How to Solve Longstanding Open

Problems In Quantum Computing

Using Only Fourier Analysis
Scott Aaronson (MIT)
For those who hate quantum:
The open problems will be in
off-white boxes like this one
Problem 1: BQP PH?
Open since Bernstein-Vazirani 1993
Natural conjecture would be that BQP.PH.
But we dont even have an oracle separation
In fact, we dont even have an oracle A such
that BQP
A
.AM
A
. (Best is BQP
A
.MA
A
)
Furthermore, until recently our only
candidate problem was a monstrosity
(Recursive Fourier Sampling)
New Candidate Problem:
Fourier Checking
Given: Oracle access to functions f,g:{-1,1}
n
9
Promised: Either
(i) All f(x) and g(x) values are drawn independently
from the Gaussian distribution N(0,1), or
(ii) The f(x)s are drawn independently from N(0,1),
and g=FT(f) is the Fourier transform of f over Z
2
n
Problem: Decide which, with constant bias.
Claim: Fourier Checking is in BQP
Conjecture: Fourier Checking is not in PH
As usual, the problem boils down to showing Fourier
Checking has no AC
0
circuit of size 2
poly(n)

Alas, all known techniques for constant-depth circuit
lower bounds (random restriction, Razborov-Smolensky,
Nisan-Wigderson) fail for interesting reasons!
Conjecture (Linial-Nisan 1989): Polylog-wise
independence fools AC
0
[recently proved by Bazzi for DNFs!]
What I want: The Generalized Linial-Nisan Conjecture.
Namely, no distribution D over {0,1}
N
such that


for all conjunctions C of polylog(N) literals, can be
distinguished from uniform (with O(1) bias) in AC
0

( )
| |
( ) 1 1
1
1
2
Pr 1
1
O O
+ s s
N
C
N
C
D
Problem 2: The Need for
Structure in Quantum Speedups
Beals et al 1998: Quantum and classical decision
tree complexities are polynomially related for all
total Boolean functions f: D(f)=O(Q(f)
6
)
But could a quantum computer evaluate an almost-
total function with exponentially fewer queries?
Suggests that if you want an exponential quantum
speedup, then you need to exploit some structure
in the oracle being queried (e.g. periodicity in the
case of Shors factoring algorithm)
Conjecture: Let Q be a T-query quantum
algorithm. Then a classical randomized algorithm
that makes T
O(1)
queries can approximate Qs
acceptance probability on most inputs xe{0,1}
n
.
Would suffice to prove that every low-degree
bounded polynomial has an influential variable:
Let p:{-1,1}
n
[-1,1] be a real polynomial of degree d.
Suppose
Let
Then there exists an i such that Inf
i
>1/poly(d).
( ) ( ) | |.
i
x
i
x p x p E Inf =
( ) ( ) | | ( ). 1
,
O = y p x p E
y x
Conjecture 2: If P=P
#P
, then P
A
=BQP
A
with
probability 1 for a random oracle A. [insert avg, i.o. to taste]
What We Know
Dinur, Friedgut, Kindler, ODonnell 2006:
Every degree-d polynomial p:{-1,1}
n
[-1,1]
with O(1) variance has a variable with
influence at least 1/exp(d). (Indeed, p is
close to an exp(d)-junta.)
ODonnell, Saks, Schramm, Servedio 2005:
Every classical decision tree of depth d has a
variable with influence O(1/d).
Can we find a fixed f (depending only on the
input length n), such that computing
given y as input is #P-complete?
Problem 3: Quantum Algorithm
for a #P-complete Problem?!?
Let f:{0,1}
n
{0,1} be efficiently computable.
( )
2

y f
( )
2

y f
( )
2

y f
Then a simple quantum algorithm outputs
each ye{0,1}
n
with probability
If even estimating is #P-complete on
average, then FBPP=FBQP P
#P
=AM.
Open Problems