Relations of Ideas are Matters of Fact: A Unified Theory of Theoretical Unification Kevin T. Kelly Department of Philosophy Carnegie Mellon University kk3n@andrew.cmu.edu Les relations d'idées sont des questions de fait: Une théorie unifiée d'unification théorique Kevin T. Kelly Department of Philosophy Carnegie Mellon University kk3n@andrew.cmu.edu I. Relations of Ideas and Matters of Fact Table of Opposites Relations of Ideas Matters of fact Analytic Synthetic Certain Uncertain A Priori A Posteriori Philosophy of Math Philosophy of Science Proofs and algorithms Confirmation Truth-finding “Rationality” Computability Probability Progress Relations of Ideas Matters of fact Analytic Synthetic Certain Uncertain A Priori A Posteriori Philosophy of Math Philosophy of Science Proofs and algorithms Confirmation Truth-finding “Rationality” Computability Probability Progress Certain Uncertain A Priori A Posteriori Philosophy of Math Philosophy of Science Proofs and algorithms Confirmation Truth-finding “Rationality” Computability Probability Progress Philosophy of Math Philosophy of Science Certain Uncertain A Priori A Posteriori Proofs and algorithms Confirmation Truth-finding “Rationality” Computability Probability Old Wine in New Bottles? Philosophy of Math Philosophy of Science Certain Uncertain A Priori A Posteriori Proofs and algorithms Confirmation Truth-finding “Rationality” Computability Probability Philosophy of Science 101 Formal Empirical Verifiable Unverifiable The given number is even. The sun will never stop rising. Wrong Twice! Formal Empirical Verifiable Unverifiable The given number is even. The current emerald is green. The given The sun computation will never will never stop rising. halt. Aristotelian Excuse Relations of Ideas Completed analysis Rational Rational ? Reason Matters of fact Unbounded experience Animal ?? Always black? Observation Leibnizian Doubts Relations of Ideas Unbounded experience Unbounded analysis ?? Matters of fact Rational ? Reason ?? Always black? Never white? Observation Bounded Kantian Synthesis Synthetic A Priori 3 3 2 4 bounded synthesis 3+2 = 5 ? Pure intuition A Posteriori Unbounded experience 5 ?? Always black? Never white? Observation Unbounded Kantian Synthesis Synthetic A Priori Unbounded Synthesis? ?? Never cross? Pure intuition A Posteriori Unbounded experience ?? Always black? Never white? Observation Kantian Antinomies Synthetic A Priori Physical continuum Geometrical continuum ?? A Posteriori Infinite? Dense? Beyond all pure intuition ?? Infinite? Never white? Dense? Beyond all possible observation Gödel and Turing Formal Problems Unbounded Input computation Empirical Problems Unbounded reality n ?? Never halts? Computer ?? Always black? Never white? Observation Wrong First Cut Formal Empirical Verifiable Unverifiable The given number is even. The current emerald is green. The given The sun computation will never will never stop rising. halt. Right First Cut Verifiable Unverifiable Formal Empirical The given number is even. The current emerald is green. The given The sun computation will never will never stop rising. halt. Unified Epistemology For both formal and empirical questions: Justification = truth-finding performance Find the truth in the best possible sense, as efficiently as possible. II. Verification Verification as Halting with the Truth Yes! Conclusion Verification as Halting with the Truth Maybe you will change your mind. Conclusion Verification as Halting with the Truth Conclusion Verification as Halting with the Truth Conclusion Verification as Halting with the Truth Conclusion Verifiability Yes! Conclusion true Conclusion false Classical Problem of Induction I won’t present blue until he says yes! Empirical demon Always green? Empirical scientist Classical Problem of Induction Always green? Classical Problem of Induction Always green? Classical Problem of Induction Always green? Classical Problem of Induction You lose if you never halt with “yes”! Always green? Classical Problem of Induction Always green? Classical Problem of Induction Yes! Always green? Classical Problem of Induction Always green? Classical Problem of Induction Always green? Classical Problem of Induction You lose anyway. Always green? “Problem” of Uncomputability Never halts? Formal demon Formal scientist Similar to the Problem of Induction? Maybe I am in an infinite loop. Or maybe I will halt later… ? Never halts? Apparently Not Similar I see your program! Never halts? 666 Apparently Not Similar Now I can cleverly calculate a priori what you will do! Never halts? 666 Apparently Not Similar I’m beaten! 666 I’m beaten! Never halts? shoots? 666 21456 chug chug chug But Maybe Similar After All… Aha! His program is also showing! Aha! His program is also showing! Never halts? shoots? 