February 10, 2006

This Week's Finds in Mathematical Physics (Week 226)

John Baez

This month I'm hanging out at CIRM, the "Centre International de Recontres Mathematiques" near Marseille. It's like a little hotel with a classroom, library and computers, set at the edge of a forest that borders the region of limestone hills and cliffs called the Calanques on the coast of the Mediterranean. All they do here is hold conferences: guests come in, listen to math talks, and eat nice French meals prepared by the overworked staff.

This February they're having a conference on logic and computation, with a strong slant towards the use of diagrams:

1) Geometry of Computation 2006 (Geocal06), http://iml.univ-mrs.fr/geocal06/

The first week they had lots of talks on "higher-dimensional rewriting", where you use diagrams to draw ways of rewriting ways of rewriting ways of... rewriting strings of symbols. This is inherently connected to n-categories, since n-categories show up whenever you consider "processes that take a process and turn it into another process" - and iterate this notion until your eyes start bugging out.

The first week was pretty cool, and I hope to talk about it someday. The second week I've been recovering from the first week. But today I'll start with some astronomy pictures and some gossip I heard about cryptography, random sequences, and logic.

Let's zoom in on NGC 1097. Here's a photograph of this galaxy taken in ultraviolet light. It's a barred spiral galaxy that's colliding with a smaller elliptical galaxy called NGC 1097A:

2) NGC 1097, photograph taken by the Galaxy Evolution Explorer, NASA/JPL-Caltech/SSC, http://www.galex.caltech.edu/MEDIA/2005-02/images.html

NGC 1097 is called a "Seyfert galaxy" because its center emits lots of radio waves, but not enough to be considered a full-fledged "quasar". As you can see, the center also emits visible light - and stuff is swirling into it!

3) Spiral galaxy NGC 1097, European Southern Observatory, http://www.eso.org/outreach/press-rel/pr-2004/pr-28-04.html#phot-35d-04

Recently the VLT has gotten a really close look at the center of NGC 1097. VLT stands for "Very Large Telescope". It's actually four 8-meter telescopes and three smaller auxiliary ones, all of which can function as a single unit, making it the biggest telescope in the world. It's run by Europeans as part of the "European Southern Observatory" - but it's based on a mountain called Cerro Paranal in the driest part of the Atacama Desert in northern Chile, which makes for wonderful viewing conditions. They had to carry the enormous mirrors across rugged desert roads to build this observatory:

4) Mirror transport, European Southern Observatory, http://www.eso.org/outreach/press-rel/pr-1997/phot-35-97.html

but the view of the night sky up there makes it all worthwhile:

5) The southern sky above Paranal, European Southern Observatory, http://www.eso.org/outreach/press-rel/pr-2005/images/phot-40b-05-normal.jpg

and now they have everything they need to take advantage of that location:

6) The VLT array on Paranal Mountain, European Southern Observatory, http://www.eso.org/outreach/press-rel/pr-2000/phot-14a-00-normal.jpg

So, what did they see when they looked at NGC 1097?

In the middle there seems to be a black hole, emitting radiation as filaments of gas and dust spiral in. This has caused a ring of new stars to form, which are ionizing the hydrogen in their vicinity:

7) Feeding the monster, European Southern Observatory http://www.eso.org/outreach/press-rel/pr-2005/phot-33-05.html

This ring is about 5,500 light-years across! That sounds big, but the galaxy is 45 million light-years away, so this is a stunningly detailed photo. So, we can examine in great detail the process by which black holes eat galaxies.

Anyway....

Talking to the logicians and computer scientists here, I'm hearing lots of gossip that I don't usually get. For example, I just learned that in 2004 there was a successful collision attack on MD5.

Huh? Well, it sounds very technical, but it boils down to this: it means somebody took a function f that is known not to be one-to-one, and found x and y such that f(x) = f(y).

You wouldn't think this would make news! But such functions, called "cryptographic hash functions", are used throughout the computer security business. The idea is that you can take any file and apply the hash function to compute a string of, say, 128 bits. It's supposed to be hard to find two files that give the same bit string. This lets you use the bit string as a kind of summary or "digest" of the file. It's also supposed to be hard to guess the contents of a file from its digest. This lets you show someone the digest of a file without giving away any secrets.

MD5 is a popular hash function invented by Ron Rivest in 1991. People use it for checking the integrity of files: first you compute the digest of a file, and then, when you send the file to someone, you send the digest via some separate channel. If they're worried that the file has been corrupted or tampered with, they compute its digest and compare it to what you sent them.

People also use MD5 and other hash functions for things like keeping passwords safe, digital fingerprinting and copy protection. To illustrate this I'll give a silly example that's easy to understand.

Suppose that this February, Alice proves that P = NP. She wants to take her time writing a nice paper about it... but she wants to be able to show she was the first to solve this problem, in case anyone else solves it while she's writing her paper.

