1 / 26

Physical Limits of Communication

Physical Limits of Communication. Madhu Sudan Microsoft Research. Joint work with Sanjeev Khanna ( U . Penn . ). TexPoint fonts used in EMF. Read the TexPoint manual before you delete this box.: A A A A A A. Reliable Communication. Sender. Receiver. Encoder. Channel. Decoder.

lily
Download Presentation

Physical Limits of Communication

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Physical Limits of Communication Madhu Sudan Microsoft Research Joint work with SanjeevKhanna(U. Penn.) Physical Communication Limits: FSTTCS TexPoint fonts used in EMF. Read the TexPoint manual before you delete this box.: AAAAAA

  2. Reliable Communication Sender Receiver Encoder Channel Decoder • Communication is Reliable • Can be communication across space (e.g. cellphones) or time (DVD). • Implicit axiom: • If Sender/Receiver are physically separated, then only finite # bits can be communicated in finite amount of time. (is finite.) • This talk: Why? Physical Communication Limits: FSTTCS

  3. Why should computer science care? • Axioms of computation: Computation is local. • Works with finite state • Operates on/based on finite number of (preselected) bits at a time. • Preselection changes locally from step to step. • Rely on communication axiom implicitly: • Why is state finite? • Why finite number of bits at a time? • If communication were different, computation should/would be too! Physical Communication Limits: FSTTCS

  4. Why is finiteness restrictive? • Physical channels are not a priori discrete. • Input to channel/Output of channel = signal. • Signal: real-valued function of a real variable • Two sources of capacity • Range of has possibilities • Domain of has possibilities. • Can these lead to? (Even when channel is sufficiently noisy) • What do physically realizable channels look like? Physical Communication Limits: FSTTCS

  5. Part I: Classical Models Physical Communication Limits: FSTTCS

  6. Continuous-valued functions [Shannon] • Say signals are discrete-time, continuous-valued: • Channel = ? • Error • Output signal • Capacity := finite? Infinite? • Analysis (two cases): • Adversarial error: Easy. • adversary can fix to be multiple of . • Capacity Physical Communication Limits: FSTTCS

  7. Continuous-valued functions (contd.) • Recall • Input: • Error • Output signal • Probabilistic error: ind., • Spirit of Shannon’s analysis: • Capacity of channel without noise = • Entropy of noise = • Capacity of noisy channel = cap of channel w/o noise – entropy of noise = Physical Communication Limits: FSTTCS

  8. Continuous-time [SP: Nyquist et al.] • Signals (input/output): • Methodology quite different: • Well-studied in classical Signal Processing (SP): • Works of [Nyquist, Shannon, Landau-Pollak-Slepian] • Many Variations: • Layperson version • Frequency spectrum of signal suffices to sample signal times. • Correct versions: • More complex (theorems + models). Physical Communication Limits: FSTTCS

  9. Continuous-time (contd.) • Actual versions: • Shannon: • Frequency spectrum finite subset of suffices to sample finitely many times. (V. weak). • Nyquist: • Frequency spectrum signal f reconstructible from • Infinite many samples! Finite version can’t work (with exact reconstruction). Physical Communication Limits: FSTTCS

  10. (My) Problems with SP axioms • Why do we need Fourier transforms? • What are these operations in time domain? • (Fourier analysis should remain analysis technique – not natural operation). • Not clean (like Shannon for discrete-time). • Frequency vs. time: • Only signal bounded in time and frequency spectrum is the zero signal • So we need to relax even bandwidth restrictions (some variations studied). • Impulse-response of low-pass filter is non-causal! • Are variations causal? Physical Communication Limits: FSTTCS

  11. Part II: Our Model: Delays Physical Communication Limits: FSTTCS

  12. Noisy and Tardy Channels • Input: • Noise: (typically small ) • Delay: (typically 1). • Output: where • Noise + Delay: • Probabilistic or Adversarial ? • If one is adversarial, does it know the other? Physical Communication Limits: FSTTCS

  13. Motivations for delay • Channels seem to do some frequency “attenuation”/”smoothing”. • Such attenuation should be expressible in time domain (impulse response). • Impulse response should be causal. • Under simplifying assumptions (response is non-negative) impulse response looks like pdf of delay. • (Making delay probabilistic necessary to introduce some uncertainty. If not, easy to invert distortion.) Physical Communication Limits: FSTTCS

