1 Additional Proofs for Joint User Grouping and Linear Virtual Beamforming: Complexity, Algorithms and Approximation Bounds Mingyi Hong, Zi Xu, Meisam Razaviyayn and Zhi-Quan Luo Abstract In this document, we provide missing proofs for some of the results in the paper “Joint User Grouping and Linear Virtual Beamforming: Complexity, Algorithms and Approximation Bounds” [1]. I. P ROOF OF C LAIM 2 Proof: Suppose on the contrary, there exists i ∈ M such that Tr[Di X∗ ] < 1. Then from the definition of Di in (6) of [1], we have: Tr[Ci,1 X∗1 ] + Tr[Ci,0 X∗0 ] < 1, which is equivalent to: P1 X∗1 [i, i] < 12 + 12 X∗0 [i, M + 1]. Due to the assumption that the inequality is strict, we can find a constant δ > 0 such that 1 1 1 ∗ X1 [i, i] + δ = + X∗0 [i, M + 1]. P 2 2 b 1 = X∗ + δei eT 0, then X b = blkdg[X∗ , X b 1 ] is also feasible for problem As a result, let X 1 i 0 (SDP1). This alternative solution achieves the following objective (a) b = Tr[R(X∗ + δei eT )] = Tr[RX∗ ] + δTr[Rei eT ] > Tr[RX∗ ], v1SDP (X) 1 i 1 i 1 (1) where in (a) we have used that fact that δ > 0 and R[i, i] > 0 (due to the strict positive definiteness of R). Clearly, the inequality (1) is a contradiction to the optimality of X∗ . In conclusion, we have 1 ∗ 1 1 X [i, i] = + X∗0 [i, M + 1], ∀ i = 1, · · · , M. P 1 2 2 (2) II. P ROOF OF A PPROXIMATION R ATIO FOR S CHEDULING P ROBLEM Proof: Our goal is to show that there exits a δ > 0 such that Prob (ℓ) (ℓ) min {(wk )H Rwk } ≥ k=1,2 1 min {Tr[RX∗k ]} α2 k=1,2 ≥ δ > 0. M. Hong, M. Razaviyayn and Z.-Q. Luo are with the Department of Electrical and Computer Engineering University of Minnesota, Minneapolis, USA. Z. Xu is with the Department of Mathematics, Shanghai University, Shanghai, China. 2 Note that we have 1 ∗ Prob ≥ min {Tr[RXk ]} α2 k=1,2 ) ! ( 1 1 Tr[R[Sk ]Yk∗ ] ≥ min {Tr[RX∗k ]} = Prob min (ℓ) k=1,2 α2 k=1,2 (t )2 (ℓ) (ℓ) min {(wk )H Rwk } k=1,2 k ≥ Prob min k=1,2 {Tr[R[Sk ]Yk∗ ]} β 1 1 1 1 ≥ min {Tr[RX∗k ]}, (ℓ) ≥ , (ℓ) ≥ α2 k=1,2 β (t )2 β (t )2 1 2 ! . We first lower bound Tr[R[S2 ]Y2∗ ] by Tr[R[S2 ]Y2∗ ] ≥ Tr[R[S2 ]X∗2 [S2 ]], where the inequality is from the optimality of Y2∗ as well as the feasibility of X∗2 [S2 ] to the problem (17) in [1]. The right hand side of the above inequality can be further lower bounded by Tr[R[S2 ]X∗2 [S2 ]] ≥ X X∗2 [i, i]λM (R) i∈S2 P λM (R) = 2 (i) (ii) P λM (R) ≥ 2 M −Q− X ! X∗0 [i, M + 1] i∈S2 (M − Q)Q P (M − Q)2 M −Q− = λM (R). M M (3) where (i) is because of the tightness of the constraint (18b) in [1]. We argue the inequality (ii) as follows. Due to Step S2) of the algorithm, {X∗0 [i, M + 1]}i∈S̄ are the smallest M − Q P ∗ elements in {X∗0 [i, M + 1]}M . Combining with the fact that M i=1 i=1 X0 [i, M + 1] = Q, we must P have that i∈S̄ X∗0 [i, M + 1] ≤ MM−Q Q, which further implies (b). Combining the above two 2 inequalities, we obtain Tr[R[S2 ]Y2∗ ] ≥ P (MM−Q) λM (R). 