Sum utility optimization in MIMO multi-user multi-cell: Centralized and distributed, perfect and partial CSIT, fast and slow CSIT

(1)

Sum Utility Optimization in MIMO Multi-User Multi-Cell:

Centralized and Distributed, Perfect and Partial CSIT,

Fast and Slow CSIT

Dirk Slock

Mobile Communication Department, EURECOM

Cloud & Fog Workshop, GdR ISIS, 20/11/2015

Dirk Slock 5G MIMO Broadcast/Interference Channels - Cloud & Fog- GdR ISIS, 20/11/2015 1/116

(2)

Topics (1)

Cloud: centralized design Fog: decentralized design Cloud design of MIMO IBC Tx

IA and ZF: feasibility, number of solutions various approaches to max WSR

Deterministic annealing: finding global optimum, counting local optima

partial CSIT

Cloud CSIT acquisition

MaMIMO analysis in M,K → ∞regime, smallN of not much use if keep single stream

(3)

Topics (2)

Fog (distributed) designs

Fog global CSIT acquisition & replicated centralized designs various distributed approaches

replacing global CSIT w local CSIT + Rx info

MISO case: MaMIMO allows for slow Rx info, not for MIMO case

convergence issues of distributed approaches (ADMMs etc), convergence speed (more alternating optimization) stochastic approaches for partial CSIT

pathwise (motivations: MaMIMO, mmWave):

non-Kronecker channel model,N may have effect MIMO notation: Nt : M , Nr :N

(4)

Outline

interference single cell: Broadcast Channel (BC)

utility functions: SINR balancing, (weighted) sum rate (WSR) MIMO BC : DPC vs BF, role of Rx antennas (IA, local optima) interference multi-cell/HetNets: Interference Channel (IC)

Degrees of Freedom (DoF) and Interference Alignment (IA) multi-cell multi-user: Interfering Broadcast Channel (IBC) Weighted Sum Rate (WSR) maximization and UL/DL duality Deterministic Annealing to find global max WSR

Max WSR with Partial CSIT CSIT: perfect, partial, LoS

EWSMSE, Massive MIMO limit, large MIMO asymptotics CSIT acquisition and distributed designs

distributed CSIT acquisition, netDoF

topology, rank reduced, decoupled Tx/Rx design, local CSIT distributed designs

Massive MIMO, mmWave : covariance CSIT, pathwise CSIT

(5)

Outline

(6)

Utility Functions

single user (MIMO) in Gaussian noise: Gaussian signaling optimal (avg. power constr.)

rate stream k : R_k = ln(1 + SINR_k)

SINR balancing: max_BFmin_kSINR_k/γ_k under Tx powerP, fairness

related: min Tx power under SINR_k ≥γ_k GREEN max Weighted Sum Rate (WSR): maxBFP

kukRk, given P weightsuk may reflect state of queues (to minimize queue overflow)

weights also allow to vary orientation of normal to Pareto boundary of rate region and hence to explore whole Pareto boundary if rate region convex

Pareto boundary: cannot increase an Rk without decreasing someR_i.

(7)

MIMO BroadCast

MIMO BC = Multi-User MIMO Downlink N_t transmission antennas.

K users withN_k receiving antennas.

Assume perfect CSI

Possibly multiple streams/user d_k. Power constraintP

Noise variance σ²= 1.

H_k the MIMO channel for userk. F_ky_k =F_kH_kPK

i=1G_is_i+F_kz_k

= FkHkGksk

| {z } useful signal

+

K

X

i=1,i6=k

FkHkGisi

| {z }

inter-user interference +Fkzk

| {z } noise

(8)

System Model (2)

Rx signal: y_k =Hkx+zk =HkPK

i=1Gisi+zk

F_k

|{z}

dk×N_k

y_k

|{z}

Nk×1

= F_k

|{z}

dk×N_k

H_k

|{z}

Nk×N_t K

X

i=1

G_i

|{z}

Nt×d_i

s_i

|{z}

di×1

+ F_k

|{z}

dk×N_k

z_k

|{z}

Nk×1

[Christensen etal:T-WCdec08]: use of linear receivers in MIMO BC is not suboptimal (full CSIT, // SU MIMO): can prefilter G_k with a d_k ×d_k unitary matrix to make

interference plus noise prewhitened channel matrix - precoder cascade of user k orthogonal (columns)

Optimal MIMO BC design requires DPC, which is significantly more complicated than BF.

