a new definition for the quantum conditional rényi...

29
A new definition for the quantum conditional R ´ enyi entropy Fr ´ ed´ eric Dupuis Aarhus University May 21, 2014 Fr ´ ed´ eric Dupuis Quantum R ´ enyi entropies May 21, 2014 1 / 29

Upload: nguyenmien

Post on 15-Aug-2018

214 views

Category:

Documents


0 download

TRANSCRIPT

A new definition for the quantumconditional Renyi entropy

Frederic DupuisAarhus University

May 21, 2014

Frederic Dupuis Quantum Renyi entropies May 21, 2014 1 / 29

The papers

This talk will contain results from four papers:Martin Muller-Lennert, Frederic Dupuis, Oleg Szehr, SergeFehr, Marco Tomamichel. On quantum Renyi entropies: anew definition and some properties. 1306.3142Mark M. Wilde, Andreas Winter, Dong Yang. Strongconverse for the classical capacity ofentanglement-breaking channels. 1306.1586Rupert L. Frank and Elliott H. Lieb. Monotonicity of arelative Renyi entropy. 1306.5358Salman Beigi. Quantum Renyi divergence satisfies dataprocessing inequality. 1306.5920

Frederic Dupuis Quantum Renyi entropies May 21, 2014 2 / 29

Entropy

Shannon entropy:

Hppq :“ ´ÿ

i

pi log pi

Immensely useful in information theory, appears everywhere.Why this function in particular?

Frederic Dupuis Quantum Renyi entropies May 21, 2014 3 / 29

Renyi’s axioms

Shannon entropy can be derived axiomatically [Renyi, 1961].Let H be a function from unnormalized probability distributionsto the reals, and suppose it satisfies:

1 Continuity: Hppq is continuous.2 Symmetry: Hppq “ Hpπppqq for any permutation π3 Normalization: Hpt1

2uq “ 1

4 Additivity: Hpp� qq “ Hppq `Hpqq

5 Mean value: Hpp� qq “ TrrpsTrrp`qs

Hppq ` TrrqsTrrp`qs

Hpqq

Then Hppq is the Shannon entropy of the distribution p.

Frederic Dupuis Quantum Renyi entropies May 21, 2014 4 / 29

Renyi’s axioms

Why? (1), (3) and (4) impose that

Hptxuq “ ´ log2 x

[Erdos 1946]. The mean-value property does the rest:

Hptp1, . . . , pnuq “ Hptp1u� tp2, . . . , pnuq

“p1

TrrpsHptp1uq `

Trrtp2, . . . , pnus

TrrpsHptp2, . . . , pnuq

“ÿ

i

piTrrps

Hptpiuq

“ ´ÿ

i

piTrrps

log2 pi.

Frederic Dupuis Quantum Renyi entropies May 21, 2014 5 / 29

Renyi’s axioms

The mean-value axiom seems oddly specific. What if we relaxit?

1 Continuity: Hppq is continuous.2 Symmetry: Hppq “ Hpπppqq for any permutation π3 Normalization: Hpt1

2uq “ 1

4 Additivity: Hpp� qq “ Hppq `Hpqq5 General Mean:gpHpp� qqq “ Trrps

Trrp`qsgpHppqq ` Trrqs

Trrp`qsgpHpqqq, where g is

some strictly monotonic function.

Frederic Dupuis Quantum Renyi entropies May 21, 2014 6 / 29

Renyi’s axioms

Different choices of g yield interesting entropies:gpxq “ x: Shannon entropygαpxq “ 2p1´αqx: Renyi entropy of order α:

Hαppq :“1

1´ αlog

1

Trrps

ÿ

i

pαi .

Frederic Dupuis Quantum Renyi entropies May 21, 2014 7 / 29

Renyi entropy

The Renyi entropy unifies a lot of other entropies:αÑ 1: Shannon entropyα “ 2: Collision entropyαÑ 8: Min-entropy: Hminppq “ ´ log 1

Trrpsmaxi pi

α “ 12: Max-entropy: Hmaxppq “ 2 log 1

Trrps

ř

i

?pi.

Frederic Dupuis Quantum Renyi entropies May 21, 2014 8 / 29

Divergence and conditional entropy

What about conditional entropy, mutual information, etc? Thesecan be defined via a notion of divergence (“relative entropy”).For Shannon entropy:

Dpp}qq :“ÿ

i

piplog pi ´ log qiq.

