You are on page 1of 13


3, MARCH 2015 1213

Beamforming for Multiuser MIMO-OFDM
Interference Channels With
Multipath Diversity
Milad Amir Toutounchian, Student Member, IEEE, and Rodney G. Vaughan, Fellow, IEEE

Abstract—This paper presents three beamforming designs for in [2] and [3]. Extended work [4] exploited uplink-downlink
multiuser multiple-input and multiple-output with orthogonal SINR duality. Joint transmit-receive optimization based on null
frequency-division multiplexing, where the transmit and receive space constraint was introduced in [6]. But none of these can be
beamformers are obtained iteratively with closed-form steps. In
the first case, the transmit (Tx) beamformers are set and the applied to the more complicated interference channel, which is
receive (Rx) beamformers are calculated. It works by projecting the subject of this paper. As a further clarification of the context,
the Tx beamformers into a null space of appropriate channels. some recent works refer to interference suppression operations
This eliminates one interference term for each user. Then the as precoder matrices and interference alignment (IA), e.g., [7],
Rx beamformer for each user maximizes its instantaneous signal- but in this paper, the signal processing operations are referring
to-noise ratio (SNR) while satisfying an orthogonality condition to
eliminate the remaining interferences. The second case is jointly to beamforming as a vector operator. Therefore, only single-
optimizing the Tx and Rx beamformers from constrained SNR stream data transmission is considered. So multi-stream data
maximization. It uses the results from the first case. The third transmission for each user, usually discussed in IA design, is
case is also for joint optimization of Tx–Rx beamformers but not used here.
combines constrained SNR and signal-to-interference-plus-noise The relationship between this work and other representative
ratio maximization. The minimum number of antennas required
is derived as part of the formulation. All cases can include a examples of single-stream transmission beamforming works is
linear constellation precoder for extracting multipath diversity. summarized as follows. In [8] and [9], the SINR maximization
Finally, the required feedback rates are derived and compared to was treated for a specific user which has multiple transmit and
existing beamforming methods. Using the standardized statistical receive antennas, while the other user—as one interferer—has
channel model for IEEE 802.11n, our simulations demonstrate single transmit and receive antennas; whereas in this paper,
faster beamforming, improved error performance, and the ability
to extract multipath diversity which is not possible in the least- all users have multiple transmit and receive antennas. The
square approach. joint transmit beamformer (Tx-BF) and receive beamformer
(Rx-BF) design for minimum SINR maximization in a MIMO
Index Terms—Mobile communications, interference channel,
beamforming, optimization, MIMO-OFDM, multipath diversity. interference channel system has been proven to be a strongly
NP-hard problem [10]. Recently, a joint leakage interference
I. I NTRODUCTION (LI) minimization and maximization of the individual signal
powers (SPs) and SINRs was reported [11] using an opti-

B EAMFORMING for a multiuser MIMO interference
channel is for communications between pairs of terminals
where there are several pairs sharing the spectrum simultane-
mization approach with a fixed point method. The numerical
results revealed that the joint LI-SP-SINR problem has superior
performance compared to the max–min SINR problem, but with
ously. Each multi-antenna transmitter strives to direct its data to much less computational complexity. The minimum number of
only one multielement receiver in the presence of interference required antennas for each user with respect to the number of
from all the other users’ transmitters [1]. This is a recent users for the joint LI-SP-SINR problem has not been addressed
subject, and it is different from multiuser uplink and downlink to date, but new results are presented in this paper.
beamforming which has been studied extensively, e.g., [2]–[6]. Various objective (cost) functions have been presented in
Of these studies, the use of an SINR constraint, where inde- interference channel systems, and the optimization criteria have
pendent data streams are transmitted from a multi-antenna base included, for example, beamforming design for mean-squared-
station (BS) to several single-antenna mobile units, is discussed error (MSE) minimization [12], weighted sum rate maximiza-
tion [7], and maximization of the sum signal power across
Manuscript received January 6, 2014; revised May 1, 2014 and August 29, the network divided by the sum interference power [13]. For
2014; accepted October 15, 2014. Date of publication October 28, 2014; date the MSE minimization in [12], the problem led to an iterated
of current version March 6, 2015. The associate editor coordinating the review second-order cone programming routine. For the weighted sum
of this paper and approving it for publication was Z. Wang.
The authors are with the School of Engineering Science, Simon Fraser rate maximization of [7], the constrained problem was con-
University, Burnaby, BC V5A 1S6, Canada (e-mail:; rvaughan@ verted to an unconstrained problem and solved by a gradient descent algorithm. Finally, the problem addressed in [13], i.e.,
Color versions of one or more of the figures in this paper are available online
at the sum signal power divided by sum interference power, was
Digital Object Identifier 10.1109/TWC.2014.2365464 simplified by using an alternating maximization method.

1536-1276 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See for more information.

along with some error performance. a derivation for 3) Each tap is modeled using the Kronecker model. the spa- The new contributions of this paper include: closed-form tial correlation (correlation matrices) and temporal correlation solutions for multiple cases of the joint beamformers designs (Doppler spectrum) for each tap are modeled independently. interference. For these three cases. setting the modulation and a calculation of some aspect of the associated communications coding: setting the data payload size and scheduling.e. the optimization of share the channel between: sounding. m = size(A. Hence. complexity of the decoder is independent of the number of As noted above. are not performance—usually an information-theoretic capacity or addressed here (such protocols are not yet developed). i.11n propagation further improve the digital communications error performance model or 2) a single-tap flat fading channel with a perfect (without compromising the sum rate performance). Finally. NO. Perfect channel knowledge is as- first two methods is lower than that of existing solutions if sumed at all the users and perfect timing is also assumed in same number of antenna is considered).1214 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS. The second case is joint maximum eigenvalue of A. This is to allow multipath diversity gain from the to compare the performance of the existing methods with our OFDM system [14]. The Rx-BF of each user. The third case is joint constrained SNR and n = size(A. has a closed. 1) and convergence. the normalized eigenvector of matrix A that corresponds to the form solution if its norm is one. The ultimate metric is a practicable new method for designing and deploying beamformers for mul- digital communications performance. 14. simulated performance the usual manner to allow the simplest linear model for the results using a standard multipath channel model showing the link. All of these new contributions come together in a beamforming problem. [x]m are the (m. The details of the communications protocol to sible to optimize this directly. In order to line (L + 1 taps) according to the IEEE 802. interchanging of channel some analogue channel performance functions is followed by information between the terminals. It can be considered with complex entries. [12]. complex conjugate transpose. The Euclidean norm of a vector x is x. i) = [A]i . In other words. n)th where only the Rx-BF at each receiver terminal nulls out all the entry of matrix A.n . and a∗ represents a This null-space assignment for Tx-BFs eliminates one term of value of a. not need instantaneous channel knowledge. i) is MATLAB notation for the ith column of matrix tion which can be solved iteratively because it has guaranteed A. and does not denote the complex conjugate of a. and [13] but at the in time and space. all solutions allow this aspect of its usage is not included in the formulation. in this case. 2). The at each receiver can be eliminated by means of orthogonal null space of matrix A is denoted by N (A). respectively. the required feedback rates are AND T HEIR S OLUTIONS computed for each approach and compared with those of ex- isting beamforming schemes. here we plumb for optimizing the analogue . It is noted that multipath diversity gain for OFDM delay profile (or spectrum (PDS)) are separable which means systems can also be obtained by using a multi-tap receive FIR each tap is modeled independently. spectrum for each tap are separable. in the K-user MIMO interference channel. A(:. 3. interference at each receiver. This precoder is a fixed matrix that does three proposed designs. and [13] if the same summarized as follows. S YSTEM M ODEL . respectively. and a new formulation for the benchmark interchanging of information bites into the capacity which is the least-square (LS) solution and simulation results for all the very quantity usually being sought by MIMO systems. although it does The IEEE 802. There are K pairs of multi-antenna ter- number of antennas is considered.. respectively. [12]. a constrained SNR maximization is sought in and matrices are denoted by boldface lower and upper case which the Tx-BFs for all the users are acquired by deploying the letters. but it is not yet pos. A(:. First. this assumption is a challenging one in the performance of the system in more realistic conditions than pre. 2) The PAS and the Doppler filter [15]. In practice. the system spatial correlation matrix. The second channel model is used only beamformer. The channel is modeled as 1) a tapped delay price of one more antenna element at each terminal. sense that the necessary continual sounding of the channels and viously considered. MARCH 2015 It is clear that there are several different approaches to the antennas. x and y are orthogonal if xH y = 0. tiuser systems.. these lines. and mth entry of vector x. time domain) diversity where the sets up a difficult interpretation for any capacity optimization. the minimum required number of antennas is derived as a part II. [A]m. this paper presents three new beamforming design The notation is conventional. Tx–Rx beamformer design for constrained SNR maximization IP is the P × P identity matrix. and yet presented beamformer design cases. This problem leads to a multi-objective optimiza. in order throughput rate. P ROBLEMS F ORMULATIONS of the formulation. VOL. Finally. E{. Superscripts T and H give the transpose and null-space of an appropriate channels matrix (described below). This problem has a solution because its Finally. The first channel model justifies both formulation can include a precoder matrix before the transmit the MIMO-OFDM configuration and deploying the multi-path beamformer and a sphere decoder (SD) following the receiver precoder in this paper. Instead. Along to focus on the beamforming problem. Our proposed methods are minals which are striving to share simultaneously the spectrum also computationally simpler than [7].} is the expected value. wmax (A) is vectors. it the minimum number of antennas required for these solutions.11n statistical propagation model assumes need knowledge of some channel characteristics for optimal [16]: 1) The power azimuth spectrum (PAS) and the power deployment. Two out of three proposed cases The communications situation as it relates to the model is require less feedback compared to [7]. is assumed that the transmit and receive correlation matrices a derivation of the required feedback rate (the rate for the are separable for each tap. as a new formulation for LI-SP-SINR. This for exploiting multipath (i. The remaining interference terms The complex conjugate of x is represented by conj(x). x ∈ Cn \ 0 means that vector x is a non-zero vector corresponding vector field is nonexpansive.e. For A ∈ Cm×n . as follows: column vectors cases. SINR maximization.

