When I find something interesting and new, I post it here - that's mostly programming, of course, not everything.

Wednesday, August 24, 2011

Either OOP or Liskov (choose one)

A while ago I read some Abadi-Cardelli, and illustrated what I learned by an example how Java does not have subsumptions principle (that is, if A is a subtype of B, an instance of A can be safely substituted everywhere where an instance of A is required).

So I had posted a blog entry with the example.

There's another post on this on blogger.

A deeper look into the issue can be found in "Is the Java Type System Sound? by Sophia Drossopoulou and Susan Eisenbach.
The paper shows that no, it is not, and also builds a sublanguage of Java where subsumption holds.

From this article I've got the following demonstration sample:


class A {}
class A1 extends A {}
class B {
char f(A x) { System.out.println("f(A." + x + ")"); return 'f'; }
int f(A1 y) { System.out.println("f(A1." + y + ")"); return 12345; }
void g(A1 z) { System.out.println("Called g"); System.out.println(f(z)); System.out.println("out of g"); }
void h(A u, A1 v) { System.out.println("Called h"); System.out.println(f(u)); System.out.println(f(v)); System.out.println("out of h"); }
}

@Test
public void testJavaSubsumption() {
new B().g(new A1());
new B().h(new A1(), new A1());
}


The output is here:

Called g
f(A1.pigi.examples.library.tests.QueriesTest$A1@6526804e)
12345
out of g
Called h
f(A.pigi.examples.library.tests.QueriesTest$A1@42b1b4c3)
f
f(A1.pigi.examples.library.tests.QueriesTest$A1@20d2906a)
12345
out of h


I wonder though how long will the Earth population hold to the popular belief that one can have both OOP, with reasonable enough "dispatch", and substitution principle. One cannot.

Questions?

P.S. Actually, the issue was discussed in details By Oleg Kiselyov a while ago: http://okmij.org/ftp/Computation/Subtyping/.

Saturday, July 30, 2011

Generic Data Model

Generic Data Model



Introduction



15 Years ago I had a problem designing a database that was supposed to hold data from various questionnaires; the questions were slightly varying every time and had no
regular structure, reflecting the weird ways of thinking of marketing people. The solution was to produce something more or less generic, storing “entities” and “attributes” separately.

(there was some discussion of xml and corba, now totally irrelevant)

Tables



The model consists of three tables: ENTITY, ATTRIBUTE, CONTEXT.


ENTITY Table



This is the main table in the database:

create table ENTITY (PK long, CLASS varchar, CREATED timestamp);

All the entities that we are going to have stored in our database have a representative in ENTITY table. An entity represents an object of our model. An object has an inalienable property, class. The term ‘type’ would probably suit better. An entity cannot change its class; actually, all the fields in entity record are read-only.

ATTRIBUTE Table



This table contains attributes of the entities. Records consist of name-value pairs and reference ENTITY table:

create table ATTRIBUTE (ENTITY long, NAME varchar, VALUE varchar);

The usage of this table is probably obvious. Note that if NAME is null, it means that the attribute contains the value of the entity itself. Two other fields are non-null.

CONTEXT Table



On one hand, this table can be considered as containing all the collections of the database; but on the other hand, the purpose of this table is wider. It contains all the naming contexts, and it can be used for navigation as well as for storing master-detail relationships.

create table CONTEXT (NAME varchar, OWNER long, MEMBER long, MEMBERID varchar);

Here NAME is the name of the context. If you are interested in collections, it is probably the name of collection owned by OWNER. OWNER is obviously the entity that owns the context, the “master”. MEMBER points to the member of the collection; in this case MEMBERID is irrelevant. But we can consider this as a Hash, in which case MEMBERID becomes a key, and MEMBER is the corresponding value. In the Naming Service paradigm, MEMBERID is the name of the entity MEMBER in the context NAME.

Conclusion



These days the idea of key-value pairs has become ubiquitous; it was not so in 1996 or in 2001 when I wrote this. So I'm dropping the argument in favor of this, now obvious, solution.

Friday, July 15, 2011

Chain Calls in Java - a better solution

A while ago I wrote "Java Solutions: inheriting chain calls"; it was about what to do if we want to make it possible to write calls like
return new StringBuffer().append("Now is ").append(new Date()).append(", and ticking...").toString();


and be able to subclass at the same time.

See, the problem is that if you setter returns this:

  MyBaseEntity withExcitingNewFeature(ExcitingNewFeature f) {
    features.add(f);
    return this;
  }


and then you subclass MyBaseEntity, like MyDerivedEntity extends MyBaseClass, you won't be able to write
  MyDerivedEntity mde = new MyDerivedEntity(entityId).withExcitingNewFeature(feature);


Since, well, the class is different.

