more general framework. in the multiple ARQ-SA scheme we allow zero. Instead. We shall see in Section V that this condition simplifies an otherwise NP-hard problem to a problem that has an optimal solution with polynomial-order complexity. As an example.e. i. to give an effective SNR of . the imposed condition under . say. clearly each ARQ subcarrier should be assigned to at least one original subcarrier. we do not allow multiple ARQ subcarriers to be assigned to any original subcarriers. (10) From (9) and (10). with SNR . The utility function to maximize is then the negative of this measure. Since our . we define if ARQ sub. such as in IEEE 802. For consingle ARQ-SA implies that we must have virtual venience of description. 3(a). 1) Single ARQ-SA: In single ARQ-SA. Equivalently. Hence. PROBLEM FORMULATION We consider the problem of finding the optimal subcarrier as(this signment for a given choice of subcarrier sets choice also fixes the parameters ).e. ARQ-SA Schemes For maximum performance gain. it is reasonable to assume that the CSI is known. (9) Two ARQ-SA schemes are possible depending on whether multiple ARQ subcarriers can be assigned to an original subcarrier. Problem Single ARQ-SA: Find that solves maximize subject to where III.e.. in all our schemes each ARQ subcarrier is assigned to exactly one original subcarrier. we explicitly. and single ARQ-SA reduces to finding an optimal permutation (not necessarily unique) that maximizes the utility function. one. the utility function is (7e) We can alternatively minimize the BLER. while We now describe two ARQ-SA schemes. are optimized via simulations. otherwise. Briefly. we could assign multiple ARQ subcarriers to boost the performance of an original subcarrier that is very weak. (11) on 2) Multiple ARQ-SA: If an original subcarrier is already very strong. an ARQ subcarrier cannot be assigned to two or more original subcarriers. we should not assign any ARQ subcarrier to it. To employ MRC directly.. this constraint is equivalent to imposing the condition . namely. or by an explicit CSI feedback. or more ARQ subcarriers to be assigned to an original subcarrier.11a system [1]. with high probability the data would be recovered. That is. each retransmitted symbol experiences an uncorrelated channel. given by (7f) be required for detection. The CSI can be estimated at the transmitter when the channel is reciprocal. if the time duration between an original transmission and its ARQ transmissions is less than the channel coherence time.: ARQ BY SUBCARRIER ASSIGNMENT FOR OFDM-BASED SYSTEMS 6007 . the effective SNR (5) can be written in vector form as (8) where are the vectors of emphasize the dependence of B. is therefore the permutation matrix. i. A. ARQ Subcarrier Assignment (ARQ-SA) To simplify the description. we have chosen the subcarrier sets such that. Note that an original subcarrier is not physically assigned to any ARQ subcarrier if the ARQ subcarrier turns out to be a virtual subcarrier. In practice. To this . consider Fig.

3. The effective SNR is the sum of the SNR of the original subcarrier and all SNRs of the assigned ARQ subcarriers . 12. 56.6008 IEEE TRANSACTIONS ON SIGNAL PROCESSING. DECEMBER 2008 Fig. VOL. NO. An ARQ subcarrier n is assigned to an original subcarrier m if a .

• order decreasingly to obtain . Algorithm 1 is given as follows. second. and no ARQ subcarriers are assigned to the first. and third original subcarriers. In Algorithm 2. For initialization. where a . the strongest ARQ subcarrier that has not been assigned so far is assigned to the original subcarrier with the smallest effective SNR. we expect that the minimum effecis increased as compared to random pairings of tive SNR ARQ and original subcarriers. • set • order decreasingly to obtain . Iteration : • assign ARQ subcarrier i. otherwise Algorithms 1 and 2 are clearly equivalent. we set the effective SNR as that of the original subcarriers. Consequently. two. multiple ARQ-SA applies for any . The ordering among subcarriers with the same value is arbitrary (this does not change the effective SNR nor the utility function). • order increasingly to obtain . Initialization with inputs : • set and . Problem Multiple ARQ-SA: Find that solves subcarriers. Notice that Algorithm 2 imitates Algorithm 1 so as to maximize the effective SNR in a greedy manner. where is the ordered index. where is the ordered index. single ARQ-SA. B.. set . and then discuss their complexity. After assignment. Without the constraint of (10). we remove the constraint (10) under multiple ARQ-SA. Iteration : • find smallest effective SNR in . where is the ordered index. Initialization with inputs : . where (12) In this case.e. respectively. . Algorithm 2 For solving problem multiple ARQ-SA. Algorithm 1 For solving problem single ARQ-SA. so that . In each iteration. Algorithm 1 produces effective SNRs that do not fluctuate significantly across . to original subcarrier . one. A. we assign the th strongest ARQ subcarrier to the th weakest original subcarrier for all . while Algorithm 2 is suboptimal. Algorithm 1 for Problem Single ARQ-SA In Algorithm 1. the effective SNR is updated in each iteration to include the contribution from the additional ARQ subcarrier. subcarrier by subcarrier. (b) Multiple ARQ subcarriers can be assigned to each original subcarrier. (a) Only one ARQ subcarrier is assigned to each original subcarrier. Algorithm 1 is optimal. and . which can be used only if An example is given in Fig. we explicitly allow multiple ARQ subcarriers to be assigned to an original subcarrier. Then. so that . a detailed discussion on their optimality will be given in the Section V. =1 =1 end. so that . unlike for Thus. Algorithm 2 for Problem Multiple ARQ-SA maximize subject to where (13) We iteratively assign ARQ subcarriers to the original subcarriers. under multiple ARQ-SA the optimization problem becomes the following. respectively. ALGORITHMS FOR ARQ-SA SCHEMES We begin by providing Algorithm 1 and Algorithm 2 to solve problem single ARQ-SA (11) and problem multiple ARQ-SA (13). 3(b). and denote its index as IV. By pairing strong ARQ subcarriers with weak original subcarriers. we order the original subcarriers increasingly and the ARQ subcarriers decreasingly according to their SNRs.