Then,Hppq :“ Dpp}1q

where 1 “ t1, 1, . . . , 1u.

Frederic Dupuis Quantum Renyi entropies May 21, 2014 9 / 29

Conditional entropy and mutual information

Conditional entropy:

HpX|Y qp :“ ´DppXY }1X � pY q “ ´minqY

DppXY }1X � qY q.

Mutual information:

IpX;Y qp :“ DppXY }pX � pY q “ minqY

DppXY }pX � qY q.

Frederic Dupuis Quantum Renyi entropies May 21, 2014 10 / 29

Renyi’s axioms for divergence

Can we generalize these? Yes, by axiomatizing the divergence:1 Continuity: Dpp}qq is continuous in p, q.2 Symmetry: Dpp}qq “ Dpπppq}πpqqq for any permutation π.3 Normalization: Dpt1u}t1

2uq “ 1.

4 Order: If p ě q, then Dpp}qq ě 0; if p ď q, then Dpp}qq ď 0.5 Additivity: Dpp� r}q � sq “ Dpp}qq `Dpr}sq for allp, q, r, s ě 0 such that q Ţ p and s Ţ r.

6 General Mean:

gpDpp� r}q� sqq “Trrps

Trrp` rsgpDpq}sqq`

Trrrs

Trrp` rsgpDpr}sqq.

Frederic Dupuis Quantum Renyi entropies May 21, 2014 11 / 29

Renyi’s axioms for divergence

Different choices of g yield corresponding divergences:gpxq “ x: Usual relative entropygpxq “ 2p1´αqx: Renyi divergence:

Dαpp}qq “1

α ´ 1log

1

Trrps

ÿ

i

pαi q1´αi .

We can define conditional entropy and mutual information thisway:

HαpX|Y qp :“ ´minqY

DαppXY }1X � qY q

IαpX;Y qp :“ minqY

DαppXY }pX � qY q.

Frederic Dupuis Quantum Renyi entropies May 21, 2014 12 / 29

Quantum divergence

Can we make this quantum? We can “quantize” the axioms. Letρ, σ be positive semidefinite.

1 Continuity: Dpρ}σq is continuous in ρ, σ.2 Symmetry: Dpρ}σq “ DpUρU :}UσU :q for any unitary U .3 Normalization: Dpr1s}r1

2sq “ 1.

4 Order: If ρ ě σ, then Dpρ}σq ě 0; if ρ ď σ, then Dpρ}σq ď 0.5 Additivity: Dpρ� τ}σ � ωq “ Dpρ}σq `Dpτ}ωq for allρ, σ, τ, ω ě 0 such that σ Ţ ρ and ω Ţ τ .

6 General Mean:

gpDpρ�τ}σ�ωqq “Trrρs

Trrρ` τ sgpDpρ}σqq`

Trrτ s

Trrρ` τ sgpDpτ}ωqq.

Frederic Dupuis Quantum Renyi entropies May 21, 2014 13 / 29

Quantum divergence

Does this isolate a unique function?For commuting ρ and σ, this must reduce to thecorresponding g-divergence.What if ρ and σ don’t commute? No, multiple definitionssatisfy the axioms. For example:

D1αpρ}σq :“1

α ´ 1log

1

TrrρsTrrρασ1´α

s

D2αpρ}σq :“1

α ´ 1log

1

TrrρsTrrpσp1´αq{4ρα{2σp1´αq{4q2s.

for α between 1 and 2.

Frederic Dupuis Quantum Renyi entropies May 21, 2014 14 / 29

Quantum divergence

How do we pick the right version?Add an axiom: If you have ideas, let me know!Next best thing: pick a version that happens to satisfy othernice properties.

Frederic Dupuis Quantum Renyi entropies May 21, 2014 15 / 29

Quantum divergence

What nice properties?Corresponds to usual quantum entropies:

For αÑ 1, get the usual relative entropy.For αÑ8, get Dmax.For αÑ 2, get the collision entropy.For α “ 1

2 , get Dmin.

Data processing: Dαpρ}σq ě DαpFpρq}Fpσqq for any CPTPmap F .Monotonicity in α: Dαpρ}σq ď Dβpρ}σq if α ď β.Duality: HαpA|Bq “ ´HβpA|Cq for 1

α` 1

β“ 2 (will be

explained later)Chain rules

Frederic Dupuis Quantum Renyi entropies May 21, 2014 16 / 29

The definition

We define:

Dαpρ}σq :“1

α ´ 1log

ˆ

1

TrrρsTr

”´

σ12α´ 1

2ρσ12α´ 1

2

¯αı˙

.

