Lecture 14 Quadratic Optimization and its SDP Relaxation • QCQP & motivation • Complexity of QCQP • SDP relaxation for QCQP • When is SDP relaxation tight? • Approximation upper and lower bounds • Proof ideas SDP Relaxation for Nonconvex QP Zhi-Quan Luo Quadratic Optimization Problem the generic QCQP can be written: minimize x†P 0x + r0 subject to x†P ix + ri ≤ 0, i = 1, . . . , m • x can be real or complex • if all Pi are (Hermitian) p.s.d., convex problem, • here, we suppose at least one Pi is not p.s.d. • can easily include linear terms • vast range of applications... 1 SDP Relaxation for Nonconvex QP Zhi-Quan Luo Example: Boolean Least Squares Boolean least-squares problem: minimize kAx − bk2 subject to x2i = 1, i = 1, . . . , n • basic problem in digital communications • could check all 2n possible values of x . . . • an NP-hard problem, and very hard in practice • many heuristics for approximate solution 2 SDP Relaxation for Nonconvex QP Zhi-Quan Luo Example: Partition Problem minimize subject to xT W x x2i = 1, i = 1, . . . , n where W ∈ S n, with W ii = 0. • a feasible x corresponds to the partition {1, . . . , n} = {i | xi = −1} ∪ {i | xi = 1} • the matrix coefficient W ij can be interpreted as the cost of having the elements i and j in the same partition. • the objective is to find the partition with least total cost • particular instance: MAXCUT (W ij ≥ 0) 3 SDP Relaxation for Nonconvex QP Zhi-Quan Luo Motivation: Transmit Beamforming • Downlink transmission: base station has K antennas; m receivers • n multicast groups, {G1, . . . , Gn}, Gk ∩ Gl = ∅ • ∪k Gk = {1, . . . , m}, Pn Gk := |Gk |, k=1 Gk = m. • wk : beamforming vector for Gk • sk : signal sent to group Gk • transmitted signal: s(t) = n X sk (t)wk k=1 4 SDP Relaxation for Nonconvex QP Zhi-Quan Luo Transmit Beamforming • assume each receiver has one antenna, with channel vector hi • For ULA, LoS, far field, † −jφi −j(K−1)φi hi = 1, e , ..., e where φi = 2π λd sin θi; Vandermonde. • signal at receiver i ∈ Gk : † s hi + v i = † sk wk hi + X † sl wl hi + v i l6=k • SINR for receiver i ∈ Gk |w†k hi|2 † 2 2 |w l6=k l hi | + σi P 5 SDP Relaxation for Nonconvex QP Zhi-Quan Luo Problem Description: Transmit Beamforming • Transmit beamforming problem: minimize transmit power, subject to QoS constraint for each receiver in each group. minimize n X kwk k k=1 subject to 2 |w†k hi|2 † 2 |w l6=k l hi | P + σi2 ≥ ci, ∀i ∈ Gk , ∀k ∈ {1, . . . , n} m minimize subject to n X kwk k k=1 † 2 |wk hi| 2 − ci X † 2 2 |wl hi| ≥ ciσi , ∀i ∈ Gk , ∀k ∈ {1, . . . , n} l6=k ⇑ separable homogeneous QCQP 6 SDP Relaxation for Nonconvex QP Zhi-Quan Luo Complexity of QCQP • QCQP is in general NP-hard: contains binary LP/QP as special cases quadratic constraint x2i = 1 ⇐⇒ binary constraint xi = ±1 • “Simple” cases of QCQP: Single group multicast transmit beamforming minimize kwk2 subject to |w†hi|2 ≥ 1, ∀i ∈ {1, . . . , n} remains NP-hard. 7 SDP Relaxation for Nonconvex QP Zhi-Quan Luo Complexity of QCQP over Rm • A known NP-complete problem Π: Given integers P a1, · · · , am, do there exist m m m binary variables {wi}i=1 ∈ {+1, −1} , such that n=1 aiwi = 0? • A polynomial reduction to the single group multicast transmit beamforming over the reals: 2 w12 + · · · + wm + minimize m w∈R m X !2 aiwi = w†Qw i=1 subject to |wi| ≥ 1, ∀i ∈ {1, . . . , m} † where a := [a1, · · · , am] , and Q := I + aa†. • “Yes” for Π iff the optimal objective of QCQP = m. 8 SDP Relaxation for Nonconvex