Multiple receive antennas cannot improve the sum rate prelog.

So what benefit can they bring?

Of course: cancellation of interference from other transmitters (spatially colored noise): not considered here.

(9)

Zero-Forcing (ZF)

ZF-BF F_1:iH_1:iG_1:i =







F₁0· · ·0 0F₂. .. ...

... . .. 0 0· · · 0F_i











 H₁ H₂ ... H_i







[G1 G2· · ·Gi] =







F₁H₁G₁ 0 · · · 0 0 F₂H₂G₂ ...

... . .. 0

0 · · · 0F_iH_iG_i







ZF-DPC (modulo reordering issues) F1:iH1:iG1:i =







F₁0· · ·0 0F2. .. ...

... . .. 0 0· · · 0Fi











 H₁ H2

... H_i







[G₁ G₂· · ·G_i] =







F₁H₁G₁ 0 · · · 0

∗ F2H2G2 ...

... . .. 0

∗ · · · ∗FiHiGi







(10)

Stream Selection Criterion from Sum Rate

At high SNR, both

optimized (MMSE style) filters vs. ZF filters optimized vs. uniform power allocation only leads to SNR¹ terms in rates.

At high SNR, the sum rate is of the form Nt

|{z}

DoF

log(SNR/Nt) +X

i

log det(FiHiGi)

| {z }

constant

+ O( 1

SNR)+O(log log(SNR))

| {z } noncoherent Tx for properly normalized ZF Rx F_i and ZF Tx G_i (BF or DPC).

(11)

Role of Rx antennas?

Different distributions of ZF between Tx and Rx give different ZF channel gains! If Rx ZF’sk streams, hence Tx only has to ZF M−1−k streams! So, number of possible solutions (assumingdk ≡1):

M

Y

k=1

(

Nk−1

X

i=0

(M−1)!

k!(M −1−k)!)

for each user, Rx can ZFk between 0 and Nk −1 streams, to choose among M−1.

Explains non-convexity of MIMO SR at high SNR.

ZF by Rx can alternatively be interpreted as IA by Tx (Rx adapts Rx-channel cascades to lie in reduced dimension subspace).

SESAM (and all existing MIMO stream selection algorithms):

assumes that all ZF is done by Tx only. Hence, Rx can be a MF, matched to channel-BF cascade.

(12)

Outline

(13)

MIMO IFC Introduction

Interference Alignment (IA) was introduced in [Cadambe,Jafar 2008]

The objective of IA is to design the Tx beamforming matrices such that the interference at each non intended receiver lies in a common interference subspace If alignment is complete at the receiver simple Zero Forcing (ZF) can suppress interference and extract the desired signal In [SPAWC2010] we derive a set of interference alignment (IA) feasibility conditions for a K-link frequency-flat MIMO interference channel (IFC) d =PK

k=1dk

MIMO Interference Channel

(14)

Possible Application Scenarios

Multi-cell cellular systems, modeling intercell

interference.

Difference from Network MIMO: no exchange of signals, ”only” of channel impulse responses.

HetNets: Coexistence of macrocells and small cells, especially when small cells are considered part of the cellular solution.

(15)

Outline

(16)

Why IA?

The number of streams (degrees of freedom (dof)) appearing in a feasible IA scenario correspond to prelogs of feasible multi-user rate tuples in the multi-user rate region.

Max Weighted Sum Rate (WSR) becomes IA at high SNR.

Noisy IFC: interfering signals are not decoded but treated as (Gaussian) noise.

Apparently enough for dof.

Lots of recent work more generally on rate prelog regions:

involves time sharing, use of fractional power.