We first introduce the notions of majorization and Schur-concavity. The actual assignments are linear in complexity. be increasingly ordered such that Theorem 1: Let and be decreasingly ordered such that . To test for Schur-concavity. 2) Algorithm 2: The initialization requires one sorting operation for . involving a simple recording of the assignment solution in memory. Definition of Majorization: For any . Theorem 2: Algorithm 1 optimally solves ARQ-SA for the utility functions (7). practically no complexity is required in determining the assignment during run time. It follows from the definition of Schur-concavity that (17) holds.A. at most 1 comparisons are resorted if we exclude . Theorem 1 shows that. which we have considered in this paper. has already been sorted except for ately. 1In implementation. we have . Theorem 2 particularizes this result to the utility functions (7). We now consider the complexity of the first step of iteration in Algorithm 2. we can assume that the original subcarriers are ordered increasingly based on their SNRs. for of ordering or sorting example. so the weakest subcarrier can be found immedi. For been sorted. Proof: By using Theorem 1. In Algorithms 1 and 2.: ARQ BY SUBCARRIER ASSIGNMENT FOR OFDM-BASED SYSTEMS 6009 • assign ARQ subcarrier to original subcarrier . if (14a) (14b) denotes a decreasing ordering such that . we show that for utility functions (7). the subscript on (15) C.4]: Let nonempty convex set and let be continuously differentiable. By definition. to find the smallest effective SNR in . since it is a fixed assignment independent of the CSI. has already and incur only a linear complexity. 1) Algorithm 1: The initialization requires two sorting operations for and . For a Schur-concave function (17) which are permutations of . we only need to remove and appropriately insert1 it back to . set • update the effective channel power as . the following two lemmas in [10] can be used. Without loss of generality. Ch. the assignment has to be recomputed during run time. 6. then is Schur-concave on . where the worst case complexity for each insertion . we say that is majorized by (or majorizes ). To this end. Algorithm 1 has a complexity of . Lemma 1 [10. we need a few results from the theory of majorization [10]. We note that the complexity the complexity scales with items is by using. Complexity For the cyclic assignment considered in the literature [15]–[17]. Hence. which is equivalent to Algorithm 1. Our main result can then be summarized as follows. Since there are a total complexity of is required. To resort . denoted as . Ch. Lemma 2 [10. Since the vector is already . Algorithm 1 solves problem single ARQ-SA (11) optimally. the optimal single ARQ-SA is to assign decreasingly ordered ARQ subcarriers to the original subcarriers. The function is Schur-concave on if and only if is symmetric on for all (16a) (16b) where denotes the partial derivative of with respect to . Algorithm 1 for Problem Single ARQ-SA In order to prove optimality. Proof: From [10. we have for any which is a permutation of . this insertion operation is carried out by updating a list of pointers that tracks the orderings.A. Alternatively.e. namely. A. 3.2]. It is thus important to consider this assignment comand consider how plexity. Using Lemma 2. Definition of Schur-Concave Functions: A real-valued funcis said to be Schur-concave on tion defined on a set if Here. the remaining steps of iteration are less complex . whenever the channel changes. For that has just been updated in the prethe effective SNR vious iteration (corresponding to the third step of iteration in Algorithm 2). merge sort [21]. It is well known that functions.2]: If is symmetric and concave on . i. 3. We also make comments on the suboptimality of Algorithm 2 for solving problem multiple ARQ-SA (13). a function is symmetric if it is invariant under permutation of its variables. we initialization require a smaller complexity of . Ch. it is sufficient to show that the functions (7) are Schur-concave.HO et al. be a symmetric. conclude that Algorithm 2 has an overall complexity of V. it then follows that these functions for any . OPTIMALITY OF PROPOSED ALGORITHMS In this section. We note that (7) are all symand are concave metric functions. for a Schur-concave function . a Schur-concave function defined on vectors from a set achieves its maximum at a vector that is majorized by all other vectors in the set. Suppose that in the initialization we also sort (equals ). even after taking into account the number is given by of comparisons and shifts required to adjust the storage of the 1 iterations that require re-sorting. output.C. As the sortings in the .. This insertion is similar to quired to appropriately insert the insertion operation in the sorting algorithm insertion sort [21]. Thus. we let .

