Andrea Goldsmith Wireless Systems Laboratory Stanford University PIMRC September 3, 2014 Future Wireless Networks Ubiquitous Communication Among People and Devices Next-generation Cellular Wireless Internet Access Sensor Networks Smart Homes/Spaces Automated Highways Smart Grid Body-Area Networks Internet of Things All this and more … Future Cell Phones Everything in one device Burden for wireless this performance is on the backbone network San Francisco BS BS LTE backbone is the Internet Internet Nth-Gen Cellular Phone System Nth-Gen Paris Cellular BS Much better performance and reliability than today - Gbps rates, low latency, 99% coverage indoors and out Careful what you wish for… Source: FCC Growth in mobile data, massive spectrum deficit and stagnant revenues require technical and political breakthroughs for ongoing success of cellular On the Horizon: “The Internet of Things” Number of Connected Objects Expected to Reach 50bn by 2020 Are we at the Shannon limit of the Physical Layer? We are at the Shannon Limit  “The wireless industry has reached the theoretical limit of how fast networks can go” K. Fitcher, Connected Planet  “We’re 99% of the way” to the “barrier known as Shannon’s limit,” D. Warren, GSM Association Sr. Dir. of Tech. Shannon was wrong, there is no limit  “There is no theoretical maximum to the amount of data that can be carried by a radio channel” M. Gass, 802.11 Wireless Networks: The Definitive Guide  “Effectively unlimited” capacity possible via personal cells (pcells). S. Perlman, Artemis. Was Shannon wrong, or now irrelevant?  Of course not  What is the flaw in their logic  Use single-user capacity formula for comparison  Let power/bandwidth grow (asymptotically large)  Aggressive frequency reuse via small cells  Space as the final frontier: (Massive) MIMO In fact, we don’t know the Shannon capacity of most wireless channels  Time-varying channels with no/imperfect CSI  Channels and networks with feedback  Channels with delay/energy/HW constraints.  Channels with interference or relays  Cellular systems  Ad-hoc/peer-to-peer networks  Multicast/common information networks Rethinking “Cells” in Cellular Small Cell Coop MIMO How should cellular systems be designed? Relay DAS Will gains in practice be big or incremental; in capacity or coverage?  Traditional cellular design “interference-limited”       MIMO/multiuser detection can remove interference Cooperating BSs form a MIMO array: what is a cell? Relays change cell shape and boundaries Distributed antennas move BS towards cell boundary Small cells create a cell within a cell Mobile cooperation via relaying, virtual MIMO, analog network coding. Are small cells the solution to increase cellular system capacity? Yes, with reuse one and adaptive techniques (Alouini/Goldsmith 1999) Area Spectral Efficiency A=.25D2p  S/I increases with reuse distance (increases link capacity).  Tradeoff between reuse distance and link spectral efficiency (bps/Hz).  Area Spectral Efficiency: Ae=SRi/(.25D2p) bps/Hz/Km2. The Future Cellular Network: Hierarchical Architecture Today’s architecture MACRO: solving • 3M Macrocells serving 5 billion users initial coverage • Anticipated issue, existing network PICO: solving street, enterprise & home coverage/capacity issue FEMTO: solving enterprise & home Picocell Macrocell coverage/capacity issue 10x Lower COST/Mbps (more with WiFi Offload) 10x CAPACITY Improvement Near 100% COVERAGE Femtocell Future systems require Self-Organization (SON) and WiFi Offload SON Premise and Architecture Mobile Gateway Or Cloud Node Installation Self Healing SoN Server Initial Measurements IP Network Self Configuration Measurement SON Server Self Optimization X2 X2 Small cell BS Macrocell BS X2 X2 SW Agent Why not use SoN for all wireless networks? TV White Space & Cognitive Radio mmWave networks Vehicle networks Software-Defined Wireless Network (SDWN) Architecture Video Freq. Allocation Vehicular Networks Security Power Control Self Healing ICIC App layer M2M QoS Opt. Health CS Threshold SW layer UNIFIED CONTROL PLANE HW Layer WiFi Cellular mmWave Cognitive Radio SDWN Challenges  Algorithmic complexity  Frequency allocation alone is NP hard  Also have MIMO, power control, CST, hierarchical networks: NP-really-hard  Advanced optimization tools needed, including a combination of centralized and distributed control  Hardware Interfaces (especially for WiFi)  Seamless handoff between heterogenous networks Massive MIMO: What is it? Dozens of devices Hundreds of BS antennas  10x more BS antennas than today, serving many users  Increases diversity/capacity/multiuser gains of MIMO  With coherent massive MIMO, all the effects of noise and small-scale fading are removed.  Performance limited by pilot contamination Can small cells reduce the impact of pilot contamination? System Model  Massive MIMO TDD cellular system  K users uniformly distributed in each cell  Uplink training with fixed set of pilot sequences  Copilot Distance D is fixed and independent of cell radius R  Pilot Contamination: Downlink SIR & capacity for k-th user in cell 1: where βs depend on the position of the users in all cells: Results  Users are uniformly distributed inside the cells  Cells are approximated with circles As the cell-size reduces, pilot contamination effect diminishes Noncoherent Massive MIMO  Obtaining CSI very challenging  What can we do without CSI?  Propose low-complexity noncoherent SIMO system  Transmitter: Constellation points are power levels  Receiver: Senses only the received power  Symbol Error Rate: SERi  e  nIi ( di )  Depends on the distance “d” between adjacent constellation points  Small “d” approximation:  Accurate for large constellations d2 I i (d )  f ( pi ) Non-coherent optimal in scaling!  