(17)

Various IA Flavors

linearIA [GouJafar:IT1210], also calledsignal spaceIA, only uses the spatial dimensions introduced by multiple antennas.

asymptoticIA [CadambeJafar:IT0808] uses symbol extension (in time and/or frequency), leading to (infinite) symbol extension involving diagonal channel matrices, requiring infinite channel diversity in those dimensions. This leads to infinite latency also. The (sum) DoF of asymptotic MIMO IA are determined by thedecompositionbound [WangSunJafar:isit12].

ergodicIA [NazerGastparJafarVishwanath:IT1012] explains the factor 2 loss in DoF of SISO IA w.r.t. an interference-free Tx scenario by

transmitting the same signal twice at two paired channel uses in which all cross channel links cancel out each other: group channel realizationsH1, H2s.t. offdiag(H2) =−offdiag(H1). Ergodic IA also suffers from uncontrolled latency but provides the factor 2 rate loss at any SNR. The DoF of ergodic MIMO IA are also determined by the decomposition bound [LejosneSlockYuan:icassp14].

realIA [MotahariGharanMaddah-AliKhandani:arxiv09], also called it signal scale IA, exploits discrete signal constellations and is based on the Diophantine equation. Although this approach appears still quite exploratory, some related work based on lattices appears promising.

(18)

IA as a Constrained Compressed SVD

(compressed) SVD:

H =F D⁰G⁰^H =F [D0] h

G G⁰⁰iH

=F D G^H ⇒ F^HHG =D F_k^H: d_k×N_k, H_ki : N_k ×M_i, G_i : M_i ×d_i F^HHG =







F₁^H 0· · · 0 0F₂^H. .. ...

... . .. 0 0· · · 0F_K^H













H11H12· · ·H1K

H21H22· · ·H2K

... . .. ...

HK1HK2· · ·H_KK













G₁0· · · 0 0G2. .. ...

... . .. 0 0· · · 0G_K







=







F₁^HH₁₁G₁ 0 · · · 0 0 F₂^HH22G2 ...

... . .. 0

0 · · · 0F_K^HH_KKG_K







F^H,G can be chosen to be unitary for IA per user vs per stream approaches:

IA: can absorb thed_k×d_k F_k^HH_kkG_k in eitherF_k^H (per stream LMMSE Rx) or G_k or both.

WSR: can absorb unitary factors of SVD of F_k^HH_kkG_k in F_k^H, G_k without loss in rate⇒ F^HHG = diagonal.

(19)

Interference Alignment: Feasibility Conditions (1)

To derive the existence conditions we consider the ZF conditions

F^H_k

|{z}

dk×Nk

H_kl

|{z}

Nk×M_l

G_l

|{z}

Ml×d_l

=0, ∀l 6=k rank(F^H_kH_kkG_k) =d_k , ∀k ∈ {1,2, . . . ,K}

rank requirement⇒ SU MIMO condition: d_k ≤min(M_k,N_k) The total number of variables in G_k is

d_kM_k −d_k² =d_k(M_k−d_k)

Only the subspace of Gk counts, it is determined up to a d_k×d_k mixture matrix.

The total number of variables in F^H_k is d_kN_k −d_k² =d_k(N_k −d_k)

Only the subspace of F^H_k counts, it is determined up to a d_k×d_k mixture matrix.

(20)

Interference Alignment: Feasibility Conditions (2)

A solution for the interference alignment problem can only exist if the total number of variables is greater than or equal to the total number of constraintsi.e.,

PK

k=1d_k(M_k−d_k) +PK

k=1d_k(N_k−d_k)≥PK

i6=j=1di dj

⇒ PK

k=1d_k(M_k +N_k−2d_k)≥(PK

k=1d_k)²−PK k=1d_k²

⇒ PK

k=1dk(Mk +Nk)≥(PK

k=1dk)²+PK k=1d_k² In the symmetric case: d_k =d,M_k =M,N_k =N:

d ≤ ^M+N_K+1

For the K = 3 user case (M =N): d = ^M₂ .

With 3 parallel MIMO links, half of the (interference-free) resources are available!

Howeverd ≤ _(K+1)/2¹ M < ¹₂M for K >3.

