Programming Homework Help

Problem 1. Word-Search: Pattern Matching Revisited

In recitation we covered an O(n lg n) FFT-based algorithm for finding the
offset in the text that gives the best match score between a pattern string
and a target text. For this problem, we are interested in finding the first
exact occurrence of the pattern in the text. Let t = t1t2 . . . tn be a target
text and p = p1p2 . . . pm a pattern, both over alphabet Σ = {0, 1} with m ≤
n. Identifying the first exact occurrence of the pattern in the text amounts to
finding the smallest j ∈ {1, 2, . . . , n−m+1} such that for 1 ≤ i ≤ m, it
holds that tj+i−1 = pi .

In this problem we will work w Z ith the word RAM model, where we can
store each integer in a single word (or “by Z te”) of computer memory. The
number of bits that can be stored in a single word is ⌈lg n ⌉. So, arithmetic
operations on words or O(lg n)-bit numbers take O(1) time. Finally, with
this model comparing two n-bit numbers takes O(n) time.

For this problem you can assume that you are provided with a black-box
prime number generator, and that it takes O(1) time to sample O(lg K)-bit
primes, for K polynomial in n. Also, you can assume that π(k) ∼ k , where
π(k) be the number of distinct primes less than k.
(a) Define fp : {0, 1} m → as fp(X) := g(X) mod p, where p is a prime and
g : {0, 1} m → is a function that converts an m-bit binary string to a
corresponding base 2 integer. Note that if X is equal to Y then fp(X) = fp(Y
). However, if X differs from Y then it can still be the case that fp(X) =
fp(Y ). We will refer to cases where the results of evaluating the function
on two different string inputs are equal as false positives

© 2022 All Rights Reserved. Programming Homework Help

Take the set P = {p1, p2 . . . pt}, where pi are all primes less than some
large integer K. Suppose we choose a prime p uniformly at random from the
set P and take m-bit strings X and Y such that X = Y . Prove that we can
bound the probability of a false positive as follows:

Hint: Consider the prime factorization of |g(X) − g(Y )|. Notice that the
number of prime factors is at most m.

Solution:

A false positive happens when m-bit strings X = Y but fp(X) = fp(Y ). This
implies that the binary integers g(X) and g(Y ) are the same modulo p, g(X)
− g(Y ) ≡ 0

mod p. This will happen if and only if p is a prime factor of g(X) − g(Y ).
Thus, we should count the number of distinct prime factors of g(X)−g(Y ).
Let g(X)−g(Y ) = q1q2 . . . qc where qi are its (not necessarily distinct)
prime factors. Clearly, qi ≥ 2, since 2 is the smallest prime. Thus, g(X) −
g(Y ) ≥ 2 c . Since g(X), g(Y ) ≤ 2 m, c < m. Hence, g(X) − g(Y ) can have
at most m prime factors, which is also the maximum number of distinct
prime factors it can have. When we sample a random prime from the set P =
p1, p2, . . . pt , the probability of us finding a prime factor of g(X) − g(Y )
and obtaining a false positive is at most:

© 2022 All Rights Reserved. Programming Homework Help

Note that the above bound is very loose, we did not even use the distinctness of the
prime factors, we can achieve a slightly better bound if we take that into
consideration. Suppose g(X)−g(Y ) = q s1 1 q s2 2 . . . qsc c . Where q1, q2, . . . are
distinct. We have g(X)− g(Y ) > q1q2 . . . qc > 2 × 3 × 4 × . . . × c = c!. Hence the
bound on c is: c! < 2 m. Using Sterling’s approximation, log(c!) = Θ(c log c), so we
have c log c < Θ(m), this implies c = O(m/ log(m)), because then c log c = m m (log
m − log log m) = m. log Using this upperbound, the probability of a false positive is:

