4 ms·

I assumed that we were talking about in place sorting. "I feel your pain." As I tried to explain, the first time I saw big-O notation taken seriously as the

by graycat 11y ago

I assumed that we were talking about in place sorting. "I feel your pain." As I tried to explain, the first time I saw big-O notation taken seriously as the main metric for evaluating the actual running time of actual code on actual data on an actual computer, I had your reaction, that the constants are important and the big-O criterion, not good. There's another point: The silent assumption of the big-O criterion is that we are interested in (A) something fundamental about algorithms, just the algorithms, and not (B) actual running time of actual code on actual data on actual computers. And big progress in algorithms is also important and, as we know, in some cases can give much better performance gains in real computing than anything a programmer can do without the big progress. And we have a great example: There were days before heap sort, shell sort, quicksort when the in place sorting algorithm people coded were bubble sort or something closely related and also O(n^2). Then in practice, the difference was just huge. Early in my career, I ran into that at Georgetown University: A prof had coded up some software for teaching statistics and had used the IBM Scientific Subroutine Package (SSP) for a most of the actual calculations. One of the SSP routines was to find ranks, and it had two loops, each essentially bubble sort. In testing, the thing ran all lunch time. Well, the data to be ranked was 32 bit integers, and an array index was a 16 bit integer. So, I was able to do a tricky overlay of two 16 bit integers on the 32 bit data, use heap sort, and get O( n ln(n) ). In actually running time on nearly any possible real data, my code as so much faster than IBM's SSP that I blew the doors off the SSP. Lesson 1: O( n ln(n) ) instead of O( n^2 ) can be much better in practice. Lesson 2: Progress in big-O in algorithms can be terrific stuff. Back to practice, early in writing the software for my startup, I tried to write a polymorphic version of heap sort. I wrote in Microsoft's Visual Basic .NET. For the data type to be sorted, I passed just the type object. The software ran correctly right away. Then I did some careful timings -- the software commonly ran ballpark 10 times longer than I would have guessed. What the heck? Well, the run time was smart enough to figure out the actual data type and do the comparison, but the overhead here was enormous. Then I learned that for polymorphic code I was supposed to write some interfaces, one for each data type to be sorted, and for sorting a particular data type pass the corresponding interface. Okay. My confusion was that that use of an interface didn't sound polymorphic to me and sounded no different than what I'd long done passing entry variables in PL/I and Fortran. Okay, for polymorphism, use interfaces. And in the code of the interface, have to work essentially without strong typing, essentially with untyped pointers. Okay -- been there, done that, in both PL/I and Fortran; I was surprised that a serious object oriented language with polymorphism would have essentially just syntactic sugar over what I'd been doing in PL/I and Fortran. Okay. Lesson learned: There are cases of hype in parts of programming languages. Heck, in PL/I my approach to sort an array of PL/I structures would be to pass a PL/I structure that had an element variable that was an entry variable to a procedure that would do the comparisons. That is, could have some conventions that would let PL/I structures be self-describing. Then could write code that was more general. Maybe under the covers, that's just what's going in the Microsoft common language run time (CLR) that Visual Basic .NET used for me without letting me know -- more syntactic sugar? So, when I rewrote my polymorphic heap sort to do the comparisons with a passed interface, the performance got okay. Still, I have non-polymorphic versions for the more important element data types, and they are faster in just they way you emphasized. Yup, been there, done that. So, I, too, was concerned about actual running time of actual data. But for the algorithm itself, I was still using just heap sort and didn't try to write a hybrid routine that sometimes might exploit, say, radix sort. On the pros and cons of using big-O for the only criterion, there's a big example in linear programming: In practice, the simplex algorithm is fantastically fast. Intuitively, in practice, the early iterations in effect focus what to do to get to the optimal solution (assuming feasibility, no unboundedness, etc.). But due to some work of Klee and Minty, the worst case of simplex is exponential and just awful. Intuitively the Klee and Minty examples trick simplex into making really bad choices! So, there was research to find a polynomial algorithm for linear programming. Eventually there was an ellipsoid algorithm that was polynomial but the constant was so high that whenever in practice ellipsoid beat simplex both ran too long to be practical. So, really, the concern about simplex and the effort for ellipsoid was as progress just in algorithms and in the big-O criterion. And in practice, right away, the constant was so large that ellipsoid was a flop. Still, in algorithms, people are straining to do better in the sense of the big-O criterion. Of course the biggest effort here is to find an algorithm that shows that P = NP, maybe the one of the most important problems in both math and computer science, based on big-O for some polynomial. With all of that, I decided in my own work, for something as simple as sorting, just to go with heap sort, maybe even a polymorphic heap sort. Indeed, with all the emphasis on polymorphism, people can't be very concerned about constants! :-)!