666 666 23455 21456 But Maybe Similar After All… I won’t halt until he halts with “yes” in my a priori simulation of him looking at me! I won’t halt until he halts with “yes” in my a priori simulation of him looking at me! I won’t halt until he halts with “yes” in my a priori simulation of him looking at me! Never halts? shoots? 666 666 23455 21456 666 23455 21456 But Maybe Similar After All… I won’t halt until he halts with “yes” in my a priori simulation of him looking at me! I won’t halt until he halts with “yes” in my a priori simulation of him looking at me! I won’t halt until he halts with “yes” in my a priori simulation of him looking at me! 666 666 23455 21456 Self-reference by Kleene fixed point theorem Never halts? shoots? 666 23455 21456 But Maybe Similar After All… I didn’t halt… 666 I didn’t halt… 23455 21456 666 I didn’t halt… Never halts? shoots? 666 23455 21456 But Maybe Similar After All… I still didn’t halt… 666 I still didn’t halt… 23455 21456 666 I still didn’t halt… Never halts? shoots? 666 23455 21456 But Maybe Similar After All… 666 I give up! You’ll never halt! 666 23455 21456 I give up! You’ll never halt! Never halts? shoots? 666 23455 21456 But Maybe Similar After All… The End The End 23455 Never halts? shoots? 666 23455 But Maybe Similar After All… Never halts? shoots? 666 But Maybe Similar After All… 666 Sucker! Never halts? shoots? 666 But Maybe Similar After All… The End Never halts? shoots? 666 But Maybe Similar After All… Never halts? shoots? But Maybe Similar After All… Never halts? shoots? Rice-Shapiro Theorem Whatever a computable cognitive scientist could verify about an arbitrary computer’s input-output behavior by formally analyzing its program… 666 Q cog-sci Rice-Shapiro Theorem …could have been verified by a behaviorist computer who empirically studies only the computer’s input-output behavior. 666 Q cog-sci Analogy Universal Formal Problem Universal Empirical Problem Unverifiable Unverifiable Computable demon can fool computable agent Empirical demon can fool ideal agent Answer runs beyond empirical experience Answer runs beyond formal experience Apparent Disanalogy Universal Formal Problem Universal Empirical Problem Input given Input never fully given Response: Digested vs. Given Universal Formal Problem Universal Empirical Problem Input given Input never fully given Input never fully digested Input never fully digested Infinitely slow-melting candy Infinitely long noodle Another Apparent Disanalogy Formal Science Empirical Science Incompleteness Error Focus on Consistency Formal Science Empirical Science Incompleteness Error Add axioms Conjecture a theory Who knows if they are Who knows if it is consistent with low-level consistent with all future combinatorics? empirical data? Assume they are and keep Assume it is and keep checking checking Thesis The problem of induction and the problem of uncomputability are similar. So the problems should be understood similarly. A House Divided •Philosophy of Science •“Problem” of induction demands a “solution”. •“Confirmation” and “Bayesian rationality” •No clear connection with truth. •Philosophy of Math •Uncomputability is a fact, not a “problem”. •Logic and algorithms •Insistence on connection with truth. III. Beyond Verifiability Drop the Halting Requirement •Ought implies can. •If you can’t halt with the truth, find the truth in the best feasible sense. Second amendment! Gun control! Q (W. James, H. Putnam, E.M. Gold, R. Jeroslow, P. Hájek) Focus on Questions Answers partition the set of possible inputs A Finite input B C D Infinite input Convergence Without Halting No possible opinion change after halting With halting 1 0 Halts At most finitely many mind-changes In the limit 1 0 Ever-closer to full confidence Gradual 1 0 Success Domain over which convergence to the truth is required A B C D Otherwise, non-convergence to falsehood is required. Success Over Binary Partitions Verify A A infallibly in the limit gradually Refute A B A Decide A A infallibly in the limit gradually B infallibly in the limit gradually B Success Over Countable Partitions Compute partial f Formal case f f=0 f=1 infallibly in the limit gradually f=2 ... Dom(f) infallibly in the limit gradually Theory selection Empirical case Catch-all T0 T1 T2 ... Retractions •A retraction occurs when an answer is “taken back”. •Retractions are what halting is supposed to prevent. •When you can’t prevent all retractions, minimize them. •Halting condition, when feasible, is special case. 2 retractions 1 0 Retraction Efficiency Retraction-efficiency = success in the limit under the least feasible retraction bound in each answer. 