when the channel has sufficiently changed. It can be easily verified that for all utility functions (7). let the amount of signalling required for Algorithms 1 and 2 for . Using standard calculus. Amount of Signalling Required For Algorithm 1. Since there are original subcarriers. while in problem multiple ARQ-SA. WE FIX M M =M =N VI. Obviously. suboptimal for all utility functions (7). This is analogous to problem multiple ARQ subcarriers with SNRs to ARQ-SA: assign original subcarriers. . multiple ARQ-SA is at least NP-hard. if we interchange the first two and so . Table I shows ARQ-SA. (FSO) given by where for single ARQ-SA and for multiple . By using Stirling’s formula [23]. For simplicity. and so for all Schur-concave functions. we define . bits of signalling are required to communicate a chosen assignment. . In [22]. is therefore not likely to be optimal for complexity of solving problem multiple ARQ-SA. and . each ARQ subcarrier can be assigned possible to any original subcarrier. Using Lemma 1. the difference in FSO . acrows of (12). and are thus Schur-concave functions. we relate this NP-hard for the utility function problem with a simpler problem that is known to be NP-hard. . . Hence. required to communicate a chosen assignment. hence more efficient algorithms are desirable. Our problem is therefore harder.6010 IEEE TRANSACTIONS ON SIGNAL PROCESSING. A useful measure of the overhead used is therefore the fractional signalling overhead in bit per subcarrier per update. Hence. the channel is quasi-static over a period of time. We remark that the longest processing time first algorithm considered in [22] is a special . Independent of the mechanism actually used. A. We use base 2 for all logarithms. are all symmetric functions of . . we propose a method to reduce the amount of signaling information required to convey the assignment. This implies that knowing the optimal solution for our problem allows the task allocation problem to be solved. Typically. the original subcarrier cannot be treated identically for some . both are equivalent if A numerical counterexample confirms that Algorithm 2 is . The assignment needs only to be updated every OFDM symbols (say). For many applications. . NO. Hence. 56. 12. Table I numerically confirms that as increases. Algorithm 2. (excluding virtual subcarriers) and there are in total possible permutations. but not vice versa. We consider the saving in FSO for using Algorithm 1 compared to Algorithm . In the task allocation problem. each ARQ subcarrier is assigned to a difARQ subcarriers ferent original subcarrier. . Let . the following task allocation problem is considered: tasks with processing time to processors. To do so. we obtain the saving as Since (18) for large . Clearly. the partial derivatives of the remaining utility functions (7) with respect to are TABLE I AMOUNT OF SIGNALLING REQUIRED IN BIT PER SUBCARRIER FOR ALGORITHMS 1–3 AND VARIOUS NUMBER OF SUBCARRIERS . DECEMBER 2008 are Schur-concave. so that the minimum SNR across all effective SNRs is maximized. VOL. GROUPING OF SUBCARRIERS The assignment obtained by Algorithm 1 or Algorithm 2 can be determined by the transmitter and made known to the receiver. since their SNRs may not be the same.44 with solute amount of signalling is still quite large for Algorithm 1 and Algorithm 2. we 2 for large and so have . we see that the abapproaches 1. B. so assign that the minimum processor time required to complete all the tasks is maximized. is on the order of hundreds in wireless local-area network applications. we have . Using Algorithm 2 leads to an ARQ-SA given by (12). However. case of Algorithm 2. Clearly. Here. which can be implemented with a . the processor are identical. There are in total bits of signalling are assignments. . . Since the task allocation problem is NP-hard [22]. we obtain cording to the definition of majorization. however. Nevertheless. If our problem is exactly equivalent to the task allocation problem for any . This shows which approaches that using Algorithm 1 can lead to a significant saving. or determined by the receiver and made known to the transmitter. For Algorithm 2. and thus we conclude that Algorithm 2 is suboptimal. Algorithm 2 for Problem Multiple ARQ-SA We now establish that problem multiple ARQ-SA is at least . so it can be easily verified that (16b) is valid.

