Post on 01-Feb-2023
Multiuser Downlink Beamforming:Rank-Constrained SDP
Prof. Daniel P. Palomar
ELEC5470/IEDA6100A - Convex Optimization
The Hong Kong University of Science and Technology (HKUST)
Fall 2020-21
Outline of Lecture
• Problem formulation: downlink beamforming
• SOCP formulation, SDP relaxation, and equivalent reformulation
• General framework: rank-constrained SDP
• Numerical results
• Summary
Daniel P. Palomar 1
Downlink Beamforming
• Consider several multi-antenna base stations serving single-antenna
users:
Daniel P. Palomar 3
Signal Model
• Signal transmitted by the kth base station:
xk (t) =∑m∈Ik
wmsm (t)
where Ik are the users assigned to the kth base station.
• Signal received by mth user:
rm (t) = hHm,κmxκm +∑k 6=κm
hHm,kxk + nm (t)
where κm is the base station assigned to user m, hm,k is the channel
vector from base station k to user m with correlation matrix Rm,k.
Daniel P. Palomar 4
SINR Constraints
• SINR of the mth user is
SINRm =wHmRmmwm∑
l 6=mwHl Rmlwl + σ2
m
.
• We can similarly consider an instantaneous version of the SINR, as
well as the SINR for a dirty-paper coding type of transmission.
• Optimization variables: beamvectors wl for l = 1, . . . , L.
• The SINR constraints are
SINRm ≥ ρm for m = 1, . . . , L.
Daniel P. Palomar 5
Formulation of Optimal Beamforming Problem
• Minimization of the overall transmitted power subject to SINR
constraints:
minimize{wl}
∑Ll=1 ‖wl‖2
subject towHmRmmwm∑
l 6=mwHl
Rmlwl+σ2m≥ ρm, m = 1, . . . , L.
• This beamforming optimization is nonconvex!!
Daniel P. Palomar 6
• Assume rank-one direct channels: Rmm = hmhHm.
• By adding, w.l.o.g., the constraint wHmhm ≥ 0, we can write the
beamforming problem as
minimize{wl}
∑Ll=1 ‖wl‖2
subject to(wHmhm
)2 ≥ ρm (∑l 6=mwHl Rmlwl + σ2
m
), ∀m
wHmhm ≥ 0.
• Observe we can write∑l 6=m
wHl Rmlwl + σ2
m =∥∥∥R̃1/2
m w̃m
∥∥∥2Daniel P. Palomar 8
where
w̃m =
w1
...
wL
1
and R̃m = diag
Rm1...
0...
RmL
σ2m
.
• The beamforming problem can be finally written as the SOCP:
minimize{wl}
∑Ll=1 ‖wl‖2
subject to wHmhm ≥ ρm
∥∥∥R̃1/2m w̃m
∥∥∥ , ∀m.
Daniel P. Palomar 9
• Let’s start by rewriting the SINR constraint
wHmRmmwm∑
l 6=mwHl Rmlwl + σ2
m
≥ ρm
as
wHmRmmwm − ρm
∑l 6=m
wHl Rmlwl ≥ ρmσ2
m.
• Then, define the rank-one matrix: Xl = wlwHl .
• Notation: observe that
wHl Awl = Tr
(wHl Awl
)= Tr
(Awlw
Hl
)= Tr (AXl) , A •Xl.
Daniel P. Palomar 11
• The beamforming optimization can be rewritten as
minimizeX1,...,XL
∑Ll=1 I •Xl
subject to Rmm •Xm − ρm∑l 6=mRml •Xl ≥ ρmσ2
m, ∀mXl � 0
rank (Xl) = 1
• This is still a nonconvex problem, but if we remove the rank-one
constraint, it becomes an SDP!
• Thus, a relaxation of the beamforming problem is an SDP which
can be solved in polynomial time.
Daniel P. Palomar 12
• An equivalent reformulation was obtained in the seminal work:
– M. Bengtsson and B. Ottersten, “Optimal and suboptimal trasmit beamforming,” in Handbook
of Antennas in Wireless Communications, ch. 18, L.C. Godara, Ed., Boca Raton, FL: CRC
Press, Aug. 2001.
– M. Bengtsson and B. Ottersten, ”Optimal transmit beamforming using semidefinite
optimization,” in Proc. of 37th Annual Allerton Conference on Communication, Control,
and Computing , Sept. 1999.
• Theorem: If the SDP relaxation is feasible, then it has at least one
rank-one solution.
• The proof is quite involved as it requires tools from Lagrange duality
and Perron-Frobenius theory on nonnegative matrices.