(21)

Interference Alignment: Feasibility Conditions (3)

The main idea of IA is to convert the alignment requirements at each RX into a rank condition of an associated interference matrix

H^[k]_I =[Hk1G1, ...Hk(k−1)G(k−1),Hk(k+1)G(k+1), ...HkkGK], that spans the interference subspace at thek-th RX (the shaded blocks in each block row).Thus the dimension of the Interference subspace must satisfy rank(H^[k]_I ) =r_I^[k]≤N_k−d_k

The equation above prescribes an upperbound forr_I^[k] but the nature of the channel matrix (full rank) and the rank requirement of the BF specifies the following lower boundrI^[k]≥max_l6=k(dl−[Ml−Nk]+).

Imposing a rankr_I^[k] onH^[k]_I implies imposing (N_k−r_I^[k])(PK l=1 l6=k

d_l−r_I^[k]) constraints at RXk. Enforcing the minimum number of constraints on the system implies to have maximum rank: r_I^[k]≤min(d_tot,N_k)−d_k

(22)

Interference Alignment: Feasibility Conditions (3)

[BreslerTse:arxiv11]: counting equations and variables not the whole story!

appears in very ”rectangular” (6= square) MIMO systems example: (M,N,d)^K = (4,8,3)³ MIMO IFC system comparing variables and ZF equations:

d = ^M+N_K+1 = ⁴⁺⁸₃₊₁ = ¹²₄ = 3 should be possible

supportable interference subspace dim. =N−d = 8-3 = 5 however, the 2 interfering 8×4 cross channels generate 4-dimensional subspaces which in an 8-dimensional space do not intersect w.p. 1 !

hence, the interfering 4×3 transmit filters cannot massage their 6-dimensional joint interference subspace into a 5-dimensional subspace!

This issue is not captured by # variables vs # equations:

d = ^M+N_K+1 only depends onM+N: (5,7,3)³, (6,6,3)³ work.

(23)

Feasibility Linear IA

We shall focus here on linear IA, in which the spatial Tx filters align their various interference terms at a given user in a common subspace so that a Rx filter can zero force (ZF) it. Since linear IA only uses spatial filtering, it leads to low latency.

The DoF of linear IA are upper bounded by the so-calledproper bound [Negro:eusipco09], [Negro:spawc10], [YetisGouJafarKayran:SP10], which simply counts the number of filter variables vs. the number of ZF constraints.

The proper bound is not always attained though because to make interference subspaces align, the channel subspaces in which they live have to sufficiently overlap to begin with, which is not always the case, as captured by the so-calledquantity bound[Tingting:arxiv0913] and first elucidated in [BreslerCartwrightTse:allerton11],

[BreslerCartwrightTse:itw11], [WangSunJafar:isit12].

The transmitter coordination required for DL IA in a multi-cell setting corresponds to the Interfering Broadcast Channel (IBC). Depending on the number of interfering cells, the BS may run out of antennas to serve more than one user, which then leads to the Interference Channel (IC).

(24)

I and Q components: IA with Real Symbol Streams

Using real signal constellations in place of complex

constellations, transmission over a complex channel of any given dimension can be interpreted as transmission over a real channel of double the original dimensions (by treating the I and Q components as separate channels).

This doubling of dimensions provides additional flexibility in achieving the total DoF available in the network.

Split complex quantities in I and Q components:

Hij =

Re{Hij} −Im{Hij} Im{H_ij} Re{H_ij}

x=

Re{x}

Im{x}

Example: GMSK in GSM: was considered as wasting half of the resources, but in fact unknowingly anticipated interference treatment: 3 interfering GSM links can each support one GMSK signal without interference by proper joint Tx/Rx design! (SAIC: handles 1 interferer, requires only Rx design).

(25)

Outline

(26)

MIMO Interfering Broadcast Channel (IBC)

proper sumDoF bound: C K ^M+N_C+1 , C cells, K users/cell

. .

BS1 .

. .

. ^U^1,1

. .

. ^U^1,K1

. . .

. .

BS𝐶 .

. .