So in my previous post I suggested to make MyBaseEntity type-dependent (that is, have a generic):
 public abstract class MyBaseEntity<T extends MyBaseEntity> 


and declare a method
   abstract T self();


So that every subclass declaration would look like this:
  public class MyDerivedEntity extends MyBaseEntity<MyDerivedEntity>


and would define

  MyDerivedEntity self() {
    return this;
  }


Yes, it works. And probably for 2008, with is now ancient history, when slow people slowly typed their slow code, not caring about the speed of development... in short, it was long ago.

Now there's a better solution. No need to redefine self() in all possible subclasses. That was just stupid. The only place where it should be defined is the base class:

  T self() {
    return (T) this;
  }


And no need to declare MyBaseEntity abstract.

There's also a bug in that old post; the generic type should have been bounded, as Ran noted in his comment.

Of course all setters should look like this:
  <T extends MyBaseEntity> T withExcitingNewFeature(ExcitingNewFeature f) {
    features.add(f);
    return self();
  }

Sunday, June 19, 2011

2 eggs and 100 floors

Everybody seems to know this "Google Interview Problem": (quoting )

"There is a 100-story building and you are given two eggs. The eggs (and the building) have an interesting property that if you throw the egg from a floor number less than X, it will not break. And it will always break if the floor number is equal or greater than X. Assuming that you can reuse the eggs which didn't break, you need to find X in a minimal number of throws. Give an algorithm to find X in minimal number of throws."


A naive programmer would first apply binary search for the first egg, then linear search for the second egg; a less naive one would try to apply Fibbonacci sequence.

Now what would happen if we just start with floor 50? We will waste the first egg, most probably, then walk up with the second one, trying each floor. What would be the number of throws in this case? We Don't Know. One would say it would be 51, but there's no reason to believe it. The only case you'll need to throw 51 times is if the first egg does not break, but we will start throwing the second egg from floor 51, 52, etc, all the way to floor 100. Would anybody do it? No.
We would divide it into halves again, right?

The other reason is that, hmm, maybe we find the right floor earlier? But what's the probability?

Right. To calculate the number of drops, we need to know the probability distribution. It is not specified in the problem. More, the problem does not even specify what exactly we are optimizing. Suppose we have a million of such buildings (and 2 million eggs, 2 per each building); what would be the good strategy? Of course we would like to minimize the average search time, not the maximum possible. If we want to apply this kind of solution to real-life problems, we need to minimize the average.

But we don't know the probability distribution! One possible (real-life) distribution is that the egg breaks at the height of an inch, so the first floor is already out of luck. But then we can modify the problem, using Faberge eggs and a 100-floor doll house.

Well, but an interviewee would not go there right away. Another good choice is to take a square root of 100, and try the first egg with the step of 10 floors, and the second egg with the step of 1 floor. A popular (and wrong) answer would give 18 to 20 as the maximum number of attempts, and 10 as an average number of attempts. Actually, the average number of tries would be 11.

The popular "right" answer that you can find on the web gives this sequence of floors to try:
14, 27, 39, 50, 60, 69, 77, 84, 90, 95, 99, 100

I will not repeat the proof here, since the proof assumes we know the problem we are solving, and I state that we don't. I mean, if we are talking about the maximum number of attempts, then 14 might be a good choice, but the problem never says this is about minimizing the maximum number of attempts, right? As somebody noticed, the minimal number of throws is either 0 (for real-life eggs and floors) or 1, for the lucky ones that are not allowed in Vegas casinos.

What I suggest to do is try to minimize the average number of throws, in the assumption of uniform distribution.

So, strictly speaking, we have values from 1 to 101, and the probability of the egg to break at floor n but not at n-1 is 1/101. Why 101? We are still not sure about floor 100, but if we had values 1..100, then it would follow that we'll have a breakage at floor 100, just because the sum of all probabilities should be 1.

Now, with this distribution, we can calculate an average number of attempts for the run of n floors (in the assumption that the egg breaks at floor n): it is (n+1)/2 - 1/n. You can calculate the formula yourself. In Scala, the calculation looks like this:

def linear(n: Int) = (n + 1.) / 2


For the first suggestion, that is, trying floor 51 and then going linear, we would have 27.(2277) attempts on average (you could also try first floor 50 and see that this is a little bit worse).

