Two Big Questions: P vs. NP and P vs. BPPcs.indstate.edu/~jkinne/talks/isu-job-talk.pdf · Conjecture: P6= NP Techniques: hierarchy theorems, others Is randomness powerful? P vs

Two Big Questions Derandomization Hierarchy Theorems

Two Big Questions:P vs. NP and P vs. BPP

Jeff Kinne

University of Wisconsin-Madison

Indiana State University, March 26, 2010

1 / 35


Computational Complexity Theory

How much time, memory space, etc. are needed to solveproblems?

Is nondeterminism powerful? P vs. NP

Conjecture: P 6= NP

Techniques: hierarchy theorems, others

Is randomness powerful? P vs. BPP, L vs. BPL

Conjecture: P=BPP, L=BPL

My work: derandomization, hierarchy theorems

2 / 35




Is nondeterminism powerful?

P vs. NP

Conjecture: P 6= NP





2 / 35





Conjecture: P 6= NP





2 / 35





Conjecture: P 6= NP





2 / 35





Conjecture: P 6= NP





2 / 35





Conjecture: P 6= NP


Is randomness powerful?

P vs. BPP, L vs. BPL



2 / 35





Conjecture: P 6= NP





2 / 35





Conjecture: P 6= NP





2 / 35





Conjecture: P 6= NP





2 / 35


Introducing two Big Questions

3 / 35


Time Complexity

Decision Problem: yes/no questions

4 / 35


Time Complexity

Poly Time (P)

Exp Time (EXP)


4 / 35


Time Complexity

Poly Time (P)

Exp Time (EXP)

factoring,NP-complete sorting

shortest path


4 / 35


Time Complexity

Poly Time (P)

Exp Time (EXP)

sorting?

factoring,NP-complete

shortest path


4 / 35


Time Complexity

Poly Time (P)

Exp Time (EXP)

sorting?

factoring,NP-complete

shortest path


4 / 35


Memory Space Complexity

Amount of working space memory needed

5 / 35



Log Space (L)

Poly Space (PSPACE)


5 / 35



Log Space (L)

Poly Space (PSPACE)


5 / 35



Log Space (L)

Poly Space (PSPACE)

factoring,NP-complete arithmetic


5 / 35



Log Space (L)

Poly Space (PSPACE)

arithmetic?factoring,

NP-complete


5 / 35


Randomized Algorithm

Bounded error: Correct with probability > 99%

6 / 35



01 10 10 1011 01 10

yes/no


6 / 35



Random Strings

good

bad

01 10 10 1011 01 10

yes/no


6 / 35



Random Strings

good

bad

01 10 10 1011 01 10

yes/no


6 / 35


Polynomial Identity Testing

p(x) = x3 · (3x − x2)2 − x4 · (2x3 + 5x) + x · (4x2 − x)3

Do all terms cancel?

7 / 35



p(x) = x3 · (3x − x2)2 − x4 · (2x3 + 5x) + x · (4x2 − x)3


7 / 35



p(x) = x3 · (3x − x2)2 − x4 · (2x3 + 5x) + x · (4x2 − x)3


7 / 35



p(x) = x3 · (3x − x2)2 − x4 · (2x3 + 5x) + x · (4x2 − x)3


7 / 35



p(x) = x3 · (3x − x2)2 − x4 · (2x3 + 5x) + x · (4x2 − x)3


7 / 35


Multi-variate Polynomial Identity Testing

p(x1, x2, x3, x4) = x54 · (x1 − x2 − x3)

30 + (x4 + x3 − x1)15 ·

(x3 − x2 + x1)20 − (x2 − x3 + x4)

20 · (x2 + x1)15



Pick point (x1, ..., x4), each xi ∈R S

p non-zero, degree d ⇒ Pr[p(x1, ..., x4) = 0] ≤ d|S |

8 / 35



p(x1, x2, x3, x4) = x54 · (x1 − x2 − x3)