Constellation design  A minimum distance criterion provides a simple design  Achieves asymptotically vanishing error probability n = 500 700 600 500 400 300 200 100 0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0  Scaling Law result  As number of receiver antennas increases, scaling law of the achievable rate without CSI is the same as that with perfect CSI: C º Cnocsi log(n) £ Ccsi log(n)  Same scaling applies to multiple single antenna TXs with 1 RX Minimum Distance Criterion – SER Results Analytical upper bound Too many receive antennas are needed! Constellation Design Optimization Rician Fading K = 0, SNR = 0 dB Constellation Design Optimization in non-asymptotic regime  Design Criterion:  All points have same right and left error exponents  Implies all point have approximately the same SER dR,1& dL,2& dR,2& p1+σ2& p2+σ2& dL,3& dR,3& p3+σ2& dL,4& p4+σ2& Distance between adjacent points increases as power increases  Robustness  Can incorporate uncertainty of statistics into design  All points must have minimum value of left and right error exponent for all possible statistics in uncertainty interval Performance The robust design is able to sustain a good performance in a mismatched channel The robust design is not much worse than the “nominal” design  Extensions  Optimized and robust constellation designs for multiuser uplink systems  Noncoherent phase receivers  Comparison of noncoherent and coherent uplink designs  Noncoherent versus coherent downlink designs  Multiuser systems Minimum number of Receive Antennas: SER= 10-4 Minimum Distance Design criterion: Significantly worse performance than the new designs. For low constellation sizes and low uncertainty interval, robust design demonstrates better performance. Extension to multiuser systems: currently working on constellation optimization Spectrum innovations beyond licensed/unlicensed paradigms Cognitive Radio Paradigms  Underlay  Cognitive radios constrained to cause minimal interference to noncognitive radios  Interweave  Cognitive radios find and exploit spectral holes to avoid interfering with noncognitive radios  Overlay  Cognitive radios overhear and enhance noncognitive radio transmissions Knowledge and Complexity Underlay Systems  Cognitive radios determine the interference their transmission causes to noncognitive nodes  Transmit if interference below a given threshold IP NCR NCR CR CR  The interference constraint may be met  Via wideband signalling to maintain interference below the noise floor (spread spectrum or UWB)  Via multiple antennas and beamforming Underlay MIMO Cognitive Radios (CR)  “Intelligent” radio (SU) coexists with licensed user (PU)  Uses MIMO technology for interference mitigation If the SU transmits in 𝑁𝑢𝑙𝑙 𝑯12 , interference to PU is zero How can the SU obtain the null space to the PU? Our approach x(t) The whole process is only based on energy measurements  Power Control     The SU-Tx intelligently choses the messages x(t) The PU-Rx notifies the PU-Tx to change its TX power (or MCS,…) Variation in the TX power of the PU-Tx is sensed by the SU-Tx The SU-Tx chooses the next message x(t) based only on the increase or decrease of the sensed power from the PU-TX.  Tracking algorithm searches “around” outdated null space, can track simultaneously with sending data. Blind Null Space Learning: Example Rayleigh fading with maximum Doppler frequency Fd. Tracking meets peak interference constraints 𝑃𝑋 : 𝑋-th percentile of decrease in interference 𝑁𝑡 = 2, 𝑁𝑟 = 1 Application to Cooperative Multipoint Out of Group Interference (OGI) mitigation Software-Defined (SD) Radio: Is this the solution to the device challenges? BT Cellular FM/XM A/D GPS DVB-H Apps Processor WLAN Media Processor Wimax A/D A/D DSP A/D  Wideband antennas and A/Ds span BW of desired signals  DSP programmed to process desired signal: no specialized HW Today, this is not cost, size, or power efficient What if we sample below the Nyquist rate? Sub-Nyquist Sampled Channels Analog Channel Message N( f ) H( f ) Encoder x(t ) y (t ) Decode r Message C. Shannon Wideband systems may preclude Nyquist-rate sampling! Sub-Nyquist sampling well explored in signal processing  Landau-rate sampling, compressed sensing, etc.  Performance metric: MSE H. Nyquist We ask: what is the capacity-achieving subNyquist sampler and communication design Filter Bank Sampling t  n(mTs ) y1[n] s1 (t )  (t ) x(t ) h(t ) t  n(mTs ) yi [n] si (t ) t  n(mTs ) sm (t ) ym [n]  Theorem: Capacity of the sampled channel using a bank of m filters with aggregate rate fs Similar to MIMO Water-filling MIMO – Decoupling over singular values Pre-whitening Sub-Nyquist Sampling Optimal  “Sparse” channel model  Capacity not monotonic Effective Bandwidth in fs for 1 branch  Capacity monotonic in fs for enough branches Sub-Nyquist Region SuperNyquist Region Sampling with Modulator+Filter (1 or more)  (t ) q(t) x(t) h(t )  p(t) y[n] s(t )  Theorem:  Bank of Modulator+FilterSingle Branch  Filter Bank t  n(mT ) s q(t) zzzz p(t) zzzz zz  y1[n] s1 (t ) zzzz s(t ) zzzz zz y[n] t  n(mTs ) equals yi [n] si (t ) t  n(mTs )  Theorem sm (t )  Optimal among all time-preserving nonuniform sampling techniques of rate fs ym [n] Summary  Existing and emerging applications will require significant breakthroughs in wireless system design  Small cells and massive MIMO are key enablers, but pose new technical challenges  Innovative techniques for cognitive radio can help alleviate spectrum deficits  An interdisciplinary design approach to hardware and system design is needed to meet future challenges