To do this, she types up a quick note explaining her solution, feeds this file into a cryptographic hash function, and posts the resulting 128-bit string to the newsgroup sci.math in an article entitled "I PROVED P = NP!" A dated copy appears on Google for everyone to see.

Now, if Bob solves the problem in July while Alice is still writing up her solution, Alice can reveal her note. If anyone doubts she wrote her note back in February, they can apply the cryptographic hash function and check that yes, the result matches the bit string she posted on Google!

For this to work, it had better be hard for a nasty version of Alice to take Bob's solution and cook up some note summarizing it whose hash function equals some bit string she already posted.

So, it's very good if the hash function is resistant to a "preimage attack". A preimage attack is where for a given x you have a trick for finding y such that f(y) = f(x).

Nobody has carried out a successful preimage attack on MD5. But, people have carried out a successful "collision attack". This is where you can cook up pairs x, y such that f(x) = f(y). This isn't as useful, since you don't have control over either x or y. But, there do exist fiendish schemes for conning people using collision attacks:

8) Magnus Daum and Stefan Lucks, Attacking hash functions by poisoned messages: "The Story of Alice and Her Boss", http://www.cits.rub.de/MD5Collisions/

On this webpage you can see a letter of recommendation for Alice and a letter granting her a security clearance which both have the same MD5 digest! It also explains how Alice could use these to do evil deeds.

For more on cryptographic hash functions and their woes, try these:

9) Cryptographic hash function, Wikipedia, http://en.wikipedia.org/wiki/Cryptographic_hash

Steve Friedl, An illustrated guide to cryptographic hashes, http://www.unixwiz.net/techtips/iguide-crypto-hashes.html

Now, if you're a mathematician, the whole idea of a cryptographic hash function may seem counterintuitive. Just for a change of pace, take another important cryptographic hash function, called SHA-1. This is a function that takes any string of up to 264 bits and gives a digest that's 160 bits long. So, it's just some function

f: S → T

from a set S of size 2(264 + 1) to a set T of size 2160.

The first set is vastly bigger. So, the function f must be far from one-to-one! So why in the world is anyone surprised, much less dismayed, when people find a way to generate two elements in the first set that map to the same element in the second?

One reason is that while the first set is much bigger than the second, the second is mighty big too!

Suppose you're trying to do a preimage attack. Someone hands you an element of T and asks you to find an element of S that maps to it. The brute-force approach, where you keep choosing elements of S, applying the function, and seeing if you get the desired element, will on average take about 2160 tries. That's infeasible.

Note that the huge size of S is irrelevant here; what matters most is the size of T.

Or, suppose you're trying a brute-force collision attack. You start looking through elements of S, trying to find two that map to the same thing. On average it will take 280 tries - the square root of the size of T. That's a lot less, but still infeasible.

(Why so much less? This is called the "birthday paradox": it's a lot easier to find two people at a big party who share the same birthday than to find someone with the same birthday as the host.)

Of course, a smarter approach is to use your knowledge about the function f to help you find pairs of elements that map to the same thing. This is what Xiaoyun Wang, Yiqun Lisa Yin, and Hongbo Yu actually did in February 2005. Based on trials with watered-down versions of SHA-1, they argued that they could do a collision attack that would take only 269 tries instead of the expected 280.

They didn't actually carry out this attack - with the computer power they had, it would have taken them 5 million years. But their theoretical argument was already enough to make people nervous!

And indeed, in August 2005, an improved version of their strategy reduced the necessary number of tries to 263 - in other words, reducing the time to just 78 thousand years, after only a few months of work.

So, people are getting wary of SHA-1. This is serious, because it's a widely used government standard. You can see the algorithm for this function here:

10) SHA hash functions, Wikipedia, http://en.wikipedia.org/wiki/SHA_hash_functions

People still hope that good hash functions exist. They hope that if a function f is cleverly chosen, f(x) will depend in a "seemingly random" way on x, so that given f(x), it's hard to compute some y with f(y) = f(x). People call this a "one-way function".

Now we're finally getting to the really interesting math. It takes work to make the concept of "one-way function" precise, but it can be done. For example, Chapter 2 of this book starts out by defining "strong" and "weak" one-way functions:

11) Oded Goldreich, Foundations of Cryptography, Cambridge U. Press, Cambridge, 2004. Older edition available at http://www.wisdom.weizmann.ac.il/~oded/frag.html

Since you can look them up and they're a bit gnarly, I won't give these definitions here. But roughly, a "strong one-way function" is a function f that satisfies two conditions:

  1. You can compute f in "polynomial time": in other words, there's an algorithm that computes it in a number of steps bounded by some polynomial in the length of the input.
  2. Given some input x, any polynomial-time probabilistic algorithm has a very low chance of finding y with f(y) = f(x). Here a "probabilistic algorithm" is just an algorithm equipped with access to a random number generator.

