Well, I didn’t think I’d write another Fibonacci A Day post — but there’s a great fibonacci function that’s defined right in the Scala documents, in their discussion of Streams.

A stream is a standard Scala data structure which is very list like. Recall that a Scala list has two parts, a head and a tail; the head is a value, and the tail is either the special empty list (Nil), or another list. For example, the list created by List(“cat”, “mouse”, “cheese”) is a list whose head is “cat”, and whose tail is a list whose head is “mouse” and whose tail is a list whose head is “cheese” and whose tail is Nil (the cheese does not quite stand alone).

Elements in a stream (the values in their heads), on the other hand, are only evaluated when they are needed. For example, getting the nth value is a stream causes the steam to evaluate its first element, its second element, etc, up to n. Once evaluated, they are remembered (that is, memoized) so they the next time the element is requested, the value has already been computed (the lookup is still O(n), but no computation is done).

Here’s a definition of a fibonacci stream, pretty much as given in the documents

lazy val _fibs: Stream[BigInt] = 
  0 #:: 1 #:: (_fibs zip _fibs.tail map { case (x, y) ⇒ x + y })

This defines an “infinite” list of fibonacci numbers — infinite in the sense that the list can be as long as we have memory for. Note that Scala allows us to have recursive data structures as well as recursive functions. Here the tail of the _fibs stream (after the 2nd element, anyway) is defined in turns of itself — we zip the stream with the tail of the stream, and then each element is the sum of the two previous elements — just as we had defined.

We can access this list in a similar way to our previous functions to get positive and negative fibonacci numbers:

def stream_fibonacci(n: Int) = n match {
    case _ if (n >= 0)     ⇒ _fibs(n) // natural numbers 
    case _ if (n % 2 == 0) ⇒ -_fibs(-n) // even, negative numbers*/
    case _                 ⇒ _fibs(-n) // odd, negative numbers 
  }

How does compare in speed with our previous functions? Calling stream_fibonacci(1000) the first time results in zipping though the stream and calculating the sums. Somewhat surprisingly, this is not more expensive than our fastest version. Afterwards, though, we just “look up” the 1001st element, which requires traversing the stream’s first 1001 elements. But this is very fast, and is about 100x times faster on my machine (see the updated benchmarks).

Of course, this comes at the expense of memory storage for the 1001 fibonacci numbers in the stream. Depending on the application, this could be just fine. Of course, fibonacci numbers get big fast, so maintaining a large stream of them might not be good (the other versions only require two or so BigInts to be held in memory during the computation).

I am taking Martin Odensky’s Functional Programming in Scala class, and discovered (or rediscovered) that the Fibonacci function is discussed in SICP (Structure and Interpretation of Computer Programs, the book that Odensky seems to be basing much of his Functional Programming in Scala course on).

Even more interesting is that a Fibonacci function which runs in O(lg(n)) steps is presented — that is, it takes about lg(n) steps to compute the nth Fibonacci number, instead of n steps. It’s given as an exercise (Exercise 1.19). Bill the Lizard has been working through the SICP exercises, and his explanation of Exercise 1.19 is the clearest I’ve found. The code below might not make much sense without looking at the SICP chapter or his explanation, but here is the code in Scala, and put in the form that I’ve been using — this computes fibonacci(n) for any integer (including negative integers). Running the benchmark indicates that this version a lot faster than the straight tail recursive version.

  def fibonacci(n: Int): BigInt = {
    val limit = if (n < 0) -n else n
    /* the fibonacci function per se — on natural numbers */
    def fib(m: Int): BigInt = {
      @tailrec def recurse(a: BigInt, b: BigInt, p: BigInt, q: BigInt, count: Int): BigInt =
        if (count == 0) b
        else if (count % 2 == 0) {
          val qq = q * q
          recurse(a, b, (p * p + qq), (2 * p * q) + qq, count / 2)
        } else {
          val aq = a * q
          recurse(b * q + aq + a * p, b * p + aq, p, q, count - 1)
        }
      (m: @switch) match {
        case 0 ⇒ BigInt(0)
        case 1 ⇒ BigInt(1)
        case _ ⇒ recurse(1, 0, 0, 1, m)
      }
    }
    n match {
      case _ if (n >= 0)     ⇒ fib(n) // natural numbers 
      case _ if (n % 2 == 0) ⇒ -fib(-n) // even, negative numbers*/
      case _                 ⇒ fib(-n) // odd, negative numbers
    }
  }

The goal of this series was to come up with a general, efficient fibonacci function that goes beyond the typical computer science introduction. The version we developed is fast, efficient with respect to memory, and general. I’d be glad to hear comments.

