MIT Computer Science & Artificial Intelligence Laboratory 1 Research in Theoretical Computer Science Madhu Sudan CSAIL

MIT Computer Science & Artificial Intelligence Laboratory 2 Overview Part I: Introduction to Theory of Computation. Part II: Perspective on (immediate) relevance. Part III: A current research direction. –Introverted Algorithms –Communication with errors: Meaning of bits

MIT Computer Science & Artificial Intelligence Laboratory 3 Part I: Introduction to Theory of CS

MIT Computer Science & Artificial Intelligence Laboratory 4 Theory of Computing Mathematical study of Computation and its consequences. Computation: Sequence of simple steps, leading to complex change in information. Measures: Efficiency of algorithm/program: –Depends on hardware and implementation. Can ask how it scales? –If I double the hardware capacity (speed/memory) *Will this increase the biggest size of problem I can solve by constant factor? (polynomial solution) *Or by additive constant? (exponential solution)

MIT Computer Science & Artificial Intelligence Laboratory 5 Theory of Computing Mathematical study of Computation and its consequences. Computation: Sequence of simple steps, leading to complex change in information. Issues: –Algorithms: Design efficient sequence of steps that produce a desired effect. What is efficient? –Complexity: When is inefficiency inherent? –Implications: What effect does (in)efficiency have on human (intelligent) interaction? Surprisingly broad in scope and impact.

MIT Computer Science & Artificial Intelligence Laboratory 6 Example: Integer Arithmetic Addition: Multiplication: Factoring: Linear!

MIT Computer Science & Artificial Intelligence Laboratory Example: Integer Arithmetic Addition: Linear! Multiplication: Factoring? x Quadratic! Fastest? Not Linear?

MIT Computer Science & Artificial Intelligence Laboratory 8 Addition: Linear! Multiplication: Quadratic! Fastest? Not-linear Factoring? Write as product of two integers (each less than ) Inverse of above problem. –Not known to be linear/quadratic/cubic. –Believed to require exponential time. Example: Integer Arithmetic

MIT Computer Science & Artificial Intelligence Laboratory 9 Algorithms: Given a task (e.g., multiplication) find fast algorithms. –First algorithm we think of may not be fastest. Complexity: Prove lower bounds on resources required to solve problem. –Is multiplication harder than addition? –Is factoring harder than multiplication? Implications: Cryptography … –Economics: Markets implement efficient computation. –Biology: Nature implements efficient computation. –Networks: Errors implement efficient computation. Fundamental quests of CS Theory

MIT Computer Science & Artificial Intelligence Laboratory 10 Long-range questions Is “P=NP?” –Formally, Is all computation reversible? (e.g., multiplication vs. factoring?) –Philosophically, can every designer (mathematician, physicist, engineer, biologist) be replaced by a computer? -(Most of us don’t expect this). -Can we factor integers efficiently? -(Hopefully, still no). -If not, can we build secure communication based on this? -Led to RSA. Still many challenges today.

MIT Computer Science & Artificial Intelligence Laboratory 11 Modern addenda to long-term quests Is the universe random? –Maybe … if so: *Can build efficient algorithms this way (modern examples due to Karger, Rubinfeld, Indyk, Kelner) *Can synchronize distributed systems (essential, as shown by Lynch et al.) *Can generate and preserve secrets (essential, as shown by Goldwasser and Micali). –Maybe not … if so *Might still look random to us, because P ≠ NP. (Long history … Blum, Micali, Yao) Is the universe quantum? Factoring easy (Shor)

MIT Computer Science & Artificial Intelligence Laboratory 12 Current quests in computation Algorithms for Massive data sets –How can we leverage the computational power of a laptop, to understand data such as the WWWMain issue: Massive data – won’t fit in our storage. –Factors in our favor: *We can perform random sampling *We don’t have to deliver “guaranteed answers” –Many Results [Karger, Vempala, Rubinfeld, Indyk] *Can tell if there’s a “trend change” [Rubinfeld et al.] *Can tell if a signal has high-intensity in some frequency. [Indyk et al.] –Underlying emphasis on Randomness.

