
By now you've had a good deal of experience with straightforward recursive problems, and we hope you feel comfortable with them. In this chapter, we present some more challenging problems. But the same leap of faith method that we used for easier problems is still our basic approach.
Sort
First we'll consider the example of sorting a sentence. The argument will be any sentence; our procedure will return a sentence with the same words in alphabetical order.
> (sort '(i wanna be your man)) (BE I MAN WANNA YOUR)
We'll use the before?
primitive to decide if one word comes before
another word alphabetically:
> (before? 'starr 'best) #F
How are we going to think about this problem recursively? Suppose that
we're given a sentence to sort. A relatively easy subproblem is to find the
word that ought to come first in the sorted sentence; we'll write earliestword
later to do this.
Once we've found that word, we just need to put it in front of the sorted version of the rest of the sentence. This is our leap of faith: We're going to assume that we can already sort this smaller sentence. The algorithm we've described is called selection sort.
Another subproblem is to find the "rest of the sentence"—all the words
except for the earliest. But in Exercise 14.1 you wrote a function
removeonce
that takes a word and a sentence and returns the sentence
with that word removed. (We don't want to use remove
, which removes
all copies of the word, because our argument sentence might include the same
word twice.)
Let's say in Scheme what we've figured out so far:
(define (sort sent) ;; unfinished (se (earliestword sent) (sort (removeonce (earliestword sent) sent))))
We need to add a base case. The smallest sentence is ()
,
which is already sorted.
(define (sort sent) (if (empty? sent) '() (se (earliestword sent) (sort (removeonce (earliestword sent) sent)))))
We have one unfinished task: finding the earliest word of the argument.
(define (earliestword sent) (earliesthelper (first sent) (bf sent))) (define (earliesthelper sofar rest) (cond ((empty? rest) sofar) ((before? sofar (first rest)) (earliesthelper sofar (bf rest))) (else (earliesthelper (first rest) (bf rest)))))[1]
For your convenience, here's removeonce
:
(define (removeonce wd sent) (cond ((empty? sent) '()) ((equal? wd (first sent)) (bf sent)) (else (se (first sent) (removeonce wd (bf sent))))))
FromBinary
We want to take a word of ones and zeros, representing a binary number, and compute the numeric value that it represents. Each binary digit (or bit) corresponds to a power of two, just as ordinary decimal digits represent powers of ten. So the binary number 1101 represents (1×8)+(1×4)+(0×2)+(1×1) = 13. We want to be able to say
> (frombinary 1101) 13 > (frombinary 111) 7
Where is the smaller, similar subproblem? Probably the most obvious thing
to try is our usual trick of dividing the argument into its first
and
its butfirst
. Suppose we divide the binary number 1101
that
way. We make the leap of faith by assuming that we can translate the
butfirst, 101
, into its binary value 5. What do we have to add for
the leftmost 1
? It contributes 8 to the total, because it's three
bits away from the right end of the number, so it must be multiplied by
2^{3}. We could write this idea as follows:
(define (frombinary bits) ;; incomplete (+ (* (first bits) (expt 2 (count (bf bits)))) (frombinary (bf bits))))
That is, we multiply the first
bit by a power of two
depending on the number of bits remaining, then we add that
to the result of the recursive call.
As usual, we have written the algorithm for the recursive case before figuring out the base case. But it's pretty easy; a number with no bits (an empty word) has the value zero.[2]
(define (frombinary bits) (if (empty? bits) 0 (+ (* (first bits) (expt 2 (count (bf bits)))) (frombinary (bf bits)))))
Although this procedure is correct, it's worth noting that a more efficient
version can be written by dissecting the number from right to left. As
you'll see, we can then avoid the calls to expt
, which are expensive
because we have to do more multiplication than should be necessary.
Suppose we want to find the value of the binary number 1101
. The butlast
of this number, 110
, has the value six. To get the value of
the entire number, we double the six (because 1100
would have the value
12, just as in ordinary decimal numbers 430 is ten times 43) and then add the
rightmost bit to get 13. Here's the new version:
(define (frombinary bits) (if (empty? bits) 0 (+ (* (frombinary (bl bits)) 2) (last bits))))
This version may look a little unusual. We usually combine the value returned by the recursive call with some function of the current element. This time, we are combining the current element itself with a function of the recursive return value. You may want to trace this procedure to see how the intermediate return values contribute to the final result.
Mergesort
Let's go back to the problem of sorting a sentence. It turns out that sorting one element at a time, as in selection sort, isn't the fastest possible approach. One of the fastest sorting algorithms is called mergesort, and it works like this: In order to mergesort a sentence, divide the sentence into two equal halves and recursively sort each half. Then take the two sorted subsentences and merge them together, that is, create one long sorted sentence that contains all the words of the two halves. The base case is that an empty sentence or a oneword sentence is already sorted.
(define (mergesort sent) (if (<= (count sent) 1) sent (merge (mergesort (onehalf sent)) (mergesort (otherhalf sent)))))
The leap of faith here is the idea that we can magically mergesort
the halves of the sentence. If you try to trace this through step by step,
or wonder exactly what happens at what time, then this algorithm may be very
confusing. But if you just believe that the recursive calls will do
exactly the right thing, then it's much easier to understand this program.