2(a). becomes Second level of assignment (independent of channel): • For each group of ARQ subcarriers assigned to a group of original subcarriers. and a fixed (arbitrary) assignment for the second level of assignment typically gives good BER performance. For simis divisible by . . Performing assignment based on these grouped subcarriers would likely result in only negligible performance loss. the minimum SNR. We note from Section II-B that the transmissions in one ARQ round are independent of other ARQ rounds. The throughput can hence be determined as (19) with probability one (20) which approaches 1 1 for large . we observe that the saving increases when but increases relatively slowly (due to the logarithm) with quickly with even for small . a second (deeper) level of assignment is carried out for every subcarrier. and hence the renewal-reward theorem [20]. For each group. of for which numerically supports the above conclusions. is i. e. respectively.. we assume By definition.: ARQ BY SUBCARRIER ASSIGNMENT FOR OFDM-BASED SYSTEMS 6011 B. The original transmission and ARQ transmission over time. This level of assignment may be performed dynamically during run time depending on the CSI.. If is much smaller than the coherence bandwidth of the channel. after jointly decoding from original DSs and incremental RSs. Simulations replicity. Let be the number of bits recovered in ARQ round .e.HO et al. Let: be the event that at least one original DS is erro• original DSs (equivalently neous. Although the technique of grouping applies to both multiple ARQ-SA and single ARQ-SA. Without loss in generality.e. if a NACK is received. Method of Grouping To reduce the number of possible subcarrier assignments. . That is. . one may use additional redundancy bits. By using Stirling’s formula again. We consider the saving in FSO with grouping compared to no grouping for large .d. 4. For each group of ARQ subcarriers that has been assigned to a group of original subcarriers. arithmetic mean. as follows. assign the ARQ subcarrier with the th smallest subcarrier index to the original subcarrier . The random variables t and T denote the end and the period of ARQ round i. after jointly decoding from ARQ RSs. . by replacing in with . we assume that veal that using the minimum SNR as the group-equivalent SNR.i. respectively. . we use a group-equivalent SNR to represent the SNR of the group. we get . Hence. For large . with the th smallest subcarrier index for Algorithm 3 reduces to Algorithm 1 when . be the event that at least one new data symbol • ARQ DSs and is erroneous. Algorithm 3 For solving problem single ARQ-SA. 4. The takes the role of a reward for utilizing throughput is given by the total number of recovered bits normalized by the total time spent. we propose Algorithm 3. replaced by their • Apply Algorithm 1. but with group-equivalent counterparts. VII. consider . . but at the expense of incurring additional signalling bandwidth. We denote the amount of signalling required for Algorithm 3 . and hence a renewal process. First-level of assignment: • Group the original and ARQ subcarrier into groups of contiguous subcarriers. after decoding from the event that the original transmission fails). . which units of time. THROUGHPUT This section obtains the throughput of the ARQ system based on the transmission structure shown in Fig. where . We recall from Section II-B that an ARQ round consists of an original transmission and an additional ARQ transmission if the original transmission is erroneous. [24] applies. From Table I. we group contiguous subcarriers and apply Algorithm 1 or Algorithm 2 on these groups. by matrix. the discrete time at the end of ARQ round . Each group uses the minimum SNR in the group as the group-equivalent SNR. Table I shows the values The throughput can now be computed if the expectations of and are known (the indexes are dropped for brevity). and using AlgoFor illustration. it is convenient to define the following error events. i. we obtain the saving as Fig. rithm 3 with To offer a strong error protection for the signalling bits. be the event that at least one original DS is • erroneous. For example. i. Let . over an infinite time horizon. we focus on the latter since our main motivation is to reduce the amount of signalling.. Let be . The duration of ARQ round is given by . in Fig. or geometric mean of the group. Since the second level of assignment is fixed and does as not require signalling. where thicker vertical lines mark the boundaries of the ARQ rounds. An example of the transmissions over time is shown in Fig. 4. the FSO is . As expected. but overall the amount of overhead remains small. To this end. Since the data symbols transmitted in different ARQ rounds are decoded independently. that each transmission takes one unit time.g. the subcarriers within each group have approximately the same SNR. the saving is zero when no grouping is carried out.

Below all possible events that result in the termination of the ARQ round. then one should transmit ARQ DSs to achieve an overall gain in throughput. For example. and we use the minimum-SNR utility function in (7a).org/software/glpk. curs. VIII. event . if Fig. 2. This implies that not sending any data in the ARQ transmission can limit throughput. . The relationship can then be represented as shown in Fig. If approaches zero. the . [15]–[17] and [20].i.gnu. full redundancy indeed poses as a severe limitation to the significant throughput gain that can otherwise be achieved with incremental redundancy..d. 12. Algorithms 1 and 2 are also implemented. and . if we employ incremental redundancy so that . all the bits in the block are recovered. the original transmission is erroneous with probability . which is typical when the average SNR is high.g. NUMERICAL RESULTS We first use a case study to illustrate the difference of our proposed algorithms with optimal solutions. we demonstrate that at high SNR. exponential distribution and are shown in Fig.e. while the ARQ subcarriers are sorted decreasingly. while the right-hand side reflects the throughput lost by the original DSs. A. 6 shows the resulting minimum effective SNRs. via the error events. This scenario throughput approaches the upper bound occurs if the effective SNR is high after applying Algorithm 1. For brevity. on the average and on the size of the subcarrier sets SNR . In the simulations in Section VIII. throughput to exceed . DECEMBER 2008 We now consider the scenario that full redundancy is used. 5.e. . ocGiven that the original transmission is erroneous. All the above four approaches clearly give a significantly higher minimum effective SNR compared to the minimum SNR of the original subcarriers (without 2http://www. For clarity. values: we obtain (21) (22) So.. which is already the largest possible must exceed throughput with full redundancy. it can be easily shown that the throughput (23) . for which the ARQ round then terminates with bility . otherwise. we have employ error detection on a per block basis.e. If the gain is larger than the loss. be the number of bits transmitted in each data symbol. Let . 5. can then take four possible or 0. . Probability tree over the original and ARQ transmissions in an ARQ round. we provide (T. Depending on whether mission will be sent. NO. The of and ARQ round begins with an original transmission. This original occurs) with probatransmission is successful (i. we leave out the details. e. 56. (25) is satisfied. Then. Specifically. the left-hand side of (25) reflects the throughput gained by the ARQ DSs after the ARQ transmission. we formulate them as mixed-integer linear programs (MILPs) and solve these MILPs using GNU Linear Programming Kit. This is a typical case considi. original DSs are successfully decoded in the original transmission. it is possible for the some ARQ DSs are sent. The complementary event is the event that all the of is denoted as . the condition (25) can be explained as follows. From Fig.. 5. Case Study To better illustrate the case study. . we drop these arguments. none of the bits in a block (of original DSs or ARQ DSs) is considered to be recovered if at least one of the data symbols is erroneous. Intuitively. i. VOL. 6. i. ..e. Fig. s=n ) indicating the duration and normalized throughput of that ARQ round. On the other hand. On the other hand. To solve problem single ARQ-SA and problem multiple ARQ-SA optimally.6012 IEEE TRANSACTIONS ON SIGNAL PROCESSING. The throughput is then upper bounded as (24) since . hence occurs and whether occurs. we evaluate the solutions numerically for a small number of subcarriers with . and an ARQ trans. or 3. the original subcarriers are sorted increasingly according to their SNRs.2 a standard linear programming software.. Since we Since QPSK modulation is used. 5. The (arbitrary) SNRs of the original and ARQ subcarriers that we consider are generated independently from an i. ered in the literature. we show the improvement in block error rate and throughput achieved. as shown in Fig. This dependence is written explicitly in (23). the throughput (20) becomes (23) The throughput depends.