Condition 2 is related to the idea of a "preimage attack". We allow the attacker to use a probabilistic algorithm because it seems this can help them, and a strong one-way function should be able to resist even really nasty attacks.

Unfortunately, nobody has proved that a one-way function exists!

In fact, the existence of a one-way function would imply that "P does not equal NP". But, proving or disproving this claim is one of the most profound unsolved math problems around. If you settle it, you'll get a million dollars from the Clay Mathematics Institute:

11) Clay Mathematics Institute, P vs NP problem, http://www.claymath.org/millennium/P_vs_NP/

But if you prove that P does equal NP, you might make more money by breaking cryptographic hash codes and setting yourself up as the Napoleon of crime.

The existence of one-way-functions is also closely related to the existence of "pseudorandom" sequences. These are sequences of numbers that "seem" random, but can be computed using deterministic algorithms - so they're not really random.

To see the relation, recall that a good cryptographic hash code shouldn't let you guess a message from its digest. So, for example, if my message is the EMPTY STRING - no message at all! - its digest should be a pseudorandom sequence.

For example, if we apply SHA-1 to the empty string we get the following 160 bits, written as 40 hexadecimal digits:

SHA1("") = "da39a3ee5e6b4b0d3255bfef95601890afd80709"

Seems mighty random to me! But of course it comes out the same every time.

Indeed, if you've ever seen a "random number generator" while messing with computers, it probably was a pseudorandom number generator. There are a few people who generate random numbers from physical phenomena like atmospheric noise. In fact, there's a website called random.org where you can get such numbers for free:

12) Random.org: random integer generator, http://random.org/

There's also a website that offers nonrandom numbers:

13) NoEntropy.net: your online source for truly deterministic numbers, http://www.noentropy.net/

But joking aside, it's a really tantalizing and famous problem to figure out what "random" means, and what it means for something to "seem" random even if it's the result of a deterministic process. These are huge and wonderful philosophico-physico-mathematical questions with serious practical implications. Much ink has been spilt regarding these, and I don't have the energy to discuss them carefully, so I'll just say some stuff to pique your curiosity, and then give you some references.

We can define a "random sequence" to be one that no algorithm can generate with a success rate better than chance would dictate. By virtue of this definition, no algorithm can generate truly random sequences. It's easy to prove that most sequences are random - but it's also easy to prove that you can never exhibit any one particular sequence and prove it's random! Chaitin has given a wonderful definition of a particular random sequence of bits called Ω using the fact that no algorithm can decide which Turing machines halt... but this random sequence is uncomputable, so you can't really "exhibit" it:

14) Gregory Chaitin, Paradoxes of randomness, Complexity 7 (2002), 14-21. Also available as http://www.umcs.maine.edu/~chaitin/summer.html

So, true randomness is somewhat elusive. It seems hard to come by except in quantum mechanics. For example, the time at which a radioactive atom decays is believed to be really random. I'll be pissed off if it turns out that God (or his henchman Satan) is fooling us by simulating quantum mechanics with some cheap pseudorandom number generator!

Similarly, we could define a "pseudorandom sequence" to be one that no efficient algorithm can generate with a success rate higher than chance would dictate.

Efficiency is a somewhat vague concept. It's popular to define it by saying an algorithm is "efficient" if it runs in "polynomial time": the time it takes to run is bounded by some polynomial function of the size of the input data. If the polynomial is

p(n) = 1000000000000000 n1000000000000000 + 1000000000000000

most people wouldn't consider the algorithm efficient, but this definition is good for proving theorems, and usually in practice the polynomial turns out to be more reasonable.

A "pseudorandom number generator" needs to be defined carefully if we want to find efficient algorithms that do this job. After all, no efficient algorithm can produce a sequence that no efficient algorithm can guess: it can always guess what it's going to do!

So, the basic idea is that a pseudorandom number generator should be an efficient algorithm that maps short truly random strings to long pseudorandom strings: we give it a short random "seed" and it cranks out a lot of digits that no efficient algorithm can guess with a success rate higher than chance would dictate.

If you want a more precise definition and a bunch of theorems, try these:

15) Oded Goldreich, papers and lecture notes on pseudorandomness, available at http://www.wisdom.weizmann.ac.il/~oded/pp_pseudo.html

Luby, M. Pseudorandomness and Cryptographic Applications. Princeton, NJ: Princeton University Press, 1996.

Unfortunately, nobody has proved that pseudorandom number generators exist! So, this whole subject is a bit like axiomatic quantum field theory, or the legendary Ph.D thesis where the student couldn't produce any examples of the mathematical gadgets he was studying. It's a risky business proving results about things that might not exist. But in the case of pseudorandom number generators, the subject is too important not to take the chance.

One of the most interesting things about pseudorandom number generators is that they let us mimic probabilistic algorithms with deterministic ones. In fact there are some nice theorems about this. Let me sketch one of them for you.