QP Zhi-Quan Luo Complexity of QCQP over Cn • Let n = 2m + 1 and let the complex-valued decision vector be w = [w0, w1, ..., wm, wm+1, . . . , w2m]T ∈ Cn. • Let us denote a := A := T a1 a2 . . . am −1m Im Im − 21 1†ma a† 0†m , , Q := AT A + I m, where 1m, 0m denote the all-one, the zero vectors in Cm. 9 SDP Relaxation for Nonconvex QP Zhi-Quan Luo • Claim: “Yes” for Π iff the following QCQP over Cn has a minimum value of n minimize n w∈C w†Qw subject to |wi| ≥ 1, i = 1, ..., n, • Since w†Qw = kAwk22 + 2m X |wi|2 ≥ 2m + 1 = n, for |wi| ≥ 1 ∀i, it follows i=0 that w†Qw = n, |wi| ≥ 1 ∀ i ⇐⇒ Aw = 0, |wi| = 1 ∀ i. • This gives rise to a set of linear equations: − w0 + wi + wm+i = 0, m m X 1 X − ai w0 + aiwi = 0. 2 i=1 i=1 i = 1, . . . , m, (1) (2) 10 SDP Relaxation for Nonconvex QP Zhi-Quan Luo • Since |wi| = 1, we can let wi/w0 = eiθi for i = 1, . . . , 2m. Using (1) we have cos θi + cos θm+i = 1, sin θi + sin θm+i = 0, where i = 1, . . . , m. These two equations imply that θi ∈ {−π/3, π/3} for all i. This in particular means that cos θi = cos θm+i = 1/2 for i = 1, . . . , m, implying that ( ! ) m m X 1 X Re − ai + aiwi/w0 = 0. 2 i=1 i=1 • Therefore, (2) is satisfied if and only if ( 1 Im − 2 m X i=1 ! ai + m X ) aiwi/w0 i=1 = Im (m X i=1 ) aiwi/w0 = m X ai sin θi = 0, i=1 which is equivalent to “Yes” for Π. 11 SDP Relaxation for Nonconvex QP Zhi-Quan Luo Convex Relaxation of QCQP Using a fundamental observation: X := xxT ⇔ X 0, rank(X) = 1, and noting xT Pix = Tr (XPi), the original QCQP: minimize xT P 0x + r0 subject to xT P ix + ri ≤ 0, i = 1, . . . , m 12 SDP Relaxation for Nonconvex QP Zhi-Quan Luo can be rewritten: minimize Tr (XP 0) + r0 subject to Tr (XP i) + ri ≤ 0, X0 rank(X) = 1 i = 1, . . . , m the only nonconvex constraint is now rank(X) = 1... 13 SDP Relaxation for Nonconvex QP Zhi-Quan Luo Convex Relaxation: Semidefinite Relaxation • we can directly relax this last constraint, i.e. drop the nonconvex rank(X) = 1 to keep only X 0 • the resulting program gives a lower bound on the optimal value minimize Tr (XP 0) + r0 subject to Tr (XP i) + ri ≤ 0, X0 i = 1, . . . , m ⇒ SDP 14 SDP Relaxation for Nonconvex QP Zhi-Quan Luo How to Generate a Feasible Solution? Let X ∗ be the optimal solution of • pick x as a Gaussian variable with x ∼ N (0, X ∗) • Since Tr (X ∗P i) + ri = E[xT P ix + ri], x will solve the QCQP “on average” over this distribution in other words: minimize E[xT P 0x + r0] subject to E[xT P ix + ri] ≤ 0, i = 1, . . . , m a good feasible point can then be obtained by sampling enough x. . . Key question: • how good is the approximate solution x? • can we bound f (x)/f ∗ by a constant? 15 SDP Relaxation for Nonconvex QP Zhi-Quan Luo Problem Description: Separable Homogeneous QCQP Let C i, Aij be Hermitian matrices, and H = R or C. minimize wi ∈HKi subject to n X † w i Ci w i i=1 † † w1 A11 w1 + w2 A12 w2 + · · · + w†n A1n wn ≥ b1 † † w1 A21 w1 + w2 A22 w2 + · · · + w†n A2n wn ≥ b2 ⇐ NP-hard .. . † † w1 Am1 w1 + w2 Am2 w2 + · · · + w†n Amn wn ≥ bm minimize n X Tr(C i W i ) i=1 subject to Tr(A11 W 1 + A12 W 2 + · · · + A1n W n ) ≥ b1 .. . ⇐ SDP relaxation Tr(Am1 W 1 + Am2 W 2 + · · · + Amn W n ) ≥ bm W i 0. Question: How good is the SDP relaxation? 16 SDP Relaxation for Nonconvex QP Zhi-Quan Luo Empirical Behavior • Measured VDSL channel data by France Telecom R&D; • 17 measured channel scenarios, 28 loops, 300m, 21.5-30 Mhz; • Compared SDP solution vs precoding (ignore crosstalk) no • SDP yields nearly doubling of minimum received signal power • SDP relaxation is tight in over 50% of instances. • SDP solved by the base station Transmit precoding for VDSL multicasting. • Question: ∃ a theoretical performance bound? 