5 6 A B Retraction Efficiency 1 0 0 1 A B A Halting-verifiable = 1-verifiable B Halting-refutable = 1-refutable 0 0 A B Halting-decidable = 0-decidable Kantian Antinomies Revisited Infinitely Finitely divisible divisible Matter Upper Complexity Bound Lim-ref Infinitely Finitely divisible divisible Lim-ver I say infinitely divisible when you let me cut. Cut allowed by demon ... ... Answer produced by scientist Lower Complexity Bound Lim-ref -Lim-ver Infinitely Finitely divisible divisible I let you cut when you say finitely divisible. Lim-ver -Lim-ref Purely Formal Analogue Lim-ref -Lim-ver Infinite domain Finite domain 666 Lim-ver -Lim-ref 23455 21456 Purely Formal Analogue Lim-ref -Lim-ver Infinite domain Finite domain Lim-ver -Lim-ref I halt on another input each time he says finitely divisible in my a priori simulation of him looking at my program. 666 23455 21456 Mixed Example: Penrose Computable Uncomputable ??? . Undergraduate subject Cognitive scientist Mixed Example: Penrose Computable Uncomputable You can verify human computability in the limit… NCNCCCCCCC… Undergraduate subject Cognitive scientist Mixed Example: Penrose Computable Uncomputable …but if humans are computers, then you are a computer, in which case you can’t verify in the limit that humans are computers. CNCNCCNCCN… Undergraduate subject Cognitive scientist Formal Complexity n 23455 21456 ... S3 S2 S1 P3 D3 D2 D1 P2 P1 Arithmetical Hierarchy Formal Complexity n 23455 21456 ... grad ref S3 P3 grad ver D3 S2 P2 lim ref lim ver grad dec D lim dec 2 halt ver S1 P1 halt ref D1 halt dec Arithmetical Hierarchy Formal Complexity n grad ref lim ver 23455 21456 S2 grad dec 3-ver P3 grad ver D3 halt-ver 1-ver S3 S2 1-dec D 2 S1 0-dec lim ref D2 halt ref 2-dec D 3 2-ver P2 ... Putnam’s n-trial predicates S3 D1 P3 3-ref P2 2-ref halt-ref P1 1-ref halt-dec Arithmetical Hierarchy Empirical Complexity grad ref lim ver S3 S2 3-ver Difference hierarchy halt-ver 1-ver S2 1-dec S1 0-dec P2 lim ref D2 halt ref S3 2-dec 2-ver D3 ... grad dec P3 grad ver D3 D2 D1 P3 3-ref P2 2-ref halt-ref P1 1-ref halt-dec Borel hierarchy Joint Complexity grad ref lim ver 23455 21456 S2 halt-ver 1-ver D3 S2 1-dec S1 0-dec P2 lim ref D2 halt ref S3 2-dec 2-ver P3 grad ver ... grad dec 3-ver Effective difference hierarchy S3 D3 D2 D1 P3 3-ref P2 2-ref halt-ref P1 1-ref halt-dec Effective Borel Hierarchy Wrong First Cut in Philosophy Empirical ... ... ... ? ? Joint 23455 21456 Formal Right First Cut in Philosophy ... ... ... Empirical Joint Formal 23455 21456 How Complexity Inverts Skepticism •Justification is best possible truth-finding performance. •To establish best possible performance, one must show that better senses of success are infeasible. •Therefore, skeptical arguments justify induction! Impossible performance ... Best possible performance IV. Bounded Rationality To be empirically ignorant is human. To be logically ignorant is irrational. ? Bayesian Philosophy of Science 101 •“Science works by deducing theoretical predictions and comparing them against the empirical data”. “Scientific Method” Empirical data t t No n Theorem Prover Q m What if Predictions are NonComputable? “Scientific Method” Empirical data t Poof! No n Q Computable Refutation with Halting Can Still be Possible! “Scientific Method” Empirical data ? ? n t Q No Non-computable Predictions (with Oliver Schulte) h Everything but h There exists predictive hypothesis {h} such that: •h is infinitely non-computable; •But a computer can refute hypothesis {h} in the halting sense! Inductive Complexity : Deductive Complexity :: Implicit Definability : Explicit Definability S2 S1 ... ... S3 P3 D3 D2 D1 S3 P2 P1 {h} S2 Typical case Learning complexity of {h} = joint complexity of {h}. h S1 P3 D3 D2 D1 P2 P1 Deductive complexity of {h} = formal complexity of h. Singleton Non-Basis Theorem h There exists h such that: S2 S1 ... ... S3 P3 D3 D2 D1 P2 P1 {h} {h} is P1 S3 S2 S1 P3 D3 D2 D1 P2 P1 h is non-arithmetical h is the mu-defined Skolem function for the Delta-2 implicit definition of arithmetical truth (Hilbert and Bernays). Cor. 1: A Foolish Consistency Q Computable scientist Inconsistency Bayesian Computable demon •Consistent computers can’t verify or refute {h} even gradually! •So computable truth-seekers should be inconsistent. Cor. 2: No Firewall •No computer architecture that insulates theorem proving from future empirical data can gradually verify or refute {h}. “Scientific Method” Empirical data n t t Theorem Prover No Q m Cor. 2: No Firewall •But a computer that has a time lag before noticing inconsistency can refute {h} with halting. “Improved Scientific Method” Empirical data Theorem Prover No n Q m V. Unification Unified Which Explanation is Right? ??? Ockham Says: Choose the Simplest! But Maybe… Gotcha! Puzzle A reliable indicator must be sensitive to what it indicates. simple Puzzle A reliable indicator must be sensitive to what it indicates. complex Puzzle But Ockham’s razor always points at simplicity. simple Puzzle But Ockham’s razor always points at simplicity. complex Puzzle How can a broken compass