As I already mentioned, a "probabilistic algorithm" is just a deterministic algorithm that's been equipped with access to a (true) random number generator. Just imagine a computer with the ability to reach out and flip a coin when it wants to. A problem is said to be in "BPP" - "bounded-error probabilistic polynomial time" - if you can find polynomial-time probabilistic algorithms that solve this problem with arbitrarily high chance of success.

It's a fascinating question whether randomness actually helps you compute stuff. I guess most computer scientists think it does. But, it's tricky. For example, consider the problem of deciding whether an integer is prime. Nobody knew how to do this in polynomial time... but then in 1977, Solovay and Strassen showed this problem was in BPP. This is one of the results that got people really excited about probabilistic algorithms.

However, in 2002, Maninda Agrawal, Neeraj Kayal and Nitin Saxena showed that deciding whether numbers are prime is in P! In other words, it can be solved in polynomial time by a plain old deterministic algorithm:

16) Anton Stiglic, Primes is in P little FAQ, http://crypto.cs.mcgill.ca/~stiglic/PRIMES_P_FAQ.html

Is BPP = P? Nobody knows! But if good enough pseudorandom number generators could be shown to exist, we would have BPP = P, since then we could use these pseudorandom numbers as a substitute for truly random ones. This is not obvious, but it was proved by Nisan and Wigderson in 1994.

Here's a great review article that discusses their result:

17) Luca Trevisian, Pseudorandomness and combinatorial constructions, available as cs.CC/0601100.

I recommend reading it along with this:

18) Scott Aaronson, Is P versus NP formally independent?, available at http://www.scottaaronson.com/papers/pnp.pdf and http://www.scottaaronson.com/papers/pnp.ps

This is a delightfully funny and mindblowing crash course on logic. It starts with a review of first-order logic and Gödel's theorems, featuring a dialog between a mathematician and the axiom system ZF+not(Con(ZF)). Here ZF is the popular Zermelo-Fraenkel axioms for set theory and not(Con(ZF)) is a statement asserting that ZF is not consistent. Thanks to Gödel's second incompleteness theorem, ZF+not(Con(ZF)) is consistent if ZF is! The mathematician is naturally puzzled by this state of affairs, but in this dialog, the axiom system explains how it works.

Then Aaronson zips on through Loeb's theorem, which is even weirder than Gödel's first incompleteness theorem. Gödel's first incompleteness theorem says that the statement

This statement is unprovable in ZF.

is not provable in ZF, as long as ZF is consistent. Loeb's theorem says that the statement

This statement is provable in ZF.

is provable in ZF.

Then Aaronson gets to the heart of the subject: a history of the P vs. NP question. This leads up to an amazing paper of Razborov and Rudich, which I'll now summarize:

19) Alexander A. Razborov and Steven Rudich, Natural proofs, in Journal of Computer and System Sciences, Vol. 55, No 1, 1997, pages 24-35. Also available at http://www-2.cs.cmu.edu/~rudich/papers/natural.ps and http://genesis.mi.ras.ru/~razborov/int.ps

The basic point of this paper is that if P is not equal to NP, as most mathematicians expect, then this fact is hard to prove! Or, as Aaron more dramatically puts it, this conjecture "all but asserts the titanic difficulty of finding its own proof".

Zounds! But let's be a bit more precise. A Boolean circuit is a gizmo built of "and", "or" and "not" gates, without any loops. We can think of this as computing a Boolean function, meaning a function of the form:

f: {0,1}n → {0,1}

Razborov and Rudich start by studying a common technique for proving lower bounds on the size of a Boolean circuit that computes a given function. The technique goes like this:

  1. Invent some way of measuring the complexity of a Boolean function.
  2. Show that any Boolean circuit of a certain size can compute only functions of complexity less than some amount.
  3. Show that the function f has high complexity.

They call this style of proof a "natural" proof.

The P versus NP question can be formulated as a question about the size of Boolean circuits - but Razborov and Rudich show that, under certain assumptions, there is no "natural" proof that P is not equal to NP. What are these assumptions? They concern the existence of good pseudorandom number generators. However, the existence of these pseudorandom number generators would follow from the fact that P is not equal to NP. So, if "P is not equal to NP" is true, it has no natural proof.

Aaronson says this is the deepest insight into the P versus NP question so far. I would like to understand it better - I explained it very sketchily because I don't really understand it yet. Aaronson recommends us to these papers for more details:

20) Alexander A. Razborov, Lower bounds for propositional proofs and independence results in bounded arithmetic, in Proceedings of ICALP 1996, 1996, pp. 48-62. Also available at http://genesis.mi.ras.ru/~razborov/icalp.ps

R. Raz, P ≠ NP, propositional proof complexity, and resolution lower bounds for the weak pigeonhole principle, in Proceedings of ICM 2002, Vol. III, 2002, pp. 685-693. Also available at http://www.wisdom/weizmann.ac.il/~ranraz/publication/Pchina.ps

S. Buss, Bounded arithmetic and propositional proof complexity, in Logic of Computation, ed. H. Schwictenberg, Springer-Verlag, 1997, pp. 67-122. Also available at http://www.wisdom/weizmann.ac.il/~ranraz/publication/Pchina.ps

(Why don't these guys use the arXiv??)