and l ≤ L + 1 indexes the delay-time bin. The frequency selective channel from the μth transmit an. . The Rrl and Ril are the receive and transmit The formulation in this section assumes K ≥ 3. then the output of the transmit  π beamformer is vi (p)s̃i (p) where vi (p) = 1 and s̃i (p) is the l RXX (D) = cos (D sin(φ)) f l (φ) dφ (2) pth element of vector s̃i . the transmit beamformers for the ith user at the pth tion coefficients between the νth and μth receiving antennas. and calculate associated digital performance From the Kronecker model.μ /λ and: Δ s̃i = Φ[si (0) · · · si (P − 1)]T . Assuming perfect OFDM symbol timing synchronization.μ if ψc is defined as the lth i =i path angle-of-departure (AoD) towards cluster c. .i (p)vi (p)si (p) + Hi. Here λ is the wavelength. The f (φ) have been computed in [17]. the where A is the normalization factor ( −π f l (φ)dφ = 1).μ = RXX (D) + jRXY (D) [17]. .μ = Hνμ (p). The details of the model are as follows. .i (p) ∈ C .11n is of the rth receiving user is denoted by the delay-time func. s̃i (p) = si (p). The closed- (1) form formulations for RXX l (D) and RXYl (D) for a Laplacian Nr ×Nt l In (1).i νμ (l) = [H l ]ν.μ = ρr. the receive For a special type of uniform linear antenna array. Null space and orthogonal basis multiuser beamforming with precoder.μ for ν = μ represents the complex correla- K = 2 is special and is discussed separately below. . channel quantities. and similarly. then after removal of π the cyclic prefix with length LCP ≥ L and after the FFT. .i √ {1.l to Fig. .i channel with perfect correlation matrix means that L = 0 and (ν. For simplicity. K}. μ) entry of it is defined as [Hr. and f l (φ) is the probability density tenna of the ith transmitting user to the νth receive antenna function for the PAS. . The | | 1  l N 2 φ−ψc c plc − channel is considered unchanging for one OFDM symbol f l (φ) = e l σc (4) A c=1 σcl and independent between OFDM symbols. respectively. .i −j2πlp/P For both channel models. Hνμ (p) := hr. lth channel tap matrix where hr. assume Φ = IP ×P l for now.i νμ (l)e l=0 Applying the receiver beamformer to all the subcarriers of all . So in this case. Hr. . 1. Later on.μ apart {0. The users’ data symbols are assumed −π  π to be mutually independent. Ri = INt .  L r.μ = [Rl ]ν. modeled as a truncated Laplacian distribution: νμ (l) where ν ∈ {1. The same steps can yi (p) = Hi. . In Fig. . . P − 1}. 1) in the simulation Section VII. the focus is beamformer design. plc is the lth path average gain for cluster c. . .i after the optimization. where of si where si is the input symbol stream of user i. . 1. the MUX block stacks P samples can be modeled by ρr. . [Rrl ]ν. Typically.l l l ν. . The single-tap flat fading Δ r.i (p)]ν. we consider RXY (D) = sin (D sin(φ)) f l (φ) dφ (3) −π an optimal Φ matrix to extract multi-path diversity. where Rr = INr . Define D = 2πdν. and r ∈ tion hr.μ and Gl ∈ Nr ×Nt The K users all have Nt transmit antennas and Nr receive C is a complex matrix with zero-mean and unit-variance antennas. This channel is denoted by G(0. Nt }. . the case for spatial correlation matrix. μ ∈ {1.i (p)vi (p)si (p) + ni (p) be followed to obtain ρi. subcarrier are written vi (p) ∈ CNt ×1 . . σcl is the lth path angular spread for cluster c and ψcl is the lth  K path angle-of-arrival (AoA) from cluster c. and all users utilize each of the P subchannels. 1. Referring and [Rrl ]ν. the corre- beamformers are ui (p) ∈ CNr ×1 for i ∈ {1. K} and p ∈ lation of the fading between two antennas spaced dν. Nc is received signal vector for the ith user can be written: the number of clusters.i l = (Rl ) r 1/2 Gl (Ril )T /2 is the r.TOUTOUNCHIAN AND VAUGHAN: BEAMFORMING FOR MIMO-OFDM CHANNELS WITH MULTIPATH DIVERSITY 1215 Fig.μ = 1 for ν = μ ν. Gaussian entries. Nr }. the channel at subcarrier p is Hr. the f l (φ) for IEEE 802.l i l ν.

