Lec12

6.897: CSCI8980 Algorithmic Techniques for Big Data September 5, 2013
Lecture 1
Dr. Barna Saha Scribe: Vivek Mishra
Overview
We introduce data streaming model where streams of elements are coming in and main memory
space is not sufficient to hold all the data. We then look at the problem of finding frequent
items deterministically. This is a rare instance of data streaming algorithms that provides non-
trivial approximation guarantee deterministically. For most algorithms as we will see the answer
is approximate (close to optimal) and randomization is crucially used. To emphasize on the need
of randomization in designing data streaming algorithms, we next show that computing distinct
items (known as F0 computation) over a stream deterministically cannot achieve any approximation
without essentially storing the entire stream. Our next goal is to analyze the algorithm for distinct
items that have been covered in Lecture 1. Towards that goal, we study some basic concentration
inequality such as, Markov inequality, Chebyshev bound and The Chernoff bound.
1 Introduction to Data Streams
In a data streaming model sequence of elements a1 a2 a3 am arrive from a domain [1, n]. Each
element ai is a tuple (j, ν) where j ∈ [1, n] is an element from the domain and ν ∈ I. For simplicity,
we can consider ν = ±1 where +1 implies j is inserted in the stream and −1 implies j is deleted.
The goal is to process these elements using space (ideally) polylog of m and n, but definitely in
sub-linear in m and n. In the basic streaming setting, only a single pass over the data is allowed.
The time to process each update must be low.
When only insertions are allowed (all ν > 0), the model is known as cash-register model. On
the other hand, if both insertions and deletions are allowed, it is called turnstile model. For any
element, we generally do not allow the number of deletions (total negative frequency) to be more
than total number of insertions of that elements. However, if we do allow such scenarios, we will
refer it as general turnstile model.
2 Finding Frequent Items Deterministically
Here we describe an algorithm by Mishra and Gries [1] to find frequent items in the cash-register
model that is when only insertions are allowed. The precise problem is as follows.
Given a sequence of m elements from [1, n] and any k ∈ N, find all the elements with frequency
more than m
k where frequency simply implies the number of times an element occurs using space
O(k) (in words) and using only a single pass. Hence in bits, the total space usage is O(k log n).
We would like to have the following guarantees.
1

• No false negative. All items that have frequency > m
k will be reported.
There might be false positives, that is elements with frequency lower than m/k may be reported.
However, if we allow two passes, in the second pass all the elements reported in the ﬁrst pass can
be checked for actual frequency and any false positive can be eliminated.
Description of the Algorithm
• Data structure: An associative array of k − 1 elements initialized to empty. An associative
array contains in each of its cell a key (the item) and a value (its count) and may be maintained
in a balanced binary search tree by its key. Whenever we see an element ai = (j, 1) we can
search in the array if j exists or not in O(log k) time.
• Procedure: Given arrival of ai = (j, 1), we search if j exists in the associative array. If the
element is already there, we increment its count. If it is not there and there is an empty cell,
we store it and make its count 1. If no space is left and the element is not there, we do not
store the element and decrement the counters of all the stored elements by 1. If a counter
reaches 0, that element is dropped from the associative array and the cell becomes empty. In
the following code, we use A to represent the associative array, and give the update process
when ai is seen in the stream.
Algorithm 1 [Process ai = (j, 1)]
Search for j ∈ A
if j is already present as key at cell A[l] then
Count[l] = Count[l] + 1 {i}ncrement its count
else if j is not present in A and ∃l such that A[l] is empty then
Insert j as key to A[l]
Count[l] = 1
else
Drop j {comment: j is not present and there is no free space}
for i = 1 to k − 1 do
Count[i] = Count[i] − 1
if Count[i] == 0 then
Drop the key associated with A[i] and mark A[i] empty
end if
end for
end if
The entire algorithm for a given k ∈ N is as follows
2.1 Analysis
We now prove the following theorem.
Theorem 1. For any given k ∈ N Algorithm 2 returns all items with frequency more than m
k .
2