The key point is that if the two smaller pieces have already been sorted,
it's pretty easy to merge them while keeping the result in order.
We still need some helper procedures. You wrote merge
in Exercise
14.15. It uses the following technique: Compare the first words of the
two sentences. Let's say the first word of the sentence on the left is
smaller. Then the first word of the return value is the first word of the
sentence on the left. The rest of the return value comes from recursively
merging the butfirst
of the left sentence with the entire right
sentence. (It's precisely the opposite of this if the first word of the
other sentence is smaller.)
(define (merge left right) (cond ((empty? left) right) ((empty? right) left) ((before? (first left) (first right)) (se (first left) (merge (bf left) right))) (else (se (first right) (merge left (bf right))))))
Now we have to write onehalf
and otherhalf
. One of the
easiest ways to do this is to have onehalf
return the elements in
oddnumbered positions, and have otherhalf
return the elements in
evennumbered positions. These are the same as the procedures odds
(from Exercise 14.4) and evens
(from Chapter 12).
(define (onehalf sent) (if (<= (count sent) 1) sent (se (first sent) (onehalf (bf (bf sent)))))) (define (otherhalf sent) (if (<= (count sent) 1) '() (se (first (bf sent)) (otherhalf (bf (bf sent))))))
Subsets
We're now going to attack a much harder problem. We want to know all the
subsets of the letters of a word—that is, words that can be formed
from the original word by crossing out some (maybe zero) of the letters. For
example, if we start with a short word like rat
, the subsets are r
, a
, t
, ra
, rt
, at
, rat
, and the empty
word (""
). As the word gets longer, the number of subsets gets bigger
very quickly.[3]
As with many problems about words, we'll try assuming that we can find the
subsets of the butfirst
of our word. In other words, we're hoping to
find a solution that will include an expression like
(subsets (bf wd))
Let's actually take a fourletter word and look at its subsets. We'll pick
brat
, because we already know the subsets of its butfirst
. Here
are the subsets of brat
:
"" b r a t br ba bt ra rt at bra brt bat rat brat
You might notice that many of these subsets are also subsets of rat
.
In fact, if you think about it, all of the subsets of rat
are
also subsets of brat
. So the words in (subsets 'rat)
are some
of the words we need for (subsets 'brat)
.
Let's separate those out and look at the ones left over:
rat subsets:
 ""  r
 a  t
 ra  rt
 at 
 rat

others:  b  br
 ba  bt
 bra  brt
 bat 
 brat

Right about now you're probably thinking, "They've pulled a rabbit out of a
hat, the way my math teacher always does." The words that aren't subsets of
rat
all start with b
, followed by something that is a
subset of rat
. You may be thinking that you never would have thought
of that yourself. But we're just following the method: Look at the smaller
case and see how it fits into the original problem. It's not so different
from what happened with downup
.
Now all we have to do is figure out how to say in Scheme, "Put a b
in front of every word in this sentence." This is a straightforward
example of the every
pattern:
(define (prependevery letter sent) (if (empty? sent) '() (se (word letter (first sent)) (prependevery letter (bf sent)))))
The way we'll use this in (subsets 'brat)
is
(prependevery 'b (subsets 'rat))
Of course in the general case we won't have b
and rat
in our
program, but instead will refer to the formal parameter:
(define (subsets wd) ;; first version (se (subsets (bf wd)) (prependevery (first wd) (subsets (bf wd)))))
We still need a base case. By now you're accustomed to the idea of using an empty word as the base case. It may be strange to think of the empty word as a set in the first place, let alone to try to find its subsets. But a set of zero elements is a perfectly good set, and it's the smallest one possible.
The empty set has only one subset, the empty set itself. What should
subsets
of the empty word return? It's easy to make a mistake here
and return the empty word itself. But we want subsets
to return a
sentence, containing all the subsets, and we should stick with returning a
sentence even in the simple case.[4] (This mistake would come from not thinking about
the range of our function, which is sentences. This is why we put
so much effort into learning about domains and ranges in Chapter
2.) So we'll return a sentence containing one (empty) word to
represent the one subset.
(define (subsets wd) ;; second version (if (empty? wd) (se "") (se (subsets (bf wd)) (prependevery (first wd) (subsets (bf wd))))))
This program is entirely correct. Because it uses two identical recursive
calls, however, it's a lot slower than necessary. We can use let
to
do the recursive subproblem only once:[5]
(define (subsets wd) (if (empty? wd) (se "") (let ((smaller (subsets (bf wd)))) (se smaller (prependevery (first wd) smaller)))))
We've already mentioned the need to be careful about the value returned
in the base case. The subsets
procedure is particularly errorprone
because the correct value, a sentence containing the empty word, is quite
unusual. An empty subset isn't the same as no subsets at all!
Sometimes you write a recursive procedure with a correct recursive case
and a reasonable base case, but the program still doesn't work. The trouble
may be that the base case doesn't quite catch all of the ways in which the
problem can get smaller. A second base case may be needed. For example, in
mergesort
, why did we write the following line?