30 + (x4 + x3 − x1)15 ·

(x3 − x2 + x1)20 − (x2 − x3 + x4)

20 · (x2 + x1)15





8 / 35



p(x1, x2, x3, x4) = x54 · (x1 − x2 − x3)

30 + (x4 + x3 − x1)15 ·

(x3 − x2 + x1)20 − (x2 − x3 + x4)

20 · (x2 + x1)15





8 / 35



p(x1, x2, x3, x4) = x54 · (x1 − x2 − x3)

30 + (x4 + x3 − x1)15 ·

(x3 − x2 + x1)20 − (x2 − x3 + x4)

20 · (x2 + x1)15





8 / 35



p(x1, x2, x3, x4) = x54 · (x1 − x2 − x3)

30 + (x4 + x3 − x1)15 ·

(x3 − x2 + x1)20 − (x2 − x3 + x4)

20 · (x2 + x1)15





8 / 35



p(x1, x2, x3, x4) = x54 · (x1 − x2 − x3)

30 + (x4 + x3 − x1)15 ·

(x3 − x2 + x1)20 − (x2 − x3 + x4)

20 · (x2 + x1)15




p non-zero, degree d

⇒ Pr[p(x1, ..., x4) = 0] ≤ d|S |

8 / 35



p(x1, x2, x3, x4) = x54 · (x1 − x2 − x3)

30 + (x4 + x3 − x1)15 ·

(x3 − x2 + x1)20 − (x2 − x3 + x4)

20 · (x2 + x1)15





8 / 35



Random Strings

good

bad

01 10 10 1011 01 10

yes/no

p(x1, x2, x3, x4)

x1, ..., x4

BPP: Bounded-error Probabilistic Poly time , BPL: log space

P?= BPP

9 / 35



Random Strings

good

bad

01 10 10 1011 01 10

yes/no

p(x1, x2, x3, x4)

x1, ..., x4

BPP: Bounded-error Probabilistic Poly time

, BPL: log space

P?= BPP

9 / 35



Random Strings

good

bad

01 10 10 1011 01 10

yes/no

p(x1, x2, x3, x4)

x1, ..., x4


P?= BPP

9 / 35



Random Strings

good

bad

01 10 10 1011 01 10

yes/no

p(x1, x2, x3, x4)

x1, ..., x4


P?= BPP

9 / 35


Nondeterministic Algorithm

10 / 35


Graph 3-Coloring

11 / 35


Graph 3-Coloring

11 / 35



NP: Nondeterministic Polynomial time

“NP-Complete” problems: 3-Coloring, TSP, Knapsack, ...

P?= NP

NP?= coNP

12 / 35



01 10 10 1011 01 10

yes/no

Proof/Certificate

graph

coloring



P?= NP

NP?= coNP

12 / 35



Potential Proofs

good

bad

01 10 10 1011 01 10

yes/no

Proof/Certificate

graph

coloring



P?= NP

NP?= coNP

12 / 35



Potential Proofs

good

bad

01 10 10 1011 01 10

yes/no

Proof/Certificate

graph

coloring



P?= NP

NP?= coNP

12 / 35



Potential Proofs

good

bad

01 10 10 1011 01 10

yes/no

Proof/Certificate

graph

coloring



P?= NP

NP?= coNP

12 / 35



Potential Proofs

good

bad

01 10 10 1011 01 10

yes/no

Proof/Certificate

graph

coloring



P?= NP

NP?= coNP

12 / 35



Potential Proofs

good

bad

01 10 10 1011 01 10

yes/no

Proof/Certificate

graph

coloring



P?= NP NP

?= coNP

12 / 35


Two Big Questions

P?= NP

Is finding proofs as easy as verifying them?

Is 3-coloring in Polynomial Time?

P?= BPP

Does randomness truly add power?

13 / 35


Two Big Questions

P?= NP

Is finding proofs as easy as verifying them?