Daniel P. Palomar 14
• Lemma: The following two problems have the same unique optimal
solution λ:
minimizeλ,{um}
∑Ll=1 λlρlσ
2l
subject to uHm
(I− λmRmm +
∑l 6=m λlρlRml
)ui = 0, ∀m
‖um‖ = 1
λm ≥ 0
and
maximizeλ
∑Ll=1 λlρlσ
2l
subject to I− λmRmm +∑l 6=m λlρlRml � 0, ∀m
λm ≥ 0.
Daniel P. Palomar 15
• The next two steps of the proof as the following:
1. show that the dual problem of the beamforming problem is exactly
the maximization in the previous lemma (this is quite simple)
2. show that the beamforming problem can be rewritten as the
minimization in the previous lemma (this is quite challenging and
requires Perron-Frobenius theory).
• At this point, one just needs to realize that what has been proved
is that strong duality holds. This implies that an optimal solution
of the original problem is also an optimal solution of the relaxed
problem; hence, the relaxed problem must have a rank-one solution.
Daniel P. Palomar 16
Soft-Shaping Interference Constraints
• Constraint to control the amount of interference generated along
some particular direction:
N∑k=1
E[∣∣hHm,kxk (t)∣∣2] ≤ τm for m = L+ 1, . . . ,M.
• Defining Sml = hm,κlhHm,κl
, we can rewrite it as
L∑l=1
wHl Smlwl ≤ τm for m = L+ 1, . . . ,M.
• We can similarly consider long-term constraints, high rank matrices
Sml, total power transmitted from one particular base station to one
external user.
Daniel P. Palomar 18
Individual Shaping Interference Constraints
• Some interference constraints only affect one beamvector and we
call them individual constraints.
• We consider the following two types of constraints:
wHl Blwl = 0 for l ∈ E
wHl Blwl ≥ 0 for l /∈ E .
• The following global null shaping constraint is equivalent to individ-
ual constraints (assuming Sml � 0):
L∑l=1
wHl Smlwl ≤ 0.
Daniel P. Palomar 19
Formulation of Optimal Beamforming Problem
• Minimization of the overall transmitted power subject to all the
previous SINR and interference constraints:
minimize{wl}
∑Ll=1 ‖wl‖2
subject towHmRmmwm∑
l 6=mwHl
Rmlwl+σ2m≥ ρm, m = 1, . . . , L∑L
l=1 wHl Smlwl ≤ τm, m = L+ 1, . . . ,M
wHl Blwl = (≥) 0, l ∈ (/∈) E1
wHl Dlwl = (≥) 0, l ∈ (/∈) E2.
• This beamforming optimization is nonconvex!!
Daniel P. Palomar 20
SDP Relaxation
• Let’s start by rewriting the SINR constraint
wHmRmmwm∑
l 6=mwHl Rmlwl + σ2
m
≥ ρm
as
wHmRmmwm − ρm
∑l 6=m
wHl Rmlwl ≥ ρmσ2
m.
• Then, define the rank-one matrix: Xl = wlwHl .
• Notation: observe that
wHl Awl = Tr (AXl) , A •Xl.
Daniel P. Palomar 21
• The beamforming optimization can be rewritten as
minimizeX1,...,XL
∑Ll=1 I •Xl
subject to Rmm •Xm − ρm∑l 6=mRml •Xl ≥ ρmσ2
m, m = 1, . . . , L∑Ll=1 Sml •Xl ≤ τm, m = L+ 1, . . . ,M
Bl •Xl = (≥) 0, l ∈ (/∈) E1Dl •Xl = (≥) 0, l ∈ (/∈) E2Xl � 0, rank (Xl) = 1 l = 1, . . . , L.
• This is still a nonconvex problem, but if we remove the rank-one
constraint, it becomes an SDP!
• Thus, a relaxation of the beamforming problem is an SDP which
can be solved in polynomial time.
Daniel P. Palomar 22
Existing Literature
• M. Bengtsson and B. Ottersten, “Optimal and suboptimal trasmit beamforming,” in Handbook
of Antennas in Wireless Communications, ch. 18, L.C. Godara, Ed., Boca Raton, FL: CRC
Press, Aug. 2001.
• M. Bengtsson and B. Ottersten, ”Optimal transmit beamforming using semidefinite optimization,”
in Proc. of 37th Annual Allerton Conference on Communication, Control, and Computing ,
Sept. 1999.
• D. Hammarwall, M. Bengtsson, and B. Ottersten, ”On downlink beamforming with indefinite
shaping constraints,” IEEE Trans. on Signal Processing , vol. 54, no. 9, Sept. 2006.
• G. Scutari, D. P. Palomar, and S. Barbarossa, “Cognitive MIMO Radio,” IEEE Signal Processing
Magazine, vol. 25, no. 6, Nov. 2008.