while the additional improvement offered by other solutions is relatively small. For multiple ARQ-SA. In our simulations. subcarriers for 2) Block Error Rate (BLER): We consider these BLERs that are used to calculate the throughput in (23): • . i. Surprisingly. 7.d. this conclusion holds typically. For clarity. simulation studies (not shown here) indicate that it also results in good error performance generally when used with other detection schemes. Since i. When . First. 6. we use a (fixed) cyclic assignment considered in the literature [15]–[17]. the BLER for the original DSs without using ARQ. For benchmarking. where an AWGN channel is used.3 as expected due to Theorem 1. the BLER for the ARQ DSs. similarly we select is made up of the that are spaced uniformly apart. Moreover. B. time domain channel taps. 2(a). Specifically. remaining subcarriers. so the transformation is equivalent to rotating the data symbol is unitary and has conin the time domain. Both algorithms perform subwhen ARQ is not used. . The transform stant-amplitude elements. For 3The other effective SNRs can be different because the utility function  only considers the minimum SNR.. Details of both bounds are provided in Appendix B. the weakest subcarriers are already significantly boosted by the strongest subcarriers using Algorithm 1. DSs. respectively. Although it has been designed for maximum likelihood decoding in [25]. ARQ is always activated. Unlike [18]. Algorithm 1 and Algorithm 2 are only about 0. We model the wireless channel i. shown as dotted lines) after applying Algorithm 1 and Algorithm 2. which is expected as fewer constraints have been imposed.5 dB away from the averaged MFB bound. The channel with used for transmission is assumed to be time invariant in each ARQ round.HO et al. Finally. based on the same frequency-selective channel model. Notice that the FFT and the IFFT cancel out. Hence. ARQ . which allows each DS and its corresponding RS to experience channels that are close to independent. given that the • original transmission fails. the SNRs of the original subcarriers are ordered increasingly. lower bounds are considered . stantially better than the BLER and provide more than 2 dB of SNR gain compared to using a cyclic assignment. • . The ARQ subcarriers are classified as incremental RSs. see Fig. illustration. 2(b). any further gain by using a more sophisticated algorithm is likely to be small. is not relevant here..e. or ARQ RSs via subcarrier sets as subcarriers see Fig. Since much of the throughput gain comes from the original DSs. This is because for wireless channels that are frequency selective. the MILP solution is marginally better than Algorithm 2. We note that using Algorithm 1 already improves the worst case SNR gain significantly. the MILP solution gives the same minimum effective SNR as Algorithm 1. obtained from using Algorithm 1 is althe BLER most the same as Algorithm 2. we employ the AWGN bound.i.e. We use the transform [25] (26) and is before applying IFFT. based on typical SNR realizations of original subcarriers and ARQ subcarriers (shown as full lines). we do not reserve stronger ARQ subcarriers only for . Moreover. From the as subcarriers remaining subcarriers. : We first consider that full redundancy is employed. ARQ). the smallest effective SNR based on MILP (at subcarrier 6) is slightly higher than that based on Algorithm 2 (at subcarrier 3). ARQ DS is not sent. the BLER for the original DSs by using ARQ. we cyclically shift the indexes of the ARQ subcarrier with respect to the original subcarriers by 16 subcarriers. since this will result in weaker and . The effective SNRs (not in decibels. Monte Carlo simulations are used to obtain the BLERs for Algorithms 1–3. while the ARQ subcarriers are ordered decreasingly. Although we study a particular case here. In this cyclic assignment. we use the averaged matched filter only for bound (MFB) to provide a lower bound. we consider a QPSK modulated system and a PT-OFDM system with subcarriers at different average SNR . the original and ARQ subcarriers do not experience fading. Performance Evaluation 1) Scenario: For our performance evaluation. Finally. where the FFT matrix. so that these subcarriers experience independent channels as much as possible. we use the ZF equalizer given by (6a). We assume that there is no interference from other data symbols and hence a matched filter is used for detection. even though only a simple ZF equalizer has been employed. Secondly. in which CSI is not exploited. For single ARQ-SA. more advanced detectors such as MLD [25] or iterative subcarrier reconstruction [19] can also be used. Lower bounds based on idealistic conditions are used to check how good our schemes perform. we select that are roughly spaced uniformly apart. From Fig. We assume an error-free channel is available to signal the ACK bit and the assignment used. ARQ is always activated. we observe that using Algorithm 3 with grouping of achieves a performance that is only a few tenths of decibels away from that of Algorithm 1 (without grouping).: ARQ BY SUBCARRIER ASSIGNMENT FOR OFDM-BASED SYSTEMS 6013 Fig. as analytical expressions for the BLERs involve order statistics and often do not yield closed-form results. moreover. Algorithm 2 performs (slightly) better than Algorithm 1. and .