Is 3-coloring in Polynomial Time?

P?= BPP

Does randomness truly add power?

13 / 35





Conjecture: P 6= NP





14 / 35


Derandomization

15 / 35


Naive Derandomization

Random Strings

good

bad

01 10 10 1011 01 10

yes/no

p(x1, x2, x3, x4)

x1, ..., x4

Try all possible random bit strings – exponentially many

16 / 35



Random Strings

good

bad

01 10 10 1011 01 10

yes/no

p(x1, x2, x3, x4)

x1, ..., x4


16 / 35



Random Strings

good

bad

01 10 10 1011 01 10

yes/no

p(x1, x2, x3, x4)

x1, ..., x4

Try all possible random bit strings

– exponentially many

16 / 35



Random Strings

good

bad

01 10 10 1011 01 10

yes/no

p(x1, x2, x3, x4)

x1, ..., x4


16 / 35


Derandomization – the Standard PRG Approach

Poly many strings to try

⇒ O(log n) seed, exp stretch

17 / 35



Random Strings

good

bad

01 10 10 1011 01 10

yes/no

p(x1, x2, x3, x4)

x1, ..., x4



17 / 35



Random Strings

good

bad

pseudo–random

01 10 10 1011 01 10

yes/no

p(x1, x2, x3, x4)

x1, ..., x4



17 / 35



Random Strings

good

bad

pseudo–random

01 10 10 1011 01 10

yes/no

PRG

p(x1, x2, x3, x4)

x1, ..., x4



17 / 35



Random Strings

good

bad

pseudo–random

01 10 10 1011 01 10

yes/no

PRG

p(x1, x2, x3, x4)

x1, ..., x4



17 / 35



Random Strings

good

bad

pseudo–random

01 10 10 1011 01 10

yes/no

PRG

p(x1, x2, x3, x4)

x1, ..., x4

Poly many strings to try ⇒ O(log n) seed, exp stretch

17 / 35



Random Strings

good

bad

pseudo–random

01 10 10 1011 01 10

yes/no

PRGhard

functionh

p(x1, x2, x3, x4)

x1, ..., x4

Poly many strings to try ⇒ O(log n) seed, exp stretch

17 / 35


A New Approach – Typically-Correct Derandomization

Seed length n, poly stretch

18 / 35



Random Strings

good

bad

pseudo–random

01 10 10 1011 01 10

yes/no

PRG

p(x1, x2, x3, x4)

x1, ..., x4


18 / 35



Random Strings

good

bad

pseudo–random

01 10 10 1011 01 10

yes/no

PRG

11 01 10

p(x1, x2, x3, x4)

x1, ..., x4


18 / 35



Random Strings

good

bad

pseudo–random

01 10 10 1011 01 10

yes/no

PRG

11 01 10

p(x1, x2, x3, x4)

x1, ..., x4


18 / 35


Standard Use of PRG’s vs. Typ-Correct

Standard Derandomization Typ-Correct Derandomization[Nisan & Wigderson, ...] [Kinne, Van Melkebeek, Shaltiel]

• Always correct • Small # mistakes• Run PRG many times • Run PRG only once• Need exponential stretch • Need only poly stretch• Conditional results • Unconditional results:

fast parallel time, streaming,communication protocols

19 / 35






19 / 35




• Always correct • Small # mistakes

• Run PRG many times • Run PRG only once• Need exponential stretch • Need only poly stretch• Conditional results • Unconditional results:


19 / 35




• Always correct • Small # mistakes• Run PRG many times • Run PRG only once

• Need exponential stretch • Need only poly stretch• Conditional results • Unconditional results:


19 / 35




• Always correct • Small # mistakes• Run PRG many times • Run PRG only once• Need exponential stretch • Need only poly stretch

• Conditional results • Unconditional results:fast parallel time, streaming,communication protocols

19 / 35






19 / 35





Conjecture: P 6= NP





20 / 35


Hierarchy Theorems

21 / 35


Hierarchy Theorems

Fix a model of computing(deterministic, randomized, nondeterministic)

Can we achieve more given more resources?