How did I calculate this? See, this strategy involves:
- throwing 1 egg;
- with the probability p1=51/101 going linear from floor 1 to floor 50;
- with the probability p2=50/101 going linear from floor 52 to floor 100.

The average number of tries is 1 + linear(51) * p1 + linear(50) * p2.

We can extend this formula for a sequence of attempts, using the following Scala code:

def forSequence(s: List[Int]) = ((0., 0) /: s.reverse) ((a,b) => { val c = a._2 + b; (a._1 * a._2 / c + 1 + linear(b) * b / c, c)})


That is, taken a list of steps, we calculate the average using recursive formula, combining the previous average with the formula for linear seach.

If we try it for the "square root search", forSequence(List(10,10,10,10,10,10,10,10,10,11)) gives 11.0; for the scientifically recommended sequence 14,13,..., forSequence(List(14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 1)) gives 10.47 as an average number of attempts. Not bad; and makes sense.

But can we do better? To figure out what's the optimal sequence, we need to apply dynamic programming... unless we know the right formula; I don't. By the way, did you notice this last jump from 4 to 1 in the "scientific" sequence? Looks suspicious; it might be better to spread this jump throughout the sequence.

My solution is this: we have the right answers for the very short sequences, like 1, 2, or 3 floors; now we will need, for each subsequent number, to find the right first step, how many floors to skip; the next step is already optimized. Since this involves a lot of backtracking, I had to introduce caching of the intermediate results, splits and averages. This is Scala, and caching function is very easy to write:


def cached[K,V](f: K => V) : K => V = {
val cache = new scala.collection.mutable.HashMap[K,V]
k => cache.getOrElseUpdate(k, f(k))
}

the f() in the code is evaluated lazily - on the need-to-know basis.

Honestly, one would need to use Y-combinator, but I made a shortcut, and wrote the solution like this:

val best: Int => (Double, Int) = cached((n: Int) =>
if (n < 4) (linear(n), 0) else
(for (k <- 2 to (n/2)) yield
(1 // for the first ball
+ linear(k) * k / n
// average number of throws if
// the first ball breaks, with the probability of this, k/n
+ best(n-k)._1 * (n-k) / n,
// avg number of throws if
// the first ball does not break, with the probability
k // this is the step at which we throw the first ball
)).min) // minimize using default pair comparator


Here best is a function; it takes building height and returns two numbers: the average number of tries and the floor to try with the first ball. It uses cache. For the case of less than 4 floors the answer is linear. Otherwise, we try steps between 2 and half the height and see which one gives the best result.

To dump all intermediate steps, I had to write this:

def bestSteps(height: Int): (Double, List[Int]) = {
val first = best(height)
(first._1, if (height < 4) Nil
else first._2 :: bestSteps(height - first._2)._2)
}


What we do here is collecting the steps; it returns the price (average number of throws) for a given building height, as well as the list of steps between floors in case the first egg does not break. The result is this:

(10.43,List(14, 12, 11, 10, 9, 9, 7, 7, 6, 5, 4, 3))


See, we beat the "scientific" (but cheap) result, and the steps suggested are more in harmony with the nature of things, are not they?

Q.E.D.

Thursday, May 05, 2011

Monad Genome, part 1

Intro



Recently I started looking into the origin of computer science monad, and found out that at least some of them originate in pretty simple things. Like segments of simplicial category Δ (for a better explanation/definition, see MacLane, p.175). Here I'll give a couple of examples, Option and Reader monads, how they derive from a very simple subcategory of Δ.

0, 1, 1+1



I'm going to introduce a very simple category, consisting of three, well, sets: empty, singleton and a union of two singletons (1+1), together with two arrows, 0 → 1 and 1+1 → 1:



There's not much to talk about this category; you can think of it as living in the category of all sets, but actually any category with finite limits and colimits contains it.

Now I'll show how it generates two popular monads.

Option Monad



Given an object X, if we add X to each part of the diagram (a), we get this:



which is exactly Option Monad.

Exception Monad



First, we multiply the diagram (a) by E:



Then we can do the same trick as in Option Monad, add an X:



This is Exception Monad.

Reader Monad



Given an X, we can build the following diagram by applying X- functor to the diagram (a):



Which is, simply speaking, just



Now this can be used to build the following diagram, for any A:



Do you see Reader diagram? It is the functor , with all the natural properties that follow from the naturality of the morphisms above.

This post is work in progress; comments are very welcome.

Thursday, April 28, 2011

Church Numerals in Scala

As everyone knows (see http://en.wikipedia.org/wiki/Church_encoding), in pure lambda one can define natural numbers out of nothing, like this:

0 ≡ λ f λ x . x
1 ≡ λ f λ x . f x
2 ≡ λ f λ x . f (f x)
3 ≡ λ f λ x . f (f (f x))
etc.

