csi acquisition strategies for interference mitigation in...

Electronic and Computer Engineering Departmentthe Hong Kong University of Science and Technology

June 18, 2014

CSI Acquisition Strategies for Interference Mitigation in Massive

MIMO Networks

CSI Acquisition Strategies for Interference Mitigation in Massive

MIMO NetworksXiongbin RAO

6/18/2014 Xiongbin RAO

ContentsContents• Motivation of Thesis Topic

• Overview: Outline of Thesis Research

• CSI Feedback Reduction for IA in MIMO Networks

• CSI Acquisition Strategy in Massive MIMO Networks

• Future Work

6/18/2014 Xiongbin RAO 2

Channel State Information (CSI) AcquisitionChannel State Information (CSI) Acquisition

Motivation


CSI Acquisition is Important to Achieve Interference Mitigation

CSI Acquisition is Important to Achieve Interference Mitigation

• Interference is a key performance bottleneck of wirelesscommunication

• By adapting the precoder / decorrelator to the channel stateinformation (CSI), many advanced techniques are proposed tomitigate the interference, e.g.,– Technique of zero forcing (ZF) [Jindal, et.al. 2006]– Technique of weighted MMSE (WMMSE) [Shi, Luo, et.al. 2011]– Technique of Interference alignment (IA) [Cadambe, Jafar, et.al. 2008]


Performance Gain of Interference Mitigation with CSI

Performance Gain of Interference Mitigation with CSI


• These interference mitigation schemes achieve dramaticperformance gains:

Acquiring CSI is CostlyAcquiring CSI is Costly


• However, these interference mitigation techniques requireCSI and acquiring CSI at the receiver CSIR or at thetransmitter CSIT is costly, especially for CSIT acquisition– E.g. how to acquire CSIT in FDD systems

• Phase I: Transmitter Tx sends pilot training symbols to probe thechannel on the forward channel

• Phase II: Receiver Rx estimates and feedbacks the channelinformation to Tx side

Tx Rxpilot training

Tx RxCSI feedback

ImplicationImplication


• Take CSI acquisition cost into consideration in theinterference mitigation designsReduce the CSI acquisition costTradeoff analysis between the interference mitigationperformance and the CSI acquisition cost.

Outline of My Thesis ResearchOutline of My Thesis ResearchOverview


Research OutlineResearch Outline


CSI feedback Reduction for IAin interference network with IA

processing[Rao, Ruan, Lau, TSP 13]

CSI Acquisition Design

Uplink CSI (future work)??[Rao, Lau,??]

from MIMO

to massive MIMO

in cellular network with IA processing

[Rao, Lau, TSP 14]

Downlink CSI (including both CSI training and feedback)[Rao, Lau, TSP 14]

CSI Acquisition Strategies

CSI Feedback Reduction for IABackground


Background of IABackground of IA• Technique of interference alignment (IA)

– First proposed by in [Maddah‐Ali, et. al. 2008]

– Basic idea: Align the interference from different Txs into a lowerdimensional subspace at the Rxs

– Performance advantage: Achieve optimal capacity scaling laww.r.t. SNR [Cadambe, Jafar, et.al. 2008]

6/18/2014 Liangzhong RUAN 11

IA in MIMO interference networkIA in MIMO interference network• Network Topology

• Signal Model

• Problem Formulation


Decorrelator Channel Precoder Transmitted signal

noise

Conventional IA ResultsConventional IA Results• When Problem 1.1 feasible,

• There are two questions associated with Problem 1.1.– Feasibility conditions: when will Problem 1.1 has solution ?

– Transceiver design: Given that Problem 1.1 is feasible, how to find thesolution?

• Limitations of these IA designs: at least full channel directioninformation (CDI) is required.


(answer: feasibility study in [M. Razaviyanyn, Z. Luo, et.al. 2012])

(answer: AILM in [Gomadam, Jafar, et.al. 2011])

DoF (d)

DoF


Question: For a given data stream requirements in the network, are there any approaches to reduce the CSI feedback overhead while still achieving IA?

