experimental quantum foundations€¦ · this is a precise sense in which experimental quantum...

Experimental quantum foundations Robert Spekkens

Solstice of foundations June 19, 2017

Realism It aims at a true description of physical objects and their attributes, and aims to provide successively better approximations to the truth over time. The realist endorses a correspondence theory of truth.

Empiricism It aims at an efficient summary of our experience. The empiricist seeks to avoid false belief by building on top of what we cannot be mistaken about, such as statements about what we’ve observed directly.

Pragmatism It drops the notion of truth as correspondence with reality altogether, and aims only to be useful to us in achieving various goals.

What does a scientific theory aim to do?

Empiricism/realism/pragmatism as a philosophy of science

vs.

Empiricism/realism/pragmatism as a methodological principle for devising new

theories

What is the historical scorecard for realism vs. empiricism vs pragmatism as methodological principles for devising new theories? -  Thermodynamics

-  The atomic hypothesis

-  Relativity theory

-  Quantum theory

Empiricist

Realist

Pragmatist

Causal structure Rep’n of symmetries

Axiomatization from pragmatic principles à experimental consequences of pragmatic principles

Pragmatic principles such as: - Second law - No superluminal signalling - Data processing inequality are unlikely to be violated, so one would like to know the scope of physical theories that respect them Variation of axioms a good way to probe alternatives to QT (contrast w/ Weinberg’s proposed modification of QT)

Experimental metaphysics à Experimental consequences of ontological principles

Provide constraint on ontological possibilities for all future theories of physics This is a precise sense in which experimental quantum foundations distinguishes itself from experiments in the rest of physics

Empiricist

Realist

Pragmatist

Device-independent

paradigm

Interventionist Causal models

Theory of Bayesian inference

Ontological models

Generalized probabilistic

theories

GPTs w/ symmetries

Thermodynamic

Process theories

Frameworks for describing theories

Information processing

Resource theories

Empiricist

Realist

Pragmatist

Causal structure Rep’n of symmetries

The framework of generalized probabilistic

theories (GPTs) See: L. Hardy, quant-ph/0101012 J. Barrett, PRA 75, 032304 (2007)

Preparation Measurement

The framework of generalized probabilistic theories

P2

P4 P3

P6 P5

P1

M1

M5

M3

M10 M8

P9 P8

P7

M2

M6

M4

M7

M9

Preparation Measurement

The framework of generalized probabilistic theories

Preparation Measurement

The framework of generalized probabilistic theories

pass or fail

Suppose there are K measurements in a tomographically complete set (pass-fail mmts from which one can infer the statistics for all mmts)

State tomography for a single qubit

Preparation Measurement

Suppose there are K measurements in a tomographically complete set (pass-fail mmts from which one can infer the statistics for all mmts)

What can we say about f?

“operational state”

The framework of generalized probabilistic theories

pass or fail

Operational states form a convex set

(w,1-w)

Convex linear

Also true for mmts in tomo. complete set, so

Closed under convex combination –> a convex set

Convex linearity implies linearity

Therefore

If f is convex linear on GPT states

Then f is linear on GPT states

Preparation Measurement

“operational effects” “operational states”

S = Convex set in R = Interval of positive cone in

S and R characterize the GPT theory!

The framework of generalized probabilistic theories

pass or fail

S = a simplex

R = the unit hypercube

s can be any probability distribution

r can be any vector of conditional probabilities

(0,1)

(1,0)

(0,1)

(1,0)

(1,1)

GPT characterization of classical theory

(0,0)

GPT characterization of classical theory

S = a simplex

R = the unit hypercube

s can be any probability distribution

r can be any vector of conditional probabilities

(0,0,1)

(1,0,0)

(0,1,0)

(0,0,1)

(1,0,0) (0,1,0)

(1,1,0)

(1,1,1)

(0,1,1)

(1,0,1)

(0,0,0)

GPT characterization of classical theory

S = a simplex

R = the unit hypercube

s can be any probability distribution

r can be any vector of conditional probabilities

GPT characterization of quantum theory

S = the convex set of such operators

R = an interval of the positive cone of such operators

s can represent any trace one positive operator

r can be any positive operator less than identity

Recall: The Hermitian operators on a Hilbert space of dimension d form a real Euclidean vector space of dimension d2

Canonical variables

known known known

Probability distributions allowed by the epistemic restriction

0 1 0 1

0 1 0 1

0 1 0 1

0 1 0 1

0 1 0 1

0 1 0 1 0 1

0 1

Nothing known

0 1

0 1

Addition is mod2

Toy theory RWS, PRA 75, 032110 (2007)

The valid space of GPT states

The valid space of GPT states

The valid space of GPT states |0〉

|1〉

|-i〉 |+i〉

|+〉

|-〉 ½I

Valid measurements:

