Abstract
Quantum-optimal discrimination between one and two closely separated light sources can be theoretically achieved by ideal spatial-mode demultiplexing, simply monitoring whether a photon is detected in a single antisymmetric mode. However, we show that for any imperfections of the demultiplexer, no matter how small, this simple statistical test becomes practically useless. While we identify a class of separation-independent tests with vanishing error probabilities in the limit of large numbers of detected photons, they are generally unreliable beyond that very limit. As a practical alternative, we propose a simple semi-separation-independent test, which provides a method for designing reliable experiments, through arbitrary control over the maximal probability of error.
Published by Optica Publishing Group under the terms of the Creative Commons Attribution 4.0 License. Further distribution of this work must maintain attribution to the author(s) and the published article's title, journal citation, and DOI.
1. Introduction
Statistical hypothesis testing is an important tool in the analysis of scientific data. A typical hypothesis testing problem in optical imaging is source discrimination, i.e., establishing whether an image originates from one or two light sources [1]. This is relevant in astronomy, e.g., for efficient exoplanet and binary stars detection [2–6], and fluorescence microscopy, e.g., for counting the exact number of molecules in a sample [7–9]. For source separations smaller than the width of the point spread function of the optical apparatus, the efficiency of source discrimination protocols based on spatially resolved intensity measurements, i.e., direct imaging, drops significantly [10,11]. In this sub-Rayleigh regime, source discrimination could be performed with the help of quantum-inspired measurement techniques [12].
Recently, motivated by the super-resolving power of spatial demultiplexing (SPADE) in the closely related problem of source separation estimation [13,14], it was shown that SPADE is also quantum-optimal in source discrimination [15], even when the sources are not point-like [16]. Furthermore, it was demonstrated that, due to the symmetry of the problem, detecting even just one photon in a fixed antisymmetric mode allows to accept one of the hypotheses with zero probability of error, leading to a near-optimal, separation-independent decision strategy [15]. These findings, however, were obtained assuming ideal measurements. In practice, the obtained results are significantly affected by experimental imperfections [17–23]. In particular, in the case of SPADE, misalignment, defects in the fabrication of the demultiplexer, and other imperfections induce a finite probability of detecting photons in the incorrect output, i.e., cross talk [24–26].
In this article, we show that cross talk has a strong impact on SPADE for discriminating between one and two equally bright sources in the practically relevant regime of small separations. In particular, we prove that the simple separation-independent test discussed above changes from being quantum-optimal in the ideal case, to be as good as flipping a coin in the presence of arbitrarily small cross talk. Moreover, we find that even though it is possible to design a class of meaningful separation-independent tests even in the presence of cross talk, the associated error probabilities are hard to predict without previous knowledge of the source separations. As an alternative, we propose a semi-separation-independent test with easily tractable maximal probability of error.
2. Hypotheses and Measurement Setting
We are interested in distinguishing between two hypotheses, H0 and H1, as illustrated in Fig. 1. According to hypothesis H1, two weak, incoherent light sources of equal brightness (e.g., faraway thermal sources) are separated by a distance $d$ in the object plane. The coordinate system is chosen in such a way that the sources’ positions are given by $\pm \vec {r}_0=\pm (d/2, 0)$. We consider a diffraction-limited imaging system with a Gaussian point spread function
so that the spatial distribution of the electromagnetic field in the image plane coming from a source at $\pm \vec {r}_0$ is given by $u_{00}(\vec {r}\mp \vec {r}_0)$ [27]. For weak sources, most of the photon detection events are single-photon events. Therefore, results in the image plane are effectively described as repeated measurements on $N$ copies of the single-photon state [13]3. Asymptotic Probability of Error
Effective source discrimination for sub-Rayleigh separations, i.e., for
requires a large number $N$ of samples. When $N\gg 1$, the probability of error minimized over all possible decision strategies for a specific measurement decays exponentially as $P_{\textrm {e}}^{\min }\sim \exp \{-N\xi \}$, where is the Chernoff exponent [28]. Here, $p(k|\textrm {H0})$ [$p(k|\textrm {H1})$] denote the probability of obtaining the measurement outcome $k$ conditioned on the hypothesis H0 [H1] being true.An emblematic measurement for source discrimination is direct imaging, which can be treated as a benchmark for new measurement schemes. The effectiveness of these measurements for discriminating between one and two weak incoherent sources could be judged based on the Chernoff exponent in Eq. (6). For ideal, i.e., continuous and noiseless, direct imaging, the latter is given by
However, in experimental settings, any measurement is unavoidably subject to apparatus misalignment, design and fabrication defects of the demultiplexer, and other imperfections. Accordingly, there is a small cross talk probability, i.e., it is possible that a measured photon is transmitted to an incorrect mode (see Fig. 1). Consequently, the real measurement basis deviates from the ideal one as
