Let's say I have a string, "hello", and I want to generate a character frequency map:
Map[Char,Int] = Map(h -> 1, e -> 1, o -> 1, l -> 2)   I could do this iteratively:
val str = "hello" var counts = new scala.collection.mutable.HashMap[Char,Int] for (i <- str) {     if (counts.contains(i))         counts.put(i, counts(i) + 1)     else         counts.put(i, 1) }   By messing around in the REPL, I've found I can do something a bit more concise and not using a mutable collection:
> str.groupBy(_.toChar).map{ p => (p._1, p._2.length)} scala.collection.immutable.Map[Char,Int] = Map(h -> 1, e -> 1, o -> 1, l -> 2)   But I don't know about the performance characteristics of groupBy() nor what is going on in the block passed to map (like what, exactly, p is).
How do I do this idiomatically using the functional paradigms in Scala?
For background, I'm just coming to Scala for the first time from Ruby. In Ruby, I would use inject but I'm not sure what the parallel way to do it in Scala is:
counts = str.each_byte.inject(Hash.new(0)){ |h, c| h[c] += 1; h} 
                p mean?groupBy takes a function which maps an elements to a key of type K. When invoked on some collection Coll, it returns a Map[K, Coll] which contains mappings from keys K to all the elements which mapped to the same key.
So, in your case, str.groupBy(_.toChar) yields a map mapping from a key k (which is a character) to a string with all the elements (characters) c such that k == c.toChar. You get this:
Map(e -> "e", h -> "h", l -> "ll", o -> "o")   A Map is an iterable of pairs of keys and values. In this case, each pair is a character and a string of elements. Calling the map operation on a Map involves mapping on these pairs - p is a pair where p._1 is a character, and p._2 is the associated string (on which you can call length, as you did above).
The above is how to do it idiomatically - using groupBy and map. Alternatively, you can use an immutable map and recursion on the string length to compute the frequencies, or an immutable map and a foldLeft.
Best to benchmark to see the differences. Here are a couple of microbenchmark for a highly-repetitive string (~3GHz iMac, JDK7, Scala 2.10.0 nightly):
object Imperative extends testing.Benchmark {   val str = "abc" * 750000    def run() {     var counts = new scala.collection.mutable.HashMap[Char,Int]     var i = 0     val until = str.length     while (i < until) {       var c = str(i)       if (counts.contains(c))         counts.put(c, counts(c) + 1)       else         counts.put(c, 1)       i += 1     }      //println(f)   } }   object Combinators extends testing.Benchmark {   val str = "abc" * 750000    def run() {     val f = str.groupBy(_.toChar).map(p => (p._1, p._2.length))   } }   object Fold extends testing.Benchmark {   val str = "abc" * 750000    def run() {     val f = str.foldLeft(Map[Char, Int]() withDefaultValue 0){(h, c) => h.updated(c, h(c)+1)}   } }   Results:
Imperative: $    103 57  53  58  53  53  53  53  53  53
Combinators: $   72  51  63  56  53  52  52  54  53  53
Fold: $  163 62  71  62  57  57  57  58  57  57
Note that changing the imperative version to use withDefaultValue:
var counts = new scala.collection.mutable.HashMap[Char,Int].withDefaultValue(0) var i = 0 val until = str.length while (i < until) {   var c = str(i)   counts.put(c, counts(c) + 1)   i += 1 }   is apparently terribly slow due to forwarding each put call:
withDefaultValue: $    133 87  109 106 101 100 101 100 101 101 Conclusion: the boxing and unboxing of characters in this case is high-enough so that the differences in performance between these approaches are hard to observe.
EDIT:
Update: You may want to use ScalaMeter inline benchmarking in place of the Benchmark trait.
If you love us? You can donate to us via Paypal or buy me a coffee so we can maintain and grow! Thank you!
Donate Us With