Monte Carlo Algorithms: Shusen Wang

Monte Carlo Algorithms
Shusen Wang
Application 1: Calculating Pi
Calculating Pi
• We already know 𝜋 ≈ 3.141592653589 …

• Pretend we do not know the value of 𝜋.
• Can we find it out (approximately) using a random number generator?
Calculating Pi
𝑦
• Assume 𝑥, 𝑦 is a point sampled from
1
the square uniformly at random.
−1
1
𝑥
−1
Calculating Pi
𝑦
1
• What is the probability that 𝑥, 𝑦 is in
−1 the circle?
1
𝑥
−1
Calculating Pi
𝑦
1
−1 the circle?
𝑥
1 • Area of the square is 𝐴2 = 24 = 4.
• Area of the circle is 𝐴4 = 𝜋𝑟 4 = 𝜋.
78 :
−1 • Probability: 𝑃 = = .
79 ;
Calculating Pi
• Suppose 𝑛 points are uniformly sampled from −1, 1 × −1, 1 .
:>
• Then, in expectation, 𝑃𝑛 = points are in the circle.
;
𝑦
1
−1 the circle?
𝑥
1 • Area of the square is 𝐴2 = 24 = 4.
• Area of the circle is 𝐴4 = 𝜋𝑟 4 = 𝜋.
78 :
−1 • Probability: 𝑃 = = .
79 ;
Calculating Pi
:>
;
𝑦
1
• Given a point 𝑥, 𝑦 , how do you know
whether 𝑥, 𝑦 is in the circle?
−1 • If 𝑥 4 + 𝑦 4 ≤ 1, then it is in the circle.
1
𝑥
−1
Calculating Pi
:>
;
𝑦
1
• We found 𝑚 points in the circle.
:>
• If 𝑛 is big, then 𝑚 ≈ .
−1 ;
1
𝑥 • Thus, 𝜋 ≈
;B
.
>
−1
Calculating Pi
:>
;
𝑦
1 • Law of large numbers:
;B
→ 𝜋, as 𝑛 → ∞.
>
−1
1
𝑥
−1
Calculating Pi
:>
;
𝑦
1 • Law of large numbers:
;B
→ 𝜋, as 𝑛 → ∞.
>
−1
𝑥
1 • Concentration bound:
;B 2
−𝜋 =𝑂 .
> >
−1
Calculating Pi
;B
𝑛 points are sampled from the square; 𝑚 are in the circle. Then 𝜋 ≈ .
>
𝑦
1
−1
1
𝑥
−1
Calculating Pi
;B
>
𝑦 Algorithm
1
1. User specifies a big 𝑛; reset
counter 𝑚 = 0.
−1
1
𝑥
−1
Calculating Pi
;B
>
𝑦 Algorithm
1
1. User specifies a big 𝑛; reset
counter 𝑚 = 0.
−1 2. For 𝑖 = 1 to 𝑛:
1
𝑥
a) Randomly generate 𝑥 ∈ −1, 1 .
b) Randomly generate 𝑦 ∈ −1, 1 .
c) If 𝑥 4 + 𝑦 4 ≤ 1, then 𝑚 ← 𝑚 + 1.
−1 ;B
3. Return 𝜋 ≈ .
>
Application 2: Buffon's Needle Problem
Buffon's Needle Problem
Buffon, 1707 – 1788 Buffon's Needle Problem

French scientist
• The parallel lines have distance 𝑑.

• Needles have length 𝑙.
𝑑

𝑙 • Needles have length 𝑙.
needle
• Randomly throw a needle; the needle
may or may not lie across a line.
𝑑

needle
𝑑
needle

needle
4M
With probability 𝑃 = , they are across.
:N
𝑑
needle
• It can be proved using integral.
𝑙
needle
4M
:N
𝑑
needle
4M
:N
𝑙 • Randomly throw a total of 𝑛 needles.