17 SDP Relaxation for Nonconvex QP Zhi-Quan Luo Problem Description: Separable Homogeneous QCQP Let C i, Aij be Hermitian matrices, and H = R or C. minimize wi ∈HKi subject to n X † w i Ci w i i=1 † † w1 A11 w1 + w2 A12 w2 + · · · + w†n A1n wn ≥ b1 † † w1 A21 w1 + w2 A22 w2 + · · · + w†n A2n wn ≥ b2 ⇐ NP-hard .. . † † w1 Am1 w1 + w2 Am2 w2 + · · · + w†n Amn wn ≥ bm minimize n X Tr(Ci W i ) i=1 subject to Tr(A11 W 1 + A12 W 2 + · · · + A1n W n ) ≥ b1 .. . ⇐ SDP relaxation Tr(Am1 W 1 + Am2 W 2 + · · · + Amn W n ) ≥ bm W i 0. Question: How good is the SDP relaxation? 18 SDP Relaxation for Nonconvex QP Zhi-Quan Luo Worst-case SDP Approximation Performance In general, SDP relaxation for separable homogeneous QCQP is not tight. Example: For any M > 0, consider υqp := min s.t. x2 + y 2 y 2 ≥ 1, x2 − M xy ≥ 1, x2 + M xy ≥ 1. • Notice that the last two constraints of QCQP imply x2 ≥ M |x||y| + 1, x2 ≥ 1, which, in light of the constraints y 2 ≥ 1, further imply x2 ≥ M + 1. ⇒ υqp ≥ M + 2. • However, for the SDP relaxation υsdp := min s.t. X(11) + X(22) X(22) ≥ 1, X(11) − M X(12) ≥ 1, X(22) + M X(12) ≥ 1, X 0. • X = I is clearly a feasible solution, implying υsdp ≤ 2. ⇒ the performance ratio υqp /υsdp is at least (M + 2)/2, which can be arbitrarily large. 19 SDP Relaxation for Nonconvex QP Zhi-Quan Luo Simple Cases 1. Ki = 1, for all i. • Then, wi is a scalar, implying W i ≥ 0 ⇔ W i = w2i for some wi. • The SDP relaxation is a LP, and is equivalent to the original nonconvex QCQP. 2. m = n = 1 • Then the separable homogeneous QCQP becomes † minimize w Cw, † subject to w Aw ≥ b. • This is a generalized eigenvalue problem. • For this case, SDP relaxation is tight. ⇒ SDP admits a rank-1 solution. 20 SDP Relaxation for Nonconvex QP Zhi-Quan Luo Unicast • When each group Gk has one receiver, then m = n, and the transmit beamforming becomes minimize subject to n X kw i k i=1 † 2 |wi hi| 2 − ci X † 2 2 |wl hi| ≥ ciσi , ∀i ∈ {1, . . . , n} l6=i • Surprise: SDP relaxation is always tight in this case! (Bengtsson-Ottersten, 2001). • Proof relies on Lagrangian duality and Perron-Frobenius theorem. • Is there a more general statement? 21 SDP Relaxation for Nonconvex QP Zhi-Quan Luo Existence of Low Rank SDP Solution • (Strengthened) Pataki’s result: The separable SDP has a solution {W ∗i }ni=1 satisfying n X r(W ∗)(r(W ∗ + 1) i i 2 i=1 ≤ m. • If W ∗i 6= 0 for all i, then r(W ∗i ) ≥ 1, implying n≤ n X r(W ∗)(r(W ∗ + 1) i i=1 i 2 ≤ m. • If in addition m ≤ n, then m = n and each W ∗i must be rank-1. • This shows why SDP relaxation is tight for the unicast case. (All feasible beamforming solutions have W i 6= 0.) 22 SDP Relaxation for Nonconvex QP Zhi-Quan Luo Yet Another Case: Vandermonde hi • For uniform linear array, line of sight and far-field propagation, hi is Vandermonde † −jφi −j(K−1)φi hi = 1, e , ..., e := a(φi), Result: If all hi = a(φi) for some φi ∈ [0, π), then the SDP relaxation of the separable homogeneous QCQP has a rank-1 solution. (m > n allowed) minimize subject to n X 2 k wk k k=1 X † 2 † 2 2 |wl hi | ≥ ci σi , ∀i ∈ Gk , ∀k ∈ {1, . . . , n} | w k hi | − c i l6=k m minimize n X Tr(W k ) k=1 subject to † Tr(W k hi hi ) − ci X † 2 Tr(W l hi hi ) ≥ ci σi , ∀i ∈ Gk , ∀k ∈ {1, . . . , n} l6=k W k 0. 23