PUBLISHED ONLINE: 23 MARCH 2015 | DOI: 10.1038/NPHYS3266

A quantum advantage for inferring
causal structure
Katja Ried1,2,3*†, Megan Agnew1,2†, Lydia Vermeyden1,2, Dominik Janzing4, Robert W. Spekkens3
and Kevin J. Resch1,2

The problem of inferring causal relations from observed correlations is relevant to a wide variety of scientific disciplines.
Yet given the correlations between just two classical variables, it is impossible to determine whether they arose from a causal
influence of one on the other or a common cause influencing both. Only a randomized trial can settle the issue. Here we consider
the problem of causal inference for quantum variables. We show that the analogue of a randomized trial, causal tomography,
yields a complete solution. We also show that, in contrast to the classical case, one can sometimes infer the causal structure
from observations alone. We implement a quantum-optical experiment wherein we control the causal relation between two
optical modes, and two measurement schemes—with and without randomization—that extract this relation from the observed
correlations. Our results show that entanglement and quantum coherence provide an advantage for causal inference.

he slogan ‘correlation does not imply causation’ is meant to holds. Despite the restrictions that quantum theory places on
capture the fact that any joint probability distribution over learning about systems—that not all observables that can be defined
two variables can be explained not only by a causal influence of a system can be measured precisely at the same time—we
of one variable on the other, but also by a common cause acting show that the ability to intervene on the early quantum system
on both1 . We here investigate whether a similar ambiguity holds for allows a complete solution of the problem. This constitutes a
quantum systems, and we show that, surprisingly, it does not. new type of tomography, which subsumes tomography of bipartite
Finding causal explanations of observed correlations is a states and tomography of processes, and promises applications
fundamental problem in science, with applications ranging from for determining whether the state evolution implemented by
medicine and genetics to economics2,3 . As a practical illustration, a given device is Markovian. We implement this new type of
consider a drug trial. Naively, a correlation between the variables tomography experimentally and obtain a complete description of
treatment and recovery may suggest a causal influence of the former the causal structure.
on the latter. But suppose men are more likely than women to seek The real surprise, however, is that even if one only has the
treatment, and also more likely to recover spontaneously, regardless ability to observe the early system, the quantum correlations hold
of treatment. In this case, gender is a common cause, inducing signatures of the causal structure—in other words, certain types
correlations between treatment and recovery even if there is no of correlation do imply causation. In a recent paper, Fitzsimons,
cause–effect relation between them. Jones and Vedral5 defined a function of the observed correlations
Unless one can make strong assumptions about the nature of which acts as a witness of causal influence, by ruling out a purely
the causal mechanisms4 , the only way to distinguish between the common-cause explanation. We here present a larger framework
two possibilities is to replace observation of the early variable with that places this result on an equal footing with an analogous result
an intervention on it. For instance, pharmaceutical companies do for witnessing common-cause relations. We exploit these facts to
not leave the choice of treatment to the subjects of their trials, but devise, for a particular class of causal scenarios, a complete solution
carefully randomize the assignment of drug or placebo. This ensures of the causal inference problem using observation alone—a task
that the administered treatment is statistically independent of any that is impossible classically. We implement a family of such causal
potential common causes with recovery. Consequently, any correla- scenarios experimentally and confirm our conclusion.
tions with recovery that persist herald a directed causal influence.
The question of whether there were in fact common causes can be The quantum causal inference problem
answered by tracking whether recovery correlates with the subjects’ The two quantum systems whose causal structure we are probing
intent to treat. Thus, the ability to intervene allows a complete solu- will be denoted A and B, with A preceding B in time. The
tion of the causal inference problem: it reveals both which variables dynamics relating them may be arbitrarily complicated, involving
are causes of which others and, via the strength of the correlations, any number of additional systems and interactions. Nonetheless,
the precise mathematical form of the causal dependencies. any non-trivial causal relation that is induced between A and B
In this article, we consider the quantum version of this causal takes one of three forms: A could be a cause of B, the two could
inference problem. The challenge is to infer, by probing the be influenced by a common cause, or there could be a mixture of
correlations between two temporally ordered quantum systems, the two causal mechanisms (either a probabilistic mixture or a case
whether one causally influences the other, whether a common cause where both act simultaneously). The three possibilities are depicted
acts on both, or whether a combination of the two possibilities in Fig. 1a,b.

1 Institute
for Quantum Computing, University of Waterloo, Waterloo, Ontario N2L 3G1, Canada. 2 Department of Physics & Astronomy, University of
Waterloo, Waterloo, Ontario N2L 3G1, Canada. 3 Perimeter Institute for Theoretical Physics, 31 Caroline St. N, Waterloo, Ontario N2L 2Y5, Canada.