Also, here are some lecture notes on Boolean circuits that might help:

21) Uri Zwick, Boolean circuit complexity, http://www.math.tau.ac.il/~zwick/scribe-boolean.html

Aaronson wraps up by musing on the possibility that the P versus NP question is independent from strong axiom systems like Zermelo-Fraenkel set theory. It's possible... and it's possible that this is true and unprovable!

So, there is a fascinating relationship between one-way functions, pseudorandom numbers, and incompleteness - but it's a relationship shrouded in mystery... and perhaps inevitably so. Perhaps it will always remain unknown whether this mystery is inevitable... and perhaps this is inevitable too! And so on.

Here's a question for you experts out there: have people studied Gödel-like self-referential sentences of this form?

The shortest proof of this statement has n lines.

I hear that people have studied ones like

The shortest proof that 0 = 1 has n lines.

and they've proved that the shortest proof of this has to keep getting longer with larger n, as long as one is working in a sufficiently powerful and consistent axiom system. This is fairly obvious if you know how Gödel's second incompleteness theorem is proved... but it's possible that some interesting, nonobvious lower bounds have been proved. If so, I'd like to know!


Addenda: Here is some discussion with Wolfgang Brand, Allan Erskine, Scott Aaronson, Aaron Bergman and an entity named tessel. I had written:

Chaitin has given a marvelous definition of a particular random sequence of bits called Ω using the fact that no algorithm can decide which Turing machines halt... but this random sequence is uncomputable, so you can't really "exhibit" it.
However, Wolfgang Brand points out this paper:

22) Cristian Calude, Michael J. Dinneen, and Chi-Kou Shu, Computing a glimpse of randomness, Experimental Mathematics 11 (2002), 361-370. Also available at http://www.cs.auckland.ac.nz/~cristian/Calude361_370.pdf

where the first 64 bits of Ω have been computed. There's no contradiction, as the paper explains - but it's fairly surprising!

Next, here is an email from Allan Erskine. As usual I will take the liberty of numbering references and putting them into my own favorite format.

Allan Erskine wrote:

I enjoyed week 226! Algorithmic complexity was the area I studied in... Your readers might find this an enjoyable read:

24) Leonid Levin, The tale of one-way functions, Problems of Information Transmission (= Problemy Peredachi Informatsii), 39 (2003), 92-103. Also available as cs.CR/0012023.

As for your "shortest proof of this statement has n lines" question, you may have noticed that Chaitin asks a very similar question about the shortest proofs that a LISP program is "elegant" (most short) and proves a strong incompleteness result with an actual 410 + n character LISP program! Crazy...

25) Gregory Chaitin, Elegant LISP programs, in People and Ideas in Theoretical Computer Science, ed. C. Calude, Springer, Singapore, 1999, pp. 32-52. Also available at http://www.cs.auckland.ac.nz/CDMTCS/chaitin/lisp.html

I replied by pointing Allan to the following old article, which is related to the idea of Chaitin's paper:

From: John Baez
Subject: Re: compression, complexity, and the universe
Date: 1997/11/20
Newsgroups: sci.physics.research, comp.compression.research

Aaron Bergman wrote:

The smallest number not expressible in under ten words
Hah! This, by the way, is the key to that puzzle I laid out: prove that there's a constant K such that no bitstring can be proved to have algorithmic entropy greater than K.

Let me take this as an excuse to say a bit more about this. I won't give away the answer to the puzzle; anyone who gets stuck can find the answer in this nice survey:

26) Peter Gacs, Lecture notes on descriptional complexity and randomness, available at http://www.cs.bu.edu/faculty/gacs/

In my more rhapsodic moments, I like to think of K as the "complexity barrier". The world seems to be full of incredibly complicated structures - but the constant K sets a limit on our ability to prove this. Given any string of bits, we can't rule out the possibility that there's some clever way of printing it out using a computer program less than K bits long. The Encyclopedia Brittanica, the human genome, the detailed atom-by-atom recipe for constructing a blue whale, or for that matter the entire solar system - we are unable to prove that a computer program less than K bits long couldn't print these out. So we can't firmly rule out the reductionistic dream that the whole universe evolved mechanistically starting from a small "seed", a bitstring less than K bits long. (Maybe it did!)

So this raises the question, how big is K?

It depends on ones axioms for mathematics.

Recall that the algorithmic entropy of a bitstring is defined as the length of the shortest program that prints it out. For any finite consistent first-order axiom system A exending the usual axioms of arithmetic, let K(A) be the constant such that no bitstring can be proved, using A, to have algorithmic entropy greater than K(A). We can't compute K(A) exactly, but there's a simple upper bound for it. As Gacs explains, for some constant c we have:

K(A) < L(A) + 2 log2 L(A) + c

where L(A) denotes the length of the axiom system A, encoded as bits as efficiently as possible. I believe the constant c is computable, though of course it depends on details like what universal Turing machine you're using as your computer.