needle
4M>
• In expectation, 𝑃𝑛 = needles lie
:N
across the lines.
𝑑
needle
4M
:N

needle
4M>
:N
across the lines.
• Actually observe 𝑚 needles across the
lines.
𝑑
4 M >
needle • If 𝑛 is big, 𝑚 ≈ .
: N
4M
:N

needle
4M>
:N
across the lines.
• Actually observe 𝑚 needles across the
lines.
𝑑
4 M >
needle • If 𝑛 is big, 𝑚 ≈ .
: N
4 M >
• Thus, 𝜋 ≈ .
N B
Researcher Year 𝑛= 𝑚= Estimate of 𝛑

Wolf 1850 5000 2532 3.1596
Smith 1855 3204 1218 3.1554
De Morgan 1860 600 382 3.137
Fox 1884 1030 489 3.1595
Lazzerini 1901 3408 1808 3.1415929
Reina 1925 2520 859 3.1795
Application 3: Area of A Region
What is the area of the grey region?
𝑦
2
2
𝑥
1
𝑦
2
2
𝑥
1
𝑦
• If a point 𝑥, 𝑦 is in the circle, it
2 must satisfy
4 4
𝑥−1 + 𝑦−1 ≤ 1.
1
𝑟=1
2
𝑥
1
𝑦
• If a point 𝑥, 𝑦 is in the circle, it
2 must satisfy
4 4
𝑥−1 + 𝑦−1 ≤ 1.
1
• If a point 𝑥, 𝑦 is in the quarter
circle, it must satisfy
𝑥 4 + 𝑦 4 ≤ 24 .
2
𝑥
1
𝑦
2 • A point 𝑥, 𝑦 in the grey region
satisfies both of
1. 𝑥−1 4 + 𝑦−1 4 ≤ 1,
2. 𝑥 4 + 𝑦 4 > 24.
1
2
𝑥
1
𝑦
2 • A point 𝑥, 𝑦 in the grey region
satisfies both of
1. 𝑥−1 4 + 𝑦−1 4 ≤ 1,
2. 𝑥 4 + 𝑦 4 > 24.
1
2
𝑥
1
𝑦
2 • Area of the square: 𝐴2 = 24 = 4.
2
𝑥
1
𝑦
2 • Area of the square: 𝐴2 = 24 = 4.
• Area of the grey region: 𝐴4 .
• A point uniformly sampled from the
square falls in the grey region w.p.
1
78 78
𝑃= = .
79 ;
2
𝑥
1
𝑦 • Uniformly sample 𝑛 points from the
square 0, 2 × 0, 2 .
2 > 78
• In expectation, 𝑛𝑃 = points fall
;
in the grey region.
2
𝑥
1
𝑦 • Uniformly sample 𝑛 points from the
square 0, 2 × 0, 2 .
2 > 78
• In expectation, 𝑛𝑃 = points fall
;
in the grey region.
• We actually observe 𝑚 points in the
1 grey region.
> 78
• If 𝑛 is big, then 𝑚 ≈ .
;
;B
• Thus, 𝐴4 ≈ .
>
2
𝑥
1
𝑦 Algorithm
2 1. User specify a big 𝑛; reset
counter: 𝑚 = 0.
2. For 𝑖 = 1 to 𝑛:
a) Randomly generate 𝑥 ∈ 0, 2 .
b) Randomly generate 𝑦 ∈ 0, 2 .
1 c) If both of the following conditions
are satisfied, then 𝑚 ← 𝑚 + 1:
i. 𝑥−1 4+ 𝑦−1 4 ≤ 1.
ii. 𝑥 4 + 𝑦 4 > 24 .
;B
3. Return area = .
>
2
𝑥
1
𝑦 Algorithm
counter: 𝑚 = 0.
2. For 𝑖 = 1 to 𝑛:
i. 𝑥−1 4+ 𝑦−1 4 ≤ 1.
ii. 𝑥 4 + 𝑦 4 > 24 .
;B
3. Return area = .
>
2
𝑥
1
𝑦 Algorithm
counter: 𝑚 = 0.
2. For 𝑖 = 1 to 𝑛:
i. 𝑥−1 4+ 𝑦−1 4 ≤ 1.
ii. 𝑥 4 + 𝑦 4 > 24 .
;B
3. Return area ≈ .
>
2
𝑥
1
Application 4: Integration
Integration
2
• We are given a function, e.g., 𝑓 𝑥 = .
2WXYZ [ ⋅ ]^_` [ 8
c
• Calculate the integral: 𝐼 = ∫d.e 𝑓 𝑥 𝑑𝑥.
Integration
2
• We are given a function, e.g., 𝑓 𝑥 = .
2WXYZ [ ⋅ ]^_` [ 8
c
• Calculate the integral: 𝐼 = ∫d.e 𝑓 𝑥 𝑑𝑥.
• If 𝑓 𝑥 is very involved, there is no way to analytically calculate the
integral.
• Using Monte Carlo to approximate the integral.
Monte Carlo Integration (Univariate)
f
Task: Given a univariate function 𝑓 𝑥 , calculate 𝐼 = ∫g 𝑓 𝑥 𝑑𝑥.
f
1. Draw 𝑛 samples from 𝑎, 𝑏 uniformly at random; denote them by

