I concur with the other commenters that it's silly for you to complain about his not comparing his online string-search algorithm against offline string-search algorithms that search an index of the text, such as suffix-tree algorithms.
As for the online algorithm issue, see my reply below.
Dual-Pivot Quicksort is a demonstration that someone didn't read the existing literature; nothing more.
It's true that dual-pivot quicksort is a step in the direction of Samplesort. (I do understand Samplesort well enough to say that.) That doesn't mean it's not a worthwhile contribution in its own right. Jon Bentley (advisor of Brian Reid, Ousterhout, Josh Bloch, and Gosling while at CMU; later at Bell Labs in its glory days) was quoted as having a substantially different opinion from yours:
http://permalink.gmane.org/gmane.comp.java.openjdk.core-libs...
> I think that Vladimir's contributions to Quicksort go way beyond
> anything that I've ever done, and rank up there with Hoare's original
> design and Sedgewick's analysis. I feel so privileged to play a very,
> very minor role in helping Vladimir with the most excellent work!
I am not sure Bentley will be persuaded by your claim that he didn't read the existing literature.
Here's an excerpt from the email I wrote to Prof. Sedgewick back then; he didn't reply, so I've no idea what he thinks about this:
"I believe that your PhD thesis studies this under the name of "double partition modification". I could find no real difference between the two algorithms. Yaroslavsky claims to achieve a factor of 0.8 improvements over the standard quicksort on the number of exchanges (neutral on comparisons), while your analysis led you to reject the technique after concluding it required many more comparisons than the standard quicksort. I've only attempted very cursorily to reconcile the math; it seems that both your thesis and Yaroslavsky arrive at 0.8n(log n) as the average number of swaps, but you have (1/3) n(log n) as the average number of swaps in the classic algorithm while Yaroslavsky has 1.0 n(log n) for the same thing, so what he sees as an improvement you viewed as a disaster. My interpretation of this may be wrong, and I haven't attempted to verify either analysis yet."
The "classic algorithm" in the above is the classic quicksort as presented by Sedgewick in the thesis, which is slightly different than e.g. the standard Java implementation. I lost interest soon thereafter and stopped investigating this, but it shouldn't be too difficult to find out who was right in that algorithm comparison, if someone's interested enough.
It's entirely possible that a modification which was a significant loss in 1978 is a significant gain in 2008.
Bentley's comments from that email don't make any sense to me. I can't understand what he was thinking, unless he's just unusually prone to flattering other people's work (I don't know Bentley personally, but I've met other people who do this, much to my irritation).
Why isn't it used everywhere that people currently use Quicksort? I've never actually heard of anyone using Samplesort except in a parallel context.
Last I heard, python's built-in sort was a type of samplesort, though, so it isn't completely unused.
Python's built-in sort is timsort, last I heard, which is a variant of mergesort that looks for existing sorted or backwards runs in order to run faster in the common cases of mostly-sorted or mostly-backwards data. It's true that Python's built-in sort was a samplesort variant until 2002, though, which I didn't know.
(It does seem like Samplesort's virtue of minimizing comparisons would be particularly desirable in an environment like Python, where every comparison potentially involves a call into slow interpreted code.)
- From an input sequence of length n, choose a random subsequence of length k.
- Sort the subsequence.
- Partition the remaining (n-k) elements of the original sequence into the (k+1) partitions given by the subsequence.
- Sort the (k+1) partitions.
In the paper, the partitioning and sorting is always done using quicksort. Any n*log(n) sorting algorithm should work though, which includes using samplesort recursively.
If you use an optimal value for k (and this might be a significant fraction of n), you can prove that lim_{n->infty} E(C_n)/log(n!) = 1, where E(C_n) is the expected number of comparisons for a sequence of length n.
Now, using samplesort recursively, quicksort is samplesort with k = 1. Double-Pivot quicksort is samplesort with k = 2.
If, on the other hand, you have a fixed "corpus" and a dynamic query, O(n) search time (this algo, purportedly) is terrible.
"This algorithm especially well suited for long S, multi-substrings, small alphabet, regex/parsing and search for common substrings."
Long source text: the O(n times m) worst-case time per search kills it. Even for a single search, O(n times m) worst case here versus O(n + m) for suffix trees.
multi-substrings: suffix trees do each substring of length m in O(m), but are not compared.
search for common substrings: again, suffix trees would be more appropriate.
In most academic papers I have read that deal with online pattern matching suffix arrays/trees are generally not compared.
EDIT: Am I missing something???
Complexity analysis according to the author: m = search term, n = text
O(m) preprocessing -- that's right, O(search term). And O(n times m) worst-case query string search, so the worst case traverses the whole text.
Now compare that to suffix trees:
O(n) preprocessing O(m) string search
where worst case complexity is linear in search term.