For K ∈ {2n + maxima and a set for the global maximum. and simplifying the notation by dropping index p: example. then ratio (SNR) of the ith user ( So P is a constrained SNR maximization formulation where i=1 maximization over a quasi-convex object function with affine s.2 2 v = 0 ⎪ ⎨ uH 1 subject of widespread interest.t. xlocm = 0. variable. pressing the K − 2 remaining interference terms. . This opti- mization problem is denoted P for the first receiver as an min(Nr ) = K.t. the beamformer vi is obtained by The number of unknown parameters and number of equations vi = N (HK+1−i.i ) if K + 1 − i < i unknown parameters but K − 1 distinct equations. . ui has closed-form solution for constrained SNR maximization (Section II-B). it can be shown that xlocm is a local maximum of P if III.1 u1 P: max u1 ∈CNr \0 uH1 u1 Designing vi and ui . O PTIMAL R X -BF S FOR C ONSTRAINED SNR Qxlocm = 0. (7) where x = u1 . MARCH 2015 the users. .1216 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS. For K ∈ {2(n + 1) : n ∈ N}. max log2 (1 + SINRi ) ui . 14. For simplicity.e. Ij = 0 j = i. problem P are: (7) is reduced to a problem with ui being the only decision ∇x L(xopt .. the vi can be found by ⎧ is in the case xlocm = 0 for xlocm from (14). respectively. where N denotes positive CNr \ 0. where Lemma 1: For constrained SNR maximization..i vi si + ui ni (5) i =i uH H H 1 H1. K. and that the interest 1 : n ∈ N}. Proof: Recall that P has sets of solutions for the local is an orthonormal basis for the null space of A.e. There are Nr ⎨ N (HK+1−i. To tackle problem (7).i+1 vi+1 . The problem is expressed 2 uH H i yi = ui Hi. after the Rx-BF) while sup. Q = H1. . (14) M AXIMIZATION W HEN THE T X -BF S A RE K NOWN qHi x locm = 0. x = 1} (9) K and min(Nt ) = K + 1. This does not have a closed-form solution for vi and ui (see [18]) different problem leads to a different solution. first we assume that one interference term is eliminated before Therefore.K vK = 0. From (13). and the rest of the receivers’ beamformer designs follow by the same methodology. . its Lagrangian function is needed: that the optimization problem: xH Qx  K−2 L(x.i ) (8) in (14) for xlocm = 0 determines the minimum number of required antennas for problem P. so ⎩ Qxlocm = 0 is counted one equation. (6) quadratic forms is a known problem with an eigen solution. (15) . 3. The linear independent constraint qualification (LICQ) holds at has closed-form solution for vi and ui .3 3 v = 0 s.1 v1 v1H HH 1. . the Karush–Kuhn–Tucker (KKT) conditions for applying the receive beamformer (Section II-A). λopt ) = 0. in order to have good detection per. But here the difference is that the P has constraints. . (11) the following optimization: ⎪ ⎪ . . λ) = − − λi x H q i . Besides a set of solutions for the global maximum. denoted by xlocm ∈ sought. x∗ if x∗ is a local solution for problem P (see Appendix A). s.1 and qi = H1. integers. Si = ui Hii vi vi Hii ui . VOL. It is recalled that Q ∈ CNr ×Nr and xlocm ∈ CNr ×1 . NO.i ) if K + 1 − i > i from qH i x locm = 0 for the linear system of (14). To solve P. to be H 2 H 2 Ij = ui Hij vj and Ni = σni ui ui (σsi = 1). The maximizing of the ratio of viH vi =1 i = 1. ⎧ H formance for all users in MIMO interference channels. xlocm ∈ CNr \ 0. 2 force the interference. Previous work [18] showed that 1 H1. These constraints  H H H where SINRi = Si /( K j=i |Ij | +Ni ). sets of solutions for local maxima. for the first user as an example. is the ⎪ ⎪ u H1. It is also de- Note that from (8) or (10). If in (14) the number of unknowns is to determine ui such that it maximizes the signal-to-noise is greater than the number of equations (Nr > K − 1). P also has tion is much easier than solving (6). We also demonstrate that the joint design of vi and (13) qHi x opt = 0.i vi si + uH H i Hi.. the Tx-BFs are found from the null space of The solution xlocm could be zero but the interest is in solutions an appropriate set of channels. The next step sired that xlocm ∈ CNr \ 0. uH i ui = 1 i = 1. (12) max SNRi = Si /Ni xH x i=1 ui . E{|si |2 } = σs2  K and E{ni nHi } = σn I. In this section. Hence.i ) if K + 1 − i = i (10) ber of distinct equations is K − 1 because rank(Q) = 1.K−1 vK−1 = 0. The num- vi = N (HK. We show here eliminated. and then the optimal Rx-BFs are where xlocm is non-zero. solving constrained SNR maximiza. . H1.t. Therefore. i. and K − 2 equations N (HK−i.1 v1 v1 H1. the mini- Δ mum numbers of receive and transmit antennas are min(Nr ) = N (A) = {x|Ax = 0. ⎪ ⎩ H  K u1 H1. K constraints is sought [19].

near-maximum likelihood (i. √ . those used in the proof of the above Lemma. α2J−1 ⎥ max |α|2 μ1 + |β|2 μ2 + 2 {conj(α)β μ1 μ2 ejϕ } Θ= ⎢ . where J ∈ {2n . An SDR detector is applied here for decoding the data by means of semidefinite relaxation codes max xH Qx x for the discrete integer least squares problem [22].. qK−2 ] ∈ C (K−2)×Nr . so min(Nt ) = K + 1. Therefore.2 v2 v2H HH 2. instead of Lemma 3: For Nr = K. with the subcarrier index. x = α[N (B)]1 + β[N (B)]2 . The closed-form global optimum of this where P and Q are positive semidefinite e.e. [20].1 ). The parameter = ye−jϕ . new problem is presented below. . . then Δ the equation in (5). αJ−1 ⎤ mulated as: 1 1 √ 1 ⎢1 α2 . H The u1 (p)y1 (p)’s are stacked for all P subcarriers and then  this received OFDM frame is decoded by SD. If Nr = K then N (B) ∈ C K×2 opt . . Therefore vi ∈ CNt \ 0. is a difficult problem (see  Appendix B). μ2 μ2 xopt = 1 − [N (B)]1 + e−jϕ [N (B)]2 As discussed in [14]. |α|2 + |β|2 = 1. only two complex numbers α and β should be (24) found.1 v1 /H1. solution if Nr = K.TOUTOUNCHIAN AND VAUGHAN: BEAMFORMING FOR MIMO-OFDM CHANNELS WITH MULTIPATH DIVERSITY 1217 On the other hand.1 = H1. {α} = 1 − y.1 (p)v1 (p)s̃1 (p) + u1 (p)n1 (p) for Nr = K. Using the Rx-BF’s unitary assumption (xH x = u2 = wmax H2.. . The optimal LCP Δ Δ matrix design for Φ is summarized as follows [14]: where ϕ = ∠[N (B)]H 1 Q[N (B)] 2 . H H  u1 = wmax H1. . Δ H • The unitary rotation matrix Θ ∈ CJ×J has a μ2 = [N (B)]2 Q[N (B)]2 . and it is under- stood (assumed) that y is real and less than unity. Recently. the closed-form global maximum decoding s1 (p) subcarrier-wise (if Φ = IP ×P ). the vi ’s are the null space of matrices with A receiver design for the two-user case.2 v2 /H2. xH Qx/xH Px. for general case Nr ≥ K. has closed-form global quadratic forms. α = 1 − y Without loss of generality assume β √ √ β is to set tr(ΘΘH ) = J. . (16) v1 = N (H2. Bx = 0 Multipath diversity can be also added to this system. (The α and β are different to which means {β} = y. K = 2. (19) 1 αJ . ⎦ s...1 )  v2 = N (H1. s.g.. (17) ing the design procedure in this section and deploying Φ. . the degrees-of-freedom (DoF) of semidefinite relaxation (SDR) has been demonstrated to show the new problem P is 2. We show that the new problem P .. if Nt > Nr or considered.t.) Δ problem (19) reduces to: • Set P = M J.t. quasi-maximum-likelihood detection based on Lemma 2: For Nr = K. is: where B = [qH H 1 . the problem P is transformed to a new optimization problem Equation (23) is the solution to the maximum of the ratio of P .1 v1 v1 H1. Recalling that for K = 2. αJJ−1 √ √ √ where {αj }Jj=1 are the roots of xJ = −1.β β ⎣ . and I = {0. with respect to x ∈ Cn \ 0.. Follow- xH x = 1.2 ) (22) Finding the global optimum Rx-BF from KKT of prob. the optimization problem (11) has no constraint. . ⎥ (25) α. Vandermonde structure Proof: Based on Lemma 2.. corresponding to two degrees of freedom (see below). Hence.. using the LCP of problem P is: matrix makes it possible to decode the data frame-wise while getting multipath diversity (up to L + 1 = 1 for H1. the fixed matrix Φ extracts maximum μ1 + μ2 μ1 + μ2 (18) multipath diversity if it is designed properly. . Hence. . SD) performance but with poly- Proof: Problem P can be written as: nomial complexity [21]. n ∈ N}. uH H H 1 (p)y1 (p) = u1 (p)H1. . . the optimization (17) is refor- ⎡1 α .1 v1  lem P. 1.2 = H2. is now dimension Nr × Nt . .2 v2  (23) 1). μ 1 = [N (B)] H 1 Q[N (B)] 1.