Background of CSI Feedback ReductionBackground of CSI Feedback Reduction• Note that we focus on CSI filtering only, which is different fromCSI quantization ([Rajesh T. K. et.al. 2010]).


1In CSI filtering, only those parts of CSI that are essential to IA design are selected to be fed back.

1

Literature WorksLiterature Works• CSI feedback reduction for IA (focusing on CSI filtering)


System ModelSystem Model

• CSI Feedback Topology

• CSI Feedback Function

• Example


Tx Rx

CSI Feedback CostCSI Feedback Cost

• How to quantify the CSI feedback cost associated with

• First order metric of CSI feedback dimension to measure the CSIfeedback cost


Problem FormulationProblem Formulation•

• Three key questions associated with Problem 1.2


Q3:FindtransceiversolutionunderpartialCSIT

Q2:IAfeasibleconditionsunderpartialCSIT

Q1:HowtoreduceCSIfeedbackforIA?,i.e.,whatis F_j

IA constraints

CSI knowledge constraint

Proposed Solution to Q1Proposed Solution to Q1

• Toy Example I

– Insights:



• Strategy I: No feedback for a subset of cross links

• Strategy II: feedback of null space for a subset of cross links.

Full CDI Proposed

138 24


• Toy Example II: Strategy III

• Toy Example III: Strategy IV



Details of Example III

Details of Example II


• CSI feedback scheme (characterized by )

– There is a 1‐1 correspondence between the feedback profile andfeedback function

• The associated feedback cost is given by (a function of )



Detailed expression of F according to L


• Challenge: Hidden CSI knowledge constraint in Problem 1.2 (theprecoders are only adaptive to partial CSIT feedback).

• We obtain sufficient conditions on the CSI feedback profile toensure that Problem 1.2 has solutions.




• Explicit sufficient conditions on under divisible cases.



Remark (Backward Compatibility with Conventional Results): If the row space of the concatenated channel matrices of all cross links are fed back, Corollary 1.1 will be reduced to conventional feasibility result (Theorem 2) in [M. Razaviyanyn, Z. Luo, et.al. 2012].


• We propose a novel transceiver algorithm under the proposedCSI feedback scheme


Q3:FindtransceiversolutionunderpartialCSIT

A short summary of the resultsA short summary of the results• Problem of IA under Partial CSIT


Q3:IAtransceiverdesignunderpartialCSIT

Q2:IAfeasibleconditionunderpartialCSIT

A2:WeestablishsufficientconditionstomakeIAfeasibleundertheproposedpartial

CSIfeedbackscheme

A3:WeextendconventionalIAtransceiverdesigns i.e.,AILMalgorithm tothecases

withonlypartialCSIT

Q1:HowtoreduceCSIfeedbackforIA?

A1:WeproposeasetofCSIfeedbackreductionstrategies

CSI Feedback Profile DesignCSI Feedback Profile Design• Problem formulation

The problem of reducing the CSI feedback overhead subject to a given datasteams requirement for the Tx‐Rx pairs , can be formulated as :


1: Problem 1.3 is very challenging due the combinatorial nature of CSI feedback profile designs ( ).

2: We propose a low‐complexity greedy algorithm to derive solution.

Min {CSI feedback Cost}

DoFs achieved

Performance‐cost Tradeoff ResultsPerformance‐cost Tradeoff Results


[K. R.T., et.al. 2010] [Rezaee, et.al., 2013]

d

Simulation ResultsSimulation Results


Baseline1: full CDI

Baseline 2:Truncated feedback

Baseline 3: CSI submatrix feedback

MIMO Cellular NetworksMIMO Cellular Networks(Thesis Chapter 4) Extension to


System ModelSystem Model• MIMO Cellular Network

– G base stations (BSs) and each BS has N antennas.– Each BS has K mobile station (MS) and each MS has M antennas.– d data streams are transmitted from the BS to each MS (Given DoF requirement).