𝑥2 , ⋯ , 𝑥> .
2
2. Calculate 𝑄> = 𝑏 − 𝑎 ⋅ ∑>mn2 𝑓 𝑥m .
>
f
3. Return 𝑄> as an approximation to the integral 𝐼 = ∫g 𝑓 𝑥 𝑑𝑥.
f

𝑥2 , ⋯ , 𝑥> .
2
>
f
f

𝑥2 , ⋯ , 𝑥> .
2
>
f
f

𝑥2 , ⋯ , 𝑥> .
2
>
f
Theory: Law of large numbers guarantees 𝑄> → 𝐼, as 𝑛 → ∞.

Monte Carlo Integration (Univariate): Example
2 c
Task: Given function 𝑓 𝑥 = , calculate 𝐼 = ∫d.e 𝑓 𝑥 𝑑𝑥.
2WXYZ [ ⋅ ]^_` [ 8
2 c
2WXYZ [ ⋅ ]^_` [ 8
1. Draw 𝑛 samples from 0.8, 3 uniformly at random; denote them by

𝑥2 , ⋯ , 𝑥> .
2 c
2WXYZ [ ⋅ ]^_` [ 8

𝑥2 , ⋯ , 𝑥> .
2
2. Calculate 𝑄> = 2.2 ⋅ ∑>mn2 𝑓 𝑥m .
>
2 c
2WXYZ [ ⋅ ]^_` [ 8

𝑥2 , ⋯ , 𝑥> .
2
2. Calculate 𝑄> = 2.2 ⋅ ∑>mn2 𝑓 𝑥m .
>
c
3. Return 𝑄> as an approximation to the integral 𝐼 = ∫d.e 𝑓 𝑥 𝑑𝑥.
Monte Carlo Integration (Multivariate)
Task: Given a multivariate function 𝑓 𝐱 , calculate 𝐼 = ∫p 𝑓 𝐱 𝑑𝐱.
𝐱 ∈ ℝN is a vector Ω is a subset of ℝN
1. Draw 𝑛 samples from set Ω uniformly at random; denote them by

𝐱2 , ⋯ , 𝐱 > .
2. Calculate 𝑉 = ∫p 𝑑𝐱.
1. Draw 𝑛 samples from set Ω uniformly at random; denote them by

𝐱2 , ⋯ , 𝐱 > .
2. Calculate 𝑉 = ∫p 𝑑𝐱.
2
3. Calculate 𝑄> = 𝑉 ⋅ ∑>mn2 𝑓 𝐱 m .
>
4. Return 𝑄> as an approximation to the integral 𝐼 = ∫p 𝑓 𝐱 𝑑𝐱.
Monte Carlo Integration: Bivariate Example
𝑦
1, if 𝑥 4 + 𝑦 4 ≤ 1;
1 • Let 𝑓 𝑥, 𝑦 = t
0, otherwise.
−1
1
𝑥
−1
𝑦
1, if 𝑥 4 + 𝑦 4 ≤ 1;
1 • Let 𝑓 𝑥, 𝑦 = t
0, otherwise.
• Let Ω = −1, 1 × −1, 1 .
−1
1
𝑥 • What is 𝐼 = ∫p 𝑓 𝑥, 𝑦 𝑑𝑥 𝑑𝑦?
−1
𝑦
1, if 𝑥 4 + 𝑦 4 ≤ 1;
1 • Let 𝑓 𝑥, 𝑦 = t
0, otherwise.
• Let Ω = −1, 1 × −1, 1 .
−1
1
𝑥 • What is 𝐼 = ∫p 𝑓 𝑥, 𝑦 𝑑𝑥 𝑑𝑦?
• 𝐼 is the area of the circle:
𝐼 = 𝜋𝑟 4 = 𝜋.
−1
1. Draw 𝑛 samples from set Ω uniformly

