Comment by Jd | Hacker News Reader

Comment by Jd | Hacker News Reader

Parent Full thread

I have mixed feelings about this. I had an interview at Palantir with a self-described "front-end" guy, who asked me to write an algorithm to find a number between zero and infinity, where you could pick a number and you would get a response that would tell you whether or not the number was lower or higher than the secret number. The dude had a particular form of approaching the problem that was not at all the fastest form, in fact, he insisted on starting at zero and using an algorithm that was "2x," which he insisted was "fast enough."

I thought that a system which started at zero and went up by multiples of two was obscenely slow, and I hinted at this, but he insisted on having things his way. In the end I'm sure he wrote me a bad review and I wasn't offered the job (I didn't really want it anyways, given my impression that the team I was interviewing for wasn't that focused).

In any case, I didn't go right now and say this is dumb, but I hinted at a better way, and even this hinting lead to some conflict. In the end I didn't get the job. I'm conflicted if this was the best approach. I could have pretended to be dumber than I actually am and gotten the job offer at least, which at this present moment I could probably use for leverage in the search for contracts or a job, but I couldn't bring myself to accept his not particularly good algorithm (full disclosure: the process was also affected by the fact that my flight was delayed from the east coast, that I got 4 hours of sleep, the fact that I insisted in wearing a tweed coat to the interview, which may have impacted ever important perception of "fit," and that the interview in question was a "surprise" interview after several other interviews).

It's a binary search in infinite space. His solution expands the problem area until the number is in the space and then shrinks the problem area until it finds the number. A better solution is to pick an arbitrarily large number and only expand if you have to--otherwise you're shrinking from the very beginning.

That's how I would have answered the question, anyway. But yes, I've dealt with interviewers like this--typically it's because they don't understand the subject matter well enough to evaluate or accept an answer they didn't prepare for.

I disagree that that solution is unequivocally better. It depends on the distribution of the input data; if the expected value of the number is small, then you're better off starting at zero than dramatically overestimating.

Of course, either way the algorithm runs in O(log n) time.

Within the context of the interview question, the solution is specifically stated to be between 0 and infinity, and you can't surmise that the number is necessarily closer to 0. This is the error that my interviewer made. He was trained to deal with things within the context of situations closer to real life rather than the theoretical box he had established for the interviewing situation.

As for my case, I did attempt to start at an arbitrarily large number (but he wouldn't let me), and then went up exponentially (i.e. x = x * x), bi-secting the space once the upper-bound was found and solving from there.

My interviewer was sadly not that bright and didn't understand my solution, forcing me to re-write it -- incidentally, lecturing me on how wonderful O(log n) is along the way.

> the solution is specifically stated to be between 0 and infinity, and you can't surmise that the number is necessarily closer to 0.

Closer to 0 than what? Infinity? By any reasonable definition, every integer is closer to 0 than infinity.

I think the problem as stated is underspecified; I disagree that increasing "exponentially" (the solution your interviewer described would increase exponentially; your version is doubly exponential, I think) is always faster. Both versions have the same worst-case running time, and if you want to talk about expected running time you need to make assumptions about the data distribution. As a counterexample, if you start with an initial guess of 2 and keep squaring, but the true number is 2^1025, then you end up doing much more work than if you had doubled instead of squaring.

I can certainly agree with you that lecturing in an interview is unproductive. Not only is it understandably off-putting, but it doesn't give the interviewer any additional input to the hiring decision and therefore is a waste of everyone's time. (Unless one of the job requirements is coping well with being lectured to, I guess.)

> By any reasonable definition, every integer is closer to 0 than infinity.

Perhaps, but originally we weren't discussing 0, we are starting with an arbitrarily large number. This problem is exacerbated when you say:

> I disagree that increasing "exponentially"... is always faster ... Both versions have the same worst-case running time

Worst case, sure, but how do you compute average? You could take a selection of n number of integers from 0 to infinity and the number of guesses until you found the correct answer, but of course you can't do this because you can't easily get a random number to test with that is between 0 and infinity. That said, it is fairly clear that x * x is faster than 2x is faster than x+1, but you can only prove this to be true once you pick an arbitrary limit less than infinity -- and the benefits of the better algorithm are only evident as you approach infinity.

Seems like a catch 22. Perhaps there is some expert in set theory as applied within computer science that can provide the appropriate formal context for determining the better algorithm in this context.

As for O notation, it is an interesting tool but one that has no application within any realm of actual programming in which I have worked, and I have never bothered to learn it simply to dazzle people in algorithm-based interviews (perhaps because I do more nuts and bolts type work rather than optimization).

From the perspective of theoretical computer science, the interviewer was correct. I'll try to sketch out why.