(<= (count sent) 1)
This tests for two base cases, empty sentences and oneword
sentences, whereas in most other examples the base case is just an empty
sentence. Suppose the base case test were (empty? sent)
and suppose we
invoke mergesort
with a oneword sentence, (test)
. We would end
up trying to compute the expression
(merge (mergesort (onehalf '(test))) (mergesort (otherhalf '(test))))
If you look back at the definitions of onehalf
and
otherhalf
, you'll see that this is equivalent to
(merge (mergesort '(test)) (mergesort '()))
The first argument to merge
is the same expression we
started with! Here is a situation in which the problem doesn't get smaller
in a recursive call. Although we've been trying to avoid complicated base
cases, in this situation a straightforward base case isn't enough. To avoid
an infinite recursion, we must have two base cases.
Another example is the fib
procedure from Chapter . Suppose
it were defined like this:
(define (fib n) ;; wrong! (if (= n 1) 1 (+ (fib ( n 1)) (fib ( n 2)))))
It would be easy to make this mistake, because everybody knows
that in a recursion dealing with numbers, the base case is the smallest
possible number. But in fib
, each computation depends on two
smaller values, and we discover that we need two base cases.
The technique of recursion is often used to do something repetitively,
but don't get the idea that the word "recursion" means
repetition. Recursion is a technique in which a procedure invokes itself.
We do use recursion to solve repetitive problems, but don't confuse the
method with the ends it achieves. In particular, if you've programmed in
other languages that have specialpurpose looping mechanisms (the ones
with names like for
and while
), those aren't recursive.
Conversely, not every recursive procedure carries out a repetition.
15.1 Write a procedure tobinary
:
> (tobinary 9) 1001 > (tobinary 23) 10111
15.2 A "palindrome" is a sentence that reads the same backward as forward.
Write a predicate palindrome?
that takes a sentence as argument and
decides whether it is a palindrome. For example:
> (palindrome? '(flee to me remote elf)) #T > (palindrome? '(flee to me remote control)) #F
Do not reverse any words or sentences in your solution.
15.3 Write a procedure substrings
that takes a word as its
argument. It should return a sentence containing all of the substrings of
the argument. A substring is a subset whose letters come
consecutively in the original word. For example, the word bat
is a
subset, but not a substring, of brat
.
15.4 Write a predicate procedure substring?
that takes two words as
arguments and returns #t
if and only if the first word is a substring
of the second. (See Exercise 15.3 for the definition of a
substring.)
Be careful about cases in which you encounter a "false start," like this:
> (substring? 'ssip 'mississippi) #T
and also about subsets that don't appear as consecutive letters in the second word:
> (substring? 'misip 'mississippi) #F
15.5 Suppose you have a phone number, such as 2235766, and you'd like to figure out a clever way to spell it in letters for your friends to remember. Each digit corresponds to three possible letters. For example, the digit 2 corresponds to the letters A, B, and C. Write a procedure that takes a number as argument and returns a sentence of all the possible spellings:
> (phonespell 2235766) (AADJPMM AADJPMN …CCFLSOO)
(We're not showing you all 2187 words in this sentence.) You may assume there are no zeros or ones in the number, since those don't have letters.
Hint: This problem has a lot in common with the subsets example.
15.6 Let's say a gladiator kills a roach. If we want to talk about the roach, we say "the roach the gladiator killed." But if we want to talk about the gladiator, we say "the gladiator that killed the roach."
People are pretty good at understanding even rather long sentences as long as they're straightforward: "This is the farmer who kept the cock that waked the priest that married the man that kissed the maiden that milked the cow that tossed the dog that worried the cat that killed the rat that ate the malt that lay in the house that Jack built." But even a short nested sentence is confusing: "This is the rat the cat the dog worried killed." Which rat was that?
Write a procedure unscramble
that takes a nested
sentence as argument and returns a straightforward sentence about
the same cast of characters:
> (unscramble '(this is the roach the gladiator killed)) (THIS IS THE GLADIATOR THAT KILLED THE ROACH) > (unscramble '(this is the rat the cat the dog the boy the girl saw owned chased bit)) (THIS IS THE GIRL THAT SAW THE BOY THAT OWNED THE DOG THAT CHASED THE CAT THAT BIT THE RAT)
You may assume that the argument has exactly the structure of these examples, with no special cases like "that lay in the house" or "that Jack built."
accumulate
for this purpose:
(define earliestword sent) (accumulate (lambda (wd1 wd2) (if (before? wd1 wd2) wd1 wd2)) sent))
[2] A more straightforward base case would be a onebit number, but we've reduced that to this more elegant base case, following the principle we discussed on page there.
[3] Try writing down all the subsets of a fiveletter word if you don't believe us.
[4] We discussed this point in a pitfall in Chapter 12.
[5] How come we're worrying about
efficiency all of a sudden? We really did pull this out of a hat.
The thing is, it's a lot slower without the let
. Adding one
letter to the length of a word doubles the time required to find its
subsets; adding 10 letters multiplies the time by about 1000.
Brian Harvey,
bh@cs.berkeley.edu