so by taking the square root. . .1 . . The mth group which is a square. . . . .K }. pm. it is equivalent to Δ . . P − 1} to index the P subcarriers. Subcarrier group- max μ1 (1 − y) + μ2 y + 2 μ1 μ2 y(1 − y) (20) ing can be represented by partitioning I into M noninter- y Δ secting subsets Im = {pm.

. of subcarriers selector matrix Ψm = IP (Im . Because μ1 > ing would be Imopt = {m − 1. which 1)M + m − 1}. where √ √ √ max f (y) = 1 − y μ1 + μ2 y. :) is a J × P permutation matrix comprising the y {pm. .j + 1}Jj=1 rows of IP . (J − 0 and μ2 > 0. (21) IP (Im . y opt = μ2 /(μ1 +μ2 ) is derived. there results 0 < y opt < 1. The optimal subcarrier group- From df /dy = 0. . M + m − 1. . . :).

satisfies the  above .

hence xopt=αopt [N (B)]1+β opt [N (B)]2 .  The complexity of SD for multiuser MIMO-OFDM interfer- Therefore. which can extract multipath diversity through constrained SNR maximization problem where the Tx-BFs are the LCP matrix. . is the same as SISO-OFDM system because Δ the null space of the appropriate channels. αopt = 1−y opt and • The LCP matrix is Φ = M T m=1 Ψm ΘΨm .assumptions about y.1 (p)v1 (p) is a scalar channel. the optimum Rx-BF is in closed-form for the ence system. h1 (p) = uH 1 (p)H1. β opt= y opt e−jϕ . Therefore.

. . . vK ). However. . N is the fixed number of iterations problem G1 as: (N ) K after which the {xi }i=1 is the approximation for {x∗i }K i=1 . . . SNR M AXIMIZATION AND SINR M AXIMIZATION then there is a Nash equilibrium (NE) for K sub-problems In the previous section.2 v2 )H . .1 K u =0 1 H1.. . . . K. .e. . . . then from the optimal solution of (21) using Lemma 3.1 Ji . which is optimization w. xi . T X -BF AND R X -BF D ESIGN FOR J OINT C ONSTRAINED . . . v1 and u1 . xi+1 . . vi+1 . . . . . .. (29) Secondly. . 1. .1 u1 v1 ∈CNt . + JK (x1 . . can odd number of users. . . . . . . li  ≤ a for col{x1 . then EAO approximates the difficult problem’s solution by simultaneous and the unique global solution of (32) is: solving of the following K problems: v1 = wmax (G1 + G2 ) = l1 (v2 . G1 : max v1H HH H 1.r. xK (30) objective optimization by the fixed point method is applied. . · · · .1 u1 u1 H1. xK } ≤ a (where the col operator concatenates vectors). each Rx-BF nulls its interference and (games). (H1. respectively.. Finally.1 v1 v1 H1.u1 ∈CNr ⎧ H v1 v1 ≤ 1. . and this procedure is deployed for all ∀ xi ∈ Ωi . 3. xi+1 . xi+1 . xi+1 . v1  = 1. xK . i = 1. (1) (0) k = 2. . (34) G: ⎪ ⎪ . .t. .xK ) (18) using the known vk . xi−1 . . v1 only. v1 Consider the following optimization problem: s. . xK ) (27) xi ∈Ωi Generally. . k = 3. . . . xK ]T and ΩJ is the feasible set. . K. . (33) min Ji (x1 . NO.1 v1 We show that the EAO algorithm can be deployed for joint v1 v1 ∈CNt and u1 design for constrained SNR problem. we are seeking a Nash equilibrium point for two s. Generally. Ji (x1 . . . xK ) users over N iterations. . .r. xi−1 . . . xK ) x∈ΩJ where (26) Δ   where x = [x1 . k ∈ K. then the ⎪ ⎪ v1H HH joint constrained SNR problem for the first user is formulated as: ⎪ ⎨ 3. . vK are assumed to be known (or simply were obtained by (18) while the Tx-BFs are the null space of fixed). channels as expressed by (8) or (10) according to an even or the process of finding the optimal Tx-BF of the first user. v1 = γ [N (D)]1 + δ [N (D)]2 (35) .1218 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS.. Let Nr = K + 1 for In the previous section. . . v1 . vi−1 . . i. the problem G w.1 [N (C1 )]1 [N (C1 )]1 H1. xK ) i = 1. be represented as: Algorithm 1 in Table I obtains the optimal Tx-BF for the first (1) (0) xi = li (x1 . Finally. there is a unique global minimizer with G2 = HH Δ H 1. . (In an alternative ⎪ ⎪ uH1 H1. .1 u3 = 0 s. .. .t. NE for these K games can then this solution is inserted to the constrained SNR objective be approximated iteratively by: function which yields the Tx beamformer.t. xi . define the In the rest of the paper. K. . . . if v2 . . ... C1 = (H1. vK are assigned arbitrarily at the first iteration. . (31) games. . are then used where li is a nonlinear function with following property: (1) for obtaining the v2 . . multi-   xi (n+1) (n+1) = li x 1 (n+1) (n) (n) . ⎪ ⎪ u1 H1. . the Rx-BFs are obtained from (N ) ≥ Ji(x1. . . . 14. xi−1 . . xi+1. Instead of G. . xi+1 . .1 u2 = 0 v2 .. In this section. MARCH 2015 IV.xi−1 .2 v2 = 0 ⎪ ⎪ This problem is maximization w. xi+1 . xi−1 . if firstly for each objective function. vi = li (v1 . . . Assume that the ⎧ H H ⎪ ⎪ v1 H2. see Appendix C. . xi−1 . J OINT R X -BF AND T X -BF FOR C ONSTRAINED Following a similar discussion as above.1 [N (C1 )]2 [N (C1 )]2 H1. .3 v3 = 0 ⎪ ⎨ . .. .t. The v1 and vk . joint Tx-BF be summarized as: and Rx-BF are designed for the constrained SNR maximization problem by using the extended alternating optimization (EAO) min J1 = −v1H (G1 + G2 )v1 algorithm for a multi-objective optimization. . . the maximum normalized eigenvector function. if for some a. . xi−1 . . xK )|xi =li (x1 . the optimal closed-form Rx-BFs problem G. notation. K (28) user at the first iteration.r. . based on some initialization vk . SNR M AXIMIZATION decision variable u1 has a solution if Nr − K + 1 ≥ 2 and closed-form solution if Nr = K + 1. (32) min J (x) = J1 (x1 . . .xi+1 . . . vK ) where here the li is where Ω1 × · · · × ΩK = ΩJ .t. xi−1.1 respect to xi for fixed x1 . ..K vK )H solving such a nonlinear constrained optimization problem is H G1 = HH Δ difficult. . . xK ) + . .t. xi . . ⎪ ⎪ H H max uH H H ⎪ v H ⎩ 1H K. Because |li | ≤ Now it can be assumed that the optimal solution of (27) can 1 then EAO can be deployed (Appendix C). ... . .K vK = 0 Denote the global unique minimizer of G1 by ⎪ ⎪ u H u =1 ⎪ ⎩ H 1 1 v1 v1 = 1.) ⎪ ⎪ H ⎪ ⎪ u 1 H 1. In this section. . V. VOL.