  14. Discrete Modelling of Continuous time To simplify our analysis, will discretize time (and signal value), but will allow encoder/decoder to choose how fine the discretization is. So 1 unit of time = M micro-intervals (each microinterval is of length 1/M). Signal value and constant within microinterval. Will ask: Does as ? Physical Communication Limits: FSTTCS

  15. Notationally: • Let • Encoding • Error ; for -fraction of ’s. • Delay = • Output = ; • Will be interested in: Physical Communication Limits: FSTTCS

  16. Questions: • Will be interested in: • Does • Might depend on whether Noise/Delay are adversarial/probabilistic. • Furthermore, if only one is adversarial, is it adaptive wrt randomness of the other? • Probabilistic Models: • Noise: Bernoulli r.v. w.p. and o.w. • Delay: Geometric r.v. with mean (So unit time delay.) Physical Communication Limits: FSTTCS

  17. Answers: • Adv. Noise + Adv. Delay: Capacity is finite. • Random Noise + Random Delays: Capacity unbounded. • Final theorem (a classification): • s.t.for random delay with adversarial/random noise of rate , provided noise independent of delay. • Otherwise, Physical Communication Limits: FSTTCS

  18. Part III: Some Proofs Physical Communication Limits: FSTTCS

  19. Adversarial Noise and Delay • General view of delays: • Think of “delay” as a queue/buffer. • Incoming bits () held in this buffer and released at time . • Analysis of Adversarial channel: • Adversary can force channel to look discrete: • Bits depart the queue at integral time units (is a multiple of ). • Number of ones departing buffers are always integral mutiples of Physical Communication Limits: FSTTCS

  20. Random Noise and Delay 0-block L= M4/5 L = M4/5 0000000 1111111 1-block 1111111 0000000 • Basic idea: Repeating bits times gives enough “signal” to overwhelm deviation due to delay/noise (especially if buffer is balanced). • Encoder: • Differential Decoding: • Compare fraction of 1’sin middle of block to end. • report 0iff increase. Physical Communication Limits: FSTTCS

  21. Random Noise and Delay (contd.) 0-block L= M4/5 L = M4/5 0000000 1111111 1-block 1111111 0000000 • Differential Decoding: • Compare fraction of 1’sin middle of block to end. • report 0iff increase. • Analysis: Chernoff bounds. • Same works if noise is adv. but ind. of delay. • Analysis: If current block is a 1 block, then fraction of 1’s output by channel in current block likely to go up relative to previous block (except when blah blah blah). • Proof: Law of large numbers (with some care). • Key calculation: variances add M^1/2 jitter, but not enough to overwhelm M^{1-\epsilon} effect of input. Physical Communication Limits: FSTTCS

  22. Other Finite Cases: • Random Delay | Adversarial Noise: • Adv. groups signal into blocks of length • At end of each block, round buffer contents to multiple of • Also zeroes out all bits that arrive & depart within same block. • Analysis: • Output process “distributionally defined” by contents of buffer at end of blocks. • Uses: “Geometric distribution is memoryless.” Physical Communication Limits: FSTTCS

  23. Other Finite Cases – II • Adv. Delay | Random Noise • Divide input into blocks; • Delay enough (of the right) bits to make sure buffer contents at end of blocks are multiples of (before noise). • Analysis: • Prove that output signal is “distributionally determined” by buffer contents at end of blocks. • Involves analysis of “signal via noise” channel. Physical Communication Limits: FSTTCS

  24. Part IV: Conclusions Physical Communication Limits: FSTTCS

  25. Physical Limits of Communication = ? • Most reasonable interpretation of nature: non-adversarial. • In such settings capacity = infinite! • Counterintuitive + Contrary to SP literature. • Did we model physics correctly?: Not sure … • Other possible explanations: • Universe is finite … (was this implicit in Shannon?) • Precise measurements are expensive (but wasn’t this taken care of?) • Some non-linearity? • No natural explanations in time domain! Physical Communication Limits: FSTTCS

  26. Thank You! Physical Communication Limits: FSTTCS

More Related