What I want to know is, how big in practice is this upper bound on K(A)? I think it's not very big! The main problem is to work out a bound on c.

Next, some email from the man himself: Scott Aaronson!

Hi John,

I just read the latest issue of TWF. I'd been waiting a long time for you to tackle computational complexity - thanks! :) Here are some comments and responses that you might find helpful.

It's not known whether P≠NP implies the existence of one-way functions (OWF's) - indeed, there's evidence that proving such an implication will be hard:

27) Adi Akavia, Oded Goldreich, Shafi Goldwasser and Dana Moshkovitz, On basing one-way functions on NP-hardness, November 22, 2005, available at http://theory.csail.mit.edu/~akavia/AGGM.pdf

What's known is that one-way functions exist iff pseudorandom generators (PRG's) do - that's a long, gnarly result of Hastad et al.:

28) John Hastad, Russell Impagliazzo, Leonid A. Levin and Michael Luby, A pseudorandom generator from any one-way function, available at http://citeseer.ifi.unizh.ch/hastad99pseudorandom.html

So when I wrote that Razborov-Rudich illustrates the "self-referentiality" of proving P≠NP, really I meant that it illustrates the self-referentiality of proving that OWF's and PRG's exist! But most of us believe not only that P≠NP, but also that OWF's exist, and that all of these questions are "morally similar" anyway... :)

You write: "It's a fascinating question whether randomness actually helps you compute stuff. I guess most computer scientists think it does."

Well, today most of us think that good pseudorandom generators exist, which (as you pointed out) implies that P=BPP. That's not to say that randomness never helps; it just doesn't provide more than a polynomial advantage.

Incidentally, it followed already from Andy Yao's work in 1982 that if REALLY good pseudorandom generators exist, then P=BPP. What Nisan and Wigderson showed in 1994 was that even if only kinda sorta good PRG's exist, then still P=BPP!

Since you wrote at length about hash functions like MD5, you might be interested in the following abstraction of the collision-finding problem. Suppose we have a two-to-one function

f: {0,1}n → {0,1}n-1,

which we can only access in a "black-box" fashion - i.e., we can feed it an input and get an output. The goal is to find distinct x and y such that f(x)=f(y). Classically, it's easy to see that ~2n/2 accesses to f are necessary and sufficient - that's just the birthday paradox you described. But what if we can access f in quantum superposition, to create states like

x |x>|f(x)>?

In that case, it turns out that one can combine the birthday paradox with Grover's quantum search algorithm, to find a collision using ~2n/3 accesses to f instead of ~2n/2. But could we do it with nO(1) accesses, or even a constant number independent of n? You might be surprised that this was open for years!

Probably my best-known result was to show that ~2n/5 accesses are needed:

29) Scott Aaronson, Quantum lower bound for the collision problem, available as quant-ph/0111102.

Subsequently Yaoyun Shi improved that to show that ~2n/3 is indeed the right answer:

30) Yauyon Shi, Quantum lower bounds for the collision and the element distinctness problems, available as quant-ph/0112086.

You ask why theoretical computer scientists don't use the arXiv. The answer is historical and cultural. Quantum computer scientists like me do post to quant-ph, as a result of hanging around physicists for so long. But in classical complexity, if you want your paper to be read, then you (1) put in on your homepage, (2) submit it to STOC or FOCS, which are the top conferences in the field (and are MUCH more important than journals), and (3) post it here:

31) Electronic Colloquium on Computational Complexity, http://www.eccc.uni-trier.de/eccc/

which is our own homegrown arXiv.

I like your finite Gödel statement! Let me restate it as follows:

G(n) = "This statement has no ZF proof of at most n symbols."

What can we say? Obviously, if ZF is consistent, then G(n) has no proof of at most n symbols. Hence G(n) is true. Furthermore, G(n) has a proof of ~2n symbols. This proof simply enumerates all proofs of ≤n symbols, and shows that none of them work.

You might wonder: does G(n) have a proof of n+1 symbols, or for that matter nO(1) symbols? This turns out to be related to the main open questions of complexity theory! In particular, if NP=coNP, then G(n) has a proof of nO(1) symbols. NP=coNP just means that, whenever a Boolean formula of size n is unsatisfiable, there's a proof of that fact of length polynomial in n. Clearly if P=NP then NP=coNP, but the converse isn't known. The prevailing belief is not only that P≠NP, but also that NP≠coNP.