. ^U^𝐶,1

. .

. ^U^C,KC

. . . . .

.

𝐇1,1,1

𝐇1,K1,1

𝐇𝐶,1,1 𝐇𝐶,𝐾𝐶,1

𝐇1,1,𝐶

𝐇1,K1,𝐶 𝐇𝐶,1,C 𝐇𝐶,K𝐶,𝐶

Cell 1

Cell C

(27)

From IA to Optimized IFC’s

from Interference Alignment (=ZF) to max Sum Rate (SR) for the ”Noisy IFC”.

to vary the point reached on the rate region boundary:

SR→ Weighted SR (WSR)

problem: IFC rate region not convex ⇒multiple (local) optima for WSR (multiple boundary points with same tangent direction)

solution of [CuiZhang:ita10]: WSR → max SR under rate profile constraint: ^R_α¹

1 = ^R_α²

2 =· · ·= ^R_α^K

K : K−1 constraints.

Pro: explores systematically rate region boundary.

Con: for a fixed rate profile, bad links drag down good links.

⇒ stick to (W)SR

(monitoring global opt issues).

Note: multiple WSR solutions

⇔ multiple IA solutions.

(28)

Max WSR Tx BF Design Outline

max WSR Tx BF design with perfect CSIT using WSR - WSMSE relation

from difference of concave to linearized concave MIMO BC: local optima, deterministic annealing Gaussian partial CSIT

max EWSR Tx design w partial CSIT Line of Sight (LoS) based partial CSIT max EWSR Tx design with LoS based CSIT

(29)

MIMO IBC with Linear Tx/Rx, single stream

IBC withC cells with a total ofK users. System-wide user numbering:

theNk×1 Rx signal at userkin cellbk is y_k=Hk,b_kgkxk

| {z } signal

+X

i6=k b_i=b_k

Hk,b_kgixi

| {z } intracell interf.

+X

j6=b_k

X

i:b_i=j

Hk,jgixi

| {z }

intercell interf.

+vk

wherexk = intended (white, unit variance) scalar signal stream,Hk,b_k = Nk×Mb_k channel from BSbk to userk. BSbkservesKb_k=P

i:b_i=b_k1 users. Noise whitened signal representation⇒ vk∼ CN(0,IN_k).

TheMb_k×1 spatialTx filterorbeamformer (BF) isgk.

Treating interference as noise, userk will apply a linearRx filterfk to maximize the signal power (diversity) while reducing any residual interference that would not have been (sufficiently) suppressed by the BS Tx. The Rx filter output isbxk=f_k^Hy_k

xbk = fk^HHk,b_kgkxk+

K

X

i=1,6=k

fk^HHk,b_igixi+fk^Hvk

= f_k^Hhk,kxk+X

i6=k

fk^Hhk,ixi+fk^Hvk

wherehk,i =Hk,b_igi is the channel-Tx cascade vector.

(30)

Max Weighted Sum Rate (WSR)

Weighted sum rate (WSR)

WSR =WSR(g) =

K

X

k=1

uk ln 1 e_k whereg={g_k}, theu_k are rate weights MMSEsek =ek(g)

1

e_k = 1 +g^H_kH^H_kR⁻¹

k H_kg_k = (1−g_k^HH^H_kR⁻¹_k H_kg_k)⁻¹ R_k =R_k+H_kg_kg_k^HH^H_k , R_k =P

i6=kH_kg_ig^H_i H^H_k +I_N_k , R_k,R_k = total, interference plus noise Rx cov. matrices resp.

MMSEek obtained at the outputxbk =f_k^Hy_k of the optimal (MMSE) linear Rx

f_k =R⁻¹_k H_kg_k .

(31)

From max WSR to min WSMSE

For a general Rx filter f_k we have the MSE e_k(f_k,g)

= (1−f_k^HH_kg_k)(1−g^H_kH^H_kf_k) +P

i6=kf_k^HH_kg_ig_i^HH^H_kf_k+||f_k||²

= 1−f_k^HH_kg_k−g^H_kH^H_kf_k+P

if_k^HH_kgig_i^HH^H_kf_k+||f_k||². The WSR(g) is a non-convex and complicated function of g.