1 (In the proof.1 v1 v1 H1. . Remark: The proposed methods here are not LS beamform- Here. v K Δ Δ where ν1 =[N(D)]H H H H H 1 H1. :). .1 u2 ) . .1[N(D)]1 . it is possible tionary algorithms. With |γ| = ω1 and algorithm is almost surely convergent to the fixed point of its |δ| = ω2 . just si is to be decoded from yi . pK (u1 . .1 v1 v1H HH 1.TOUTOUNCHIAN AND VAUGHAN: BEAMFORMING FOR MIMO-OFDM CHANNELS WITH MULTIPATH DIVERSITY 1219 TABLE I E XTENDED A LTERNATING O PTIMIZATION (EAO) FOR J OINT T X -BF AND R X -BF IN C ONSTRAINED SNR M AXIMIZATION P ROBLEM W ITH K U SERS Δ where D = [(HH H H H 2. uH 1 u1 = 1. ⎦ + ni 2 u1 ∈CNr uH 1 K j=1 Hi.t. uK ). ω12 + ω22 ≤ 1.K vK ] ⎣ . (36) (n+1) ui (n) (n) = qi v 1 .K vK ]. . . the ω1 and ω2 are the solution of: corresponding nonexpansive vector field f :   (n) (n) (n) max ω12 ν1 + ω22 ν2 + 2ω1 ω2 ν3 v i = pi u 1 . −1 H K σ 2 ⎣ . . .1 u1 u1 H1.t. . .ω2   s. ⎦ + HH i Hi Hi ni u1 = wmax⎝⎝ Hi.t. then by applying LS: Problem G2 has a unique global solution w.) Algorithm 2 for Tx-BF and Rx-BF design for joint con- v1 = ω1 [N (D)]1 + ω2 e−jθ [N (D)]2 (38) strained SNR maximization and SINR maximization is sum- marized in Table II.j vj vjH HH i. to obtain u1 from each users’ SINR maximization (problem G2 ) Proof: From (1). u1 . (HK. u K ω1 .j vj vjH HH n ⎠ i. . where θ = ∠[N (D)]H H H 1 H1.1 uK ) ]. . which is the approach of Section II.1 u1 u1 H1.j + 2 I H1. the notation is: Δ  T Δ u1 uH H H H 1 H1. . . H −1 H . .. ⎦ = Hi Hi Hi yi = ⎣ . uK )) ω1 ν1 ν3 = wmax (37) ω2 ν3 ν2 Δ where the vector field f defined by f = [f1 . .1 u1 u1 H1.. . . Δ both G1 and G2 applied for all K users. the following iterative Denote the ith row of Ti = (HH −1 H i Hi ) Hi by ti = Ti (i.r. ⎡ ⎤ ⎡ ⎤ ⎛⎛ ⎞−1 ⎞ ŝ1 s1  .1 [N (D)]2 . By KKT. uTK real. yi = [Hi. expansive [18]. . .1 u1 s1 G2 : max   . . ν3 = |[N(D)]1 H1. (39) Δ Let Hi = [Hi. . . . fK ]T is non- Without loss of generality. the received vector yi can be also by knowing v1 from (38): written as: uH H H ⎡ ⎤ 1 H1. . · · · .j + σn σs2 I u1 (41) sK s.1 v1 · · · Hi.1[N(D)]2 . . The Tx-BF design with LS can be obtained only by evolu- fixed (cf. are x = uT1 . . instead of obtaining u1 from problem G1 while v1 is ing. .1 v1 · · · Hi.. ν2 = [N(D)]2 H1. (38)).1 ⎠ ! "# $ σs ŝK sK ei j=1 (40) (42) Theorem (Proof in [18]): With the unique global solution for In the interference channel. the unique global solution of (36) is:     Δ fi = qi (p1 (u1 . .1[N(D)]2 |.

obtained by deleting row j and column i of A. For Nr = Nt = K: In this section.K feedback from one receiver node to transmitter nodes and dashed-dotted arcs s. the complexity 1  K % %2 was demonstrated to be lower than existing systems.. and K − 1 Tx-BFs should be fed back from each where A{ji} . K(K − 1) nel sounding is an associated issue. The dashed arcs a represent the Tx-BF which are feedback from receiver nodes to transmitter nodes and dashed-dotted arcs b represent Tx-BF which are feedback among receiver nodes. 3. Fig. This is important for number is quantized by 2q bits information.. NO.. VOL. denotes the submatrix of A receiver node to the other receiver nodes (arcs b in Fig. 2). vi  ≤ 1 (43) c show the Rx-BF from one receiver node to K − 1 receiver nodes. vi  ≤ 1 (45) the receiver nodes to the transmitter nodes (shown as arcs a in Fig... 2 depicts all the feedback required for K = 3 in our Solving the optimization problem (45) is possible by using beamforming method.K |det(Hi )| j=1 (i. the feedback of the presented beamforming is compared with existing interference channel beamforming Ti = H−1 i (44) schemes. the Tx-BF design with LS is formulated as: Fig. 3. and is extended to OFDM via scaling by P . MARCH 2015 TABLE II T X -BF AND R X -BF D ESIGN FOR J OINT C ONSTRAINED SNR-SINR BY EAO Fig. B EAMFORMING S CHEMES FOR Fig. the data describing the Tx-BF) should be fed back from s. As there are K users.. This is exactly the same for maximization of From the discussion in the Introduction and the start of the sum signal power across the network divided by the sum Section II. an issue in multi-user beamforming is the amount of interference power formulated in [13] and solved by the alter- information required to be exchanged among receivers and nating maximization technique. K 3 + K 2 complex numbers evolutionary algorithms for example. (8) and (10).) channels. As an example of existing approaches. Feedback graphs for the proposed constrained SNR maximization with known Tx from nullspace when K = 3. 2). 3 illustrates all the feedback required for the beamforming I NTERFERENCE C HANNELS design in [12]. rate maximization discussed in Section V of [7]. For the iterative weighted sum transmitters.t.. The analysis is for flat channels.. Feedback graphs for all joint Tx–Rx beamforming designs. to be fed back.. By some matrix manipulations. Generally. arcs a ). K Tx- M ETHOD IN C OMPARISON W ITH OTHER BFs (arcs b ). the Tx-BF design with LS complexity and performance are competing factors in K-user reduces to: interference channels. The dashed arcs a represent the all channels which are feedback from K − 1 receiver min max ti tH i nodes to one receiver node and dotted arcs b represent Tx-BF which are vi i=1. 14. Then it is desired to minimize E(ti ni nH H i ti ) which is the power of the ith row of ei for the si th symbol. which bites into the payload capacity. the MSE-based transceiver design in [12] requires K(K − 1) channels (i. VI.. If a complex designs presented here are closed-forms. .e. and it is not addressed here. (The chan. the cofactor of A. and K − 1 Rx-BFs (arcs c ). F EEDBACK R ATE OF P ROPOSED B EAMFORMING the data describing the sounded channel state. Nr = K).t. In the previous section.e. it is evident that K Tx-BFs 2 {ji} vi i=1. K Tx-BFs from receiver nodes to transmitter nodes. min max %det Hi % From problem P . 2. But all of the beamformer should be fed back (Nt = K + 1. then 2(K 3 + K 2 )q implementation. It is emphasized that the feedback rate.1220 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS.  feedback bits are needed for each update of the weights.

