Extreme dimensional compression with quantum modelling

by   Thomas J. Elliott, et al.

Effective and efficient forecasting relies on identification of the relevant information contained in past observations – the predictive features – and isolating it from the rest. When the future of a process bears a strong dependence on its behaviour far into the past, there are many such features to store, necessitating complex models with extensive memories. Here, we highlight a family of stochastic processes whose minimal classical models must devote unboundedly many bits to tracking the past. For this family, we identify quantum models of equal accuracy that can store all relevant information within a single two-dimensional quantum system (qubit). This represents the ultimate limit of quantum compression and highlights an immense practical advantage of quantum technologies for the forecasting and simulation of complex systems.



There are no comments yet.


page 4


Thermal Efficiency of Quantum Memory Compression

Quantum coherence allows for reduced-memory simulators of classical proc...

Robust inference of memory structure for efficient quantum modelling of stochastic processes

A growing body of work has established the modelling of stochastic proce...

Quantum coarse-graining for extreme dimension reduction in modelling stochastic temporal dynamics

Stochastic modelling of complex systems plays an essential, yet often co...

Strong and Weak Optimizations in Classical and Quantum Models of Stochastic Processes

Among the predictive hidden Markov models that describe a given stochast...

Quantum adaptive agents with efficient long-term memories

Central to the success of adaptive systems is their ability to interpret...

Optimality in Quantum Data Compression using Dynamical Entropy

In this article we study lossless compression of strings of pure quantum...

From Quantum Source Compression to Quantum Thermodynamics

This thesis addresses problems in the field of quantum information theor...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

Technical Appendix

Truncated dual Poisson processes. As noted above, remains infinite at any level of discretisation, as there is no maximum that the processes cannot exceed, and no merging of different into the same causal state. However, keeping this infinite overhead of states provides very little additional predictive power at large , as the probability of the process reaching these states is small, and the difference in conditional probabilities between and is monotonically decreasing towards zero with increasing . We therefore introduce a truncated form of the process, where there is a designated ‘terminal’ state at , such that all states at are merged down into this state. This terminal state must have transition probabilities that are a weighted average of all the merged states; when an event happens the model transitions to as before, but now on non-events the system remains in the state, as opposed to advancing further.

There is a level of subjective choice in how is selected. Appropriate methods can be for example based on the fidelity of the conditional distributions at larger , or on the probabilities of reaching such states. For concreteness, we pick a straightforward criterion:


That is, is the first state for which the probability of reaching said state is less than a fraction of the probability of decay in the first timestep. We here use .

Quantum model construction. As described in the main text, a unitary and set of quantum memory states that satisfy Eq. (Extreme dimensional compression with quantum modelling) form a quantum causal model of the dual Poisson process at a particular set of parameters. Here we show that the memory states given by Eq. (4) form such a set of states, and give the corresponding . This will prove our main result.

We begin by postulating a unitary operator that is stipulated to act in the following manner on two (non-orthogonal) states that we refer to as ‘generator’ states:


with , and we refer to as the ‘reset’ state. We define the overlap of the two generator states , noting that it depends on the size of the timesteps . Without loss of generality, we can enforce that this quantity be both real and positive. Analogous to current systematic approaches for constructing quantum causal models Binder et al. (2018); Liu et al. (2019) we utilise the relation , which arises from the properties of unitary operators. From this, we obtain


Armed with this, we can now express the generator states in the computational basis of a qubit . Without loss of generality, we can assign


We see that after emitting a 1, the memory always transitions to the reset state ; it defines the state to which the memory returns after the occurence of an event. This thus corresponds to the memory state . The remaining can be obtained by applying the unitary to the reset state times and post-selecting on the probe being measured as after each application: , where is the projector onto the non-event subspace of the probe. Accounting for normalisation, we obtain


By inserting Eq. (Technical Appendix) into Eq. (9) we recover the the memory states as prescribed in Eq. (4). We can similarly express in the computational basis using Eqs. (6) and (Technical Appendix):




The remaining two columns describing the action and are not uniquely defined by the model, and remain a free choice provided the unitarity of is upheld. We can also construct a pair of Kraus operators that describe the effective evolution of the memory conditional on the observed output and may be used to update the memory when tracking an external system that behaves according to the process. These operators are defined according to , where the states in this expression belong to the probe subspace Nielsen and Chuang (2000); they can be readily obtained from Eqs. (Technical Appendix) and (Technical Appendix).