Inspired by [Christensen:TW1208], we introduced

[Negro:ita10],[Negro:ita11] an augmented cost function, the Weighted Sum MSE,WSMSE(g,f,w)

=

K

X

k=1

uk(wk ek(fk,g)−lnwk) +λ(

K

X

k=1

||g_k||²−P) whereλ= Lagrange multiplier and P = Tx power constraint.

After optimizing over the aggregate auxiliary Rx filters f and weightsw, we get the WSR back:

min

f,w WSMSE(g,f,w) =−WSR(g) +

constant z }| {

K

X

k=1

u_k

(32)

From max WSR to min WSMSE (2)

Advantage augmented cost function: alternating optimization

⇒ solving simple quadratic or convex functions minwk

WSMSE ⇒ w_k = 1/e_k minfk

WSMSE ⇒ f_k= (X

i

H_kgig^H_i H^H_k+I_N_k)⁻¹H_kg_k ming_k WSMSE ⇒

gk= (P

iuiwiH^H_i fif_i^HHi+λIM)⁻¹H^H_kfkukwk

UL/DL duality: optimal Tx filter g_k of the form of a MMSE linear Rx for the dual UL in which λplays the role of Rx noise variance andukwk plays the role of stream variance.

(33)

Optimal Lagrange Multiplier λ

(bisection) line search onP_K

k=1||g_k||²−P = 0 [Luo:SP0911].

Orupdated analytically as in [Negro:ita10],[Negro:ita11] by exploitingP

kg_k^H^∂WSMSE_∂g∗

k = 0.

This leads to the same result as in [Hassibi:TWC0906]: λ avoided byreparameterizing the BF to satisfy the power constraint: g_k =q _P

PK

i=1||g⁰_i||²g⁰_k with g_k⁰ now unconstrained SINRk = |f_kH_kg⁰_k|²

PK

i=1,6=k|f_kHkg⁰_i|²+_P¹||f_k||²PK

i=1||g⁰_i||² . This leads to the same Lagrange multiplier expression obtained in [Christensen:TW1208] on the basis of aheuristic that was introduced in [Joham:isssta02] as was pointed out in [Negro:ita10].

(34)

From WSR to WSSINR

The WSR can be rewritten as WSR =WSR(g) =

K

X

k=1

u_k ln(1 + SINR_k) where 1 + SINR_k = 1/e_k or for general f_k :

SINR_k = |f_kH_kg_k|² PK

i=1,6=k|f_kH_kg_i|²+||f_k||² . WSR variation

∂WSR =

K

X

k=1

u_k

1 + SINR_k ∂SINR_k

interpretation: variation of a weighted sum SINR (WSSINR) The BFs obtained: same as for WSR or WSMSE criteria.

But this interpretation shows: WSR = optimal approach to the SLNR or SJNR heuristics.

WSSINR approach = [KimGiannakis:IT0511] below.

(35)

[KimGiannakis:IT0511]: Difference of Convex Fns

Let Q_k =g_kg^H_k be the transmit covariance for stream k ⇒ WSR =

K

X

k=1

u_k[ln det(R_k)−ln det(R_k)]

w R_k =H_k(P

iQ_i)H^H_k +I_N_k,R_k =H_k(P

i6=kQ_i)H^H_k +I_N_k. Consider the dependence of WSR onQk alone:

WSR =ukln det(R⁻¹

k Rk)+WSR_k, WSR_k =

K

X

i=1,6=k

uiln det(R⁻¹_i Ri) where ln det(R⁻¹

k R_k) is concave in Q_k and WSR_k is convex in Qk. Since a linear function is simultaneously convex and concave, consider the first order Taylor series expansion inQ_k aroundQb (i.e. allQb_i) with e.g. Rb_i =R_i(Q), thenb

WSR_k(Q_k,Q)b ≈WSR_k(Qb_k,Q)b −tr{(Q_k −Qb_k)Ab_k} Ab_k =− ∂WSR_k(Qk,Q)b

∂Qk

Qbk,Qb

=

K

X

i=1,6=k

u_iH^H_i (Rb⁻¹_i −Rb⁻¹_i )H_i

(36)

[KimGiannakis:IT0511] (2)

Note that the linearized (tangent) expression forWSR_k constitutes a lower bound for it.