• CSI Feedback Topology

• Challenge


There are both intra‐cell and inter‐cell interference in cellular networks, which makes the dependencies of IA precoderson CSIs, very complicated.

k‐th MS of BS j

Problem FormulationProblem Formulation• Two Stage Precoding at BS

• Problem formulation


IA constraints

CSI knowledge constraint

Proposed SchemeProposed Scheme• Regarding Problem 2.1, we shall answer the following questions:


Q3:HowtofindtransceiversolutionunderpartialCSIT


Q1:HowtoreduceCSIfeedbackforIA?,i.e.,what

is F_jk

A1:CSIfeedbackprofilewithCSIfeedbackcost .

A2:NecessaryConditions

SufficientConditions

A3:A novel transceiver algorithm‐‐‐‐Extended from AILM‐‐‐‐Adapt to partial CSI in MIMO cellular network

CSI Feedback Profile DesignCSI Feedback Profile Design• Problem formulation

– The problem of reducing the CSI feedback cost subject to agiven d data steams requirement for each MS, can beformulated as :

– We replace (2.2) with its sufficient conditions , and propose alow complexity achievable solution that satisfies the sufficientconditions


(Suff. cond)Upper Bound

CSI Feedback Profile DesignCSI Feedback Profile Design

– We replace (2.2) with its necessary conditions and derivean lower bound on the optimal feedback dimension

– We show that as , and hence the proposed solutionis asymptotically optimal.


(Nece. cond)Lower Bound

SummarySummary• We extend the framework to achieve CSI feedback reduction forIA from MIMO interference networks to MIMO cellularnetworks.

• The framework consists of the following components:– A set of CSI feedback reduction strategies for MIMO cellularnetwork with IA processing

– IA feasibility study under partial CSIT– IA transceiver algorithm under partial CSIT

• We post the problem of CSI feedback cost minimization subjectto a given DoFs requirement, and we propose a low‐complexityasymptotical optimal solution.


CSIT Acquisition in FDD Massive MIMOCSIT Acquisition in FDD Massive MIMO

From MIMO to Massive MIMO


Related Literature WorksRelated Literature Works• Group I: Conventional (LS or MMSE) based CSIT training andfeedback ([Biguesh, et.al. 2006]):– Require that the training and feedback overhead scales with the numberof antennas at the BS, O(M).• (a): The BS sends a sequence of training pilots X;• (b): The i‐th MS observes Yi, and estimate Hi via LS.• (c) The MS feedbacks estimated Hi to the BS side.

• Group II: Recent works of compressive sensing (CS) basedschemes, i.e., replace the LS with CS‐sparse recovery ([Bajwa, et.al.2010])– They consider to point‐to‐point scenarios.


BS MSpilot training

CSI feedbackBS MS

BackgroundBackground• Multi‐usermassive MIMO channels are jointly sparse

– Sparsity due to limited local scatters at BS (e.g., measurement report [Sayeed,et.al. 2006])

– Structured sparsity (joint) due to some physical features:• Rich local scatterers at MS• Shared local scatterers between different MS (e.g., measurement report[Hoydis, et.al., 2008])


TargetExploit not only the sparsity but also the joint sparsity in the channels to further enhance the CSI training and feedback efficiency!

Proposed Scheme in Multi‐user Massive MIMOProposed Scheme in Multi‐user Massive MIMO


System ModelSystem Model• Network Topology

– 1 BS and K MSs where BS and MS have M antennas (M is large) N antennas.

• Channel Model

• Assumption of Statistical Sparsity Information


Zero coefficientsNon-zero coefficients

System ModelSystem Model


1) Observation: the pilot training and feedback overheads both scale as the training length O(T).

2) It is very desirable to reduce T by exploiting the underlying joint channel sparsities.


Question: How to jointly recover the channel based on compressed measurements ?

CSIT Recovery

Proposed Joint CSIT RecoveryProposed Joint CSIT Recovery• Proposed Joint‐OMP Algorithm

– Extended from conventional OMP and can exploit the joint sparsitystructures in the user channel matrices to enhance the CSITestimation performance.

– Algorithm flow

– Complexity analysis: Consider for simplicity, overallcomplexity of the J‐OMP is

– Information requirements: Only the channel sparisitystatistics is required.