MIT Computer Science & Artificial Intelligence Laboratory 13 Part II: Perspective of theory

MIT Computer Science & Artificial Intelligence Laboratory 14 History of theoretical CS 1930s: Turing – invented Turing machine. –Universality: One machine implements all algorithms. –Why? To model thought/reasoning/logic * theorems and proofs –Became foundation of modern computers (von Neumann) 1960s: Non-trivial algorithms: –Peterson – BCH decoder –Cooley-Tukey – FFT –Dijkstra – shortest paths 1970s: NP-completeness, Cryptography, RSA. 1990s: Internet algorithms (Yahoo!, Akamai, Google).

MIT Computer Science & Artificial Intelligence Laboratory 15 Theory vs. Practice Theoretical Perspective –Focus on Long-term time horizon; not very close attention to current nature of: *Hardware *Domain-specific information *Solution feasibility Why should you care (today?) –Lessons learned from past are useful (theories more important than theorems). –Good insight into problems of the future. –Occasionally … solutions useful today!

MIT Computer Science & Artificial Intelligence Laboratory 16 Part III: Recent Research Problems, Solutions

MIT Computer Science & Artificial Intelligence Laboratory 17 Part IIIa: Introverted Algorithms

MIT Computer Science & Artificial Intelligence Laboratory 18 Sublinear time algorithms [R. Rubinfeld, P. Valiant] Typical Algorithmic Tasks. –Given x, compute some f(x) in time |x|. Linear time! Modern challenges: –Data too “massive” to allow time |x| to process it. –Can we do much faster? –Allow “randomness” in algorithms. –Allow some “approximation error”.

MIT Computer Science & Artificial Intelligence Laboratory 19 Motivations Internet Traffic –Suppose we maintain vast amounts of logs of internet traffic through a router. –Was there a major shift in the nature of requests within the last hour (perhaps a denial of service attack). Disease Patterns –Suppose we have data for spread of a disease. –What are causal factors. … Theme: Data Abundant; Processing bottleneck

MIT Computer Science & Artificial Intelligence Laboratory 20 “Introverted Algorithms” New Area : Many Problems, Few Tools [P. Valiant]: Symmetric Approximation Properties of Distributions “Intrinsic properties” “Uniform a—m” = “Uniform n—z” yes ? no Distribution Space Invariant under renaming

MIT Computer Science & Artificial Intelligence Laboratory 21 “Introverted Algorithms” New Area : Many Problems, Few Tools [P. Valiant]: Symmetric Approximation Properties of Distributions yes ? no Distribution Space Reals α β “Intrinsic properties” “Uniform a—m” = “Uniform n—z” Invariant under renaming

MIT Computer Science & Artificial Intelligence Laboratory 22 “Introverted Algorithms” New Area : Many Problems, Few Tools [P. Valiant]: Symmetric Approximation Properties of Distributions yes ? no Distribution Space Reals α β continuous Includes: approximating Entropy, Statistical (L1) Distance, Support Size, Information Divergences, other L c distances, weighted distances … “Intrinsic properties” “Uniform a—m” = “Uniform n—z” Invariant under renaming

MIT Computer Science & Artificial Intelligence Laboratory 23 New Contribution Entropy Approximation: β? Statistical Distance: β? g u a c d g u e n α/β [BDKR ’02] n 2α/3β [RRSS ’07] n [B ’01] n 1/2 [BFRSW ’00] nα/βnα/β n Two Components of a Solution: An Upper Bound (Algorithm) A Lower Bound (Impossibility Proof) d g u e

MIT Computer Science & Artificial Intelligence Laboratory 24 New Contribution Entropy Approximation: β? Statistical Distance: β? nα/βnα/β n Canonical Tester Canonical Testing Theorem: Determining the sample complexity of property testing is now a question of algorithm analysis “If the Canonical Tester does not work, nothing does.” Both an upper and a lower bound —What’s the algorithm?