nightcracker 11y ago
I'm impressed you can write such a long story yet address none of my points. You should consider becoming a politician. The only relevant piece of information I could find in your anecdote is this: > I assumed that we were talking about in place sorting. If we limit ourselves to in-place sorting algorithms then we still are left with block sort, pdqsort, introsort and heap sort, all of which run in O(n log n). My argument still holds.
- graycat 11y ago
  I wrote that long post trying to explain my attitude on sorting software from 50,000 or so feet up and, in particular, why I have been happy just to use heap sort and forget about looking for anything faster on average. I was reluctant to read the documentation in your GitHub submission if only because in my software development I have never needed or used GitHub. Also, I was reluctant to believe that there could be any version of quicksort that would have worst case running time of O( n ln(n) ). Why reluctant to believe? First, sure, the usual version of quicksort has us select a pivot value for a partition from a "median of three" keys from that partition. Second, sure: For a good pivot value, we want essentially the median of the keys in the partition we are trying to split into two partitions. And, sure, the median of three approach to selecting a pivot value is an approximation to the median we really want and, thus, a relatively good pivot value. So this approach to selecting a pivot value promises on average to get faster average running times. Fine. Faster on average is good. Terrific. But this pivot selection rule says next to nothing about worst case running time, and, for claiming worst case running time of O( n ln(n) ), this is a biggie issue. Third, with this median of three approach to selecting pivot values, we can think of bad arrays of input keys that will cause quicksort to run in time no faster than O( n^2 ). Then maybe we can think of other ways to select pivot values that result in a new version of quicksort, a version that runs in O( n ln(n) ) again on the bad arrays. But with this change, here's the problem: We are lacking a proof that the worst case running time of the new version of quicksort has worst case running time of O( n ln(n) ). That is, our new version of quicksort is good for some old cases of bad arrays but may have created for itself some new cases of bad arrays. So, with no proof, the new version of quicksort may also have worst case running time of O( n^2 ) like the old version did. The main difference might be that the bad arrays for the new version are more complicated to construct than the bad arrays for the old case. But I just looked at the file README.MD you suggested. I see two issues: First Issue: > When a new pivot is chosen it's compared to the greatest element in the partition before it. If they compare equal we can derive that there are no elements smaller than the chosen pivot. When this happens we switch strategy for this partition, and filter out all elements equal to the pivot. If I understand this statement correctly, I suspect that for some arrays of input keys, the resulting algorithm won't actually sort the keys. That is, we won't have a sorting algorithm. Second Issue: > While this technically still allows for a quadratic worst case, the chances of it happening are astronomically small. For proving worst case running time of O( n ln(n) ), "astronomically small" is irrelevant. Maybe your remark "astronomically small" is just a side comment and not relevant to PDQSORT having worst case running time of O( n ln(n) ), but from your documentation I can't easily tell. > A bad partition occurs when the position of the pivot after partitioning is under 12.5% (1/8th) percentile or over 87,5% percentile - the partition is highly unbalanced. When this happens we will shuffle four elements at fixed locations for both partitions. This effectively breaks up many patterns. If we encounter more than log(n) bad partitions we will switch to heap sort. For this and the associated documentation, I don't see a solid argument that PDQSORT has worst case running time of O( n ln(n) ). For some work, a big-O guarantee on worst case performance is a biggie, as in crucial to human life. Such a guarantee is essentially a theorem in applied math called computer science. A good example of how to prove such theorems is, of course, in D. Knuth, The Art of Computer Programming, Volume 3, Sorting and Searching. Again, I have been happy just to use heap sort and forget about looking for anything faster on average.