• Z.-Q. Luo and T.-H. Chang, “SDP relaxation of homogeneous quadratic optimization: approxima-
tion bounds and applications,” Convex Optimization in Signal Processing and Communications,
D.P. Palomar and Y. Eldar, Eds., Cambridge University Press, to appear in 2009.
Daniel P. Palomar 23
Curiosity: Multicast Downlink Beamforming
• This small variation was proven to be NP-hard:
– N. D. Sidiropoulos, T. N. Davidson, Z.-Q. (Tom) Luo, “Transmit Beamformingfor Physical Layer Multicasting”, IEEE Trans. on Signal Processing , vol. 54,no. 6, June 2006.
Daniel P. Palomar 24
Separable SDP
• Consider the following separable SDP:
minimizeX1,...,XL
∑Ll=1 Cl •Xl
subject to∑Ll=1 Aml •Xl = bm, m = 1, . . . ,M
Xl � 0, l = 1, . . . , L.
• If we solve this SDP, we will obtain an optimal solution X?1, . . . ,X
?L
with an arbitrary rank profile.
• Can achieve a low-rank optimal solution? (recall that for the
beamforming problem we want a rank-one solution)
Daniel P. Palomar 26
Rank-Constrained Solutions
• A general framework for rank-constrained SDP was developed in
– Yongwei Huang and Daniel P. Palomar, “Rank-Constrained Separable Semidef-inite Programming for Optimal Beamforming Design,” in Proc. IEEE Inter-national Symposium on Information Theory (ISIT’09), Seoul, Korea, June28 - July 3, 2009.
– Yongwei Huang and Daniel P. Palomar, “Rank-Constrained Separable Semidef-inite Programming with Applications to Optimal Beamforming,” IEEE Trans.on Signal Processing , vol. 58, no. 2, pp. 664-678, Feb. 2010.
• Theorem: Suppose the SDP is solvable and strictly feasible. Then,
it has always an optimal solution such that
L∑l=1
rank2 (Xl) ≤M.
Daniel P. Palomar 27
Purification-Type AlgorithmAlgorithm: Rank-constrained solution procedure for separable SDP
1. Solve the SDP finding X1, . . . ,XL with arbitrary rank.
2. while∑Ll=1 rank2 (Xl) > M , then
3. decompose Xl = VlVHl and find a nonzero solution {∆l} to
L∑l=1
VHl AmlVl •∆l = 0 m = 1, . . . ,M.
4. choose maximum eigenvalue λmax of ∆1, . . . ,∆L
5. compute Xl = Vl
(I− 1
λmax∆l
)VHl for l = 1, . . . , L.
6. end while
Daniel P. Palomar 28
A Simple Consequence
• Observe that if L = 1, then rank (X) ≤√M (real-valued case by
[Pataki’98]).
• Furthermore, if M ≤ 3, then rank (X) ≤ 1:
minimizeX�0
C •X
subject to Am •X = bm, m = 1, . . . ,M.
• Corollary: The following QCQP has no duality gap:
minimizex
xHCx
subject to xHAmx = bm, m = 1, 2, 3.
Daniel P. Palomar 29
Rank-One Solutions for Beamforming
• Our result allows us to find rank-constrained solutions satisfying∑Ll=1 rank2 (Xl) ≤M .
• But, what about rank-one solutions for beamforming?
• Recall that to write the beamforming problem as an SDP we used
Xl = wlwHl .
• Corollary: Suppose the SDP is solvable and strictly feasible, and
that any optimal primal solution has no zero matrix component.
Then, if M ≤ L + 2, the SDP has always a rank-one optimal
solution.
Daniel P. Palomar 30
• The following conditions guarantee that any optimal solution has no
zero component:
L ≤ M
−Aml � 0, ∀l 6= m, m, l = 1, . . . , L
bm > 0, m = 1, . . . , L.
• These conditions are satisfied by the beamforming problem:
minimizeX1,...,XL
∑Ll=1 I •Xl
subject to Rmm •Xm − ρm∑l 6=mRml •Xl ≥ ρmσ2
m, m = 1, . . . , L∑Ll=1 Sml •Xl ≤ τm, m = L+ 1, . . . ,M
Xl � 0, l = 1, . . . , L.
Daniel P. Palomar 31
Contribution Thus Far
• Original result:
– original seminal 2001 paper by Bengtsson and Ottersten showed
strong duality for the rank-one beamforming problem with M = L
– the proof relied on Perron-Frobenius nonnegative matrix theory
and on the uplink-downlink beamforming duality
– it seemed that strong duality was specific to that scenario.