4 Max Planck Institute for Intelligent Systems, Spemannstraße 38, 72076 Tübingen, Germany. †These authors contributed equally to this work.



© 2015 Macmillan Publishers Limited. All rights reserved

Quantum circuits depicting these causal to them as interventionist schemes. particle in one dimension. The measurement on C is projective. are the conventional depiction of causal structure in the causal there is a broad class of schemes that share this feature. An example of a family of quantum circuits the same variable’s value after the intervention. Observational scheme. nodes represent quantum systems and directed edges represent causal © 2015 Macmillan Publishers Limited. Their defining characteristic structures. Directed acyclic graphs. 1b).nature. The (conditioned on the measurement outcome) is the same as the case of pure identity corresponds to a purely cause–effect connec. 1c. nor exactly according to their choice. with probability 1 − p. and boxes represent is that the information they yield about a variable’s value before operations: gates (green). no quantum observational scheme can be a precise cause–effect connection from a common-cause connection. the probability distribution over phase space describing unknown quantum operation drawn from a one-parameter family: what is known about the particle before the measurement identity. analogue of a completely passive observation because learning the 2 NATURE PHYSICS | ADVANCE ONLINE PUBLICATION | www.3 . Interventionist scheme.ARTICLES NATURE PHYSICS DOI: 10. The quantum interventionist scheme we use Interventionist versus observational probing schemes mirrors a classical perfectly randomized trial: we first measure the The randomized trial discussed in the introduction is an example early system (A). Completely passive observation is an example of such functionality of the unknown circuit fragment that relates A and B a scheme. In fact. for a classical (the dashed region in Fig. A complete solution of the causal inference problem specifies On the other hand. For instance. E E whereas D is prepared in states drawn from a tomographically complete set. we aim to specify the structure. partially randomized. and which therefore—classically—cannot reveal the causal circuits depicted in Fig.1038/NPHYS3266 a b B c B a c FE B u m B B u m D B F E D A A t l ? Gate s k A |Φ+〉 C D E t l B B b B u m C s k B C E D State E E preparation A A s k ? A |Φ+〉 C HWP Filter QWP LCR PBS ppKTP APD Lens 50:50 NPBS B B B Figure 2 | Two schemes for probing causal relations and experimental set-up. a purely common-cause connection (middle). For instance. The unknown circuit fragment enclosed in the dashed box can be probed by two schemes. We aim to discriminate for optical elements: half-wave plate (HWP). The outputs B and C are both subjected to tomographically complete sets of measurements. c. where wires represent quantum systems. in Fig. temporally ordered quantum systems: (top to bottom) cause–effect. pre-measurement state as they do about the post-measurement this can be achieved by specifying the identity of the gates in the state. then one can still learn something about the effectiveness of the drug2 . All rights reserved . the distinction between observational and interventionist schemes in the quantum sphere. On of a classical experimental scheme that allows one to distinguish a the other hand. we define the class of observational schemes not just the causal structure but also the functional relationship as those that yield precisely the same information about the that holds between each system and its causal parents. the three possible causal relations that may hold between a pair of liquid crystal retarder (LCR). The outputs B and C are both subjected to |Φ+〉 tomographically complete sets of measurements. 1b. a. In a drug trial. if the treatment is assigned to patients neither completely E (dashed green box) is either identity (top). c. then reprepare it in a state selected at random. Subsequently. Experimental set-up including polarization-entangled photon source and probabilistic swap gate. Notation Figure 1 | The quantum causal inference problem. More generally. A system A is prepared in a maximally entangled state Nonetheless. state preparations (orange) and the operation of the intervention is different from the information they yield about discarding a system (black). with probability p. but there are many others. A b. one of which is turned on with probability p. for that range over the three possible causal relations. A and E are subjected to an uniform. quarter-wave plate (QWP). but rather is probabilistic mixture of the two (bottom). fixing the preparation on D. a measurement that reveals its position A particular example of our causal inference problem is depicted while randomizing its momentum is clearly not completely passive. and every other case We take the property of informational symmetry to also define corresponds to a hybrid of the two causal structures (bottom). A A Lowercase variables denote settings and outcomes of these interventions. distribution describing what is known about the particle after tion between A and B (top). polarizing beamsplitter (PBS). The gate acting on A and example. a. and swap. we refer inference literature2. swap (middle) or a at random. b. the case of pure swap corresponds to the measurement. periodically poled KTP crystal (ppKTP). common cause or a combination of both. as long as the initial distribution over momentum is with an ancillary system E. and non-polarizing beamsplitter (NPBS). where avalanche photodiode (APD).

