Introduction to quantum information processing qic 710 cs 667 ph 767 co 681 am 871
Download
1 / 17

Introduction to Quantum Information Processing QIC 710 / CS 667 / PH 767 / CO 681 / AM 871 - PowerPoint PPT Presentation


  • 158 Views
  • Uploaded on

Introduction to Quantum Information Processing QIC 710 / CS 667 / PH 767 / CO 681 / AM 871. Lecture 16 (2011). Richard Cleve DC 2117 cleve@cs.uwaterloo.ca. Distinguishing between two arbitrary quantum states. Holevo-Helstrom Theorem (1).

loader
I am the owner, or an agent authorized to act on behalf of the owner, of the copyrighted work described.
capcha
Download Presentation

PowerPoint Slideshow about 'Introduction to Quantum Information Processing QIC 710 / CS 667 / PH 767 / CO 681 / AM 871' - aoife


An Image/Link below is provided (as is) to download presentation

Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author.While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server.


- - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - -
Presentation Transcript
Introduction to quantum information processing qic 710 cs 667 ph 767 co 681 am 871

Introduction to Quantum Information ProcessingQIC 710 / CS 667 / PH 767 / CO 681 / AM 871

Lecture 16 (2011)

Richard Cleve

DC 2117

cleve@cs.uwaterloo.ca



Holevo helstrom theorem 1
Holevo-Helstrom Theorem (1)

Theorem: for any two quantum states  and , the optimal measurement procedure for distinguishing between them succeeds with probability ½ + ¼|| − ||tr (equal prior probs.)

Proof* (the attainability part):

Since  − is Hermitian, its eigenvalues are real

Let + be the projector onto the positive eigenspaces

Let − be the projector onto the non-positive eigenspaces

Take the POVM measurement specified by+and − with the associations +   and −  

* The other direction of the theorem (optimality) is omitted here


Holevo helstrom theorem 2
Holevo-Helstrom Theorem (2)

Claim: this succeeds with probability ½ + ¼|| −||tr

Proof of Claim:

A key observation is Tr(+−−)( −) = || −||tr

The success probability is ps=½Tr(+ ) + ½Tr(−)

& the failure probability is pf=½Tr(+ ) + ½Tr(−)

Therefore, ps−pf=½Tr(+−−)( −) =½|| −||tr

From this, the result follows


Purifications ulhmann s theorem
Purifications & Ulhmann’s Theorem

Any density matrix , can be obtained by tracing out part of some larger pure state:

a purification of 

Ulhmann’s Theorem*: Thefidelitybetween  andis the maximum of φψtaken over all purifications ψ and φ

* See [Nielsen & Chuang, pp. 410-411] for a proof of this

Recall our previous definition of fidelity as

F(, )= Tr√1/21/2||1/21/2||tr


Relationships between fidelity and trace distance
Relationships between fidelity and trace distance

1 − F(,)  ||−||tr  √1 − F(,)2

See [Nielsen & Chuang, pp. 415-416] for more details


Introduction to quantum information processing qic 710 cs 667 ph 767 co 681 am 871

Entropy and

compression


Shannon entropy
Shannon Entropy

Let p=(p1,…, pd) be a probability distribution on a set {1,…,d}

Then the (Shannon) entropy of p is H(p1,…, pd)

Intuitively, this turns out to be a good measure of “how random” the distribution p is:

vs.

vs.

vs.

H(p) = log d

H(p) = 0

Operationally, H(p) is the number of bits needed to store the outcome (in a sense that will be made formal shortly)


Von neumann entropy
Von Neumann Entropy

For a density matrix , it turns out that S() =− Tr log is a good quantum analog of entropy

Note:S() = H(p1,…, pd), where p1,…, pd are the eigenvalues of  (with multiplicity)

Operationally, S() is the number of qubits needed to store  (in a sense that will be made formal later on)

  • Both the classical and quantum compression results pertain to the case of large blocks of n independent instances of data:

  • probability distribution pn in the classical case, and

  • quantum state n in the quantum case


Classical compression 1
Classical compression (1)

Let p=(p1,…, pd) be a probability distribution on a set {1,…,d} where n independent instances are sampled:

( j1,…, jn) {1,…,d}n (dnpossibilities, nlogd bits to specify one)

Theorem*: for all > 0, for sufficiently large n, there is a scheme that compresses the specification to n(H(p) + )bits while introducing an error with probability at most 

Intuitively, there is a subset of {1,…,d}n, called the “typical sequences”, that has size 2n(H(p) +)and probability 1 −

A nice way to prove the theorem, is based on two cleverly defined random variables …

* “Plain vanilla” version that ignores, for example, the tradeoffs between nand


Classical compression 2
Classical compression (2)

Define the random variable f :{1,…,d} R as f ( j ) =− log pj

Note that

Thus

Define g:{1,…,d}n R as

Also, g( j1,…, jn)


Classical compression 3
Classical compression (3)

By standard results in statistics, as n , the observed value of g( j1,…, jn) approaches its expected value, H(p)

More formally, call ( j1,…, jn){1,…,d}n-typical if

Then, the result is that, for all > 0,for sufficiently large n,

Pr[( j1,…, jn) is-typical] 1−

  • We can also bound the number ofthese-typical sequences:

  • By definition, each such sequence has probability  2−n(H(p) +)

  • Therefore, there can be at most 2n(H(p) +) such sequences


Classical compression 4
Classical compression (4)

In summary, the compression procedure is as follows:

The input data is ( j1,…, jn) {1,…,d}n, each independently sampled according the probability distribution p=(p1,…, pd)

The compression procedure is to leave ( j1,…, jn) intact if it is -typical and otherwise change it to some fixed -typical sequence, say, ( j,…, j) (which will result in an error)

Since there are at most 2n(H(p) +)-typical sequences, the data can then be converted inton(H(p) + )bits

The error probability is at most , the probability of an atypical input arising


Quantum compression 1
Quantum compression (1)

The scenario:n independent instances of a d-dimensional state are randomly generated according some distribution:

φ1 prob.p1

  

φr  prob.pr

0 prob.½

+ prob.½

Example:

Goal: to “compress” this into as few qubits as possible so that the original state can be reconstructed with small error in the following sense …

  • -good:

  • No procedure can distinguish between these two states

    • compressing and then uncompressing the data

    • the original data left as is

  • with probability more than ½ + ¼ 


Quantum compression 2
Quantum compression (2)

Theorem: for all > 0, for sufficiently large n, there is a scheme that compresses the data to n(S() + ) qubits, that is 2√ -good

Express  in its eigenbasis as

Define

For the aforementioned example, 0.6n qubits suffices

The compression method:

With respect to this basis, we will define an -typical subspace of dimension 2n(S() +) =2n(H(q) +)


Quantum compression 3
Quantum compression (3)

1 prob.q1

  

d prob.qd

Define typ as the projector into the -typical subspace

By the same argument as in the classical case, the subspace has dimension 2n(S() +) andTr(typ n)  1−

The -typicalsubspace is that spanned by

where ( j1,…, jn) is -typical with respect to (q1,…, qd)

This is because  is the density matrix of


Quantum compression 4
Quantum compression (4)

Proof that the scheme is 2√ -good:

(This is to be inserted)