Now, dropping constant terms, reparameterizing Q_k =g_kg^H_k and performing this linearization for all users,

WSR(g,g) =b

K

X

k=1

u_kln(1+g^H_kH^H_kRb⁻¹_k H_kg_k)−g_k^H(Ab_k+λI)g_k+λP. The gradient of this concave WSR lower bound is actually still the same as that of the original WSR or of the WSMSE criteria! Allows generalized eigenvector interpretation:

H^H_kRb⁻¹_k H_kg_k = 1 +g^H_kH^H_kRb⁻¹_k H_kg_k uk

(bA_k +λI)g_k or henceg⁰_k =V_max(H^H_kRb⁻¹_k H_k,Ab_k+λI)

which is proportional to the ”LMMSE” g_k,

with max eigenvalue σ_k =σ_max(H^H_kRb⁻¹_k H_k,Ab_k +λI).

(37)

[KimGiannakis:IT0511] = Optimally Weighted SLNR

Again, [KimGiannakis:IT0511] BF:

g⁰_k =Vmax(H^H_kRb⁻¹_k H_k,

K

X

i=1,6=k

u_iH^H_i (Rb⁻¹_i −Rb⁻¹_i )H_i+λI) This can be viewed as an optimally weighted version ofSLNR (Signal-to-Leakage-plus-Noise-Ratio) [Sayed:SP0507]

SLNR_k = ||H_kg_k||² P

i6=k||H_ig_k||²+P

i||g_i||²/P vs SINR_k = ||H_kg_k||²

P

i6=k||H_kg_i||²+P

i||g_i||²/P SLNR takes as Tx filter

g⁰_k =V_max(H^H_kH_k,X

i6=k

H^H_i H_i+I)

(38)

[KimGiannakis:IT0511] Interference Aware WF

Let σ_k⁽¹⁾=g_k⁰^HH^H_kRb⁻¹_k Hkg⁰_k and σ_k⁽²⁾=g_k⁰^HAbkg⁰_k. The advantage of this formulation is that it allows straightforward power adaptation: substituting g_k =√

p_kg⁰_k yields

WSR =λP+

K

X

k=1

{u_kln(1 +p_kσ_k⁽¹⁾)−p_k(σ_k⁽²⁾+λ)}

which leads to the following interference leakage aware water filling

p_k = u_k σ⁽²⁾_k +λ

− 1 σ⁽¹⁾_k

!+

.

For a given λ,gneeds to be iterated till convergence.

And λcan be found by duality (line search):

minλ≥0max

g λP+X

k

{u_kln det(R⁻¹

k R_k)−λp_k}= min

λ≥0WSR(λ).

(39)

High/Low SNR Behavior

At high SNR, max WSR BF converges to ZF solutions with uniform power

g_k^H =f_kH_kP_(fH)^⊥ H k

/||f_kH_kP_(fH)^⊥ H k

||

whereP_X^⊥=I−PX andPX=X(X^HX)⁻¹X^H projection matrices

(fH)_k denotes the (up-down) stacking of f_iH_i for users i = 1, . . . ,K,i 6=k.

At low SNR, matched filter for user with largest||H_k||₂ (max singular value)

(40)

Deterministic Annealing

Athigh SNR:max WSR solutions are ZF. When ZF is possible (IA feasible), multiple ZF solutions typically exist.

Homotopy on the MIMO channel SVD:

H_ji=

d

X

k=1

σ_jiku_jikv^H_jik+t X

k=d+1

σ_jiku_jikv^H_jik

The IA (ZF) condition for rank 1 linki−j can be written as σjif^H_j ujiv^H_jig_i = 0

Two configurations are possible: f^H_j uji= 0 or v^H_jig_i = 0

Either the Tx or the Rx suppresses one particular interfering stream These different ZF solutionsare thepossible local optima for max WSR at infinite SNR. By homotopy, this remains the number of max WSR local optima as the SNR decreases from infinity. As the SNR decreases further, a stream for some user may get turned off until only a single stream remains at low SNR. Hence, the number of local optima reduces as streams disappear at finite SNR.

