Skip to content
GitLab
Explore
Sign in
Primary navigation
Search or go to…
Project
D
ds2-notes
Manage
Activity
Members
Labels
Plan
Issues
Issue boards
Milestones
Code
Merge requests
Repository
Branches
Commits
Tags
Repository graph
Compare revisions
Deploy
Releases
Model registry
Analyze
Model experiments
Help
Help
Support
GitLab documentation
Compare GitLab plans
GitLab community forum
Contribute to GitLab
Provide feedback
Keyboard shortcuts
?
Snippets
Groups
Projects
Show more breadcrumbs
datovky
ds2-notes
Commits
defbcae6
Commit
defbcae6
authored
4 years ago
by
Parth Mittal
Browse files
Options
Downloads
Patches
Plain Diff
wrote count-min and the AMS estimator for distinct
parent
9744cd1b
No related branches found
No related tags found
No related merge requests found
Changes
1
Show whitespace changes
Inline
Side-by-side
Showing
1 changed file
streaming/streaming.tex
+130
-4
130 additions, 4 deletions
streaming/streaming.tex
with
130 additions
and
4 deletions
streaming/streaming.tex
+
130
−
4
View file @
defbcae6
...
@@ -43,7 +43,7 @@ us with a small set $C$ containing $F_k$, and hence lets us solve the frequent
...
@@ -43,7 +43,7 @@ us with a small set $C$ containing $F_k$, and hence lets us solve the frequent
elements problem efficiently.
elements problem efficiently.
\algo
{
FrequencyEstimate
}
\algalias
{
Misra/Gries Algorithm
}
\algo
{
FrequencyEstimate
}
\algalias
{
Misra/Gries Algorithm
}
\algin
the data stream
$
\alpha
$
, the target for the estimator
$
k
$
\algin
the data stream
$
\alpha
$
, the target for the estimator
$
k
$
.
\:\em
{
Init
}
:
$
A
\=
\emptyset
$
.
\cmt
{
an empty map
}
\:\em
{
Init
}
:
$
A
\=
\emptyset
$
.
\cmt
{
an empty map
}
\:\em
{
Process
}
(
$
x
$
):
\:\em
{
Process
}
(
$
x
$
):
\:
:If
$
x
\in
$
keys(
$
A
$
),
$
A
[
x
]
\=
A
[
x
]
+
1
$
.
\:
:If
$
x
\in
$
keys(
$
A
$
),
$
A
[
x
]
\=
A
[
x
]
+
1
$
.
...
@@ -93,12 +93,138 @@ $\vert C \vert = \vert$keys($A$)$\vert \leq k - 1$, and a key-value pair can
...
@@ -93,12 +93,138 @@ $\vert C \vert = \vert$keys($A$)$\vert \leq k - 1$, and a key-value pair can
be stored in
$
\O
(
\log
n
+
\log
m
)
$
bits.
be stored in
$
\O
(
\log
n
+
\log
m
)
$
bits.
\qed
\qed
\subsection
{
The Count-Min sketch
}
\subsection
{
The Count-Min Sketch
}
We will now look at a randomized streaming algorithm that solves the
frequency estimation problem. While this algorithm can fail with some
probability, it has the advantage that the output on two different streams
can be easily combined.
\algo
{
FrequencyEstimate
}
\algalias
{
Count-Min Sketch
}
\algin
the data stream
$
\alpha
$
, the accuracy
$
\varepsilon
$
,
the error parameter
$
\delta
$
.
\:\em
{
Init
}
:
$
C
[
1
\ldots
t
][
1
\ldots
k
]
\=
0
$
, where
$
k
\=
\lceil
2
/
\varepsilon
\rceil
$
and
$
t
\=
\lceil
\log
(
1
/
\delta
)
\rceil
$
.
\:
: Choose
$
t
$
independent hash functions
$
h
_
1
,
\ldots
h
_
t :
[
n
]
\to
[
k
]
$
, each
from a 2-independent family.
\:\em
{
Process
}
(
$
x
$
):
\:
:For
$
i
\in
[
t
]
$
:
$
C
[
i
][
h
_
i
(
x
)]
\=
C
[
i
][
h
_
i
(
x
)]
+
1
$
.
\algout
Report
$
\hat
{
f
}_
a
=
\min
_{
i
\in
t
}
C
[
i
][
h
_
i
(
a
)]
$
.