help you find something unless you already know where it is? complex Bayesian Explanation Ignorance (over simple vs. complex) + Ignorance (over parameter settings) = Knowledge (that simple data can’t be produced by the complex theory). Simple q q q q q q q q Complex = The Old Paradox of Indifference Ignorance (over blue, non-blue) + Ignorance (over ways of being non-blue) = Knowledge (that the truth is blue as opposed to any other color) Blue Bayesian Not-Blue = The Old Paradox of Indifference Ignorance (over blue, non-blue) + Ignorance (over ways of being non-blue) = Knowledge (that the truth is blue as opposed to any other color) Blue Bayesian Not-Blue And how does he explain formal simplicity? Better Idea: Retraction Efficiency Truth Better Idea: Retraction Efficiency Truth Curve Fitting Data = arbitrarily precise intervals around Y at specified values of X. Question: assuming that the truth is polynomial, what is the degree? Curve Fitting Demon shows flat line until convergent method concludes flat. Zero degree curve Curve Fitting Demon shows flat line until convergent method concludes flat. Zero degree curve Curve Fitting Then switches to tilted line until convergent method concludes linear. First degree curve Curve Fitting Then switches to parabola until convergent method concludes quadratic … Second degree curve Simplicity The current simplicity of answer H = the number of distinct answers the demon can force an arbitrary convergent method to produce before producing H. Game-theoretic, topological invariant. Cubic Linear Constant 0 Quadratic 1 2 In Step with the Demon There yet? Maybe. Cubic Linear Constant Quadratic In Step with the Demon There yet? Maybe. Cubic Linear Constant Quadratic In Step with the Demon There yet? Maybe. Cubic Linear Constant Quadratic In Step with the Demon There yet? Maybe. Cubic Linear Constant Quadratic Ahead of the Demon There yet? Maybe. Cubic Linear Constant Quadratic Ahead of the Demon I know you’re coming! Cubic Linear Constant Quadratic Ahead of the Demon Maybe. Cubic Linear Constant Quadratic Ahead of the Demon !!! Hmm, it’s quite nice here… Cubic Linear Constant Quadratic Ahead of the Demon You’re back! Learned your lesson? Cubic Linear Constant Quadratic Ockham Violator’s Path Cubic Linear Constant Quadratic Ockham Path So Ockham is uniquely retraction efficient! Cubic Linear Constant Quadratic Same Argument Works For •Fewer causes •Unified causes •More Conserved quantities •Hidden particles •Fewer free parameters Formal Curve Fitting M is given the index of a computable polynomial function on the reals. M must converge to that function’s polynomial degree. Finite stage of effective decimal expansion M “Computational Experience” M also has to refute each polynomial degree with a special halting state for that degree. Finite stage of effective decimal expansion Definitely not quadratic! Halting state 2 M Formal Curve-Fitting Demon For j = k to n-1: act like a polynomial of degree j until: M refutes degree k-1 and M concludes that I am a polynomial of degree j; set j = j+1; act like a polynomial of degree n. Polynomial degree? 21456 M D(M, k, n) Formal Ockham’s Razor •M should not conclude that D(M, k, n) has degree > k when M’s experience has refuted at most degree k. •If M is Ockham, M retracts at most n - j times on D(M, k, n) on experience refuting degree < k. •Else, M retracts at least n – j + 1 times. Q Formal Causal Inference •Given an index of a computable joint probability distribution whose conditional independence relations are represented by a causal network, compute the equivalence class of causal networks representing the distribution. •N-choose-2 retractions required. •Empirical complexity = # of edges in graph •Ockham’s razor = assume no more causes than necessary! X Causal network? Y W Z W Formal Causal Inference •Given an index of a computable joint probability distribution whose conditional independence relations are represented by a causal network, compute the equivalence class of causal networks representing the distribution. •N-choose-2 retractions required. •Empirical complexity = # of edges in graph •Ockham’s razor = assume no more causes than necessary! X Causal network? Y W Z W Unified Account of Unification •When halting is dropped as a condition for success, a penchant for elegance is truth-conducive (retractionefficient) in both formal and empirical inquiry. Q Discussion 1. Uncomputability is similar to the problem of induction. 2. In both cases, halting with the truth is infeasible. 3. Without the halting requirement, skeptical arguments justify inductive reasoning by establishing efficiency. 4. Truth-finding efficiency explains Ockham’s razor both in empirical and formal reasoning. 5. Insistence on separation of formal and empirical reasoning restricts the power of computable science. Q THE END