Contrast with

D1αpρ}σq :“1

α ´ 1log

ˆ

1

TrrρsTrrρασ1´α

s

˙

D2αpρ}σq :“1

α ´ 1log

ˆ

1

TrrρsTrrpσp1´αq{4ρα{2σp1´αq{4q2s

˙

.

Frederic Dupuis Quantum Renyi entropies May 21, 2014 17 / 29

Properties

Properties of the new Renyientropy

Frederic Dupuis Quantum Renyi entropies May 21, 2014 18 / 29

Corresponds to known entropies

For αÑ 1, get the usual relative entropy.For αÑ 8, get Dmax.For αÑ 2, get the collision entropy.For α “ 1

2, get Dmin.

Frederic Dupuis Quantum Renyi entropies May 21, 2014 19 / 29

α “ 8: Max-relative entropy/min-entropy

Recall that

Dmaxpρ}σq “ logmintλ : ρ ď λσu

“ log›

›σ´

12ρσ´

12

8.

Contrast with α-divergence:

Dαpρ}σq “α

α ´ 1log

ˆ

1

Trrρs1{α

›σ

12α´ 1

2ρσ12α´ 1

2

α

˙

.

Frederic Dupuis Quantum Renyi entropies May 21, 2014 20 / 29

α “ 12: Min-relative entropy/Max-entropy

Recall that

Dminpρ}σq “ 2 log

ˆ

1

TrrρsTr

pσ12ρσ

12 q

12

ı

˙

Contrast with α-divergence:

Dαpρ}σq :“1

α ´ 1log

ˆ

1

TrrρsTr

”´

σ12α´ 1

2ρσ12α´ 1

2

¯αı˙

Frederic Dupuis Quantum Renyi entropies May 21, 2014 21 / 29

α “ 2: Collision entropy

Collision entropy:

D2pρ}σq “ log1

TrrρsTr

pσ´1{4ρσ´1{4q2‰

.

The associated conditional entropy is used in proofs of privacyamplification, decoupling, etc.Contrast with α-divergence:

Dαpρ}σq :“1

α ´ 1log

ˆ

1

TrrρsTr

”´

σ12α´ 1

2ρσ12α´ 1

2

¯αı˙

Frederic Dupuis Quantum Renyi entropies May 21, 2014 22 / 29

Data processing inequality

We have thatDαpρ}σq ě DαpFpρq}Fpσqq

for any CPTP map F and any 12ď α ď 8.

For entropies, this means that

HαpA|BCq ď HαpA|Bq.

Frederic Dupuis Quantum Renyi entropies May 21, 2014 23 / 29

Duality

For any pure ρABC , we have that

H1pA|Bqρ “ ´H1pA|Cqρ.

Frederic Dupuis Quantum Renyi entropies May 21, 2014 24 / 29

Duality

Recall

HminpA|Bqρ “ ´minσDmaxpρAB}1A � σBq

HmaxpA|Bqρ “ ´minσDminpρAB}1A � σBq.

For any pure state ρABC ,

HminpA|Bqρ “ ´HmaxpA|Cqρ.

Frederic Dupuis Quantum Renyi entropies May 21, 2014 25 / 29

Duality

We can generalize this:

HαpA|Bqρ “ ´HβpA|Cqρ,

where 1α` 1

β“ 2.

0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5

Frederic Dupuis Quantum Renyi entropies May 21, 2014 26 / 29

Chain rules

We can even prove chain rules for these:

HαpAB|Cq ě HβpA|BCq `HγpB|Cq,

where α, β, γ ą 1 and αα´1

“ββ´1

`γγ´1

.If α ą 1 and either β ą 1 and γ ă 1 or vice-versa, then:

HαpAB|Cq ď HβpA|BCq `HγpB|Cq.

Let me know if you have a use for this!

Frederic Dupuis Quantum Renyi entropies May 21, 2014 27 / 29

Conclusion

New definition of quantum Renyi divergenceLots of cools propertiesMilan Mosonyi’s talk: application to hypothesis testingKoenraad Audenaert’s talk: generalization of α entropies

Frederic Dupuis Quantum Renyi entropies May 21, 2014 28 / 29

Thank you

Thank you!

Frederic Dupuis Quantum Renyi entropies May 21, 2014 29 / 29