At our BACAT meeting yesterday, there was a question, for typed lambda, how can one express natural numbers, and also how can one interpret them in a Cartesian-closed category?

The answer to this question turned out to be trivial, but we also had managed to express it in Scala:
class TypedLambda[T] {
type TT = T => T
type Church = TT => TT
val cero: Church = f => x => x
def succ(n: Church) = f => n(f) compose f
val uno = succ(cero)
val dos = succ(uno)
val tres = succ(dos)
}


Looks neat, eh?

The alternative approach can be found in Jim McBeath's post.

Sunday, January 02, 2011

Notes on my relations with CT

Zaki suggested me to write something on this topic...

So, what happened to me... In several stages.

1. I'm 19, a student in Leningrad University (that's USSR, now called Russia), and I attend a special course on homological algebra. Professor Yakovlev starts it with a short intro into category theory, a topic nobody heard of. I was overwhelmed. The thing that generalizes groups and sets and whatever. Diagrams, wow. Commutative diagrams, diagram chasing, wow. The rest was not so interesting; abelian categories, exact diagrams, hom, tensor product, ext and tor... oh whatever. In a couple of years I decided not to go into algebra.

2. I am 24, working as a programmer, trying to pick up some "science". Somewhere (where?) I see a title of a paper, "Applying Category Theory to Computer Science". Wow. My eyes are open. Here it is; database structures, they are just pullbacks. Programs, they are diagrams, are no they? Okay, nobody hears me; and more, there's no way I can get the paper. So just have to guess.

3. I am 27; me and my friends are camping somewhere in the forests of Novgorod region. Vodka, talks... I mention category theory as a good tool for perceiving comp. sci. Andrey responds by inviting me to a category seminar that they have at Technological Institute in Leningrad. Wow! I join them.

4. Several years of studying categories and toposes; MacLane's Categories for the Working Mathematician - I get the microfilm, and print it, samizdat style, in tinest possible print, for the participants. It costs 10 roubles each (like 3 bottles of vodka, if you know what I mean).

5. Andrey's friend is married to a relative of Vonnegut. So this friend, he is an otkaznik, and his American wife stays with him for several years, raising a kid, painting, just being patient. Meanwhile, after one of her trips back to US she brings Johnstone's Topos Theory... and on another trip she brings "Lolita". I devour both books. Have made hand copies of Topos Theory. Twice.

6. Andrey suggests to try to calculate topologies (and, of course, logics), over some specific finite categories; he is specifically interested in Delta3, the initial segment of simplicial category. I write code first in Basic, then figure out that the calculation for delta3 would take 2 weeks. Rewrite it into Fortran; now it's just a couple of days. So I rewrite the core parts (like enumerating subsets, calculating cones and limits) into Assembler. We have the list of 31 Grothendieck topologies over delta3 in just four hours. Meanwhile we calculate topologies for simple categories, like linear segments, commutative squares... a strange thing is discovered, on many occasions the number of topologies is 2^n, where n is the number of objects. That's for posets.

7. I prove it that for finite posets this is the case, and a little bit more. Then I spend months trying to find a typewriter with English letters (KGB is watching); used one in Hungarian trade mission, and one in the Institute of Hydrology (where global warming was then starting). Then I spend some time figuring out how to send the paper, then a KGB colonel suggests a solution, and I use it. It's published in Cah.Topo.Geom.Diff. The hardest part was deciphering French written notes on proofs.

8. I meet Yakovlev; for him, it is all a) "abstract nonsense", and b) it can't be that the result is not known yet.

9. I generalize this stuff to Karoubian categories over Boolean toposes (Booleanness is necessary) and send it to Johnstone. Johnstone, it turned out, was working on a similar problem, from the other end. So I spend the next 15 years trying to get the proof in the opposite direction - that if... then the category is Karoubian. That's how I did not do anything else. At least I tried.

10. The seminar meanwhile had died out - the guy, Stepanov (can I call him professor? something like that... he was not), had married and moved to Novorossiysk, a pretty criminal place; there he died in strange circumstances.

11. Now here in California, the word "monad" rings the bell. So I honesty try to put it in the proper perspective, writing some kind of "tutorials" and "talks", writing again the code that calculates stuff, opening presheaf.com, and ready to talk to anybody willing to listen.

12. And now we have BACAT, a category seminar where we can discuss things. And this is great.

Followers

Subscribe To My Podcast

whos.amung.us