My work: hierarchy theorems for randomized algorithms

22 / 35


Hierarchy Theorems

Fix a model of computing

(deterministic, randomized, nondeterministic)



22 / 35


Hierarchy Theorems




22 / 35


Hierarchy Theorems




22 / 35


Hierarchy Theorems



n time


22 / 35


Hierarchy Theorems



n time

n2 time


22 / 35


Hierarchy Theorems



n time

n2 timen3 time


22 / 35


Hierarchy Theorems



n time

n2 time

Poly Timen3 time


22 / 35


Hierarchy Theorems



n timePoly time =


22 / 35


Hierarchy Theorems



log n space

log2 n space

log3 n space


22 / 35


Hierarchy Theorems



log n spacelog3 n space =


22 / 35


Hierarchy Theorems



log n space

log2 n space

log3 n space


22 / 35


Hierarchy Theorems




22 / 35


Hierarchy Theorems




22 / 35





Conjecture: P 6= NP





23 / 35


Hierarchy Theorems for Deterministic Algorithms

24 / 35



A2 A3 ...A1

All Algorithms time n

24 / 35



A2 A3 ...

x1

x2

x3

...

A1

All Algorithms

AllI

nput

s

time n

24 / 35



A2 A3 ...

A1(x1) A2(x1) A3(x1)

A1(x2) A2(x2) A3(x2) ...

A1(x3) A2(x3) A3(x3)

x1

x2

x3

...

A1

All Algorithms

AllI

nput

s

...

time n

24 / 35



A2 A3 ... D

A1(x1) A2(x1) A3(x1)

A1(x2) A2(x2) A3(x2) ...

A1(x3) A2(x3) A3(x3)

x1

x2

x3

...

A1

All Algorithms

AllI

nput

s

...

time ntime n2

24 / 35



A2 A3 ... D

A1(x1) A2(x1) A3(x1)

A1(x2) A2(x2) A3(x2) ...

A1(x3) A2(x3) A3(x3)

x1

x2

x3

...

A1

All Algorithms

AllI

nput

s

...

time ntime n2

24 / 35



A2 A3 ... D

A1(x1) A2(x1) A3(x1) ¬A1(x1)

A1(x2) A2(x2) A3(x2) ... ¬A2(x2)

A1(x3) A2(x3) A3(x3) ¬A3(x3)

...

x1

x2

x3

...

A1

All Algorithms

AllI

nput

s

...

time ntime n2

24 / 35



n time

n2 time

Poly Timen3 time

25 / 35



log n space

log2 n space

log3 n space

25 / 35


Hierarchy Theorems for Nondeterministic Algorithms?

26 / 35


Hierarchy Theorems for Nondeterministic Algorithms?

A2 A3 ... D

A1(x1) A2(x1) A3(x1) ¬A1(x1)

A1(x2) A2(x2) A3(x2) ... ¬A2(x2)

A1(x3) A2(x3) A3(x3) ¬A3(x3)

...

x1

x2

x3

...

A1

All Nondet. Algorithms

AllI

nput

s

...

time ntime n2

26 / 35


Hierarchy Theorems for Nondeterministic Algorithms

27 / 35



A1 Dtime ntime n2

27 / 35



A1 D

x1

0x1

00x1

...

0`−1x1

0`x1

0`−2x1

inpu

ts

` ≈ 2|x1|

time ntime n2

27 / 35



A1 D

A1(x1)

A1(0x1)A1(00x1)

...

A1(0`−1x1)

x1

0x1

00x1

...

0`−1x1

A1(0`x1)0`x1

A1(0`−2x1)0`−2x1

inpu

ts

` ≈ 2|x1|

time ntime n2

27 / 35



A1 D

A1(x1)

A1(0x1)A1(00x1)

...