Finally, we verify that this model produces the correct survival probability for the process Eq. (1). This is found from the probability of recovering a contiguous string of 0s after starting from the reset state:


Thus, our construction faithfully replicates the process, and may be used to track the dynamics of any dual Poisson process, at any level of discretisation, thus proving our main result. Notably, our quantum models are free from the need to introduce a truncation at long times as was done in the classical case.

As would be expected, the quantum memory states Eq. (9) and Eqs. (Technical Appendix) and (Technical Appendix) depend on the particular parameters defining the specific dual Poisson process to be modelled. Nevertheless, once initialised in a particular memory state corresponding to our observed past (which, being a qubit state, can always be prepared with at most three rotations from – corresponding to rotations of a qubit around the and axis of the Bloch sphere Vatan and Williams (2004)), the model operates by repeated applications of the same unitary , each followed by measurement and reset of the ancilla to simulate the future statistics. Being a two-qubit unitary, can always be synthesised by at most fifteen rotations from and three CNOT gates Kraus and Cirac (2001); Vatan and Williams (2004), irrespective of the parameters.

Considering current state-of-the-art ion trap experiments with gate fidelities in excess of 99.9% for two qubit gates, and 99.99% for single qubit gates Harty et al. (2014); Ballance et al. (2016)

, we can naively estimate their fidelity in implementing

as being at least 99.5% when multiplying the fidelity in implementing the elementary gates (i.e., ). This estimate is likely too pessimistic, as it neglects that some single qubit gates are applied in parallel to different qubits, and that the errors can act in opposing directions to partially cancel each other out. Nevertheless, we can use this as a back-of-the-envelope benchmark for implementability. Using the parameters of Fig. 2 () with , from Eqs. (Technical Appendix) and (Technical Appendix) we obtain



indicates a ’free’ value to be chosen subject to the constraint that the columns form mutually orthogonal vectors. By inspection, we can see that any implementation of this with at least 99.5% fidelity must capture the salient features to a high degree of accuracy. At this value of

the quantum scaling advantage is already clearly visible, and so is practical to verify with current ion trap experiments.

Calculating statistical complexity and quantum statistical memory. Prior work on the computational mechanics of renewal processes Marzen and Crutchfield (2015) has established that the steady-state probabilities of the causal states are proportional to their survival probabilities. That is, , where . For dual Poisson processes,


Thus, we have that and


where . Further, using Eq. (4) with we obtain


which may be straightforwardly diagonalised to find the two eigenvalues

, and hence calculate .


This work was funded by the Lee Kuan Yew Endowment Fund (Postdoctoral Fellowship), grant FQXi-RFP-1809 from the Foundational Questions Institute and Fetzer Franklin Fund (a donor advised fund of Silicon Valley Community Foundation), Singapore Ministry of Education Tier 1 grant RG190/17, National Research Foundation Fellowship NRF-NRFF2016-02, and National Research Foundation and Agence Nationale de la Recherche joint Project No. NRF2017-NRFANR004 VanQuTe . T.J.E., C.Y., and F.C.B. thank the Centre for Quantum Technologies for their hospitality.