Common support identification

LS estimation based on estimated support

Individual support identification


Question: What is the performance of the proposed CSIT estimation scheme

E.g., ?

Performance AnalysisPerformance Analysis• Define the following support recovery events:

• We obtain a bound on the normalized mean absolute error ofCSIT as below.


InsightsInsights


Insights: a larger size of common support can benefit the CSIT estimation performance!

Simulation ResultsSimulation Results


SummarySummary• We consider CSIT acquisition {includes both CSI training andfeedback} in multi‐user FDD massive MIMO systems

• We propose an CS‐enabled CSI acquisition scheme to exploit thejoint sparsity in the user channel matrices.– A {CSIT measurement and feedback} scheme in which user measures thechannel locally and feedback the measurements to the BS side for jointCSIT recovery.

– A joint‐OMP algorithm to conduct the CSIT recovery at the BS.

• We obtain simple analytical results showing that the jointchannel sparsity structures can be exploited to enhance theCSIT estimation performance.


CSIR Acquisition in Uplink Massive MIMOCSIR Acquisition in Uplink Massive MIMO

Future works


System ModelSystem Model• Network Topology

– 1 BS and K MSs where BS and MS have M antennas N antennasrespectively. Suppose that both M and K are large.

• Model of Sparse Active Users

• Mission: Uplink CSIR Estimation


A. Sparse user due to random access or bursty transmissions of user packets.

B. Temporary correlations between active user sets due to burstyfeatures of user packets.

A. The MSs transmit a sequence of uplink pilots.

B. The BS estimate the uplink CSI based on the channel observations.

Compressiveuplinkpilots

...

MS

JointRecoveryof

Activeusers

inactiveuser

BS

PublicationsPublications• Journal

1. X. Rao, V.K.N. Lau, "Minimization of CSI Feedback Dimension for Interference Alignment in MIMOInterference Multicast Networks," submitted to IEEE Transactions on Information Theory, 2014.

2. X. Rao, V.K.N. Lau, "Distributed Compressive CSIT Estimation and Feedback for FDD Multi‐userMassive MIMO Systems," IEEE Transactions on Signal Processing, vol. 62, no. 12, pp. 3261‐3271,June 15, 2014.

3. X. Rao, V.K.N. Lau, "Interference Alignment with Partial CSI Feedback in MIMO Cellular Networks,"IEEE Transactions on Signal Processing, vol.62, no.8, pp. 2100‐2110, April 15, 2014.

4. X. Rao, L. Ruan, V.K.N. Lau, "CSI Feedback Reduction for MIMO Interference Alignment," IEEETransactions on Signal Processing, vol. 61, no. 18, pp. 4428‐4437, Sept. 15, 2013.

5. X. Rao, L. Ruan, V.K.N. Lau, "Limited feedback design for interference alignment on MIMOinterference networks with heterogeneous path loss and spatial correlations," IEEE Transactionson Signal Processing, vol. 61, no. 10, pp. 2598‐2607, May 15, 2013.

6. L. Ruan, V.K.N. Lau, X.Rao, "Interference Alignment for Partially Connected MIMO CellularNetworks," IEEE Transactions on Signal Processing, vol.60, no.7, pp. 3692‐3701, July 2012.

• Conference1. X. Rao, V.K.N. Lau, "Partial CSI Feedback Design for Interference Alignment in MIMO Cellular

Networks," in Proc. IEEE ICASSP, May 2014.2. X. Rao, V.K.N. Lau, X. Kong, "CSIT Estimation and Feedback Design for FDD Multi‐user Massive

MIMO Systems," in Proc. IEEE ICASSP, May 2014.3. X. Rao, L. Ruan, V.K.N. Lau, "Limited feedback design for interference alignmenton MIMO

interference networks with asymmetric path loss and spatial correlations," in Proc. IEEE ICUFN,July 2012, pp. 446‐451.



Thank you!(Any questions?)

csi acquisition strategies for interference mitigation in...

Documents