Transcript A Brief Survey of Quantum Computing

EECS 598:
Background in Theory of
Computation
Igor L. Markov and John P. Hayes
http://vlsicad.eecs.umich.edu/Quantum/EECS598
Outline
On computational models …
 More on terminology and notation
 Representing states in a computer
 Tensor products (again)
 Postulates of Q.M. specified for Q.C.
 Handling the probabilistic nature of quantum
computation
 Entanglement

Parameters of algorithms…

Execution Time
– Wall-clock time
– Asymptotic time

Memory (size)
 Ease of programming
 Worst, best, average, amortized, typical
 Multiple objectives -> often no single winner
– Cost trade-offs
Principle of Invariance

Challenges for analysis of [time] complexity
– Different hardware executes different #instr./sec
– Need a uniform measure of time complexity

Solution: measure constant-time steps
– Will be off by at most a constant

Principle of invariance
– Two implementations of an algorithm (when executed
on two actual computers) will not differ in time
complexity by more than a constant
– This is not a theorem !
(unlike, e.g., mathematical induction)
Comparing algorithms


POI: A similar situation with memory
POI gives hope to compare algorithms,
but does not provide a good mechanism
 What we want:
– Equivalence relations for algorithms (“as fast”)
– Order relations (“as fast or faster”)

Possible ideas:
– Count elementary operations
– Count elementary units of memory
– Are all operations counted alike?
Elementary Operations

What are elementary operations?
– Operations whose execution time is bounded by a constant
that does not depend on input values
– Actual “seconds per operation” may be disregarded
– But the very selection of elem. ops is still hardware-dependent !!!

What about +, -, *, /, <, >, =, ==, etc. of integers (or doubles)?
– Yes, if integers have a bounded number of bits, e.g., 32 (or 64)
– No, if the number of bits is not bounded


What about function calls?
What about memory accesses, e.g., a[i] ?
Computational Models

A computational model is determined by
– data representation and storage mechanisms
– available elementary operations

Most popular examples
–
–
–
–
–

Sequential and combinational Boolean circuits – EECS 270, 478
[Non-] Deterministic Finite Automata (DFA/NFAs) – EECS 376(476),478
Push-down automata – EECS 376(476)
Turing machines – EECS 376(476)
C programs, C++ programs – EECS 280, 281(380), 477
Computation models originate in technologies, physics, biology, etc
–
–
–
–
Parallel and distributed computing
Optical and DNA computing
Analog computing
Quantum computing
Measures of Algorithm Efficiency
Based on Allowed Inputs

Each set of input data gives a new measure of efficiency !
– Best cases
– Worst cases
– Representative / typical inputs (“benchmarks”)
 application-specific and domain-specific

Averaged measures (give “expected efficiency”)
– Consider different inputs in independent experiments
– Average over a particular distribution of inputs
 Some inputs may happen more frequently than others
 Application-specific
– Formal “average case”
 Averaged over the uniform distribution of all allowed inputs
– Empirical evaluation by sampling
 Generate random samples of the target distribution
 Much faster than enumeration. Theorem: results are very close.
Asymptotic Complexity


Recall that resource consumption is typically described by
positive monotonic functions
Asymptotic complexity measures (“on the order of”)
– Main idea:
 Given f(n) and g(n), does the difference grow with n or stay the same ?
Difference: f(n)/g(n). Only pay attention to the “limit behavior” @ n∞
 If f(n)/g(n) const as n∞ then f(n) is at least as good as g(n)
if f(n)/g(n) ≤ const as n∞ then f(n) is at least as good as g(n)
 If f(x) is at least as good as g(x) and g(x) is at least as good as f(x),
then the two functions are equivalent
– Coarser than counts of elementary “things” (larger equiv. classes)
 e.g., “200 N steps” and “3 N + log(N) steps” are now equivalent
– Much greater hardware independence
Complexity of Problems

In a given computational model,
for a given problem…
consider best possible algorithms
– In fact, whole equivalence classs of algos
in terms of asymptotic worst-case complexity

Call that the complexity of the problem
 Can often consider the same problem
in multiple computational models
– Will they have the same complexity?
Problem reductions

Reduction of problem X to problem Y
– Every instance of problem X translated into an instance
of problem Y
– Can apply any algorithm to solve Y
– Every solution to a translated instance
translated back

Complexity of problem X is no more than sum of
– Complexity of translating instances
– Complexity of Y
– Complexity of translating solutions back
Simulating Computational Models



Write a C program that simulates
a Turing machine
Write a C interpreter as a program for a Turing machine 
The computational models are equivalent
– Need to be careful about bit-sizes of basic types

Modern Church-Turing thesis
– Whatever you can do in terms of computation,
can be done on a Turing machine “as efficiently”
– “as efficiently” allows poly-time reductions (next slide)

How would you prove/disprove this?
NP-complete problems



P is the class of problems whose solutions can be found in poly-time
NP is the class of problems whose solutions can be verified in
polynomial time
Polynomial-time reductions
– Translations take poly-time



NP-complete problems are those to which every other problem in NP
reduces in poly-time
NP=P? is an open question
Orphan problems:
– Those in NP\P but not NP-complete, assuming NP!=P
– Number factoring
– Graph auto-morphism
So, how could you disprove the
modern Church-Turing thesis ?

Find a problem that has different worst-case asymptotic
complexity
– In terms of Turing machines
– In terms of, say, quantum computers


Need to define a computational model first!
Back to Quantum Mechanics
– Measurements are not deterministic
– So, need to include randomized algorithms
– BPP (answers can be correct with probability ½+e)

More complexity classes:
– PSPACE
– Decidable problems
Back to Quantum Mechanics

Details of computational model should follow
from the postulates of quantum mechanics
 Postulate 1: State Space (complex, Hilbert)
– Will need only finite-dim vector spaces

Postulate 2: Evolutions = Unitaries
– Think of matrices

Postulate 3: Quantum Measurement
– Two forms: Hermitian observables and ortho-normal
decompositions

Postulate 4: Composite systems
– Tensor products
Computing with functions

Functions form a Hilbert space
–
–
–
–
It is infinite-dimensional, much harder to study
Inner product is defined via integrals
Examples of Hermitian ops: differential operators
Analogues of matrices: “kernels”


Not nearly as nice as finite-dim matrices
How do we get finite-dim spaces?
– Restrict allowed states to a finite-dim subspace

Special case: restrict functions to finite sets
– All notions map to common finite-dim notions
Tensor products

A state of a quantum system is f(x)
 A state of a composite system is f(x,y)
 We would like some construction that would give
us the space of f(x,y) from spaces of g(x) and h(y)
 Tensor product
– Take bases: g1(x), g2(x),… and h1(x), h2(x),…
– Pair-wise product will be a basis of 2-var funcs

Tensor product of operators