Actually I can say more than that: if NE=coNE, then G(n) has a proof of nO(1) symbols. Here NE is Nondeterministic Exponential-Time, or the exponential-time analogue of NP. (See my Complexity Zoo:

32) Scott Aaronson and Greg Kuperberg, Complexity Zoo, http://qwiki.caltech.edu/wiki/Complexity_Zoo

for lots more about this and other classes.) coNE is just the complement of NE. If NP=coNP then NE=coNE, but again the converse isn't known.

What do these weird exponential classes have to do with G(n)? The point is that to describe an integer n takes only log(n) symbols, not n symbols. So when we ask for a proof of G(n) with at most n symbols, from "log(n)'s standpoint" we're actually asking for an exponentially long proof.

At this point I'll conjecture that I have an "if and only if" characterization: if G(n) has a proof of nO(1) symbols, then NE=coNE.

Incidentally, your "finite Gödel statement" is closely related to the Hartmanis-Stearns time hierarchy theorem, which launched computational complexity in the mid-1960's. The time hierarchy theorem is a finite version of Turing's undecidability theorem, where instead of asking whether a given Turing machine ever halts, you ask whether it halts after at most T steps. Solving this problem turns out to take slightly more than T steps. So in particular, you can solve more problems in n2 time than in n time, more in n3 time than in n2 time, and so on.

Let me know if you have other questions.

Best,
Scott

Next, here is my reply to a post on sci.physics.research from a mysterious entity named tessel... again, I'll edit it a bit:

tessel@um.bot wrote:
On Sat, 11 Feb 2006, John Baez mentioned that md5sum was "broken" about a year ago. I just wanted to add:

1. If I am not mistaken, sha-1 and md5sum are different algorithms (IIRC, both are known to be insecure).

Yeah, I said SHA-1 and MD5 are different, and I said they were both vulnerable to collision attacks. MD5 is very vulnerable in practice, while the vulnerability of SHA-1 is still theoretical: you'd have to have big computers or lots of time or another clever idea to exploit it. (Guess who's likely to have all three!)

For more on this subject, try this:

33) Arjen K. Lenstra, Further progress in hashing cryptanalysis, February 26, 2005, available at http://cm.bell-labs.com/who/akl/hash.pdf

2. The latest versions of the open source utility gpg supports a more secure algorithm, SHA-512, which AFAIK has not been broken; see

34) Tony Stieber, GnuPG Hacks, Linux Journal, March 2006.

3. Even insecure checksum utilities are probably better than none at all. Indeed, checking the given example:

gpg --print-md md5 letter_of_rec.ps order.ps
A2 5F 7F 0B 29 EE 0B 39  68 C8 60 73 85 33 A4 B9
A2 5F 7F 0B 29 EE 0B 39  68 C8 60 73 85 33 A4 B9
Oh NOOOOOO!!! But wait, there's more:
gpg --print-md sha1 letter_of_rec.ps order.ps
0783 5FDD 04C9 AFD2 8304  6BD3 0A36 2A65 16B7 E216
3548 DB4D 0AF8 FD2F 1DBE  0228 8575 E8F9 F539 BFA6

gpg --print-md RIPEMD160 letter_of_rec.ps order.ps
9069 8ACC 6D67 6608 657B  9C26 F047 59A1 DC0E 6CA1
C1BB DE12 B312 EAAD DD3D  D3B8 4CA1 CB1B BA47 DD13
Ah HAAAAAA!!! Gotcha, Alice!

JB wrote:

These are huge and wonderful philosophico-physico-mathematical questions with serious practical implications.
You mean the Weyl curvature hypothesis? :-/
Heh, no - I mean stuff like whether there's such a thing as a provably good cryptographic hash code function, or cipher.
But while we should never neglect incompleteness entirely, I was fascinated to discover from my readings a few years back that even first order logic has its fascinations!

35) Joel Spencer, The Strange Logic of Random Graphs, Springer, Berlin, 2001.

Here's a thought: "Everyone knows" that if on day D, mathematician M is studying an example of size S in class C, he is more likely to be studying a "secretly special" representative R than a generic representative G of size S. Why? Because the secretly special reps show up in disguise in other areas, and M was probably hacking through the jungle from one of those places when he got lost and ate a poisoned cache.

Interesting.
Finally, here's my reply to an article on sci.math.research by Aaron Bergman:
Aaron Bergman wrote:
John Baez wrote:
In fact, the existence of a one-way function would imply that "P does not equal NP". But, proving or disproving this claim is one of the most profound unsolved math problems around.
This brings to mind a question I was wondering about. Given an NP (-complete?) problem, is it ever possible to engineer a (partial) differential equation, the solution of which, if known, would solve the NP problem?

I realize this is vague. The general thought was whether a continuous dynamical system can somehow be more computationally powerful than something discrete.

I was very interested in this question back in college, and I wrote a paper about it:

36) John Baez, Recursivity in quantum mechanics, Trans. Amer. Math. Soc. 280 (1983), 339-350.

where I showed that the time evolution for bunch of PDE's - including the wave equation, the Klein-Gordon equation, and Schroedinger's equation for finitely many electrically charged point particles - is "computable" in a precise sense.