The polarization degrees of freedom of a pair of photons to allow a family of unitaries that coherently interpolate between constitute the pair of systems. for instance. scheme. If the structure is It is useful to adopt a distinct notation for the versions of A before purely common cause (as in Fig. In the interventionist scheme. if the initial state of our (preparations. Conversely. ECB|D must therefore satisfy the additional constraint that C cannot Instead of a full specification of ECB|D . and by u and m for B. However.2)% with our theoretical L(H) is the space of linear operators on the Hilbert space H. the map will informationally complete sets of measurements on B and on C have the form and an informationally complete set of preparations on D. one is often interested causally depend on D: tracing out B from ECB|D leaves a map EC|D in more coarse-grained information. where dC is the measurement on C. we fit the experimental data to a map of the form of are further discussed in the Supplementary Methods. Circuit fragments that do not fall into one of the standard classes complementary. predictions. B. we can write is reprepared with t = s and l = k. Figure 4a shows our best estimate. We term unaffected. pfit . both bipartite states and unipartite processes. 2c. We denote these by C and D the form respectively.13 .11 and others12. we are limited to ce ECB|D = ρC ⊗ EB|D 0 (2) performing an informationally complete set of measurements on B and on C. in the same way that downconversion pair are then subjected to a probabilistic swap gate: informationally complete sets of preparations and measurements with probability p the modes are exchanged. An even more general Fig. channels.NATURE PHYSICS DOI: 10. 2a. the experimental data in neither states nor processes (as in Fig. and if the measurement is projective considered in the context of alternative formulations of quantum and obeys the standard state update rule (wherein the system is theory by a number of authors. obtains about the post-measurement state of A. is subjected to a polarization measurement. 2. so that together they completely characterize photon’s polarization. but also describes more general possibilities. The first photon of the output. otherwise they are allow conventional tomography of states and processes. the causal map can describe objects that are l eigenstate of the t observable. state on C. subsumes bipartite states and processes as special cases. if be completely solved in the interventionist scheme. whereas D can only have a directed causal influence on B. after Supplementary Methods. identity and swap. if the example of Fig. the map will have and after the measurement on it6 . by performing the structure is purely cause–effect (as in Fig. Therefore. is subjected to a final this scheme causal tomography as it achieves a complete solution measurement of its polarization before both photons are detected of the causal inference problem. We denote these by s and k respectively for equation (2) and TrB (ρCB ) = (1/dC )1C in equation (1). using the experimental set-up shown in the same time. for instance. The two cc ECB|D = ρCB TrD (1) probing schemes are illustrated in Fig. 3a and are found preserving map of the form ECB|D : L(HD ) → L(HC ⊗ HB ).7 ± 0. as a function of NATURE PHYSICS | ADVANCE ONLINE PUBLICATION | www. then one obtains the combs7 . D. We will show that the quantum causal inference problem can describing a state on CB and the trace operation on D. For example. 3} and Data analysis in the interventionist scheme outcomes {+1. Note. where interventionist scheme is sufficient to tomographically reconstruct |H i(|V i) denotes horizontal (vertical) polarization. on the other hand. 2a schemes as limiting cases. 1b middle). For an informationally complete set of measurements we use the three Pauli observables. 1] interpolates between the We implement the one-parameter family of circuits introduced in extreme cases given in equations (1) and (2). All rights reserved . observables. Its properties scheme. In the interventionist scheme. and the sets of measurements on B and C span this measurement is followed by a random repreparation of the L(HB ) and L(HC ). the standard state update rule implies that D effect. 1c is modified Fig. more general scheme that includes conventional tomography The circuit fragment represented by the dashed box in Fig. we require. system D is prepared in the More generally.b. The map Poissonian noise in the observed count numbers. In the the structure is a probabilistic mixture of common cause and cause– observational scheme. C. or measurements) have been previously system A is completely mixed. if this case is the conditional probability distribution P(km|lstu). as shown in data obtained in the experiment that implements the interventionist the Supplementary Methods. 1b top). 2a. implements a physical process that takes one input. equation (3). by a completely positive and trace. −1}. cc ECB|D = pECB|D + (1 − p)ECB|D ce (3) Experiment where the mixing parameter p ∈ [0. that is. which ranges through the possible causal structures as form arises if cause–effect and common-cause contributions act at we vary the parameter p. the value of the that produces a fixed state ρC regardless of the input at D. 1b bottom). so the experimental data takes the form P(km|su). Such a scheme is Note that the causal map ECB|D incorporates as special cases consequently in the observational class. quantum conditional same information about the pre-measurement state of A as one 3 © 2015 Macmillan Publishers Limited. probability p in the mixture. One of the the map ECB|D for systems of any dimension. where to achieve an average fidelity of (97. hence measurement settings {1. and produces We apply our scheme to tomographically reconstruct ECB|D from outputs B and C. whereas the preparation of D is determined by the describing a quantum channel from D to B and a normalized outcome of the measurement on C. The key is that the set of preparations on which the photon is denoted D. as noted earlier. process matrices9 . ρC = (1/dC )1C in and the outcome. which in this case are qubits. specifying which observable is measured. as depicted in the dimension of the Hilbert space of C. D span L(HD ). operator tensors8 . We use downconversion √ to create entangled The conditional probability distribution P(km|lstu) obtained in the photon pairs in the state |8+ i = (1/ 2)(|H i|H i + |V i|V i). To estimate p in the interventionist EC|D = ρC TrD . This is proved in the photons. Fig. Considering that the map ECB|D in coincidence. Photon D and the other photon from the the input–output functionality of the map. initial state on C be completely mixed. To ensure that projective measurements yield purely Each measurement is described by two classical variables: the observational schemes. that is. It can therefore be represented.1038/NPHYS3266 ARTICLES value of one quantum observable randomizes the values of other. We term such an object a causal map.nature. The resulting maps are presented in Fig. In the observational scheme. The quoted error on the fidelity reflects the effect of however.b. that the output C precedes the input D in time. under different names: quantum left in the pure state in which it was found). so that C can only be related to B by a common cause. 1c. that the measurement setting. causal tomography constitutes a Mathematical form of the unknown circuit fragment novel.

