Data Rate Theorem for Stabilization over Time-Varying Feedback Channels

Transcription

1 Data Rate Theorem for Stabilization over Time-Varying Feedback Channels Workshop on Frontiers in Distributed Communication, Sensing and Control Massimo Franceschetti, UCSD (joint work with P. Minero, S. Dey and G. Nair) California Institute for Telecommunications and Information Technology

2 Motivation

3 Abstraction DYNAMICAL SYSTEM OBSERVER/ ESTIMATOR CHANNEL CONTROLLER/ ACTUATOR Quality channel Time

4 Packet loss point of view

5 Rate process: Rate Process Packet loss point of view R k = { w.p. 1 p 0 w.p. p 0 Time There is a critical dropout probability p for estimation (Sinopoli et al 2004) and control (Schenato et al 2007, Gupta et al 2007): p < 1 max i λ i 2.

6 Rate limited feedback point of view

7 Rate limited feedback point of view Rate process: R k = R Rate Process R Time Data rate theorem (Tatikonda-Mitter 2002): R > u m u log 2 λ u

8 Can we merge the two approaches?

9 Can we merge the two approaches? Stochastic time-varying rate limited feedback: rate process {R k } i.i.d. according to a random variable R. Rate Process R 1 R 3 R 2 Special cases: R is deterministic, R is 0 or Time

10 Problem formulation R k x k+1 = Ax k + Bu k + v k y k = Cx k + w k y k OBSERVER/ ESTIMATOR C O D E R D E C O D E R CONTROLLER/ ACTUATOR u k A set of unstable eigenmodes λ 1 1,, λ n 1. Problem: conditions on R to achieve second moment stability: sup k E [ x k 2] <.

11 Problem formulation Distributions of x 0 and system disturbances (v and w): Unbounded support A bounded moment > 2 (e.g. Gaussian distribution)

12 Problem formulation Distributions of x 0 and system disturbances (v and w): Unbounded support A bounded moment > 2 (e.g. Gaussian distribution) Rate process: At time k, coder and decoder have knowledge of {R i } k i=0. The case of R k = 0 can be thought as a packet loss with acknowledgment at the transmitter.

13 Problem formulation Distributions of x 0 and system disturbances (v and w): Unbounded support A bounded moment > 2 (e.g. Gaussian distribution) Rate process: At time k, coder and decoder have knowledge of {R i } k i=0. The case of R k = 0 can be thought as a packet loss with acknowledgment at the transmitter. Coder has access to more information than the decoder (Classical information structure).

14 Related works Decoding Disturbances Time-varying errors rate Tatikonda-Mitter No Bounded No (Tran AC 02) Nair-Evans No Unbounded No (SIAM J.C. 04) Yuksel-Basar Yes Unbounded No (CDC 05) Martins-Dahleh-Elia No Bounded Yes (Tran AC 06) Sahai-Mitter Yes Bounded Yes (Tran IT 06) This work No Unbounded Yes

15 Summary of results Scalar case: Necessary and sufficient condition for second moment stabilizability.

16 Summary of results Scalar case: Necessary and sufficient condition for second moment stabilizability. Vector case: we sandwich the stabilizability region between two polytopes. Necessity: The outer polytope has a special geometric structure. Sufficiency: The inner and outer polytopes coincide in some special cases.

17 Scalar case R k x k+1 = λx k + u k + v k y k = x k + w k y k OBSERVER/ ESTIMATOR C O D E R D E C O D E R CONTROLLER/ ACTUATOR u k Unstable system, so λ 1. Theorem 1. Necessary and sufficient condition for stabilization in the mean square sense is that [ ] λ 2 E < R

18 h Scalar case: E λ 2 i 2 2R At each time step, uncertainty volume λ 2, 1 2 2R If the ratio, on average, is 1, then the information sent cannot compensate the system dynamics. < 1

19 h Scalar case: E λ 2 i 2 2R At each time step, uncertainty volume λ 2, 1 2 2R If the ratio, on average, is 1, then the information sent cannot compensate the system dynamics. < 1 Deterministic rate: we recover the data rate theorem R > log 2 λ.

20 h Scalar case: E λ 2 i 2 2R At each time step, uncertainty volume λ 2, 1 2 2R If the ratio, on average, is 1, then the information sent cannot compensate the system dynamics. < 1 Deterministic rate: we recover the data rate theorem R > log 2 λ. Packet loss: we recover the critical dropout probability [ ] λ 2 E 2 2R = p λ 2 + (1 p) λ r < 1 p < 1 λ 2, as r

21 Scalar case: proofs Necessity: recursion Thus, Using the entropy power inequality we establish the following [ ] λ E[x 2 2 k] E E[x 2 k 1] + const., sup k 2 2R [ ] λ E[x 2 2 k] < E < R Sufficiency: Design an adaptive quantizer, avoid saturation, high resolution through successive refinements.

22 Successive refinements Divide time into cycles of fixed duration τ (of our choice) Observe the system at the beginning of each cycle and send an initial estimate During the remaining of a cycle, refine the initial estimate. Quantize Quantize 1 2 τ 1 τ τ + 1 2τ 1 Refinements Refinements Number of bits per cycle is a random variable, dependent on the rate process

23 Successive refinements Quantizer can be generated recursively from a 2-bit quantizer, as follows 2-bit quantizer bit quantizer ρ ρ 4-bit quantizer ρ 2 ρ ρ ρ 2 ρ is a parameter chosen dependent on the distribution of the disturbances.