Each of these are de- and outdoors). there is no channel coding.2) ZF In this section numerical experiments are described for vali. typical office. in [7].i Hi. Any column of ViZF is a ZF solution. mean angles of departure (AoD)... This difference is because the for ZF is: exchange of average SNR is not required. Strictly.j + 2 I Hi. Moreover multipath  T  diversity for OFDM transmission is possible by deploying the ViZF = N HT1. The other schemes do not have this capability because their Rx-BFs are also the decoder. The ZF uation of BER performance. The maximum multipath diversity for LCP. channel models [16] are designed for indoor WLAN for band. For simplicity. Hi. and in terms of performance. The IEEE 802. Our leakage interference (LI) minimization and gradient based sum- two beamforming methods require lower feedback compared rate maximization [7] are considered here as benchmarks to to these other methods if K + 1 antennas are considered for compare with our proposed EAO method. large office. 2-joint Tx-BF and Rx-BF (one tap for model A.i · · · HTi−1. and mean angles of arrival (AoA) are all set according to the [16] standard. The taneous SNR maximization while the Tx-BFs are fixed and channel models comprise a set of 6 profiles. and K Tx-BFs among receiver nodes are required to be sent. S IMULATION viZF = arg max Vi (:. at frequencies of 2 and 5 GHz. c) c=1. and then applying a fixed communications configuration allows a fair performance This problem has been also solved by gradient based method comparison between the differently optimized beamformers. have a very large range of average ⎛ ⎛ ⎞−1 ⎞ SNRs.i σs j=i yet possible in general as discussed.k vk = 0 for i = k. The IEEE 802. The legend refers to: 1-Rx-BF design for constrained instan- widths of up to 100 MHz. all the users use QPSK in the eval. The Φ = IP ×P OFDM is achieved with maximum likelihood decoding and means that there is no change invoked from the LCP matrix with uniform power delay profile (PDP). Nevertheless.11n standard characterizes MIMO channels The bit error rate performance of the optimized beam- for Wireless Local Area Networks (WLAN).i Vi (:. labeled A to F Tx-BFs null one interference term. c) HH ZF i. and 9 to 18 taps for models B-F).i⎠ communications behavior should be optimized. and for both clusters 1 and 2: average path gains. optimizing (48) with the analogue objective functions. we only present the final capacity (efficiency) over a range of average SNRs. residential/small 3-joint Tx-BF and Rx-BF design for constrained instanta- office. For brevity. joint Tx–Rx BF design for Table III summarizes the feedback bits for all the cases.TOUTOUNCHIAN AND VAUGHAN: BEAMFORMING FOR MIMO-OFDM CHANNELS WITH MULTIPATH DIVERSITY 1221 TABLE III F EEDBACK R ATE FOR D IFFERENT B EAMFORMING S CHEMES IN I NTERFERENCE C HANNELS TABLE IV benefit of LCP in the multiuser MIMO-OFDM interference S IMULATION S ETUP PARAMETERS channel. angular spread (AS) at the receiver and at the transmitter. λ/2 element spacing at the transmit and receive antennas. (Rayleigh result here: channels.) Similarly. This channel description is still far from truly realistic (in particular the antenna aspects) but it is nevertheless a standard allowing simulation results to be repeatable. nullspace criteria imposes that Nt ≥ Nr K [23]. cover the scenarios of flat fading.i (46) fixed precoder matrix. picted with and without the LCP matrix Φ.. (47) dating the analysis. 4 versus average SNR (σs2 /σn2 ).11n formers is depicted in Fig. By this nullspace allocation. for example. As discussed in the Introduction. but this is not i.i HTi+1.11n channel model B is used with the following settings: 3 Hz maximum Doppler shift for all paths with Bell Doppler spectrum. To determine the action at the transmitter side.j vjSR vjSR H HH i. The zero-forcing (ZF) Tx-BF. however to increase ZF performance (ZF with selection): ZF H VII. residential. 15 ns rms delay spread. It can be digital communications performance is a tricky aspect of link shown that the EAO over K games can be applied to sum- optimization and using a single modulation cannot create high rate maximization as well. The Tx-BF design all systems at each terminal.i · · · HTK. the IEEE 802. . and large space (indoors neous SNR and SINR maximization.size(Vi . The simulation parameters are summarized in Table IV. which design for constrained instantaneous SNR maximization.. the digital  K  2 σn ⎠ viSR= wmax ⎝HH ⎝ Hi.

3. NO. 5 illustrates that it also has that our proposed method is five times faster for M = 4 and better sum-rate performance and also lower BER (not shown K = 3 (Fig. Tx-BF and Rx-BF design (approach 3) has the best perfor. From This simulation also demonstrates that for this MIMO channel Fig.11n channel model B. However. 6. The BER performance of ZF with selection.1222 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS. SD decoding at the receiver increases with the constellation size Fig. then the joint single-objective Tx-BF and Rx-BF design proposed methods have been compared with existing known (approach 2). The BER performance comparison of gradient method and proposed former design for K = 4 users with IEEE 802. Fig. sum rate maximization by gradient descent method discussed The key benefit of the presented beamforming approaches is in [7]. VOL. and finally the individual design (approach 1).o) deploying LCP precoder and sphere decoder with IEEE 802. for K = 3 and G(0. The channel for system performance. here) K performance. Finally it is recalled that the complexity of this simulation is flat fading in time domain denoted by G(0. with sum rate objective function. Fig. MARCH 2015 Fig. LI and sum-rate maximization. 7. symbol-wise and frame-wise detection at the receiver have the average path gain of 0 dB. 1) channels. Nt ).11n channel model A. mean angle of departure is 45◦ and over IEEE 802. The required CPU times (not shown) indicate LS with GA for K = 4 users. Aver. 2. The BER performance and computational complexity of mance. The BER performance of approaches 1. tational complexity of all closed-form methods presented here aging over 100 realizations. their computational simplicity. 7). The computational complexity of this gradient method is ing execution times with finding the Tx-BF for LS criteria with O(N K 2 M 3 ) where M = max(Nr . methods. The joint multi-objective mean angle of arrival is 45◦ . 3 for K = 3 with (w) or Fig. The sum rate in bits/s/Hz is computed by Robustness to imperfect channel information can be read- i=1 log2 (1 + SINRi ).8 dB. Here. In summary. angular spreads AS = 40◦ at the same result. channel model B. 5. Fig. 7 compares the sum rate maximization by EAO with but not with the number of antennas. 4. LI and proposed ap- without (w. the performance loss is only 0. The sum rate performance of the approach 3 and LS transmit beam. EAO. the compu- an evolutionary algorithm (GA) at a given average SNR. 6.11n ily gauged in the . the channel is the IEEE 802. This is quantified from compar. 14. From Fig. such as ZF. 4 shows the BER for K = 3 users transmitter and receiver. Fig.11n proaches 2 and 3 for K = 3 and G(0. 1). approaches 2 and 3 have better performance than ZF for WLAN with 9 paths. 7. 1) channels. approach 3 is five times faster than is O(N KM 3 ). using the LCP matrix still improves the with selection and also much better than LI.

unit . which is a flat-fading MIMO channel with Ĥi.j + (1 − ρ2 )W.usual way by modeling the channel with channel model A.j = ρHi. where W is a zero mean.