1)% (p = 0).00 0. we estimated the phase shift to be φ = −0. we note that the expected Choi states would appear almost indistinguishable from the reconstructed ones depicted here. (98.00 0.07)% and (96. are compared to the expected Choi states by computing the fidelity F ≡ (Tr τ 1/2 τfit τ 1/2 )2 (ref. In the case p = 0.25 Figure 3 | Reconstruction of the causal map. 1 − pexp ) to decide whether identity or swap is implemented. one can simply revise the Choi states expected from each causal pstructure to include it.25 0.012.38 ± 0. the state prepared on CE contained an inadvertent phase shift: instead of |8+ i as planned.25 0.26 .00 −0. In light of these high fidelities. 4a: as pexp reaches the probabilistic mixture. by dividing each measurement into 25 1. the observational scheme does therefore.07)%.00.25 0.25 0. 〈HV | HV V 〈H V| | HV . only 96% of the input photons the implemented value of only 0. The relative Data analysis in the observational scheme frequency of swap realized over the course of such a measurement.25 0. The rms not allow a complete tomographic reconstruction of the map deviation between pexp and pfit due to this effect was estimated by ECB|D in an arbitrary causal scenario. follows a binomial distribution about pexp . we expect E CB|D = (1/2)1C ⊗ I B|D . | 〈HH VH| 〈HH V| H| 0. The case p = 1/2 produces an equal mixture of the two. denoted τfit . (98.09)%. The arrays represent the real (top) and imaginary (bottom) components of the density matrices. comparable to the observed value.25 0. sampling from (pexp .00 0. the value that the experiment sought to implement.00 −0. without the Monte Carlo to be 0.nature. Unlike the interventionist scheme. with probability of common cause p = 0 (left). respectively.25 −0.25 0. which is defined as the tripartite state τCBD ≡ (E CB|D0 ⊗ 1D )(|8+ iD0 D h8+ |). All rights reserved . 16). intervals and. Reconstructions based on the interventionist scheme (a) and the observational scheme (b). emerge in the correct output path. |HHHi = |HiC |HiB |HiD .25 0. which the observational scheme matches at (98.00 0. Based on the correlations between C and B in the purely common-cause case.25 −0.00 0. that is.6 ± 0. In the experiment√that implemented the interventionist scheme. The interventionist scheme achieves fidelities of (98. This is reflected in Fig. This accounts for the non-zero imaginary components in the Choi states in the interventionist © 2015 Macmillan Publishers Limited.25 −0.25 −0.5 p=1 0. for each interval. Our scheme A small systematic error in pfit arises owing to the finite contrast is shown to extract p with high accuracy.25 −0.5 p=1 0.25 −0.31 ± 0.25 0.25 0.00 −0. with blue representing positive values and red negative values. it is close to (1/ 2)(|Hi|Hi + eiφ |Vi|Vi)..38 ± 0.00 0.25 0.3)% (p = 1) to the expected states. p = 1/2 (middle) and p = 1 (right).25 HH 〉 |H HHV H〉 〉 .25 b p=0 p = 0.1038/NPHYS3266 a p=0 p = 0.024.00 〉 −0. for three different causal structures (shown in Fig.17π. When p = 1.25 −0. whereas pexp = 0 implements This deviation is due mostly to the way we implemented identity with 99% contrast. the best-fitting pfit is only 0. The basis is ordered as CBD. The tomographically reconstructed states. which corresponds to a Choi state τCBD = (1/2)1C ⊗ |8+ iBD h8+ |. 1c). randomizing repreparation.ARTICLES NATURE PHYSICS DOI: 10..00 0. which has Choi state τCBD = |8+ iCB h8+ | ⊗ (1/2)1D . we expect E CB|D = |8+ iCB h8+ |TrD . with an rms deviation from of the swap gate: when pexp = 1. We show the Choi representation of the maps14.. This is because.25 −0.2 ± 0. with the identity map I B|D from B to D.96. the state prepared on D is the same 4 NATURE PHYSICS | ADVANCE ONLINE PUBLICATION | www.25 0.00 0.2)% (p = 1/2) and (96.25 0. pexp . This phase shift is no impediment to causal inference.3 ± 0..15.