This seemed interesting at the time because Pour-El and Richards had claimed that the behavior of the wave equation in 3+1 dimensions was uncomputable. Their result is correct, but it's an artifact of using a funny space of solutions! If you use a sensible Hilbert space of solutions for which time evolution is unitary, then time evolution is also computable.

Hmm, I see that someone has rediscovered this a few years ago:

37) Klaus Weihrauch and Ning Zhong Is wave propagation computable or can wave computers beat the Turing machine? Proc. Lond. Math. Soc. 85 (2002) 312-332. Abstract available at http://www.lms.ac.uk/publications/proceedings/abstracts/p1364a.html

I wonder if they cite my paper.

Anyway, I came away feeling that time evolution for any PDE people actually care about is computable once you define your concepts correctly. This is simply a way of saying that you can numerically compute the solutions to any desired accuracy: it's sort of obvious, modulo the crucial technical details.

But a more interesting question would be the one you ask, where mere "computability" is replaced by something more refined, like "computable up to a specified accuracy in polynomial time".

Another addendum: here's a new article on the problems with finding a cryptographic hash function that's hard to crack:

38) Susan Landau, Find me a hash, AMS Notices 53 (March 2006), 330-332. Also available at http://www.ams.org/notices/200603/fea-landau.pdf and http://www.ams.org/notices/200603/fea-landau.ps

Landau points out that there's no good mathematical theory backing up the most popular hash functions. "There are hash functions based on hard mathematical problems, making them likely to be secure, but these hash function are inefficient and not used in practice."

Next, here's yet another email from Scott Aaronson, dated July 22, 2006:

Hi John,

A while ago you asked whether anyone had studied sentences of the form

P(n) = "This statement has no proof at most n symbols long."

Clearly P(n) has a proof with ~2n symbols. I emailed you conjecturing that P(n) has no proof much shorter than that.

Well, I just came back from a complexity meeting in Prague, where Pavel Pudlak gave a talk about exactly this sort of question. It turns out that Harvey Friedman made the same conjecture I did in the 1970's, but his conjecture was soon shown to be FALSE! Indeed, the number of symbols needed to prove P(n) grows only a little bit faster than n itself.

Let me sketch why, since I'm guessing this will be as surprising to you as it was to me.

To prove P(n), it clearly suffices to show that, if our proof system is inconsistent, then any proof of 0=1 must contain statements with more than n symbols. We can do that using a technique invented by Tarski in the 1930's for "defining truth" - i.e., for recursively building up a one-to-one correspondence between syntax and semantics. More concretely, let Tk(A) be a function that takes as input a string A with £k symbols. We want Tk to return 1 if A encodes a true statement and 0 otherwise. So for all k Tk+1( "not" + A ) := 1 if Tk(A)=0, or 0 otherwise.

Tk+1( A + "and" + B ) := 1 if Tk(A)=1 and Tk(B)=1, or 0 otherwise.

Tk+1( "There exists an x such that" + A ) := 1 if there exists an x* such that Tk(A[x=x*])=1, or 0 otherwise.

and so on. Of course I'm skipping lots of technicalities, like what I mean by A[x=x*].

Anyway, one can prove that, if Tn(A)=Tn(B)=1, and if C is derivable from A and B via a first-order inference rule, then Tn(C)=1 as well. Also, if A is an axiom of our proof system, then (reasoning within the system) A is true, and hence Tn(A)=1.

It follows that, as long as we restrict ourselves to statements of length £n, no contradiction can be proved. Furthermore, we can formalize the whole argument using only O(n) symbols (or maybe O(n log(n)), depending on the proof system).

So what's the catch? Why can't we generalize this argument to statements of all lengths, thereby contradicting the Incompleteness Theorem? Because, when we examine the proof that the syntax correctly models the semantics, we'll always find that it involves statements with more than n quantifiers. I.e. to prove consistency for statements with n quantifiers, we need at least n+1 quantifiers.

So, is there any way to salvage a finitary Gödel's Theorem? There probably is! In particular, consider the following statement:

P(n,k) = "This statement has no proof at most n symbols long, involving statements with at most k quantifiers."

The new conjecture would be that, for every fixed k, P(n,k) has no proof with less than 2O(n) symbols involving statements with at most k quantifiers. According to Pavel, this conjecture is still open. Indeed, it's not even known whether P(n) itself has a proof with less than 2O(n) symbols, involving statements with at most k quantifiers for some fixed k.

Best,
Scott


Anyone who considers arithmetical methods of producing random digits is, of course, in a state of sin. - John von Neumann

He is the Napoleon of crime, Watson. He is the organizer of half that is evil and of nearly all that is undetected in this great city. He is a genius, a philosopher, an abstract thinker.... - Sherlock Holmes


© 2006 John Baez
baez@math.removethis.ucr.andthis.edu