yet we can additionally send ARQ DSs with an error probability of only around 0. for the ARQ DSs ARQ DSs and ARQ RSs. 8. using grouping of can be a good compromise. which gives good performance with small overhead. 56. : Next. To obtain further improvement at high SNR. 12. see Fig. 9 illustrates the throughput obtained with and without ARQ. Since . In our simulations. We now optimize these So far we have fixed parameters with the use of incremental redundancy to maxi- Fig. M N M N . To improve throughput in this regime. compared to in Fig. 2(a). This sugis adequate for this scenario. For the multiple ARQ-SA scheme. the original DSs are received with low error probability. CONCLUSION We propose two ARQ schemes based on subcarrier assignment for general OFDM systems with a possible pretransform. we vary the parameters in steps of four and obtain the maximized throughput as shown in Fig. With full redundancy. we propose a suboptimum algorithm. can be realized at high SNR. VOL. DECEMBER 2008 Fig. Fig. the BLER is about 1 dB away from Algorithm 1 but is still about 1 dB better than using cyclic permutation. However. 7. 8. BLER performance of the original DSs using incremental redundancy. we consider in Fig. We now consider the BLER by using Algorithm 3 with . should be This implies that the number of incremental RSs reduced. where = 24. 9. In particular. gests that using grouping of if further reduction of signalling is desired. incremental redundancy is strictly used (instead of full redundancy). The performance of Algorithm the original DSs 2 is very close to that of Algorithm 1 and is not shown. where the optimized parame. = 32. At a sufficiently high SNR of 12 dB.06. but this has little effect at high SNR since the original DSs can usually still be recovered. full redundancy should preferably be used to ensure reliable packet recovery. an option is to increase the number of ARQ transmissions to more than one. We observe that a signifiters are indicated as cant gain. when redundancy is limited. Fig. the proposed algorithms become more important in maintaining a reasonable system performance. . these We transmit parameters have been optimized to maximize the throughput. 8. these performance gaps between different algorithms widen for in Fig. NO. BLER performance using full redundancy. IX. = 8. is also shown. Hence. Throughput using Algorithm 3 with subcarrier grouping of G = 2. for example. the gaps between cyclic permutation and proposed algorithms are more significant. 9. as explained in Section VIII-B-3). We focus on using Algorithm 3 with . incremental redundancy must therefore be used. we note that this improvement is limited by the upper bound given in (24) for SNR dB. we propose an optimum subcarrier assignment algorithm that optimizes the class of Schur-concave utility functions. For the single ARQ-SA scheme. 8 shows that using more signalling for the assignment leads to better performance. 7. However. but at the expense of incurring higher delay. We first consider the BLER for . On the other hand. in the low-SNR regime. In order to .6014 IEEE TRANSACTIONS ON SIGNAL PROCESSING. 3) Throughput: Fig. M = N = 64. = 64. 7. As in Fig. we observe that a substantial improvement is obtained compared to when ARQ is not used. mize the throughput for SNR dB. An illustrative case of the BLER Pr( j ) when with ARQ DSs are sent. up to about 4 dB.