The signature of the causal structure is contained.6 10 state has this pattern of correlations. |1i} the eigenstates of σ3 .043 according black points marking the optimal pfit for each pexp .8 Pattern of correlations Cause–effect Common-cause explanation? explanation? 0.2 causal inference problem (up to one choice of sign). as discussed in the text. a fact sometimes described as the nonexistence of an ‘antisinglet’ state.0 6 distribution over k (respectively m) is uniform for all values s (respectively u). All rights reserved . No bipartite 0.0 0. Entanglement and repreparations of D span L(HD ).6 0. Common-cause mechanisms it requires more measurements. An explicit proof is provided in the Supplementary Methods. scheme (a) or the observational scheme (b). perfect negative correlation for all three observables can only be explained by a common ln(χ2) cause—namely. 2). as demonstrated by the example in Table 1. (s. This is because in NATURE PHYSICS | ADVANCE ONLINE PUBLICATION | www. There is no channel that produces this pattern: it would constitute 0. tage in causal inference. Our experiment implements such a mixture (where the 0. and cause–effect. Despite this fact.032 from the implemented values. pfit √ 33 . one the classical context without additional assumptions. this result can clearly discern a narrow valley around pfit = pexp . with correlation coefficients Csu ≡ p(k = m|su) − p(k 6 = m|su) ∈ {±1}. Similarly. the advantage arises only for entangled out that one can perfectly distinguish any unitary process from any states that have negative partial transpose and for coherent channels pure maximally entangled bipartite state. using the interventionist finding an rms deviation of 0.0 pexp and confirms that we can obtain a complete solution of the causal inference problem: Fig. either positively or negatively.6 11. 0. Figure 4b shows observational schemes.8 a universal NOT gate. The example from Table 1 makes use of entangled states and coher- (Fewer values of pexp were sampled in the interventionist scheme because ent channels.nature. they do not together span the coherence are therefore necessary for achieving the quantum advan- operator space L(HC ⊗ HD ). This is not an accident.8 1.024 in Discussion the interventionist scheme and 0.1038/NPHYS3266 ARTICLES a 1. when the measurements are implemented on the input and output of the identity channel E(·) = 1(·)1.NATURE PHYSICS DOI: 10. the correlations that are observed between A and cause or a purely cause–effect relation between two qubits.2 0.4 0. they B in the observational scheme may still contain a signature of the are also sufficient. The average fidelity of (97. which is not a completely positive map. thereby affording a complete solution of the open quantum system is Markovian or not17–24 . (3. If we are promised either a purely common- Nonetheless. although pare’) produce the same patterns of correlations under observational the measurements on C span the operator space L(HC ) and the schemes—they do not reveal the causal structure. For one to infer both the probability and the exact nature of the process instance. σ2 and σ3 are the Pauli matrices and |9 i ≡ (1/ 2)(|0i|1i ± |1i|0i). Every row of the table can be explained in this fashion. cause. positive correlation for all three observables admits of a cause–effect explanation—namely. In this case. The generally large to a Monte Carlo simulation.0 particular choice of process and state removes the sign ambiguity). 1). they can provide a test of whether the dynamics of a given and the bipartite state. in the product of the correlation coefficients. scenario is a probabilistic mixture of these two possibilities. as measured by the logarithm of the least-squares residue χ 2 with before.7 ± 0.0 8. u) ∈ {(1.75 Table 1 | Signatures of causal structure accessible by the observational scheme. that is. 0. with a root mean square (rms) deviation of 0.0 +1 +1 −1 −1 No ρ = |9 + ih9 + | 0. when the measurements are implemented on two qubits prepared in the singlet state ρ = |9 − ih9 − |.0 pexp Suppose the same Pauli observable is measured on two qubits A and B. (2.) that prepare separable states and cause–effect mechanisms that implement entanglement-breaking channels (‘measure and repre- as the one found in the measurement on C. Moreover. Colour encodes the quality source of noise is again the statistical fluctuations in pexp discussed of the fit.0 0. 