Any commuting set of canonical variables

0 1 0 1

0 1 0 1

0 1 0 1

Valid GPT effects

GPT characterization of convex closure of toy theory

GPT characterization of boxworld (Popescu-Rohrlich box correlations)

Quantum

Boxworld

Convex hull of toy theory Generic GPT

Interesting question about a given GPT:

Does it satisfy No-restriction hypothesis: the space of GPT effects in a theory include all effects that assign positive probabilities to every GPT state in the theory

GPT characterization of convex closure of toy theory

Dual of effect space

Dual of state space

Convex theories with maximal dual cone

C* algebraic theories

Quantum theory

Classical theory

Classical Statistical Theories with epistemic restriction

Boxworld

Toy theory

Deviations from quantum theory in the landscape of

generalized probabilistic theories: Direct constraints from experimental data

Joint work with: Matthew Pusey Mike Mazurek Kevin Resch

Determining GPT from infinite-run experimental statistics

Preparation Measurement

pass or fail

Pre

para

tions

Measurements

p(0|P2,M4) =⇣1 P (2)

2 · · · P (k)2

⌘·⇣M (1)

4 · · · M (k)4

⌘

Determining GPT from infinite-run experimental statistics

Preparation Measurement

pass or fail

Determining GPT from infinite-run experimental statistics

GPT states GPT effects

Preparation Measurement

pass or fail

Use singular value decomposition: k = rank of data matrix

Quantum

Boxworld

Convex hull of toy theory Generic GPT

GPT states GPT effects

Determining GPT from finite-run experimental statistics

Raw data Because of statistical

noise, the matrix of raw data is always full rank

Determining GPT from finite-run experimental statistics

=

Raw data

Find GPT model of best fit of

rank k

=

Determining GPT from finite-run experimental statistics

Raw data

Find GPT model of best fit of

rank k for Poissonian noise This is the “weighted low-rank approximation problem”

and factorizing appropriately, minimize

In variation over Kij satisfying

=

Determining GPT from finite-run experimental statistics

Raw data

Find GPT model of best fit of

rank k

Experimental set-up

Heralded SinglePhoton Source

State Preparation

PBS

Measurement

GT-PBS

Coupler

Mirror

IF

HWP

QWP

Dh

PPKTP

Dt

Drmeascompprep

Experimental data 100 measurements on 100 preparations

2 3 4 5 6 7 8 9 107000

8000

9000

10000

2 3 4 5 6 7 8 9 10103

104

105

106

107

10899% rangeFit

Model Rank

(a)

(b) (c)

50 100

Test

Measurement

50

10050 100

Pre

para

tion

Training

Measurement

0.0

0.2

0.4

0.6

0.8

1.0

χ2statistic

Characterize quality of fit by Â2 statistic

Characterize tradeoff between quality of fit and overfitting by Akaike information criterion

Characterize tradeoff between quality of fit and overfitting by Akaike information criterion

2 3 4 5 6 7 8 9 100

1

Rel

ativ

e M

odel

Lik

elih

ood

Model Rank

Score(rank-k model) = Exp{-1/2 [AIC(rank-k model) - AIC_min]}, and the relative likelihood of each rank-k model is its score divided by the sum of all the model scores. AIC_min is the minimum AIC value among the 9 model ranks I'm considering, and in this case it is the AIC for the rank-4 model.

rank 4: 0.9998 rank 5: 1.99£10-4

rank 6: 1.6£10-13 all others: <10-25

Use a GPT of rank 4!

Rank-4 GPT of best fit for the experimental data 100 measurements on 100 preparations

Rank-4 GPT of best fit for the experimental data 100 measurements on 100 preparations

2 3 4 5 6 7 8 9 107000

8000

9000

10000

2 3 4 5 6 7 8 9 10103

104

105

106

107

10899% rangeFit

Model Rank

(a)

(b) (c)

50 100

Test

Measurement

50

10050 100

Pre

pa

ratio

nTraining

Measurement

0.0

0.2

0.4

0.6

0.8

1.0χ2statistic

1006 measurements on 6 preparations 6 measurements on 1006 preparations

More experimental data

1006 measurements on 6 preparations 6 measurements on 1006 preparations

More experimental data

VSmin

/VSmax

= 0.968± 0.001

1006 measurements on 1006 preparations

Rank-4 GPT of best fit for the experimental data

Experimental constraints on violations of Tsirelson bound

Empiricist

Realist

Pragmatist

Causal structure Rep’n of symmetries

Experimental constraints on violations of Tsirelson bound

Some morals of the story

The GPT framework provides a means of analyzing experimental data that does not presume the correctness of quantum theory. Use it for any experiment that seeks to look for deviations from QT! Tomography for states and measurements can be achieved in a bootstrap manner Don’t worry only about underfitting. Worry also about overfitting.