Chen. Thus. Feb. 43. K. Vancouver. we assume that ARQ is always activated and we take the expectation of the error probability over the fading channel. [12] E. COM-25. Wolter. Falconer. R. 56th IEEE Veh. 1979. Consider a unitary transform with constant-amplitude ele. can be . Commun. Wicker.. Wu. 3. A matched filter is used riers with SNRs at the receiver. Mag. (CrownCom). 6. 1999. [7] D. Jun.. Jul. Orlando. which is fixed treated as a single fixed term. Boche.org/ftp/Specs/htmlinfo/36211. For QPSK modulation. and S. our ARQ schemes can be generalized easily to any number of retransmissions. Signal Process. pp. Benyamin-Seeyar. 51.. Tolhuizen and S. suppose that the subcarrier set is available to provide redundancy for the original DSs. IEEE 802. vol. the BLER is BLER (28) . Apr. Mag. Jorswieck and H. pp. May 1977. In this second ARQ transmission. 2381–2401. Even though in this paper we restrict to the case of at most one retransmission. Sep.. Conf. APPENDIX B BOUNDS FOR ARQ IN FADING CHANNELS Typically. Theory. Thus. 2006.” IEEE Trans. for all . we consider subcarrier grouping techniques by grouping contiguous subcarriers and performing subcarrier assignment on these groups. 2007. transmitted over the original subcarriers and the ARQ subcar. K.16: Current performance benchmarks and future potential. “Broadband wireless access with WiMax/802.” in Proc. Ho. and R. W. [9] A. we have included the additional contribution of the second ARQ transmission. pp. past assignments are fixed before the current assignments are carried out. the subcarrier assignment problem. Sindhu. “MIMO transceiver design via majorization theory. Due to causality. Sun. C. Pandharipande and S. 9. 129–136. Sep. “Frequency domain equalization for single-carrier broadband wireless system. which equivalently collects the SNR over all original and ARQ subcarriers. the expectation is performed over method can be used to obtain numerical results by averaging the term within the expectation operator over realizations of generated by Monte Carlo simulations. Pandharipande and C. Heikkinen. Physical Channels and Modulation. [2] A. 331–551. we modified it for ARQ systems in a fading channel.211. [8] A. 2003. BLER ACKNOWLEDGMENT The authors gratefully acknowledge discussions with L.htm [4] P.” IEEE Trans. Serbetli. 3. which has led to significant throughput gains over a wide range of SNR. the effective SNR at subcarrier becomes (27) Compared to (5). the equivalent SNR for becomes . “Optimum DMT-based transceivers for multiuser communications. [6] Y. vol. Theory. except that the channel is always fixed as the average SNR.HO et al. Since any bit error constitutes a block error and there are 2 bits in a block.” IEEE Trans. 2003. D. Inf. 2096–2100. no. 4–5. 1–5. Ghosh.. pp. Costello. vol. no. Sep. Dec.0.. where bound. Cognitive Radio Oriented Wireless Networks Commun. 2007 [Online]. 553–701. vol. APPENDIX A ARQ-SA FOR TWO OR MORE ARQ TRANSMISSIONS We show that the subcarrier assignment problem for the original DSs for two ARQ transmissions is fundamentally the same as for one ARQ transmission. 2002. Eidson. “Retransmission error control with memory. vol. Inf. pp. New York: Academic. Consequently. Canada. L. A semianalytical Here. “Spectrum pool reassignment for a cognitive OFDM-based relay system. Inequalities: Theory of Majorization and Its Applications. 44. either of given (27) or of selecting given selecting (5). vol. and B. “Optimal subchannel assignment in a two-hop OFDM relay. Palomar and Y.: ARQ BY SUBCARRIER ASSIGNMENT FOR OFDM-BASED SYSTEMS 6015 keep the amount of feedback required to communicate the assignment low. Hence. [11] D. 3GPP TS 36. Conf.” IEEE Trans. Jiang. i. “Majorization and matrix-monotone functions in wireless communications. and M. (SPAWC). Hottinen and T. is fundamentally the same problem and can be solved similarly. Marshall and I. To this end. S. Commun. “On some properties of Walsh-Hadamard transformed OFDM. 2005. but in closed form. the BER is . Consider that a second ARQ transmission is activated.” IEEE Commun.3gpp. Rev. Ariyavisitakul. Cioffi. 2nd Int. Inf. FL. no. Oct.” IEEE Commun. [13] D. Palomar. pp. be the SNRs corresponding to . [14] A.e. Trends Commun. Here. 3GPP Std. 51. pp. 2531–2560.. To obtain a lower ments. It is obtained similarly as the MFB.. and let Let be the assignment of the ARQ subcarriers for the th original subcarrier. 2002. IEEE 8th Workshop Signal Process. a constant. Dasgupta. Adv. 2038–2046. 4. pp. . pp. the BLER provided by the AWGN bound is given by . Theory. After MRC. 58–66. A. Hagenauer. Hence.” in Proc. [10] A. Helsinki. vol. [3] Evolved Universal Terrestrial Radio Access (E-UTRA). 473–479. [5] J. the MFB gives a lower bound for an error probability without ARQ for a AWGN channel. “Applications of error-control coding. Trends Commun. Finland.” Found. H. REFERENCES [1] Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications: High Speed Physical Layer in the 5 GHz Band. 8.11a Standard. such as (26). 2007. Ho. and so both problems can be solved similarly. 2007. respectively. the equivalent SNR for is . J. “Joint Tx-Rx beamforming design for multicarrier MIMO channels: A unified framework for convex optimization. D. Andrews. similar to in (5). without loss of generality we assume that only the first . 40. pp. Numerical results have shown an improvement of the error performance even when limited redundancy is available for ARQ. J. This conclusion also holds for more than two ARQ transmissions by treating all previous assignments as fixed (and similarly if we consider the assignment for the ARQ DSs instead of the original DSs). G. The AWGN bound provides a looser bound. Technol. BC.e. This symbol is symbol is transmitted. Olkin. B. P. J. and S.” Found. Available: http://www. J. Imai.” in Proc. 1998. i.0. Lagunas. vol. Wireless Commun. vol.