3)}. and the The causal inference schemes described here promise extensive systems in question are qubits. with {|0i. 3b shows causal maps reconstructed from observational data. and fit to a mixed causal only slightly worse than with the interventionist scheme. Therefore. analysis recovers the correct value for the probability.2 0.1)% is on par Figure 4 | Indicators of causal structure determined by interventionist and with the results from the interventionist scheme.6 0. For simplicity. For causal structure.4 −1 +1 −1 +1 E(·) = σ2 (·)σ2 No −1 −1 +1 +1 E(·) = σ3 (·)σ3 No −1 −1 −1 −1 No ρ = |9 − ih9 − | 0. with probability pexp .032 using the observational scheme.4 |8± i ≡ 1/ 2(|0i|0i ± |1i|1i) are the Bell states.4 0. Notation: σ1 . The main structure with probability of common cause pfit . ln(χ2) 0. This shows that our demonstrates a quantum advantage. C11 · C22 · C√ ± 5 © 2015 Macmillan Publishers Limited. as we show in the Supplementary Methods. if the causal that are Choi-isomorphic to such states. then the observational scheme allows applications in experiments exhibiting quantum effects.5 C11 C22 C33 5v Cvv +1 +1 +1 +1 E(·) = 1(·)1 No pfit +1 −1 −1 +1 E(·) = σ1 (·)σ1 No 0. and suppose the outcomes k and m are found to be perfectly correlated. which can account for an rms deviation of 0. It turns higher-dimensional systems. Considering that causal inference values of χ 2 are due to uncertainty in the values of pexp implemented about two variables based on observation alone is impossible in during each measurement. We probe a probabilistic mixture of common the best fit for the mixing parameter p based on observational data. we also assume that the marginal b 1.8 1. thereby identifying the causal structure. therefore.2 −1 +1 +1 −1 No ρ = |8− ih8− | +1 −1 +1 −1 No ρ = |8+ ih8+ | 0.

M. Richardson. G. travelling anticlockwise. Reichenbach. When the controlled LCR implements the identity. This asymmetry results in the birefringence affecting the two 25. Weinstein. we can then reprepare it with another QWP and HWP non-completely positive maps. S. D.. 285–290 (1975). Phys. Bull. 97. Rev. H. P. There Characterization. D. S. paths differently. Rivas. downconverted pairs at 809. 77 (2014). S. entered. Janzing. J. Tollaksen. Rep. photon polarizations from D → F and E → B with probability p. both paths Lett. S. 052130 (2013). J-S.5 nm in a 10 mm periodically poled KTP (ppKTP) Soc. 86. 1092 (2012). & Scheines. T. T. 2c (refs 34. 6117–6133 (2004). Rev. Quantum process tomography and Lindblad estimation of a The experiments proceed as follows. A 77. Our inference schemes may also help to detect initial correlations between system and environment. J. Phil. with a transverse displacement. we separate the pump from the Phys. 7. M. Fitzsimons. L. |Ai polarization. Chiribella. we prepare the 11. 77. Methods 6. Flammia. Soc. R. Nature Phys. Lett. Rev. of California Press.. 2315–2323 (1994). B. & Sun. Huelga. The clockwise path encounters the LCRs after passing through Robust randomized benchmarking tests for non-unitality and both waveplates.. interferometer. Reduced dynamics need not be completely positive. © 2015 Macmillan Publishers Limited. If there is no phase difference between the two paths. 101. Costa. & Brukner. et al. & Wong. which act as a reservoir coincidence count rate at B and F is approximately 2. G. Rev.. Phys. J. Pechukas. 10002 (2012). Choi. 062115 (2010). & Spekkens. P. J. A 88. M.. Phys. fidelity of 98. We implement 19. M. 042113 (2008). Carteret. Cory. Lett. H-P.. et al. 012316 (2006). and |Li = (1/ 2)(|H i − i|V i). whilst the π phase shift implements 22. J. Tang. A 370. A 80. distribution (p. Havel. 8. Quantum tomography for measuring that only one particular eigenstate can pass. M. Liu.ARTICLES NATURE PHYSICS DOI: 10. the state of the system at some time produced the requisite state. value of the parameter p. What can be inferred from observational data when the (Cambridge Univ. Á. Phys. Jozsa. 121. Europhys. if unaccounted for. O. Fidelity for mixed quantum states. & Vaidman. Y. Kim. Theoretical framework for splitting into two components that travel in opposite directions in the Sagnac quantum networks. This probabilistic switching is implemented using a variable liquid crystal 931–934 (2011). 14. Light in modes D and E is then sent into the gate 31. Phys. Quantum Fisher information flow and the opposite side. Prediction. & Schölkopf. D. A 81. 