and s. qH i x=0 plying an LCP matrix. and here our simulations demonstrate that for the more realistic In short. the second algorithm may be better than the third − 2⎝ αj Qaj ⎠ ⎝ |αj |2 ⎠ algorithm and also it requires less feedback. xi+1 . and min Ji (x1 . the algorithms comprise iterative proce. compared max |αj |2 aH H j Qaj + 2 αj αs aj Qas α1 .j . when the same number of antennas is considered. the solution for KKT of problem P is x = r −K+2 tional complexity is less than existing beamforming methods. we get for the second design (approach 2). The results of this paper can also be viewed as some quantification of the trade-offs of between algorithmic −K+2 Nr simplicity. xK ) x∈ΩJ Δ (A-5) The active set A(x) at any feasible x is defined as: A(x) = E ∪ The idea of EAO is to replace this difficult joint optimization of {i ∈ I|Ci (x) = 0}. xK ) + · · · + JK (x1 . With a unit norm for the transmit and Δ receive beamformers. The simplicity of the presented algorithms comes at the & ∗ ' price of one more antenna element at each terminal. (43). Consider the sum of objective functions x Ji (x): s.j while the actual channel is Hi. .i+1 vi+1 . · · · .s>j to existing methods. Consequently the ∇Ci (x) are tie-point. . IEEE 802. if a solution is found.. and this must be checked. xi ∈Ωi . it is not known if it is the global in strongly idealized channels (uniform power delay profile). c. xi . For the first user (say).i+1 random matri. . there are closed forms for α1 = α and α2 = β. .11n channel ⎛ ⎞⎛ ⎞ −K+2 Nr −K+2 Nr model). A lower feedback rate =⎝ |αj |2 ⎠ λi q i (A-3) than existing beamforming methods is a feature of the first two j=1 i=1 algorithms. Otherwise. S UMMARY AND C ONCLUSION The KKT for problem P is: Three new beamforming algorithms are presented for a 2 K−2 multiuser MIMO-OFDM interference channel which can also 2x(xH Qx) − 2Qx(xH x) = (xH x) i=1 λi qi (A-2) develop multipath diversity using the known technique of ap. 2⎝ αj aj⎠⎝ |αj |2 aH ∗ H j Qaj +2 αj αs aj Qas ⎠ For quasi-realistic channels (exponential power delay profile. performance of less than 10−4 at SNR = 10 dB for ρ ≥ 0.TOUTOUNCHIAN AND VAUGHAN: BEAMFORMING FOR MIMO-OFDM CHANNELS WITH MULTIPATH DIVERSITY 1223 variance complex Gaussian random matrix. . . Define B = [qH H 1 .t. a bit error linearly independent. LICQ holds at x∗ if the set J over the sub-problems: of active constraint gradients {∇Ci (x∗ )|i ∈ A(x∗ )} is linearly independent. Ci (x) = 0 i∈E Ci (x) ≥ 0 i ∈ I. A PPENDIX A (A-4) should be solved. A PPENDIX B G LOBAL O PTIMUM OF P ROBLEM P IS A H ARD P ROBLEM VIII. . for the case Nr = K (Section III).j are G(0. It has the lowest feedback ⎛ ⎞2 rate. But it has worse performance compared to the other two −K+2 Nr  K−2 algorithms and some existing designs. The beamformers ces are assumed independent and the vi+1 are non-zero random are obtained from Ĥi.11n channel models. which is hard. xi−1 .αNr −K+2 j=1.. the computa. Therefore. or a local optimum for problem P. By definition. xK ) (A-6) ∇Ci (x) = qi .s>j Kronecker antenna correlations. j=1 j=1. Problem (11) only has equality constraints. qK−2 ] ∈ C (K−2)×Nr . Consider the following E XTENDED A LTERNATING O PTIMIZATION (EAO) optimization problem: The EAO is a general form of AO for multi-objective op- min f (x) timization problem. the addition of the LCP matrix problem P is transformed to a new optimization problem: (prior to the Tx beamformer) still improves the error perfor- −K+2 Nr mance. . Let aj be an or- dures with closed-form steps. . in beamforming j=1 for the MIMO-OFDM interference channel. Nr−K+2 Nr −K+2 & ' ing design.995 when Hi. allowing a fast solution. and this problem is simpler. feedback rate. . LICQ H OLDS FOR P ROBLEM P It is shown that if x∗ is a local solution to non-convex A PPENDIX C problem (11) then LICQ holds at x∗ . j=1 αj aj where αj = 0 are obtained from: It is shown that the third algorithm—joint constrained SNR and ⎛ ⎞⎛ ⎞ SINR maximization—outperforms the least-square beamform. It is known that the LCP matrix improves the error performance Moreover. the H1. Therefore. with a much lower computational time. LICQ holds for problem (11). as in the IEEE 802. (A-1) min J (x) = J1 (x1 . Nr − no derivative or Lagrangian multiplier is needed. . · · · . this is a difficult situation. K N + 2. By assuming uH 1 u1 = 1.t.. . As a vectors so qi = H1. 1) channels. a minimum number of antennas.. . |αj |2 = 1 (A-4) the capability of extracting multipath diversity. .f. . Because thonormal basis for the null space of B where j = 1. The first algorithm j=1 j=1 is the simplest in terms of complexity.