A1(0`−1x1)

x1

0x1

00x1

...

0`−1x1

A1(0`x1) ¬A1(x1)0`x1

A1(0`−2x1)0`−2x1

inpu

ts

6=` ≈ 2|x1|

time ntime n2

27 / 35



A1 D

A1(x1) A1(0x1)

A1(0x1) A1(02x1)A1(00x1) A1(03x1)

... ...A1(0`−1x1) A1(0`x1)

x1

0x1

00x1

...

0`−1x1

A1(0`x1) ¬A1(x1)0`x1

A1(0`−2x1) A1(0`−1x1)0`−2x1

inpu

ts

6=

=========

=========

` ≈ 2|x1|

time ntime n2

27 / 35



A1 D

A1(x1) A1(0x1)

A1(0x1) A1(02x1)A1(00x1) A1(03x1)

... ...A1(0`−1x1) A1(0`x1)

x1

0x1

00x1

...

0`−1x1

A1(0`x1) ¬A1(x1)0`x1

A1(0`−2x1) A1(0`−1x1)0`−2x1

Assume A1 the same

inpu

ts

6=

=========

=========

as D on all inputs

` ≈ 2|x1|

time ntime n2

27 / 35



A1 D

A1(x1) A1(0x1)

A1(0x1) A1(02x1)A1(00x1) A1(03x1)

... ...A1(0`−1x1) A1(0`x1)

x1

0x1

00x1

...

0`−1x1

A1(0`x1) ¬A1(x1)0`x1

A1(0`−2x1) A1(0`−1x1)0`−2x1

Assume A1 the same

inpu

ts

6=

=========

=========

===

===

===

===

===

===

as D on all inputs

` ≈ 2|x1|

time ntime n2

27 / 35



¬A1(x1)

A1 D

... ...

x1

0x1

00x1

...

0`−1x1

¬A1(x1)0`x1

0`−2x1

Assume A1 the same

inpu

ts

6=

=========

=========

===

===

===

===

===

===

as D on all inputs

` ≈ 2|x1|

¬A1(x1)¬A1(x1) ¬A1(x1)

¬A1(x1)¬A1(x1)

¬A1(x1)¬A1(x1)¬A1(x1)

¬A1(x1)¬A1(x1)

time ntime n2

27 / 35



¬A1(x1)

A1 D

... ...

x1

0x1

00x1

...

0`−1x1

¬A1(x1)0`x1

0`−2x1

Assume A1 the same

inpu

ts

6=

=========

=========

===

===

===

===

===

===

as D on all inputs

` ≈ 2|x1|

¬A1(x1)¬A1(x1) ¬A1(x1)

¬A1(x1)¬A1(x1)

¬A1(x1)¬A1(x1)¬A1(x1)

¬A1(x1)¬A1(x1)

time ntime n2

27 / 35



n time

n2 time

Poly Timen3 time

28 / 35



log n space

log2 n space

log3 n space

28 / 35


Hierarchy Theorems for Randomized Algorithms?

What if Pr[A1(x1) = “yes”] ≈ .5?

Then D does not have bounded error, not valid

29 / 35



A2 A3 ... D

A1(x1) A2(x1) A3(x1) ¬A1(x1)

A1(x2) A2(x2) A3(x2) ... ¬A2(x2)

A1(x3) A2(x3) A3(x3) ¬A3(x3)

...

x1

x2

x3

...

A1

All Randomized Algorithms

AllI

nput

s

...

time ntime n2



29 / 35



A2 A3 ... D

A1(x1) A2(x1) A3(x1) ¬A1(x1)

A1(x2) A2(x2) A3(x2) ... ¬A2(x2)

A1(x3) A2(x3) A3(x3) ¬A3(x3)

...

x1

x2

x3

...