121. Measuring non-Markovianity of processes with controllable second LCR is included and set to perform the identity operation for system-environment interaction. Phase-stable source of G . S. References 1. Phys. & Zyczkowski. which would extract the 17. Trans. the photon pair is entangled in polarization.. Completely positive linear maps on complex matrices. F. 2955–2961 (2004). D’Ariano. R. Appl. The experiment requires a gate G that can faithfully transmit the photon 18. can lead to errors in the characterization of processes25–30 . Fiorentino. P. Huelga. so the net effect of the gate is the swap. Incoherent noise random number generator in LabView. H. Rev. F. Rev. & Plenio. 150402 (2008). B.. accepted 29 January 2015. Opt. Mod. J.. J. Altepeter. T. Phys. quantification and detection. Rivas. 3. J. X-M. which are adjusted so 15... Emerson. J. Preprint at http://arxiv. C. M. better: interferometer. K. Chem. Mooij. X. Lu. Phys. Wang. dynamics. Howard. we measured in a completely non-destructive manner. Pearl. Wallman. F.. R. pick up the same phase shift. R. F. Assuming that the photon 30. J. The switching rate of the LCR is 5 Hz and data is 28. one path picks up a π phase shift with respect to Lett. Our some environmental degrees of freedom. Rev. We detect in coincidence to ensure that the source Phys. University of Washington. In one arm. centre wavelength 405 nm propagates through a polarizing beamsplitter (PBS). compensation. zero phase shift implements the identity. & Furuta. M.. G. Single World Intervention Graphs (SWIGs): We produce polarization-entangled photons using parametric downconversion in A Unification of the Counterfactual and Graphical Approaches to Causality a nonlinear crystal embedded in a Sagnac interferometer31–33 . polarization using a PBS preceded by a HWP and a QWP. T. 13. J. Ancilla-assisted quantum process tomography. J. the other. et al. Glymour. J. σ3 . Oreshkov. 042310 (2006). 41. Experimental control of the transition from Markovian to the swap. B. 6 NATURE PHYSICS | ADVANCE ONLINE PUBLICATION | www. et al. & Perinotti. Lett.000 Hz. Y. multiple-time measurements in quantum mechanics. √ |Di = (1/ 2)(|H i + |V i). 2000). non-Markovian dynamics of open quantum systems. On exiting the order. A 82. 1060–1062 (1994). 042103 (2010). Rev. we measure the polarization of C. Quantum correlations with no causal each component enters the crystal with horizontal polarization. et al. The Direction of Time (Univ. Physical accessibility of passes through the PBS. Prog. Two half-wave plates at 45◦ are inserted into each path of the 24. and F is directly detected. Rev. Phys. We use quantum state tomography to characterize the source and find an average 052110 (2009). J. & Lo Presti. yielding 25 samples for every transform. Multiple-time states and single-mode fibres after passing through a bandpass filter to reduce background. All rights reserved . M. Barnhill. Boulant.5% with |8+ i. if there is a π phase difference. 1 − p). Terno.. Each component produces degenerate type-II phase-matched 8. Quantum dynamics as an analog of conditional probability. C. which generates samples from the and quantum information processing. 105. S. Our results suggest several interesting avenues for future 2. Rev.. 2013). crystal. Quantum correlations which imply causation. P. A 23. Towards a formulation of quantum theory as a exact entangled state can be set using a HWP in the pump beam and a causally neutral theory of Bayesian inference. Assessing non-Markovian quantum eigenvalue of the desired Pauli observable while leaving the photon intact. 022339 (2009). G. B.. Á.nature.1038/NPHYS3266 a non-Markovian evolution. implements the phase gate. Phys. & Plenio. QWP and PBS. S. Quantum non-Markovianity: this with the displaced Sagnac interferometer shown in Fig. M. & Emerson. Lett. it exits at 21. M. Laine. whereas the anticlockwise path encounters the LCRs between non-Markovianity in quantum devices. J. Entanglement and non-Markovianity of light exits the interferometer at the same side of the beamsplitter at which it quantum evolutions. Leifer. Preprint at http://arxiv. There is a half-wave plate (HWP) placed on one side of the crystal so that 9. 4195–4198 (2001). Popescu. Am. published online 23 March 2015 which. 