At intermediate SNR, the number of streams may also be larger than the DoF though.

(41)

Deterministic Annealing (2)

Homotopy for finding global optimum: atlow SNR, noise dominates interference ⇒ optimal: one stream per power constraint,matched filter Tx/Rx. Gradually increasing SNR allows lower SNR solution to be in region of attraction of global optimum at next higher SNR.

Phase transitions: add a stream.

As a corollary, in the MISO case, the max WSR optimum is unique, since there is only one way to perform ZF BF.

SNR 0 SNR1 SNR2 SNR3

{d }k {d }=_k1 {d }+1k {d }=_k2 {d }+1_k1 {d }=_k3 {d }+1_k2

(42)

Difference of Convex Functions vs Majorization

Difference of Convex functions: linearize convex part in terms of Tx covariance matricesQ_k to make it concave

afterwards work with BF inQ_k =gkg^H_k

but the linearization inQ_k does not correspond to second-order Taylor series or any precise development ingk

other interpretation: majorization: replace cost function to be maximized by one below it that touches the original one in one point [Stoica:SPmagJan04]

specifically: matrix version of x−1−ln(x)≥0, x>0 :

Itakura-Saito distance in AR modeling (x = ratio of true spectrum and AR model spectrum)

majorized cost function can be optimized with any parameterization

(43)

MIMO Tx/Rx Design w Other Utility Functions

in all cases (e.g. also SINR balancing), Rx filter = LMMSE

Tx filter = LMMSE in dual uplink

influence of precise utility function is in the design of theactual &

dual stream powers and noise variances

(44)

Outline

(45)

Mean and Covariance Gaussian CSIT

Mean information about the channel can come from channel feedback or reciprocity, and prediction, or it may correspond to the non fading (e.g. LoS) part of the channel (note that an unknown phase factor e^j^φ in the overall channel mean does not affect the BF design).

Covariance information may correspond to channel estimation (feedback, prediction) errors and/or to information about spatial correlations. The separable (or Kronecker) correlation model (for the channel itself, as opposed to its estimation error or knowledge) below is acceptable when the number of propagation paths Np becomes large (NpMN) as possibly in indoor propagation.

Given only mean and covariance information, the fitting maximum entropy distribution is Gaussian.

(46)

Mean and Covariance Gaussian CSIT (2)

Hence consider

vec(H)∼ CN(vec H

,C^T_t ⊗Cr) orH=H+C^1/2r HeC^1/2_t whereC^1/2r ,C^1/2_t are Hermitian square-roots of the Rx and Tx side covariance matrices

E (H−H)(H−H)^H = tr{C_t}Cr

E (H−H)^H(H−H) = tr{C_r}Ct

and the elements of He are i.i.d. ∼ CN(0,1). A scale factor needs to be fixed in the product tr{C_r}tr{C_t}for unicity.

In what follows, it will also be of interest to consider the total Tx side correlation matrix

R_t = E H^HH=H^HH+ tr{C_r}C_t . Gaussian CSIT model could be considered an instance of Ricean fading in which the ratio tr{H^HH}/(tr{C_r}tr{C_t}) = Ricean factor.

(47)

Location Aided Partial CSIT LoS Channel Model

Assuming theTx disposes of not much more than the LoS component information, model

H=hrh^H_t (θ) +He⁰

whereθ is the LoS AoD and the Tx side array response is normalized: ||h_t(θ)||²= 1.

Since the orientation of the MT is random, model the Rx side LoS array response h_r as vector of i.i.d. complex Gaussian

h_r i.i.d. ∼ CN(0,_µ+1^µ ) and

He⁰ i.i.d. ∼ CN(0,_µ+1¹ _M¹) , independent of hr, where the matrix He represents the aggregate NLoS components.