  • Thompson et al. (2018) J. Thompson, A. J. P. Garner, J. R. Mahoney, J. P. Crutchfield, V. Vedral, and M. Gu, Physical Review X 8, 031013 (2018).
  • Loomis and Crutchfield (2018) S. Loomis and J. P. Crutchfield, arXiv:1808.08639 (2018).
  • Liu et al. (2019) Q. Liu, T. J. Elliott, F. C. Binder, C. Di Franco, and M. Gu, Physical Review A 99, 062110 (2019).
  • Ghafari et al. (2018) F. Ghafari, N. Tischler, J. Thompson, M. Gu, L. K. Shalm, V. B. Verma, S. W. Nam, R. B. Patel, H. M. Wiseman, and G. J. Pryde, arXiv:1812.04251 (2018).
  • Monras and Winter (2016) A. Monras and A. Winter, Journal of Mathematical Physics 57, 015219 (2016).
  • Cabello et al. (2016) A. Cabello, M. Gu, O. Gühne, J.-Å. Larsson, and K. Wiesner, Physical Review A 94, 052127 (2016).
  • Cabello et al. (2018) A. Cabello, M. Gu, O. Gühne, and Z.-P. Xu, Physical Review Letters 120, 130401 (2018).
  • Warszawski and Wiseman (2019) P. Warszawski and H. M. Wiseman, arXiv:1905.10935 (2019).
  • Gu et al. (2012) M. Gu, K. Wiesner, E. Rieper, and V. Vedral, Nature Communications 3, 762 (2012).
  • Mahoney et al. (2016) J. R. Mahoney, C. Aghamohammadi, and J. P. Crutchfield, Scientific Reports 6, 20495 (2016).
  • Palsson et al. (2017) M. S. Palsson, M. Gu, J. Ho, H. M. Wiseman, and G. J. Pryde, Science Advances 3, e1601302 (2017).
  • Garner et al. (2017) A. J. P. Garner, Q. Liu, J. Thompson, V. Vedral, et al., New Journal of Physics 19, 103009 (2017).
  • Aghamohammadi et al. (2017) C. Aghamohammadi, J. R. Mahoney, and J. P. Crutchfield, Scientific Reports 7 (2017).
  • Elliott and Gu (2018) T. J. Elliott and M. Gu, npj Quantum Information 4, 18 (2018).
  • Elliott et al. (2019) T. J. Elliott, A. J. P. Garner, and M. Gu, New Journal of Physics 21, 013021 (2019).
  • Khintchine (1934) A. Khintchine, Mathematische Annalen 109, 604 (1934).
  • Crutchfield and Young (1989) J. P. Crutchfield and K. Young, Physical Review Letters 63, 105 (1989).
  • Shalizi and Crutchfield (2001) C. R. Shalizi and J. P. Crutchfield, Journal of Statistical Physics 104, 817 (2001).
  • Crutchfield (2012) J. P. Crutchfield, Nature Physics 8, 17 (2012).
  • Cassandra et al. (1994) A. R. Cassandra, L. P. Kaelbling, and M. L. Littman, in