An even tighter bound is possible, however, they do not improve the asymptotic
running time of our algorithm in the subsequent parts.
(b) Let X(j) be a length-m substring of the target text that starts at position j, for a given
j ∈ {1, 2, . . . , n − m + 1}. Design a randomized algorithm that given fp(Y ) and
fp(X(j)) determines if there is a match between the pattern and the target text for a
given offset j ∈ {1, 2, . . . , n − m + 1}.
Solution:

Recall that a Monte Carlo algorithm is guaranteed to terminate quickly on any input,
however it has a probability of returning the wrong result. Whereas a Las Vegas
algorithm is guaranteed to be correct, but only terminates in some expected time, ie. in
some cases it can take a long time. For this problem, we will present both types of
algorithms. Monte Carlo version:

© 2022 All Rights Reserved. Programming Homework Help

If: fp(Y ) = fp(X(j)), return “A match” Else: return “Not a match” Analysis: Notice that
when fp(Y ) = fp(X(j)), X = Y , so when our algorithm yields a negative result, it is
always correct. If our algorithm yields a positive result, as we have concluded in part
(a), the probability that it is a false positive is at most m/

Thus, the overall error probability of our algorithm is at most m/t. The algorithm only
compares two numbers fp(Y ) and fp(X(j)), which are both at most p − 1. If we can
bound p to be poly(n,m), then it only takes O(log n) bits to represent both fp(Y ) and
fp(X(j)), and comparing them would take O(1) time in our model. Hence we have a
Monte Carlo algorithm that runs in O(1) time with an error probability of at most m/t.

Las Vegas version:

If: fp(Y ) = fp(X(j)), check the strings g(X(j)) and g(Y ), if they are equal, return “A
match”

Else: return “Not a match”

Analysis: The difference between this algorithm and the previous one is that when
the finger prints fp(Y ), fp(X(j)) are the same, we must make sure the strings g(Y )
and g(X(j)) are actually the same. There is no other way to guarantee this other than
to check them directly. Since g(Y ) and g(X(j)) are m-bit numbers, this takes O(m)
time.

© 2022 All Rights Reserved. Programming Homework Help

If g(Y ) and g(X(j)) are actually the same, any algorithm must take O(m) to verify
this with 100% certainty (since we need to at least look at all the bits). So in this
sense, our randomized algorithm isn’t doing any worse. If g(Y ) and g(X(j)) are not
the same, from part (a), we know that the probability of a false positive is m/t, which
means that our algorithm will compare two m-bit strings in O(m) time but
eventually conclude that they are not the same. Thus, the total expected running
time when g(Y ) = g(X(j)) is O(1 + m2 /t). This will be O(1) if we let K be, for
example, Θ(m3 ), in which case, by the prime number theorem, t = Θ(m3/ log(m)).
Notice that fp(Y ) and fp(X(j)) are still O(log n) in this case, so comparing them still
takes O(1) time. Thus, this is an algorithm that runs in expected O(1) time, and
always outputs the correct answer. It turns out that the Las Vegas version is better
suited for the subsequent parts.

(c) Design a formula that given g(X(j)) computes g(X(j+1)), where X(j) is a length-m
substring of the target text that starts at position j, for a given j ∈ {1, 2, . . . , n−m+1}.
Use it to compute fp(X(j + 1)) from fp(X(j)). Note that the formula should depend on
X, j, and m.

Solution: We regard g(X(j)) as an m-bit binary string with the left most bit being the
most significant, ie. g(X(j)) = 2m−1Xj+2m−2Xj+1+. . . 2 0Xj+m−1. g(X(j+1)) is
basically g(X(j)) shifted right by 1 digit, discarding the original most significant bit Xj
in the process, while adding a new least significant bit Xj+m. Hence we have the
following relation between g(X(j)) and g(X(j + 1)):
so

© 2022 All Rights Reserved. Programming Homework Help

 We only need to compute 2 m−1 mod p once and save it for all future uses,
subtracting this constant takes O(1) time in our model as long as p is poly(n,m).
Multiplying by 2 is equivalent to adding two log p bit numbers, which can be done in
O(1) time. Finally, adding Xj+m clearly takes constant time. Thus it takes O(1) time
to compute fp(X(j + 1)) from fp(X(j)).