1.35). B 575. W. S. interferometer through the PBS. Chem. of information about the state of the system at some earlier time Received 20 June 2014. Rev. 1956). The output of the gate in mode B is detected using a HWP. M. 73. R. N. we obtain the same statistics as if 16. When it 26. Rev. Phys. Piilo.. & Robins. A 74. The LCR is controlled by the 27. The 10. Zscheischler. E-M.. observational class and what can be inferred from it? How does one 4. B-H. Phys. New J. Linear |Ri = (1/ 2)(|H i + i|V i). quarter-wave plate (QWP) in one of the exiting photon paths. downconverted light using a dichroic mirror. Causation. observable in different runs of the experiment. F.3773 (2014).org/abs/1412. state is not pure and the channel is not unitary? What is the most 3. wave plates to transmit either one or the other eigenstate of a given Pauli Phys. A 73. generalize such schemes from pairs of systems to arbitrary numbers Distinguishing cause from effect using observational data: Methods and benchmarks. 10.. Eisert.. but also by single-photon detectors and coincidence logic with a window of 3 ns. L. of systems? 5. A 79. N. By alternating the settings of the experimentally the matrix elements of an arbitrary quantum operation. Spirtes. Oeckl. faster. Phys. D’Ariano. 90. The photons are coupled into 12. polarization-entangled photons using a polarization Sagnac interferometer. The coincidence detection is performed using t2 is influenced not only directly by the state at time t1 . J.. 318–324 (2003). 050403 (2010). Aharonov. After preparing the entangled state on solid-state qubit. 29. 7. C. Reasoning and Inference research. t0 —the common cause. Cubitt.2731 (2013). 193601 (2003). the other 094001 (2014). modes C and E. M. are two distinct paths in the interferometer: one travelling clockwise. A ‘‘general boundary’’ formulation for quantum mechanics and Our measurement√ set consists of horizontal √ |H i and vertical |V i quantum gravity. V. Jones. so the net effect of the gate is the identity. M. Rev. Phys. & Cirac. R. If light is incident on both input ports of the interferometer. & Vedral. We measure Algebr. Simpler. A 10 mW laser with (CSSS. D. √ = (1/ 2)(|H i − |V i). J. and Search general sort of quantum measurement scheme that falls into the (MIT Press. Wolf. the non-Markovian processes of open systems. the 20. Nature Commun. & Breuer. 33 (2006). The operator tensor formulation of quantum theory. Causality: Models. Lett. 3385–3417 (2012). Press. Quantum process tomography of the quantum Fourier accumulated for 5 s for each measurement setting. M. in the desired state for mode D. Phys. Peters. Leifer. maximally entangled state |8+ i. J. 2000). Hardy.. C. retarder (LCR).org/abs/1302. whose birefringence can be controlled by an external voltage. the two waveplates. Measure for the non-Markovianity of polarizations directly from D → B and E → F with probability 1 − p and swap the quantum processes.

R. Kwiat. D. O’Brien. K.R. M. Nagata. Canada Research Competing financial interests Chairs.J.A. Science 316. & Takeuchi. and K.S. A. Biggerstaff. and R. J.. A.A. L. conceived the original idea for the project. 7 © 2015 Macmillan Publishers Limited. M. and R. Cluster-state quantum computing enhanced Author contributions by high-Fidelity generalized measurements.V. Sasaki.J.. NATURE PHYSICS | ADVANCE ONLINE PUBLICATION | www. J. M.1038/NPHYS3266 ARTICLES 32. Herbst. K. A. Research at The authors declare no competing financial interests..R. Acknowledgements Correspondence and requests for materials should be addressed to K. T.R.J. and K. T.NATURE PHYSICS DOI: 10. Phys.A.R.S.. N. M. D. J. Express 15. Mod. standard quantum limit with four-entangled photons. M. 15377–15386 (2007). Fedrizzi.R. Opt.nature. G. Poppe. analysed the results. and by the Province of Ontario through the Ministry of Research and Innovation. M. S. 257–266 (2000).. developed the 240504 (2009). and R. project and the theory. & White. D.W. We thank J.. and M.S. G. K. Lett. Industry Canada and the Canada Foundation for Innovation (CFI).W. Lavoie for valuable discussions. performed the numerical calculations. Opt. P. Donohue and J. Reprints and permissions information is available online at www. R. Okamoto. 103. 34. 726–729 (2007). Grover’s search performed the experiment. All rights reserved . Perimeter Institute is supported by the Government of Canada through Industry Canada A wavelength-tunable fiber-coupled source of narrowband entangled photons. and R. Additional information Supplementary information is available in the online version of the paper. R.R.R. 47. & Zeilinger.S..W. K.. designed the experiment. Mitchell. and L. Schwindt.A. wrote the first 35..J.W. D. Mazurek for his assistance in preparing the figures. P. T.. et al. K. 33.. algorithm: An optical approach. Beating the draft of the paper and all authors contributed to the final version. A.. K. This research was supported in part by the Natural Sciences and Engineering Research Council of Canada (NSERC).