\endalgo
We will now look at a randomized streaming algorithm that performs the same task
Note that the algorithm needs
$
\O
(
tk
\log
m
)
$
bits to store the table
$
C
$
, and
$
\O
(
t
\log
n
)
$
bits to store the hash functions
$
h
_
1
,
\ldots
h
_
t
$
, and hence
uses
$
\O
(
1
/
\varepsilon
\cdot
\log
(
1
/
\delta
)
\cdot
\log
m
+
\log
(
1
/
\delta
)
\cdot
\log
n
)
$
bits. It remains to show that it computes
a good estimate.
\endchapter
\lemma
{
$
f
_
a
\leq
\hat
{
f
}_
a
\leq
f
_
a
+
\varepsilon
m
$
with probability
$
\delta
$
.
}
\proof
Clearly
$
\hat
{
f
}_
a
\geq
f
_
a
$
for all
$
a
\in
[
n
]
$
; we will show that
$
\hat
{
f
}_
a
\leq
f
_
a
+
\varepsilon
m
$
with probability at least
$
\delta
$
.
For a fixed element
$
a
$
, define the random variable
$$
X
_
i :
=
C
[
i
][
h
_
i
(
a
)]
-
f
_
a
$$
For
$
j
\in
[
n
]
\setminus
\{
a
\}
$
, define the
indicator variable
$
Y
_{
i, j
}
:
=
[
h
_
i
(
j
)
=
h
_
i
(
a
)
]
$
. Then we can see that
$$
X
_
i
=
\sum
_{
j
\neq
a
}
f
_
j
\cdot
Y
_{
i, j
}$$
Note that
$
\E
[
Y
_{
i, j
}
]
=
1
/
k
$
since each
$
h
_
i
$
is from a 2-independent family,
and hence by linearity of expectation:
$$
\E
[
X
_
i
]
=
{
\vert\vert
f
\vert\vert
_
1
-
f
_
a
\over
k
}
=
{
\vert\vert
f
_{
-
a
}
\vert\vert
_
1
\over
k
}$$
And by applying Markov's inequality we obtain a bound on the error of a single
counter:
$$
\Pr
[
X
_
i >
\varepsilon
\cdot
m
]
\geq
\Pr
[
X
_
i >
\varepsilon
\cdot
\vert\vert
f
_{
-
a
}
\vert\vert
_
1
]
\leq
{
1
\over
k
\varepsilon
}
\leq
1
/
2
$$
Finally, since we have
$
t
$
independent counters, the probability that they
are all wrong is:
$$
\Pr\left
[
\bigcap
_
i X
_
i >
\varepsilon
\cdot
m
\right
]
\leq
1
/
2
^
t
\leq
\delta
$$
\qed
\section
{
Counting Distinct Elements
}
We continue working with a stream
$
\alpha
[
1
\ldots
m
]
$
of integers from
$
[
n
]
$
,
and define
$
f
_
a
$
(the frequency of
$
a
$
) as before. Let
$
d
=
\vert
\{
j : f
_
j >
0
\}
\vert
$
. Then the distinct elements problem is
to estimate
$
d
$
.
\subsection
{
The AMS Algorithm
}
Define
${
\tt
tz
}
(
x
)
:
=
\max\{
i
\mid
2
^
i
$
~divides~
$
x
\}
$
(i.e. the number of trailing zeroes in the base-2 representation of
$
x
$
).
\algo
{
DistinctElements
}
\algalias
{
AMS
}
\algin
the data stream
$
\alpha
$
, the accuracy
$
\varepsilon
$
,
the error parameter
$
\delta
$
.
\:\em
{
Init
}
: Choose a random hash function
$
h :
[
n
]
\to
[
n
]
$
from a 2-independent
family.
\:
:
$
z
\=
0
$
.
\:\em
{
Process
}
(
$
x
$
):
\:
:If
${
\tt
tz
}
(
h
(
x
))
> z
$
:
$
z
\=
{
\tt
tz
}
(
h
(
x
))
$
.
\algout
$
\hat
{
d
}
\=
2
^{
z
+
1
/
2
}$
\endalgo
\lemma
{
The AMS algorithm is a
$
(
3
,
\delta
)
$
-estimator for some constant
$
\delta
$
.
}
\proof
For
$
j
\in
[
n
]
$
,
$
r
\geq
0
$
, let
$
X
_{
r, j
}
:
=
[
{
\tt
tz
}
(
h
(
j
))
\geq
r
]
$
, the
indicator that is true if
$
h
(
j
)
$
has at least
$
r
$
trailing
$
0
$
s.