    Proceedings of the Twelfth AAAI National Conference on Artificial Intelligence

    (AAAI Press, 1994), pp. 1023–1028.
  • Kaelbling et al. (1996) L. P. Kaelbling, M. L. Littman, and A. W. Moore, Journal of artificial intelligence research 4, 237 (1996).
  • Haslinger et al. (2010) R. Haslinger, K. L. Klinkner, and C. R. Shalizi, Neural Computation 22, 121 (2010).
  • Marzen et al. (2015) S. E. Marzen, M. R. DeWeese, and J. P. Crutchfield, Frontiers in Computational Neuroscience 9, 105 (2015).
  • Li et al. (2008) C.-B. Li, H. Yang, and T. Komatsuzaki, Proceedings of the National Academy of Sciences 105, 536 (2008).
  • Kelly et al. (2012) D. Kelly, M. Dillingham, A. Hudson, and K. Wiesner, PloS one 7, e29703 (2012).
  • Park et al. (2007) J. B. Park, J. W. Lee, J.-S. Yang, H.-H. Jo, and H.-T. Moon, Physica A: Statistical Mechanics and its Applications 379, 179 (2007).
  • Clarke et al. (2003) R. W. Clarke, M. P. Freeman, and N. W. Watkins, Physical Review E 67, 016203 (2003).
  • Palmer et al. (2000) A. J. Palmer, C. W. Fairall, and W. A. Brewer, IEEE Transactions on Geoscience and Remote Sensing 38, 2056 (2000).
  • Varn et al. (2002) D. P. Varn, G. S. Canright, and J. P. Crutchfield, Physical Review B 66, 174110 (2002).
  • Ghafari Jouneghani et al. (2017) F. Ghafari Jouneghani, M. Gu, J. Ho, J. Thompson, W. Y. Suen, H. M. Wiseman, and G. J. Pryde, arXiv:1711.03661 (2017).
  • Binder et al. (2018) F. C. Binder, J. Thompson, and M. Gu, Physical Review Letters 120, 240502 (2018).
  • Aghamohammadi et al. (2016) C. Aghamohammadi, J. R. Mahoney, and J. P. Crutchfield, Physics Letters A 381, 1223 (2016).
  • Riechers et al. (2016) P. M. Riechers, J. R. Mahoney, C. Aghamohammadi, and J. P. Crutchfield, Physical Review A 93, 052317 (2016).
  • Suen et al. (2017) W. Y. Suen, J. Thompson, A. J. P. Garner, V. Vedral, and M. Gu, Quantum 1, 25 (2017).
  • Thompson et al. (2017) J. Thompson, A. J. P. Garner, V. Vedral, and M. Gu, npj Quantum Information 3, 6 (2017).
  • Aghamohammadi et al. (2018) C. Aghamohammadi, S. P. Loomis, J. R. Mahoney, and J. P. Crutchfield, Physical Review X 8, 011025 (2018).
  • Yang et al. (2018) C. Yang, F. C. Binder, V. Narasimhachar, and M. Gu, Physical Review Letters 121, 260602 (2018).
  • Nielsen and Chuang (2000) M. A. Nielsen and I. Chuang, Quantum Computation and Quantum Information (2000).
  • Smith (1958) W. L. Smith, Journal of the Royal Statistical Society. Series B (Methodological) pp. 243–302 (1958).
  • Marzen and Crutchfield (2015) S. E. Marzen and J. P. Crutchfield, Entropy 17, 4891 (2015).
  • Marzen and Crutchfield (2017) S. Marzen and J. P. Crutchfield, Journal of Statistical Physics 168, 109 (2017).
  • Allen et al. (2017) J.-M. A. Allen, J. Barrett, D. C. Horsman, C. M. Lee, and R. W. Spekkens, Physical Review X 7, 031021 (2017).
  • Pearl (2000) J. Pearl, Causality: models, reasoning and inference, vol. 29 (Springer, 2000).
  • Woods et al. (2018) M. P. Woods, R. Silva, G. Pütz, S. Stupar, and R. Renner, arXiv:1806.00491 (2018).
  • Negnevitsky et al. (2018) V. Negnevitsky, M. Marinelli, K. Mehta, H.-Y. Lo, C. Flühmann, and J. P. Home, Nature 563, 527 (2018).
  • Renner (2008) R. Renner, International Journal of Quantum Information 6, 1 (2008).
  • Del Rio et al. (2011) L. Del Rio, J. Åberg, R. Renner, O. Dahlsten, and V. Vedral, Nature 474, 61 (2011).
  • Brandao and Datta (2011) F. G. S. L. Brandao and N. Datta, IEEE Transactions on Information Theory 57, 1754 (2011).
  • Wang and Renner (2012) L. Wang and R. Renner, Physical Review Letters 108, 200501 (2012).
  • Tomamichel and Hayashi (2013) M. Tomamichel and M. Hayashi, IEEE Transactions on Information Theory 59, 7693 (2013).
  • Åberg (2013) J. Åberg, Nature Communications 4, 1925 (2013).
  • Horodecki and Oppenheim (2013) M. Horodecki and J. Oppenheim, Nature Communications 4, 2059 (2013).
  • Halpern et al. (2015) N. Y. Halpern, A. J. P. Garner, O. C. O. Dahlsten, and V. Vedral, New Journal of Physics 17, 095003 (2015).
  • Tomamichel (2015) M. Tomamichel, Quantum Information Processing with Finite Resources: Mathematical Foundations, vol. 5 (Springer, 2015).
  • Korzekwa et al. (2016) K. Korzekwa, M. Lostaglio, J. Oppenheim, and D. Jennings, New Journal of Physics 18, 023045 (2016).
  • Regula et al. (2018) B. Regula, K. Fang, X. Wang, and G. Adesso, Physical Review Letters 121, 010401 (2018).
  • Vatan and Williams (2004) F. Vatan and C. Williams, Physical Review A 69, 032315 (2004).
  • Kraus and Cirac (2001) B. Kraus and J. I. Cirac, Physical Review A 63, 062309 (2001).
  • Harty et al. (2014) T. P. Harty, D. T. C. Allcock, C. J. Ballance, L. Guidoni, H. A. Janacek, N. M. Linke, D. N. Stacey, and D. M. Lucas, Physical Review Letters 113, 220501 (2014).
  • Ballance et al. (2016) C. J. Ballance, T. P. Harty, N. M. Linke, M. A. Sepiol, and D. M. Lucas, Physical Review Letters 117, 060504 (2016).