(d) Suppose that X(j) and Y differ at every string position. Give the best upper bound
you can on the expected number of positions j such that fp(X(j)) = fp(Y ), where j ∈ {1,
2, . . . , n − m + 1}

Solution: Let Ck be the indicator variable that fp(X(k)) = fp(Y ). Then, the total number
of false positives F P for all the positions is:

By linearity of expectation, this is equal to:

Since we are told that g(X(j)) = g(Y ) for all j, using the result of part (a), E[Ci ] ≤
m/t for every i. Hence,

© 2022 All Rights Reserved. Programming Homework Help

(e) Using parts above, design a randomized algorithm that determines if there is a
match between a pattern and a target text in O(n+m) expected running time. The
algorithm should always return the correct answer.

Solution: We will determine K (the upper bound on p) at the end, for now we will
treat it as a variable. Our algorithm works as follows: 1. sample a random prime
from the range 1 to K. This can be accomplished using either the blackbox described
in the introduction of the problem in O(1) time. Or we can do it ourselves as
follows:

First we generate a random number from 1 to K, not necessarily prime, this should
take O(1) time in our model if K is poly(n,m). Typically, pseudo random numbers
are generated with “seeds” using modular arithmetic, but these can all be done in
O(1) time. We note here that deterministic primality testing algorithms exist that run
in polylog(n) time. That is to say, given any number less than n, we can determine
whether it is prime or composite in polylog(n) time. Thus, to pick a random prime
number from the set, we random sample a number, use the deterministic primality
testing algorithm, and repeat until we obtain a prime number. From the prime
density theorem, the probability of us finding a prime number is at least π(K)/K = 1/
log(K). So we need to try expected O(log(K)) times before we actually find a prime
number. This brings our total runtime to log(K) × polylog(K) = polylog(K). 2.
Compute fp(Y ) and fp(X(1)). Both of them are m-bit numbers, arithmetic
operations will take O(m) time.

3. Using fp(X(1)), compute all fp(X(j)) using the formula from part (c). We argued
that each additional value only takes O(1) time to compute, and there are O(n − m +
1) = O(n) of them in total, so this step takes O(n).

© 2022 All Rights Reserved. Programming Homework Help

4. Compare fp(Y ) to each of fp(X(j)) computed above, using the Las Vegas version
of the algorithm in part (b). Notice that the algorithm terminates as soon as we
encounter the first exact match. Thus, evevn if the original text contains many exact
matches, we will spend at most O(m) time on a position where g(X(j)) = g(Y ). For
all other cases, we may then assume g(X(j)) = g(Y ). We can then use the result from
part (d). The total work to check is O(n) +E[F P]×O(m). Since for every false 2
positive, we need to do O(m) work. This gives us a run time of O(n + nm t ).
Correctness: It is quite clear that the result of the algorithm is always correct, since
we used the Las Vegas version of the algorithm in part (b). Whenever the algorithm
returns a match at position k, we know that it is indeed a match because we checked
the m-bit strings directly. Likewise, whenever the algorithm reports no match, it is
also guaranteed to be correct. Thus, the first instance of a positive result is the
position of the first exact match. 2

Run time analysis: The total runtime of the algorithm is O(polylog(n)+m+n+ nm t )

2 O(m + n + nm log(K) or K ). We want to make the term O(nm2 log(K)/K) small.
For example, we can pick K = Ω(nm2 log(nm2 )), so that O(nm2 log(K)/K) = O(1),
and hence arriving at a O(n + m) algorithm. In fact, we can make K arbitrarily large,
as long as K is poly(n,m), so that operations on log(K) bit numbers can be done in
O(1) time.

(f) Provide a bound for the probability that the running time is more than 100 times
the expected running time.
Solution: Let T be the random variable that describes the running time of the
algorithm. By Markov’s inequality, we have

© 2022 All Rights Reserved. Programming Homework Help