𝑦 at random; denote them by
1
𝑥2 , 𝑦2 , ⋯ , 𝑥> , 𝑦> .
−1
1
𝑥
−1

1
𝑥2 , 𝑦2 , ⋯ , 𝑥> , 𝑦> .
2. Calculate 𝑉 = ∫p 𝑑𝑥 𝑑𝑦 = 4. (It is the
−1 area of set Ω.)
1
𝑥
−1

1
𝑥2 , 𝑦2 , ⋯ , 𝑥> , 𝑦> .
1
𝑥
2
3. Calculate 𝑄> = 𝑉 ⋅ ∑>mn2 𝑓 𝑥m , 𝑦m .
>
−1

1
𝑥2 , 𝑦2 , ⋯ , 𝑥> , 𝑦> .
1
𝑥
2
3. Calculate 𝑄> = 𝑉 ⋅ ∑>mn2 𝑓 𝑥m , 𝑦m .
>
4. Return 𝑄> as an approximation to the
−1
integral 𝜋 = ∫p 𝑓 𝑥, 𝑦 𝑑𝑥 𝑑𝑦.
Application 5: Estimate of Expectation
Expectation
• Let 𝑋 be a 𝑑-dimensional random vector.

• Let 𝑝 𝐱 be a probability density function (PDF).
Expectation

• Property: ∫ℝ~ 𝑝 𝐱 𝑑𝐱 = 1.
Expectation

2
• E.g., PDF of uniform distribution is 𝑝 𝑥 = , for 𝑥 ∈ 0, 𝑡 .
•
Expectation

2
• E.g., PDF of uniform distribution is 𝑝 𝑥 = , for 𝑥 ∈ 0, 𝑡 .
•
2 [„… 8
• E.g., PDF of univariate Gaussian is 𝑝 𝑥 = exp − .
• 4: 4• 8
Expectation

• Let 𝑓 𝐱 be any function of vector variable.
• Expectation: 𝔼‡~‰ 𝑓 𝑋 = ∫ℝ~ 𝑓 𝐱 ⋅ 𝑝 𝐱 𝑑𝐱.
Monte Carlo Estimate of Expectation
Task: Estimate the expectation 𝔼‡~‰ 𝑓 𝑋 = ∫ℝ~ 𝑓 𝐱 ⋅ 𝑝 𝐱 𝑑𝐱.

1. Draw 𝑛 random samples from the probability distribution 𝑝(𝐱);

denote them by 𝐱2 , ⋯ , 𝐱 > .

2
2. Calculate 𝑄> = ∑>mn2 𝑓 𝐱 m .
>

2
2. Calculate 𝑄> = ∑>mn2 𝑓 𝐱 m .
>
3. Return 𝑄> as an approximation to 𝔼‡~‰ 𝑓 𝑋 .
Monte Carlo and Beyond
Monte Carlo
• The term “Monte Carlo method"

was firstly introduced in 1947 by
Nicholas Metropolis.
Reference
• Metropolis. The beginning of the Monte

Carlo method. Los Alamos Science, 125–130,
Casino de Monte-Carlo, Monaco 1987.
• Monte Carlo refers to algorithms that rely on repeated random

sampling to obtain numerical results.
• The output of Monte Carlo algorithms can be incorrect.
• In all of our examples, the algorithms’ outputs are incorrect.
• But they are close to the correct solution.
• Monte Carlo refers to algorithms that rely on repeated random

sampling to obtain numerical results.
• The output of Monte Carlo algorithms can be incorrect.
• Las Vegas algorithms are those always produce the correct answers.
• E.g., random quicksort.
• Atlantic City algorithms are polynomial-time randomized algorithms
that answer correctly w.p. greater than 75%.
Question
𝑦
2
2
𝑥
0 1
What are the areas of the red and blue regions?
𝑦
0
𝑥
1 2
Thank you!