• Current result:
– applies to a much wide problem formulation of separable SDP
– in the context of beamforming, it allows for external interference
constraints
– proof based purely on optimization concepts (no Perron-Frobenius,
no uplink-downlink duality).
Daniel P. Palomar 32
Separable SDP with Individual Constraints
• Consider the following separable SDP with individual constrains:
minimizeX1,...,XL
∑Ll=1 Cl •Xl
subject to∑Ll=1 Aml •Xl = bm, m = 1, . . . ,M
Bl •Xl = (≥) 0, l ∈ (/∈) E1Dl •Xl = (≥) 0, l ∈ (/∈) E2Xl � 0, l = 1, . . . , L.
• By directly invoking our previous result we get:L∑l=1
rank2 (Xl) ≤M + 2L.
• Can we do better?
Daniel P. Palomar 34
Low-Rank Solutions
• Theorem: Suppose the SDP is solvable and strictly feasible. Then,
it has always an optimal solution such that
L∑l=1
rank (Xl) ≤M.
• Observe that the main difference is the loss of the square in the
ranks. This implies that solutions will have higher ranks.
• Can we do better?
Daniel P. Palomar 35
• We can sharpen the previous result is we assume that some of the
matrices (Bl,Dl) are semidefinite.
• Theorem: Suppose the SDP is solvable and strictly feasible, and
that (Bl,Dl) for l = 1, . . . , L′ are semidefinite. Then, the SDP has
always an optimal solution such that
L′∑l=1
rank2 (Xl) +
L∑l=L′+1
rank (Xl) ≤M.
• Observation: individual constraints defined by semidefinite matrices
do not contribute to the loss of the square in the rank.
Daniel P. Palomar 36
A Simple Consequence
• Corollary: The following QCQP has no duality gap (assuming B
and D semidefinite):
minimizex
xHCx
subject to xHAmx = (≥) bm, m = 1, 2, 3
xHBx = (≥) 0xHDx = (≥) 0.
Daniel P. Palomar 37
Rank-One Solutions for Beamforming
• What about rank-one solutions for beamforming?
• Recall that to write the beamforming problem as an SDP we used
Xl = wlwHl .
• Corollary: Suppose the SDP is solvable and strictly feasible, and
that any optimal primal solution has no zero matrix component.
Then, the SDP has a rank-one solution if one of the following hold:
– M = L
– M ≤ L+ 2 and (Bl,Dl) are semidefinite for l = 1, . . . , L.
Daniel P. Palomar 38
Simulation Setup
• Scenario: single base station with K = 8 antennas serving 3 users
at −5◦, 10◦, and 25◦ with an angular spread of σθ = 2◦.
• SINR constraints set to: SINRl ≥ 1 for l = 1, 2, 3.
• External users from another coexisting wireless system at 30◦ and
50◦ with no angular spread.
• Soft-shaping interference constraints for external users set to: τ1 =
0.001 and τ2 = 0.0001.
Daniel P. Palomar 40
Simulation #1
• With no soft-shaping interference constraints (15.36 dBm):
−90 −60 −30 0 30 60 90
−45
−40
−35
−30
−25
−20
−15
Radiation pattern at the transmitter
azimut in degrees
dB
W
Daniel P. Palomar 41
Simulation #2
• With soft-shaping interference constraints (18.54 dBm):
−90 −60 −30 0 30 60 90
−50
−45
−40
−35
−30
−25
−20
−15
−10
Radiation pattern at the transmitter
azimut in degrees
dB
W
Daniel P. Palomar 42
Simulation #3
• With soft- and null-shaping interference constraints (at −20◦, 30◦
and 50◦, 70◦) (20.88 dBm):
−90 −60 −30 0 30 60 90
−70
−60
−50
−40
−30
−20
−10
Radiation pattern at the transmitter
azimut in degrees
dB
W
Daniel P. Palomar 43
Simulation #4
• With soft- and null-shaping interference constraints (at −20◦, 50◦, and 70◦ with
τ1 = 0.00001, τ2 = 0, τ3 = 0.000001) as well as a derivative null constraint at
70◦ (15.66 dBm):
−90 −60 −30 0 30 60 90 −80
−70
−60
−50
−40
−30
−20
Radiation pattern at the transmitter
azimut in degrees
dB
W
Daniel P. Palomar 44
Summary
• We have considered the optimal downlink beamforming problem
that minimizes the transmission power subject to:
– SINR constraints for users within the system
– soft-shaping interference constraints to protect users from coex-
isting systems
– individual null-shaping interference constraints
– derivative interference constraints.
• The problem belongs to the class of rank-constrained separable SDP.
• We have proposed conditions under which strong duality holds as
well as rank reduction procedures.
Daniel P. Palomar 45