24 Successive refinements: example Suppose we need to quantize a positive real value x. At time k = 1, suppose R 1 = 1. x R With one bit of information, the decoder knows that x > 0

25 Successive refinements: example At time k = 2, suppose R 2 = 2: coder and decoder partition the real axis according to a 3-bit quantizer. x ρ ρ R However, coder and decoded only label the partitions in the positive real line (2 bits suffice). After receiving 01, the decoder knows that x [ 1 2, 1). The initial estimate x [0, ) has been refined. The scheme works as if we had known ahead of time that R 1 + R 2 = 3.

26 Sufficiency: analysis Quantize Quantize 1 2 τ 1 τ τ + 1 2τ 1 Refinements Refinements We find a recursion for E[x 2 kτ ] E[x 2 kτ] const Thus, for τ large enough ( E [ λ 2 2 2R ]) τ E[x 2 (k 1)τ ] + const. [ ] λ 2 E 2 2R < 1 const ( E [ ]) λ 2 τ < R

27 Vector case R k x k+1 = Ax k + Bu k + v k y k = Cx k + w k y k OBSERVER/ ESTIMATOR C O D E R D E C O D E R CONTROLLER/ ACTUATOR u k n unstable sub-mode associated to eigenvalues λ 1 1, λ 2 1,, λ n 1. Sub-mode i has multiplicity m i. (A,B) is reachable and (C,A) observable.

28 Vector case: necessity Theorem 2. Necessary condition for stabilizability in the mean square sense is that (log 2 λ 1,...,log 2 λ n ) R n + are inside the region determined by the following set of inequalities n i s i log 2 λ i < s i 2 i=1 [ log 2 E ] 2 P 2 i s R i where s = [s 1,, s n ] T 0, s i {0,...,m i }, i = 1,, n.,

29 Vector case: necessity Example 1. Suppose so m 1 = 2 and m 2 = 1. A = log λ 2 < 1 2 log 2 E λ λ λ 2 ˆ2 2R, Not stabilizable 2 log λ 1 + log λ 2 < 3 2 log 2 E h2 2 3 Ri log 2 λ log λ 1 < log 2 E ˆ2 R log 2 λ 1

30 Special case: deterministic rate We recover the necessity of the data rate theorem 2log λ 1 + log λ 2 < R. Polyhedron reduces to a hyperplane R Not stabilizable log 2 λ log 2 λ 1 R 2

31 Special case: packet erasure We recover the necessity on the critical dropout probability p < 1 max i λ i 2. Polyhedron reduces to a hypercube log 2 1 p Not stabilizable log 2 λ log 2 λ log 2 1 p

32 Vector case: sufficiency From the scalar case result we can achieve the two red points Not stabilizable 0.12 log 2 λ log 2 λ 1

33 Vector case: sufficiency We can then time share between the two red points Not stabilizable 0.12 log 2 λ Stabilizable region log 2 λ 1 Can we do better?

34 Sufficiency: rate allocation Theorem 3. Sufficient condition for stabilizability in the mean square sense is that (log 2 λ 1,...,log 2 λ n ) R n + are inside the convex hull of the regions determined by the following n inequalities [ λ i 2 ] E < 1, i = 1,...,n, 2 2 α i(r) R where the rate allocation vector α(r) := [α 1 (R),...,α n (R)] T satisfies α i (r) [0, 1] r m i α i (r) N n i=1 α i(r) 1 for all possible values r that R can take (excluding 0).

35 Sufficiency: rate allocation Example 2. Suppose the rate process is given by R = { 6 w.p. 1 p 0 w.p. p, and, as in the previous example, A = λ λ λ 2. There are four (dominant) possible allocations α(r) = [α 1 (R),α 2 (R)] T : [α 1 (6), α 2 (6)] T { [6 ] T 6, 0, [ 4 6, 2 ] T [ 2, 6 6, 4 ] T [, 0, 6 ] } T, 6 6

36 Sufficiency: example Not stabilizable ˆ4 6, 2 T 6 log 2 λ T ˆ0, 6 ˆ2 6, 4 T Stabilizable region 0.02 ˆ6 6, 0 T log 2 λ 1 One point on the dominant face of the pentagon is optimal

37 Sufficiency: remarks One point on the dominant face of the pentagon is optimal If R = j r w.p. 1 p 0 w.p. p, scheme asymptotically optimal as r Not stabilizable region 0.12 log 2 λ Stabilizable region log 2 λ 1 Scheme is optimal if pentagon reduces to hypercube or hyperplane

38 Remarks Above theorems provide polytopic inner and outer bounds to stabilizability region Not stabilizable log 2 λ Inner bound Outer bound 0.04 Stabilizable region log 2 λ 1 Improved coding scheme for the case of an erasure channel.

39 Concluding Remarks Summary: Unified approach to information-theoretic and packet loss models. second moment stabilization with stochastic time-varying rates. Scalar case: complete characterization. Vector case: geometric structure, scheme is optimal in some limiting cases. Future directions: Close the gap in the vector case Model decoding errors