MIT Computer Science & Artificial Intelligence Laboratory 25 The Canonical Tester (a,b,b,a,a,a,f,e,e,e) yes no threshold: 3 estimate high frequencies constrain low frequencies ∩ {yes,no} “If the Canonical Tester does not work, nothing will” If the k-sample Canonical Tester with threshold O( ) does not correctly distinguish  β+ε, then no tester can distinguish  β-ε in k  /n o(1) samples. log n 22  is ( ,  )-weakly continuous: if |d 1 -d 2 |<  then |  (d 1 )-  (d 2 )|<  abcde… ….4 <.3.3  yes ? no 

MIT Computer Science & Artificial Intelligence Laboratory 26 Part IIIb: Robust Intelligent Communication

MIT Computer Science & Artificial Intelligence Laboratory 27 Intelligence and Interaction [Juba & S.] Typical communication “protocols” non-robust. –Depend on perfect understanding between sender and receiver. Require universal adoption of fixed standards. Is this essential? Why? –To reduce human oversight in critical tasks. –E.g., Cars that exchange information, hospitals exchanging medical records. –Heterogeneity leads to violation of “standards”. Technical issues: –Classical communication suppresses/fears intelligence of communicators. Need new models, methods to exploit intelligence of sender & receiver.

MIT Computer Science & Artificial Intelligence Laboratory 28 Modelling the Problem Alice wishes to send algorithm A to Bob –Both know programming; but do so in different languages. –Can she send him the algorithm? Theorem: Not possible to do this unambiguously. Implications: Perfect understanding impossible in evolving settings (when two communicators evolve).

MIT Computer Science & Artificial Intelligence Laboratory 29 Modelling the Problem Alice wishes to send algorithm A to Bob –Both know programming; but do so in different languages. –Can she send him the algorithm? Theorem [Juba & S.]: Not possible to do this unambiguously. Implications: Perfect understanding impossible in evolving settings (when two communicators evolve) – What should we do?

MIT Computer Science & Artificial Intelligence Laboratory 30 Communication & Goals Communication is not an end in itself, it is a means to some (selfish, verifiable) end. –Bob must be trying to use Alice to some benefit *E.g., to alter the environment (remote control) *To learn something (intellectual curiosity). Test Case: Bob (weak computer) tries to communicate with Alice (strong computer) to use her computational abilities. Theorem [Juba & S.]: Bob can use Alice’s help to solve his problem iff problem is verifiable (without common prior background).

MIT Computer Science & Artificial Intelligence Laboratory 31 Examples Bob uses Alice to determine which programs are viruses. –Undecidable problem. Bob can not verify. –Eventually he will make an error. Bob uses Alice to break cryptosystem. –He knows when he has broken in. Should do so. –In the process of doing so he learns Alice’s language (and realizes he is learning). Bob uses Alice to add integers. –Can verify – so he won’t make mistakes. –But probably won’t learn her language.

MIT Computer Science & Artificial Intelligence Laboratory 32 Implications Architecture for communicating computers: –Each interface should have a dedicated “interpreter” –Interpreter is constantly in mode of checking and adapting. Will future of communication look like this? –Answer in 20 years …

MIT Computer Science & Artificial Intelligence Laboratory 33 Recap … Why is Theory Important? Lessons learned from past are useful (theories more important than theorems). –Message of FoxConn Algorithms Course! Good insight into problems of the future. Occasionally … solutions useful today! –RSA, Akamai (CSAIL has more royalties from theory than all other sources put together)!

MIT Computer Science & Artificial Intelligence Laboratory 34 Thank You!

MIT Computer Science & Artificial Intelligence Laboratory 35 Property Testers yes ? no abcde… …  (a,b,b,a,a,a,f,e,e,e) (  10)  tester yes no  sampling (probability > ⅔

MIT Computer Science & Artificial Intelligence Laboratory 36 The Canonical Tester yes ? no abcde… …  (a,b,b,a,a,a,f,e,e,e) (  10)  tester yes no  sampling (probability > ⅔ threshold: 3 estimate high frequencies abcde… ….4 <.3.3 constrain low frequencies  ∩ {yes,no} yes ? no 