Now define
$$
Y
_
r
=
\sum
_{
j : f
_
j >
0
}
X
_{
r, j
}
$$
How is our estimate related to
$
Y
_
r
$
? If the algorithm outputs
$
\hat
{
d
}
\geq
2
^{
a
+
1
/
2
}$
, then we know that
$
Y
_
a >
0
$
. Similarly, if the
output is smaller than
$
2
^{
a
+
1
/
2
}$
, then we know that
$
Y
_
a
=
0
$
. We will now
bound the probabilities of these events.
For any
$
j
\in
[
n
]
$
,
$
h
(
j
)
$
is uniformly distributed over
$
[
n
]
$
(since
$
h
$
is
$
2
$
-independent). Hence
$
\E
[
X
_{
r, j
}
]
=
1
/
2
^
r
$
. By linearity of
expectation,
$
\E
[
Y
_{
r
}
]
=
d
/
2
^
r
$
.
We will also use the variance of these variables -- note that
$${
\rm
Var
}
[
X
_{
r, j
}
]
\leq
\E
[
X
_{
r, j
}^
2
]
=
\E
[
X
_{
r, j
}
]
=
1
/
2
^
r
$$
And because
$
h
$
is
$
2
$
-independent, the variables
$
X
_{
r, j
}$
and
$
X
_{
r, j'
}$
are independent for
$
j
\neq
j'
$
, and hence:
$${{
\rm
Var
}}
[
Y
_{
r
}
]
=
\sum
_{
j : f
_
j >
0
}
{
\rm
Var
}
[
X
_{
r, j
}
]
\leq
d
/
2
^
r
$$
Now, let
$
a
$
be the smallest integer such that
$
2
^{
a
+
1
/
2
}
\geq
3
d
$
. Then we
have:
$$
\Pr
[
\hat
{
d
}
\geq
3
d
]
=
\Pr
[
Y
_
a >
0
]
=
\Pr
[
Y
_
a
\geq
1
]
$$
Using Markov's inequality we get:
$$
\Pr
[
\hat
{
d
}
\geq
3
d
]
\leq
\E
[
Y
_
a
]
=
{
d
\over
2
^
a
}
\leq
{
\sqrt
{
2
}
\over
3
}
$$
For the other side, let
$
b
$
be the smallest integer so that
$
2
^{
b
+
1
/
2
}
\leq
d
/
3
$
. Then we have:
$$
\Pr
[
\hat
{
d
}
\leq
d
/
3
]
=
\Pr
[
Y
_{
b
+
1
}
=
0
]
\leq
\Pr
[
\vert
Y
_{
b
+
1
}
-
\E
[
Y
_{
b
+
1
}
]
\vert
\geq
d
/
2
^{
b
+
1
}
]
$$
Using Chebyshev's inequality, we get:
$$
\Pr
[
\hat
{
d
}
< d
/
3
]
\leq
{{
\rm
Var
}
[
Y
_
b
]
\over
(
d
/
2
^{
b
+
1
}
)
^
2
}
\leq
{
2
^{
b
+
1
}
\over
d
}
\leq
{
\sqrt
{
2
}
\over
3
}$$
\qed
The previous algorithm is not particularly satisfying -- by our analysis it
can make an error around
$
94
\%
$
of the time (taking the union of the two bad
events). However we can improve the success probability easily; we run
$
t
$
independent estimators simultaneously, and print the median of their outputs.
By a standard use of Chernoff Bounds one can show that the probability that
the median is more than
$
3
d
$
is at most
$
2
^{
-
\Theta
(
t
)
}$
(and similarly also
the probability that it is less than
$
d
/
3
$
).
Hence it is enough to run
$
\O
(
\log
(
1
/
\delta
))
$
copies of the AMS estimator
to get a
$
(
3
,
\delta
)
$
estimator for any
$
\delta
>
0
$
. Finally, we note that
the space used by a single estimator is
$
\O
(
\log
n
)
$
since we can store
$
h
$
in
$
\O
(
\log
n
)
$
bits, and
$
z
$
in
$
\O
(
\log
\log
n
)
$
bits, and hence a
$
(
3
,
\delta
)
$
estimator uses
$
\O
(
\log
(
1
/
\delta
)
\cdot
\log
n
)
$
bits.
\endchapter
This diff is collapsed.
Click to expand it.
Preview
0%
Loading
Try again
or
attach a new file
.
Cancel
You are about to add
0
people
to the discussion. Proceed with caution.
Finish editing this message first!
Save comment
Cancel
Please
register
or
sign in
to comment