Onizawa. D. degree in electrical and computer engineering from the University of Iowa.. 1996. Personal. Tokyo. and the M. Netw.” in Proc. [19] C.S.S. no. and M. he has been a full Professor and Chairman of the Signal Processing Systems Group. “A maximal ratio combining frequency diversity ARQ scheme for OFDM signals. TU/e. degrees from the Department of Electrical Engineering. Deuermeyer... Suwon. Gallager. and has received around 40 U. Madhukumar. (first-class honors) and M. Xin. 2001. Eindhoven. [17] X. Eindhoven. Ho. degree in electrical and computer engineering and in mathematics and the Ph.” in Proc.E. Eindhoven. F. Technol.S. respectively. M. Stegun.D. Since 1999.” IEEE Trans. and 2002. “Enhanced HARQ scheme based on rearrangement of signal constellations and frequency diversity for OFDM systems. A.” in Proc.. The Netherlands. pp. Conf.. 17–19. pp. South Korea. 12. Hongming Yang received the B. 1. in 2007. Inf. Singapore. L. A. Jul. Conf. and L. Ahag.6016 IEEE TRANSACTIONS ON SIGNAL PROCESSING. Conf. Sep. 2842–2854. Indian Institute of Science. is author of Digital Baseband Transmisison and Recording (Norwell. [18] L. Jun. Ashish Pandharipande received the B. degree from the Department of Electrical and Computer Engineering. 1972. Gui. 1982. vol. Morikura. recording systems. Tuninetti. in 2000. Bergmans (SM’91) received the Elektrotechnisch Ingenieur degree (cum laude) and the Ph. “Iterative detection for pretransformed OFDM by subcarrier reconstruction. IEEE Wireless Commun. China. 2005. May 2004. Signal Process. vol. IEEE Int. 2.” in Proc. Aug. degree at Eindhoven University of Technology. Since 2001.D. Boston. 53. degree from Eindhoven University of Technology. He currently is conducting joint research with Philips Research Laboratories. and A. Conf. Reading. degree from Eindhoven University of Technology (TU/e). Caire and D. His research interest lies in adaptive wireless communications and signal processing for multicarrier and space–time communications. “Space-time-frequency coded OFDM over frequency-selective fading channels.” IEEE Trans. He has held visiting positions with AT&T Laboratories and the Department of Electrical Communication Engineering. pp. New York: Dover. [25] Z. Cai. He is now pursuing the Ph. respectively. MA: Addison-Wesley. vol. Tallahassee.. Mar. Song. he has been with the Institute for Infocomm Research. In 1988 and 1989. 2nd ed. 2465–2476. M. [21] D. India.Eng. 1998. working on signal-processing techniques and IC architectures for digital transmission and recording systems. NO. 2001. [16] M. “Performance studies of a VSF-OFCDM system using a symbol relocated scheme during retransmission. His research interest lies in signal processing for digital communications. and illumination systems. and M. Y. in 1998. in 1982 and 1987. sensor signal processing. The Art of Computer Programming. He is currently a Senior Scientist with Philips Research. Peng. 528–532.D. he was with Philips Research Laboratories. 1971–1988. Discrete Stochastic Processes. and the Ph. Knuth. K. respectively. Beijing. [22] B. Giannakis. During his doctoral work. 2. Indoor Mobile Radio Commun. Wan. [20] G. National University of Singapore. Singapore. 1998.. 56. and signal-processing applications.” SIAM J. vol. vol. and W. “Improved HARQ scheme using channel quality feedback for OFDM systems. Liu. Discrete Meth. Jan W. he has been a Postdoctoral Researcher with the University of Florida. H. Bangalore. he conducted joint research with Philips Research Laboratories. Z. 500–504. patents. Since then. vol. Hyderabad. 47. G. in 2000 and 2003. Yan. Iowa City. degree in electronics and communications engineering from Osmania University.” IEEE Trans. vol. 2. K. Tsinghua University. Gidlund and P. 50. The Netherlands.D. he was an Exchange Researcher with Hitachi Central Research Labs. His research interests are in the areas of cognitive wireless networks. 1138–1143. M. no. Takanashi. respectively. From 1982 to 1999. Sun. in 1999 and 2001. Technol. May 2004. 10. Alg. pp. 59th IEEE Veh. 4. 2002. S. 190–196. Lei. VOL. Chin. Eindhoven. Eindhoven. Friesen. T. He has published extensively in refereed journals. Eindhoven. pp. Signal Process. 1996). “Scheduling to maximize a multiprocessor system. Langston. B. The Netherlands. “The throughput of hybrid-ARQ protocols for the Gaussian collision channel. Eindhoven. Theory. Eds. Chin Keong Ho received the B. 3. Y. DECEMBER 2008 [15] T. pp. He received the M. S. multicarrier and MIMO wireless communications. [23] Handbook of Mathematical Functions. 2005. and a Senior Researcher with Samsung Advanced Institute of Technology. MA: Kluwer Academic. National University of Singapore. Japan. pp. [24] R. in 2005. Abramowitz and I. vol. P. and M.S. Mizoguchi.Eng. degrees from the Department of Electronic Engineering. Kumagai. . 59th IEEE Veh. pp.E. MA: Kluwer Academic. Oct. The Netherlands. and G.