------------------

1. In the initial probing for the upper bound, there is no point to grow faster than 2x each time

If we double the probe each time, we'll find the upper bound in O(log N) time. Then, we'll need O(log N) additional time to find the real answer. That makes the entire algorithm O(log N).

Suppose instead, in the initial probing we grow the bound faster, say by squaring each time. We'll find the upper bound faster, but we'll still need additional O(log N) time to find the real answer. So, we didn't really make the algorithm (asymptotically) faster - it is still O(log N).

(I glossed over some details in the explanation, but even if you work out the math exactly - which is not that hard to do - the conclusion holds.)

------------------

2. There is no point starting from a number greater than 1

There is no way to pick a "good" starting number - should it be 1,000? 1,000,000,000? 10^100? (10^100)^100? You might as well start from 1. That way, you guarantee that you'll find small numbers fast and large numbers still in an asymptotically optimal time.

------------------

As someone with fairly strong theoretical computer science backgound, I can see the intended meaning behind the interviewer's question and answer. But, it is a theoretical question. There definitely are a lot of highly valuable software developers out there who couldn't answer it.

Given that is a theoretical question that you are attempting to fit into an actual algorithm to ostensibly solve a real problem, doesn't the actual performance of the algorithm matter? I can see that within the context of theory of computer science there may not be a difference between the speed of the two algorithms, since both are "asymptotically optimal" (a term I'm not sure I entirely grok), yet within virtually any defined set of random numbers less than infinity the algorithm I provided will perform substantially faster.

I suppose I'm at a loss for words, since as a seasoned software developer without a deep background in theoretical computer science, it seems that I am being told to disregard my intuition based on experience in favor of a theoretical model that seems to ignore actual problem sets. Am I missing something?

I made a python script to time the different approaches:

https://gist.github.com/3839551

AFAICT, the parent and the interviewer is right because both methods run in roughly the same time because the time of the binary search is orders of magnitude greater than the upper bound finding part. I might have made one or more mistakes though.

If I have time later today I'll extend your snippet, but my inclination is that the differences in speed become apparent only with very large numbers. I still suspect my algorithm is about 40-50% faster if you start with a googol.

It seems that the problem involves two parts: 1) To find an upper bound 2) Then divide the remaining region in halves until the number is found.

The first observation that I have is that given that the secret number s is chosen, the first step can be completed arbitrarily quickly. One could use a function that rises arbitrarily fast. Imagine for example the function taking k to the Ackerman function A(2,2,k). That rises so fast it's incomprehensible, but really one could easily produce a function which rises faster still (the algorithm that picks s first!). The problem is, though, of how fast a fixed algorithm is for random s. If s is truly chosen at random from the positive integers, this leads to problems. Fix your putative algorithm. Suppose for the moment that it starts at 0 (it is not going to matter where it starts) What is the probability that the kth number that your function spits out is less than a random integer? 100% After all, how many integers are greater than any given integer?

Therefore, no growth function is any better on average at finding the upper bound than any other. Therefore step 1) is an insoluble problem. The problem should have been specified in some other way in order for it to make sense.

However, the second step involves log_2(n) time (in the worst-case-scenario and still O(log n) in general) where n is the upper bound --the output of step 1-- which means that the time to complete the algorithm is (Time of step 1 to find n) + O(log n). IF the problem made sense and it was the case that step 1) were soluble, then it would matter how fast step 1 is countered by the degree to which step 1 tends to overshoot the secret number --because overshooting by k has the penalty of log(k) extra operations.

Is there are framework in which question 1 makes sense? It would make sense if there were a given probability distribution on the integers (a function on the positive integers whose sum over all integers = 1, for example choose the function f(n) = 6/(n^2 * pi^2)) A probability distribution gives you a way of answering the question: "what proportion of the positive integers are greater than k?"

Excellent response. I think the framing element that was missing is if the random numbers were truly random.

My intuition says that they should almost always run at about the same time if you don't know anything about the distribution of guesses. Your algorithm has the advantage that it will find an upper bound for the number much faster than the interviewer's algorithm. HOWEVER, having found that upper bound, your algorithm will probably be much farther away from the actual number than the interviewer's, so you have to do that much more narrowing down. What it comes down to is being lucky that you're close to the number in your initial guess, otherwise, you're still talking logarithmic time to either find an upper bound or (in your case) lower bound for the number.

This was my conclusion, that although the two approaches had the same average big O time, my intuition said it was somewhat faster to start from some number larger than 0.

In retrospect, I might have mentioned this and then gave the interviewer's preferred answer in code, just to keep things simpler.

Reply on news.ycombinator.com