A1


AllI

nput

s

...

time ntime n2



29 / 35



Random Strings

good

bad

01 10 10 1011 01 10

yes/no


30 / 35



A2 A3 ... D

A1(x1) A2(x1) A3(x1) ¬A1(x1)

A1(x2) A2(x2) A3(x2) ... ¬A2(x2)

A1(x3) A2(x3) A3(x3) ¬A3(x3)

...

x1

x2

x3

...

A1


AllI

nput

s

...

time ntime n2



31 / 35



A2 A3 ... D

A1(x1) A2(x1) A3(x1) ¬A1(x1)

A1(x2) A2(x2) A3(x2) ... ¬A2(x2)

A1(x3) A2(x3) A3(x3) ¬A3(x3)

...

x1

x2

x3

...

A1


AllI

nput

s

...

time ntime n2



31 / 35



A1 D

A1(x1) A1(0x1)

A1(0x1) A1(02x1)A1(00x1) A1(03x1)

... ...A1(0`−1x1) A1(0`x1)

x1

0x1

00x1

...

0`−1x1

A1(0`x1) ¬A1(x1)0`x1

A1(0`−2x1) A1(0`−1x1)0`−2x1

inpu

ts

6=

=========

=========

` ≈ 2|x1|

time ntime n2

Make sure D has bounded error – 1 bit of advice

32 / 35



A1 D

A1(x1) A1(0x1)

A1(0x1) A1(02x1)A1(00x1) A1(03x1)

... ...A1(0`−1x1) A1(0`x1)

x1

0x1

00x1

...

0`−1x1

A1(0`x1) ¬A1(x1)0`x1

A1(0`−2x1) A1(0`−1x1)0`−2x1

inpu

ts

6=

=========

=========

` ≈ 2|x1|

time ntime n2


32 / 35



A1 D

A1(x1) A1(0x1)

A1(0x1) A1(02x1)A1(00x1) A1(03x1)

... ...A1(0`−1x1) A1(0`x1)

x1

0x1

00x1

...

0`−1x1

A1(0`x1) ¬A1(x1)0`x1

A1(0`−2x1) A1(0`−1x1)0`−2x1

inpu

ts

6=

=========

=========

` ≈ 2|x1|

time ntime n2

Make sure D has bounded error

– 1 bit of advice

32 / 35



A1 D

A1(x1) A1(0x1)

A1(0x1) A1(02x1)A1(00x1) A1(03x1)

... ...A1(0`−1x1) A1(0`x1)

x1

0x1

00x1

...

0`−1x1

A1(0`x1) ¬A1(x1)0`x1

A1(0`−2x1) A1(0`−1x1)0`−2x1

inpu

ts

6=

=========

=========

` ≈ 2|x1|

time ntime n2


32 / 35



Yes, for algorithms with 1 bit of advice!

My work [ Kinne, Van Melkebeek ]Memory Space hierarchies: randomized, quantum, ...

33 / 35





33 / 35




n time

n2 time

Poly Timen3 time


33 / 35




log n space

log2 n space

log3 n space


33 / 35




log n space

log2 n space

log3 n space


33 / 35




log n space

log2 n space

log3 n space

My work [ Kinne, Van Melkebeek ]

Memory Space hierarchies: randomized, quantum, ...

33 / 35




log n space

log2 n space

log3 n space


33 / 35





Conjecture: P 6= NP





34 / 35


The End, Thank You!

Slides available at:http://www.kinnejeff.com/GoSycamores/

(or E-mail me)

More on my research (slides, papers, etc.) at:http://www.kinnejeff.com/

35 / 35


The End, Thank You!

Slides available at:http://www.kinnejeff.com/GoSycamores/

(or E-mail me)

More on my research (slides, papers, etc.) at:http://www.kinnejeff.com/

35 / 35

Documents

Two Big Questions: P vs. NP and P vs. BPPcs.indstate.edu/~jkinne/talks/isu-job-talk.pdf · Conjecture: P6= NP Techniques: hierarchy theorems, others Is randomness powerful? P vs