and R. pp. Wong. 291–301. x∗2 .: . . In another notation.. [8] K. . Stacey. solution or stationary solution of the original problem (A-5) is an open problem. beamforming designs in multi-user MIMO-OFDM interference channels. A. Tehran. vol. Tolli. then from 10. Numerical Anal. . 51.. 7. 2004. pp. . K. Tao. . vol.-K. . and T. . 1. . x2 . to the KKT vol. . vol. New York. no. 2011. pp. Heath. no. He is an NSERC Researcher at Energy Aware in fall and Process. and vol. . . Burnaby. K (A-7) no. no.” IEEE Trans. Δ [13] O. and M.-Q. Available: http://dx. The Theory of Matrices. bio-signal processing from the Iran University Jan. Bourdoux and N. pp. Li. . “Adaptive antennas at the mobile and base stations in an OFDM/TDMA system. .-Q. no. Latva-aho.r. “Symbol-wise beamforming for x1= l1 (x2 . xi+1 . in Proc. . K. Gantmacher. 9. If for some a.-F. no. Cheng. multiuser uplink/downlink systems. He [2] M. Mogensen. [19] S. is Nash Equilibrium (NE) for K games.lK (x1. pp. . vol. 2002. [16] E. Sung. Vaughan.. Next Generation Wireless LANs. L. . Toutounchian and R. . . xK−1 )) [12] H. J1 (x∗1 . Spencer.” in Proc. A. Wireless Commun. x∗2 . . 1–5. K. Press. [Online]. xK−1 ) ! "# $ design for a multi-user MIMO interference channel. x∗K−1 . 2009. 461–471. . Y. 1993–1998. . 1802–1816. Swindlehurst. “Cooperative algorithms for MIMO interfer. 1959. Jan. M. the M. . .” in Proc. and I. “The feasibility of interference = fK−1 (x1 . 416–427. His interests include optimization theories for continuous [5] G. Signal Labs in fall 2013. “Solution of the multiuser downlink beam. Ng. Signal Process. Toutounchian and R. to be published. . “Multi-objective optimization J2 (x∗1 . Wang. pp.t. . . xK no. Cambridge. Sep. Sep.” SIAM J. Symp.” in Proc.. . He was a Visiting Graduate Researcher at Adaptive [4] M. 59. B. . and Z. IEEE ICC. Iran. U.” IEEE Trans. Wireless Commun. 5. IEEE VTC Fall. 2004.” IEEE the set of nonlinear equations (A-8) is: Trans.Sc. 1. . K-user MIMO interference channel using multi-objective optimization.-H. . 3. . xK−2 . de- gree in telecommunication engineering from Sharif R EFERENCES University of Technology. Peters. 3. 473–483. .-J. no. Juntti.” IEEE Trans. 2007. 14. vol. 2010. xK−1 = lK−1 (x1 .. Li.K. 1. . .. 4639–4655. and Z. xi−1 .” IEEE Trans.1017/CBO9780511541032 [17] L. Commun. xi−1 . pp. multiuser MIMO-OFDM downlink systems. Sep. . “Space–time interference cancellation in x = f (x). Heath. . Wireless Commun. (n+1) (n+1) (n+1) (n) (n) Signal Process. VOL. [22] X. USA: Chelsea. vol. Perahia and R. lK (x1. 53. xK−1 ]T and f = [f1 . Δ Δ [14] Z. 53. . MARCH 2015 where x1 . Technol. during fall 2012.” IEEE Trans.-S. . 4811–4822. Feb. NY. Sep. 57. . Δ [9] A. xK Jun. Dai.. “MSE-based transceiver designs for ! "# $ the MIMO interference channel. Kisialiou. no. no. .. One approximation for ece. X. Jan. . fK−1 ]T so that OFDM with maximum multipath diversity and coding gains. Semidefinite Relaxation Codes for the Discrete Integer Least Squares Problem. Technol. [21] M. xi = li (x1 . and P. xi−1 . 54. . x∗K ) for multiuser MIMO interference channels. Pers. Liu. Letzepis. He was an Intern at Powertech Tx–Rx beamformers in MIMO downlink channel. For x1 . pp. 587–592.. Haardt. . Boyd and L. x3 . 5755–5760. Jan. xK−1 . Commun. El Ayach. 2004. Technol. . (A-9) MIMO-OFDM systems. . 2. Giannakis. 3. Boche. Fraser University. “Joint design of Systems Laboratory. Indoor Mobile Radio Commun.umn. “Linear precoder designs for K-user interference channels. Tang and R. Niroo Research Insti- vol. 2324–2334. . 49. Nov. xi+1 . 55. .-H. NO. Kisialiou and Z. . Peters and R. . UCLA. Veh. “Iterative multiuser uplink and downlink master’s degree. xi and be expressed by: [7] H. K. and M.” IEEE Trans. [Online]. pp. vol. . Veh. 2002. pp. S. . . is currently working toward the Ph. xK } ≤ U. 11. Park. . M. [15] T.-K. vol. “A constructive proof of the Brouwer fixed-point theorem and computational results. .” IEEE Trans. . pp. xi+1 . 206–218.. After his [3] M. . . T. . . a (where the col operator concatenates vectors). no. li  ≤ a for col{x1 . 2005. xi−1 . xK−1 )) = f1 (x1 . x∗2 . wireless telecommunication engineering with Simon Technol. x2 . Xin. from 2005 to 2008. x∗ the fixed point of f . in 2005.. . x∗K ) ≤ J2 (x∗1 . Luo. 1. [11] M.Sc. . Dec. Canada. . xK pp. obtained by EAO. BC. . he was a Research Member at the beamforming under SINR constraints.” IEEE Trans. “SINR-based transceiver design in the . Dec. Luo. 13. Schubert and H.1224 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS. R. “Zero-forcing methods for   downlink spatial multiplexing in multiuser MIMO channels. . 2001. . Communication Department. Vandenberghe. . and R. 3480–3489.html obtaining x∗ is Gauss–Seidel iteration: [23] Q. 2010. pp. pp. JK x∗1 . . . . winter 2014–2015. Signal Process. . . degree in forming problem with individual SINR constraints. 2009. .. . . 4. 9. Schumacher. 9. Schubert and H. [18] M. pp. Khaled. . no.. Δ [10] Y. Pollok. Pedersen. 2013. pp. 171–174. . Zheng. xi = li x 1 . ing. vol. 18–28. vol. . . . Boche. Nov.doi. Milad Amir Toutounchian received the B. . 195–206. IEEE 13th Int. pp. 4309–4321. Jul. 8. . . . .. . . B. IEEE 78th VTC Fall. xK−1 ) (A-8) alignment over measured MIMO-OFDM channels. W. degree (with honors) in electrical engineer- ence channels. Veh. no. .” IEEE Trans. xK (A-11) [24] R. May 2008.” IEEE Trans. Liu. . Let the optimization problem (A-6) have a unique global SDMA based on a null-space constraint. 1. 2005. . vol. “Joint TX–RX optimisation for MIMO- fixed). no. Wong. W. x∗K−1 . x∗K ) ≤ J1 (x1 .K. xK−1 . Luo.. 2008. Convex Optimization. Shen. and G. ..” IEEE Trans.. Technol. . 12. Codreanu. xi+1 .” IEEE Trans. . . Y. 2011. minimizer w. 57. . Cambridge. “Throughput maximization in linear non-convex/convex problems and algorithms for cellular network enhancement. . pp. Veh.-K. M. Vaughan. xK .. vol. “Efficient implementation of quasi- (A-10) maximum-likelihood detection based on semidefinite relaxation. Luo. 9. Cowley. xK are assumed to be known (or [6] A. Kellogg. S. Y. the NE of K-games.” IEEE Trans. Yorke. “Max–min fairness linear transceiver x2 = l2 (x1. . Available: http://www.   [20] F. define: Commun. xK−1 )) = f2 (x1.. x∗K ≤ JK x∗1 . Relating the point x∗ . Cambridge Univ. and J. and N. in 2002 and [1] S. .. Murch. vol. Letaief. xK ) i = 1. . no. Veh. 2003.. pp. Lee. Heath. .: Cambridge Univ. . . 60.” IEEE Trans. Press.-Q.. “Linear constellation precoding for Denote x = [x1 .. 1–5. 2.” . “From antenna spacings Brouwer’s fixed point theorem [24] ∃{x∗i }K i=1 such that: to theoretical capacities—Guidelines for simulating MIMO systems. lK (x1 . tute. G. Mar.. 12. xK−1 ) MIMO-OFDM transceivers in the presence of co-channel interference ! "# $ and spatial correlation.D. of Science and Technology. x∗K ) Sep. 2010. 52. Lee. . . vol. A. and X.

He worked with the New Zealand Post Office (now Telecom NZ Ltd. and continues as the New Zealand URSI Commission B (Fields and Waves) Representative. New Zealand. signal pro- cessing. all in electrical engineering. Burnaby. His current research involves acoustic and radio wave propagation. He was an Associate Editor (compact antenna elements and arrays. he became a Professor of electrical engineering and the Sierra Wireless Chair in Communications at the School of Engineering Science.K. and diversity/MIMO techniques) for the IEEE T RANSACTIONS and has guest-edited Special Issues. an URSI Correspondent. design. Canada. communications techniques. He was an URSI Young Scientist in 1982 for Fields and Waves and in 1983 for Electromagnetic The- ory. In 2003. In 2006. multipath propagation. and in 2011 on review panels for the Canadian Communications Research Centre and for the ICT Centre of the CSIRO. He is a 2004 Fellow of the BC Advanced System Institute.TOUTOUNCHIAN AND VAUGHAN: BEAMFORMING FOR MIMO-OFDM CHANNELS WITH MULTIPATH DIVERSITY 1225 Rodney G. mobile an- tennas. Christchurch.).D. he served on the international panel for reviewing the funding and knowledge base for ICT in the U. and the theory. Denmark. Canada. He is a Registered Professional Engineer in the Province of British Columbia. and an IEEE Distinguished Lecturer. Aalborg. . Simon Fraser University. de- gree from Aalborg University. and evaluation of arrays and multiport antennas for communications and industrial food heating. Vaughan received the Bachelor’s and Master’s degrees from the University of Canterbury. and Industrial Research Limited. the New Zealand Department of Scientific and In- dustrial Research. BC. and the Ph.