by graycat 11y ago

nightcracker 11y ago

I'm impressed you can write such a long story yet address none of my points. You should consider becoming a politician. The only relevant piece of information I could find in your anecdote is this: > I assumed that we were talking about in place sorting. If we limit ourselves to in-place sorting algorithms then we still are left with block sort, pdqsort, introsort and heap sort, all of which run in O(n log n). My argument still holds.

graycat 11y ago
I wrote that long post trying to explain my attitude on sorting software from 50,000 or so feet up and, in particular, why I have been happy just to use heap sort and forget about looking for anything faster on average. I was reluctant to read the documentation in your GitHub submission if only because in my software development I have never needed or used GitHub. Also, I was reluctant to believe that there could be any version of quicksort that would have worst case running time of O( n ln(n) ). Why reluctant to believe? First, sure, the usual version of quicksort has us select a pivot value for a partition from a "median of three" keys from that partition. Second, sure: For a good pivot value, we want essentially the median of the keys in the partition we are trying to split into two partitions. And, sure, the median of three approach to selecting a pivot value is an approximation to the median we really want and, thus, a relatively good pivot value. So this approach to selecting a pivot value promises on average to get faster average running times. Fine. Faster on average is good. Terrific. But this pivot selection rule says next to nothing about worst case running time, and, for claiming worst case running time of O( n ln(n) ), this is a biggie issue. Third, with this median of three approach to selecting pivot values, we can think of bad arrays of input keys that will cause quicksort to run in time no faster than O( n^2 ). Then maybe we can think of other ways to select pivot values that result in a new version of quicksort, a version that runs in O( n ln(n) ) again on the bad arrays. But with this change, here's the problem: We are lacking a proof that the worst case running time of the new version of quicksort has worst case running time of O( n ln(n) ). That is, our new version of quicksort is good for some old cases of bad arrays but may have created for itself some new cases of bad arrays. So, with no proof, the new version of quicksort may also have worst case running time of O( n^2 ) like the old version did. The main difference might be that the bad arrays for the new version are more complicated to construct than the bad arrays for the old case. But I just looked at the file README.MD you suggested. I see two issues: First Issue: > When a new pivot is chosen it's compared to the greatest element in the partition before it. If they compare equal we can derive that there are no elements smaller than the chosen pivot. When this happens we switch strategy for this partition, and filter out all elements equal to the pivot. If I understand this statement correctly, I suspect that for some arrays of input keys, the resulting algorithm won't actually sort the keys. That is, we won't have a sorting algorithm. Second Issue: > While this technically still allows for a quadratic worst case, the chances of it happening are astronomically small. For proving worst case running time of O( n ln(n) ), "astronomically small" is irrelevant. Maybe your remark "astronomically small" is just a side comment and not relevant to PDQSORT having worst case running time of O( n ln(n) ), but from your documentation I can't easily tell. > A bad partition occurs when the position of the pivot after partitioning is under 12.5% (1/8th) percentile or over 87,5% percentile - the partition is highly unbalanced. When this happens we will shuffle four elements at fixed locations for both partitions. This effectively breaks up many patterns. If we encounter more than log(n) bad partitions we will switch to heap sort. For this and the associated documentation, I don't see a solid argument that PDQSORT has worst case running time of O( n ln(n) ). For some work, a big-O guarantee on worst case performance is a biggie, as in crucial to human life. Such a guarantee is essentially a theorem in applied math called computer science. A good example of how to prove such theorems is, of course, in D. Knuth, The Art of Computer Programming, Volume 3, Sorting and Searching. Again, I have been happy just to use heap sort and forget about looking for anything faster on average.