Monte Carlo Algorithms: Shusen Wang

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Monte Carlo Algorithms: Shusen Wang

Uploaded by

Copyright:

Available Formats

Monte Carlo Algorithms

• We already know 𝜋 ≈ 3.141592653589 …

Buffon, 1707 – 1788 Buffon's Needle Problem

• The parallel lines have distance 𝑑.

• The parallel lines have distance 𝑑.

• The parallel lines have distance 𝑑.

• The parallel lines have distance 𝑑.

𝑙 • Randomly throw a total of 𝑛 needles.

𝑙 • Randomly throw a total of 𝑛 needles.

𝑙 • Randomly throw a total of 𝑛 needles.

Researcher Year 𝑛= 𝑚= Estimate of 𝛑

1. Draw 𝑛 samples from 𝑎, 𝑏 uniformly at random; denote them by

1. Draw 𝑛 samples from 𝑎, 𝑏 uniformly at random; denote them by

1. Draw 𝑛 samples from 𝑎, 𝑏 uniformly at random; denote them by

1. Draw 𝑛 samples from 𝑎, 𝑏 uniformly at random; denote them by

Theory: Law of large numbers guarantees 𝑄> → 𝐼, as 𝑛 → ∞.

1. Draw 𝑛 samples from 0.8, 3 uniformly at random; denote them by

1. Draw 𝑛 samples from 0.8, 3 uniformly at random; denote them by

1. Draw 𝑛 samples from 0.8, 3 uniformly at random; denote them by

Task: Given a multivariate function 𝑓 𝐱 , calculate 𝐼 = ∫p 𝑓 𝐱 𝑑𝐱.

Task: Given a multivariate function 𝑓 𝐱 , calculate 𝐼 = ∫p 𝑓 𝐱 𝑑𝐱.

1. Draw 𝑛 samples from set Ω uniformly at random; denote them by

Task: Given a multivariate function 𝑓 𝐱 , calculate 𝐼 = ∫p 𝑓 𝐱 𝑑𝐱.

1. Draw 𝑛 samples from set Ω uniformly at random; denote them by

1. Draw 𝑛 samples from set Ω uniformly

1. Draw 𝑛 samples from set Ω uniformly

1. Draw 𝑛 samples from set Ω uniformly

1. Draw 𝑛 samples from set Ω uniformly

• Let 𝑋 be a 𝑑-dimensional random vector.

• Let 𝑋 be a 𝑑-dimensional random vector.

• Let 𝑋 be a 𝑑-dimensional random vector.

• Let 𝑋 be a 𝑑-dimensional random vector.

• Let 𝑋 be a 𝑑-dimensional random vector.

Task: Estimate the expectation 𝔼‡~‰ 𝑓 𝑋 = ∫ℝ~ 𝑓 𝐱 ⋅ 𝑝 𝐱 𝑑𝐱.

Task: Estimate the expectation 𝔼‡~‰ 𝑓 𝑋 = ∫ℝ~ 𝑓 𝐱 ⋅ 𝑝 𝐱 𝑑𝐱.

1. Draw 𝑛 random samples from the probability distribution 𝑝(𝐱);

Task: Estimate the expectation 𝔼‡~‰ 𝑓 𝑋 = ∫ℝ~ 𝑓 𝐱 ⋅ 𝑝 𝐱 𝑑𝐱.

1. Draw 𝑛 random samples from the probability distribution 𝑝(𝐱);

Task: Estimate the expectation 𝔼‡~‰ 𝑓 𝑋 = ∫ℝ~ 𝑓 𝐱 ⋅ 𝑝 𝐱 𝑑𝐱.

1. Draw 𝑛 random samples from the probability distribution 𝑝(𝐱);

• The term “Monte Carlo method"

• Metropolis. The beginning of the Monte

• Monte Carlo refers to algorithms that rely on repeated random

• Monte Carlo refers to algorithms that rely on repeated random

You might also like