The Series:

1. A Fibonacci A Day — Introduction
2. A Fibonacci A Day — Recursive Version
3. A Fibonacci A Day — Special Pi Day Post
4. A Fibonacci A Day — Iterative Version
5. A Fibonacci A Day — Out of Bounds
6. A Fibonacci A Day — Speedy Recursion
7. A Fibonacci A Day — Golden Ration Version
8. A Fibonacci A Day — Memoization
9. A Fibonacci A Day — Benchmarking
10. A Fibonacci A Day — Extending to Negative Numbers
11. A Fibonacci A Day — Fibonacci in O(lg(n)) Steps
12. A Fibonacci A Day — Using Streams and a Fast Memoization Method

We are now confident that the fast recursive version is the best of the algorithms we have explored — it is fast, and can handle large values of the Fibonacci function.

I haven’t been happy, though, with returning a negative -1 when the function is called with a negative number — this feels like a kind of silent error we were trying to avoid in going to BigInts. Fortunately, it turns out that there is a natural extension of the Fibonacci sequence to negative numbers. If the nth Fibonacci number is defined as F(n) = F(n-2) + F(n-1), then we can rearrange this to F(-n) = (-1)ⁿF(n).

So, this is the final version of the Fibonacci function, that supports any integer, up to the the memory size of the computer, computed quite quickly. Much better than our simple recursive or iterative version, which were ripe for error and inefficiency.

 def fibonacci(n: Int): BigInt = {
    val limit = if (n < 0) -n else n
    /* the fibonacci function per se — on natural numbers */
    def fib(m: Int): BigInt = {
      @tailrec def recurse(i: Int, u: BigInt, v: BigInt): BigInt =
        if (i == limit) v
        else recurse(i + 1, v, u + v)
      (m: @switch) match {
        case 0 ⇒ BigInt(0)
        case 1 ⇒ BigInt(1)
        case _ ⇒ recurse(1, 0, 1)
      }
    }
    n match {
      case _ if (n >= 0)     ⇒ fib(n) // natural numbers 
      case _ if (n % 2 == 0) ⇒ -fib(-n) // even, negative numbers*/
      case _                 ⇒ fib(-n) // odd, negative numbers
    }
  }

I hope you’ve enjoyed these series.

The following table describes the limits of the various Fibonacci algorithms I’ve created during this series. I ran them through a very simple benchmarking routine, and empirically tested them for the maximum size they can handle. Each benchmark was run 1000 times, except the first algorithm, which was run just once — we already know it runs too slow. These were run on a MacBook Air (Late 2010).

Log Recursive (BigInt)0.0082Unlimited (except for memory)

Here is a table comparing the versions which allow really big fibonacci numbers, calculating F(20000) 100 times each.

F(20000) is a really big integer, and a lot of this time is creating and destoying the memory associated with creating the integers. The thing to notice is that the iterative and the tail recursive versions take approximately the same amount of time — the recursive version actually runs a little more quickly than the iterative version. And the Log recursive version of the same is two orders of magnitude faster — — the stream version is yet another order faster, but it may use too much memory. Note that even the first use is faster than the iterative and tail recursive versions!

Still, there’s really no reason not to use the BigInt, log recursive version of this function — it’s both fast and accurate.

(April 2, 2013: Updated to include log recursive version.

(April 13, 2013: Updated to include < a href="stream”>http://willwhim.wpengine.com/2013/04/13/a-fibonacci-a-day-using-streams-and-a-fast-memoization-method/”>stream version; rerun benchmarks for F(20000); and update conclusions.