Algorithm 2 Mishra-Gries Algorithm
Initialize A of size k − 1 to empty
for i = 1 to m do
Process(ai)
end for
return All the elements in A
Proof. Clearly, the number of items having frequency more than m
k is at most k −1 because stream
size is m. Let fj denote the actual frequency of item j and let ˆfj be the frequency of the item j as
observed in A at the end of the processing. If j is not in A, we assume ˆfj = 0. We therefore return
all elements with ˆfj > 0.
Note that we increment frequency for j only if we see the actual item. Hence ˆfj ≤ fj. We want to
find out how low it can be. If we never drop the element then ˆfj = fj. Otherwise, either at some
point when j occurs, array A is full and j is not already there. Or because it is dropped from array
due to its count being decremented to 0.
Note that whenever an item is dropped (due to no space or count getting decremented to 0), there
are other distinct k − 1 elements whose count also gets decremented. Here we view the count of an
element on arrival is increased to 1, but it is dropped to 0 if it cannot be stored. Therefore, there
can be at most m
k steps on which element counts are decremented, k distinct elements at one shot.
The reasoning being the stream size is m and all frequencies are non-negative.
For each of these events, the difference between the computed frequency and the actual frequency
can increase by 1 and hence altogether we have ˆfj ≥ fj − m
k for all j ∈ [1, n].
Therefore, for all j ∈ [1, n] if fj > m
k then ˆfj > 0 and hence it is stored in the array at the end and
will be reported.
In most data streaming algorithms, one cannot achieve any non-trivial approximation determinis-
tically. In the following section, we come back to counting distinct items and show no deterministic
algorithm is possible that in o(m) space can give an exact count for distinct elements.
3 Lower Bound for Deterministic Computation of Distinct Ele-
ments
We prove the following theorem here.
Theorem 2. There exists no deterministic algorithm that returns the exact count for distinct items
in stream of size m in o(m) bits.
Proof. We will prove this by the by contradiction. Suppose it is possible to have an exact estimate
of distinct elements using space o(m) bits. Let R be such an algorithm. Since R uses o(m) bits,
the number of different possible configurations that R can maintain is at-most 2o(m). We let n = m
and consider all the different streams which have exactly m
2 distinct elements. How many such
3

streams are possible ? Clearly, the number of such streams is at least
m
m/2
≈
me
m/2
m
2
= (2e)
me
2 = 2Θ(m)
where the second inequality comes from Stirling’s approximation.
Since the number of such streams is more than the number of available configurations of R, there
must exist two streams y, y , y = y such that both have the same configuration, that is, R(y) =
R(y ) and the distinct items in y are not identical to distinct items of y . We now consider two
streams Y1 = y + y and Y2 = y + y where + represents concatenation here, that is stream Y1
y is followed by y and in stream Y2, y is followed by y. Since R(y) = R(y ), we must have
R(y + y) = R(y + y ). Therefore R will return same distinct elements counts for both Y1 and Y2
which is wrong because the number of distinct elements in Y1 is m
2 where for Y2 it is > m
2 .
The above proof can be extended to show that there does not exist any deterministic algorithm
with space o(m) that returns a count of distinct items within a multiplicative factor less than 2.
Similarly, one can show that no exact randomized algorithm can exist either in o(m) space.
Surprisingly, when we allow both approximation and randomization, the space usage can be dras-
tically reduced (next lecture).
4 Basic Concentration Inequalities
Here we study three basic concentration inequalities which bound deviation from expectation.
1. Markov inequality ( The 1st moment inequality)
2. Chebyshev inequality( The 2nd moment inequality)
3. The Chernoff Bound
Theorem 3 (Markov Bound). For any positive random variable X, and for any t > 0
Pr (X ≥ t) ≤
E[x]
t
(1)
Proof.
E[x] =
x
x · Pr(X = x)
=
x<t
x · Pr(X = x) +
x≥t
x · Pr(X = x)
≥ 0 + t ·
x≥t
Pr(X = x)
= t · P(X ≥ t)
4

Theorem 4 (Chebyshev Inequality). For any random variable X and for any t > 0
Pr(|X − E[x]| ≥ t) ≤
V ar(x)
t2
(2)
Proof.
Pr(|X − E[x]| ≥ t)
= Pr([X − E[x]]2
≥ t2
)
≤
E (X − E[x])2
t2
=
V ar(X)
t2
Theorem 5 (The Chernoff Bound). Let X1, X2...Xn be n independent Bernoulli random variables
with Pr(Xi = 1) = pi. Let X = Xi. Hence,
E[X] = E Xi = E [Xi] = Pr(Xi = 1) = pi = µ(say).
Then the Chernoff Bound says for any > 0
Pr(X > (1 + )µ) ≤
e
(1 + )
µ
and
Pr(X < (1 − )µ) ≤
e−
(1 − )1−
µ
When 0 < < 1 the above expression can be further simplified to
Pr(X > (1 + )µ) ≤ e
−µ 2
3 and
Pr(X < (1 − )µ) ≤ e
−µ 2
2
Hence
Pr(|X − µ| > µ) ≤ 2e
−µ 2
3
References
[1] Jayadev Misra and David Gries. Finding repeated elements. Sci. Comput. Program.,
2(2):143152, 1982.
5

Lec12

More Related Content

What's hot (20)

Viewers also liked (19)

Similar to Lec12 (20)

More from Atner Yegorov (20)

Recently uploaded (20)

Lec12