2 Using the same argument, we can lower bound Tr[R[S1 ]Y1∗ ] by: Tr[R[S1 ]Y1∗ ] ≥ QMP λM (R). Combining the above two estimates, we have that min {Tr[R[Sk ]Yk∗ ]} ≥ min{Q2 , (M − Q)2 } k=1,2 P λM (R). M Note that we further have Tr[RX∗1 ] ≤ λ1 (R)Tr[X∗1 ] = λ1 (R)QP Tr[RX∗2 ] ≤ λ1 (R)Tr[X∗2 ] = λ1 (R)P (M − Q) which implies mink=1,2 {Tr[RX∗k ]} ≤ λ1 (R)P min{Q, M − Q}. Consequently, when choosing β λ1 (R)P min{Q, M − Q} λ1 (R)M = = Q 2 2 α2 λM (R) min{Q, M − Q} min{Q , (M − Q) }P M λM (R) we have Prob min k=1,2 {Tr[R[Sk ]X∗k ]} β < min {Tr[RX∗k ]} α2 k=1,2 = 0. 3 Going through similar steps as in Step 2)–Step 3) in the proof of Theorem 1, the final ratio is α2 = 8M λ1 (R) ln(12 max{Q, M − Q}). min{Q, M − Q}λM (R) (4) This completes the proof. III. P ROOF OF P ROPOSITION 2 Proof: Below we show the case when R is diagonal. The claim is proved by using a polynomial time reduction from the equal partitioning with equal cardinality problem. Let M be a even number and set Q = M2 . Given a vector c ∈ RM consists of positive elements P c1 , · · · , cM , let C = M partitioning with equal cardinality problem finds i=1 ci > 0, the equal P M C an index set I with |I| = 2 such that 2 = i∈I ci . Let R = diag(c). We claim that determining if problem (CP2) can achieve an optimal value of CP is NP-hard. Let I denote the set such that I = {i : ai = 1} with |I| = M2 . Then the 2 objective of (CP2) can be written as min X i∈I ci |w1,i |2 , X j ∈I / X X ci , P cj |w2,j |2 ≤ min P cj , i∈I j ∈I / where the inequality is achieved by setting |w1,i |2 = P, |w2,i|2 = 0, ∀ i ∈ I, |w1,i|2 = 0, |w2,i |2 = P, ∀ i ∈ / I. Clearly, checking if we can find an index set I so that this problem can achieve an objective value of CP is equivalent to finding a subset I with |I| = M2 satisfying 2 P C = i∈I ci , which is exactly the equal partitioning with equal cardinality problem. 2 The case when R is of rank 1 can be shown similarly. IV. P ROOF OF P ROPOSITION 4 Proof: Assume that fi = 1, ∀i. It suffices to show that checking the feasibility problem wkH E(ggH )wk ≥t 2 + wH diag (E[ggH ]) w k=1,2 σn k k |w1,i |2 P0 E[|fi |2 ] + σv2 ≤ ai P, |w2,i |2 P0 E[|fi |2 ] + σv2 ≤ (1 − ai )P, i = 1, · · · , M min M X ai = Q, (5) ai ∈ {0, 1}, i = 1, · · · , M, i=1 is NP-hard. Let us also consider the following system parameters: E[|gi |2 ] = ci > 0, P0 = σv2 = 1, P = 2, Q= M , 2 t= 1 , 2 σn2 = X ci /2. i Using the above parameters, one can easily show that problem (5) is feasible iff the following NP-complete equal partition with equal cardinality problem is feasible: P P Find an index set I ⊆ {1, 2, . . . , M} with |I| = M2 such that i∈I ci = i∈I¯c ci . Such equivalence completes the proof. 4 R EFERENCES [1] M. Hong, Z. Xu, M. Razaviyayn, and Z.-Q. Luo, “Joint user grouping and linear virtual beamforming: Complexity, algorithms and approximation bounds,” 2012, To appear, IEEE Journal on Selected Areas in Communications, Special issue on Virtual Multiantenna Systems.