where $C$ stands for the (unitary) cross talk matrix, which is close to unity and $D$ restricts the number of measured modes [24]. We remark that the restriction of measurement to a finite number of modes requires renormalization of the probabilities in Eq. (12) within the chosen subset of possible outcomes. To quantify the severity of imperfections, one can use the cross talk strength $\epsilon ^2$, defined as the mean absolute value square of the off-diagonal elements of the cross talk matrix [24]: One can see that the cross talk strength $\epsilon ^2$ describes the average probability of cross talk from one measurement mode to another.We now proceed to assess the impact of cross talk on SPADE. We focus on the practically relevant regime of small separations, $x\ll 1$, and weak cross talk $\epsilon \ll 1$. As it will become apparent, in this regime, the optimal probability of error for ideal implementation of SPADE in Hermite–Gauss modes is achieved already with measurement restricted to modes with $n,m\in \lbrace 0,1\rbrace$. Therefore, for simplicity, from now on, we assume $D=2$. Accordingly, we can approximate the Chernoff exponent in Eq. (6) by a series expansion in these parameters. Analogously to previous findings for separation estimation [26], we find that the expansion depends on the ratio $x/\epsilon$ (see Supplement 1):
4. Practicality of Statistical Tests
The optimal decision strategy for a specific measurement is given by the likelihood-ratio test [30], according to which H1 is accepted if and only if
where $N_{k}$ is the number of events in which the outcome $k$ was obtained. Unfortunately, the likelihood ratio test has some drawbacks. Due to its assumption of a fixed separation, the latter has to be first estimated from the data, using, e.g., the method of moments [25]. Errors in this estimation inevitably deviate this decision strategy from its optimality, and more importantly can lead to underestimation of the probability of error, which is in addition hard to calculate. Finally, the optimal test in Eq. (17) requires measuring in many modes, which is not always feasible in experiments.A seemingly more practical test was introduced in Ref. [15]. Let us denote by $p_x$ the probability of measuring a photon in the mode $v_{10}$ conditioned on the separation $x$. The key observation behind this test is that, assuming ideal measurements, under H0 this probability takes the value $p_0=|C_{10,00}|^2=0$, while under H1 one has $p_x> 0$. In other words, photons can be measured in mode $v_{10}$ only if hypothesis H1 is true. This leads to the following test:
where whenever equality holds, one assumes H0. One can easily calculate that for such a test, the probability of accepting hypothesis H1 whenever H0 is true takes the value $\alpha (N)=0$, since when hypothesis H0 is true, it is impossible to detect photons in the mode $v_{10}$. Therefore, whenever H0 is true, the test in Eq. (18) always results in accepting H0 [15]. One can also find the probability of accepting hypothesis H0 when H1 is true: $\beta (N) = \left (1 - p_x\right )^N$, which is simply the probability of not detecting any photons in the mode $v_{10}$ across all $N$ trials. Clearly, both $\alpha (N)$ and $\beta (N)$ vanish for large photon numbers $N$ regardless of $x$, and therefore, so does the total probability of error. We thus have a test that is simple, separation-independent, and requires measuring in only one mode. Moreover, for small separations $x \ll 1$, this simple test is also quantum optimal, as we can show that Eq. (18) is in fact equivalent to Eq. (17) in this regime (see Supplement 1). Unfortunately, as we will now prove, this test goes from optimal to completely useless in the presence of any amount of cross talk, no matter how small.To see this, we observe that even for very weak cross talk $\epsilon \ll 1$, the probability $p_0$ of measuring a photon in mode $v_{10}$ under hypothesis H0 is no longer zero. As long as $p_0$ is non-zero, however small, the asymptotic probability of error is drastically changed. It is easily seen that $N_{10}$ for the test in Eq. (18) is a random variable with binomial distribution, with probabilities given by
The failure of the test in Eq. (18) does not necessarily disqualify all separation-independent tests. In fact, it is possible to design a class of separation-independent tests for source discrimination with imperfect demultiplexers that yields $\lim _{N\to \infty }P_{\textrm {e}}(N)=0$ (see Supplement 1). Such tests take the following form:
where $\zeta (N)>0$ are $x$-independent functions increasing faster than $\sqrt {N}$, but slower than $N$. The probabilities of error for such tests can be calculated analogously using a binomial distribution as in the case of Eq. (20) (see Supplement 1). Note that the natural generalization of the test in Eq. (18) given by Eq. (21) with $\zeta (N)=0$ results in the suboptimal $\lim _{N\to \infty }P_{\textrm {e}}(N)=1/4$ (see Supplement 1). Unfortunately, the family in Eq. (21) appears only marginally more practical than the original test in Eq. (18), as the corresponding rates of convergence of the probability of error to zero vary strongly with $x$, from nearly optimal to orders of magnitude worse [see Fig. 3(a)]. Accordingly, despite the tests being $x$-independent, a sensible estimation of their probability of error, for a given number $N$ of detected photons, would require some prior knowledge of $x$, severely undermining their practical utility.These findings motivate us to search for more practical tests for source discrimination. To achieve this goal, we go back to the optimal likelihood-ratio test in Eq. (17). Given that most information on the difference between the image of one source (H0) and that of two closely separated ones (H1) is contained in mode $v_{10}$, we rewrite Eq. (17) in terms of only two outcomes: photon in mode $v_{10}$ and photon in any other mode. This yields
Solving for $N_{10}$ and expanding to second order in $x$ and $\epsilon$, we obtain where $\gamma = 1 - \mathcal {O}(\epsilon ^2)$. The modified test in Eq. (23) is particularly appealing since (for small separations and weak cross talk) it inherits the optimality of the likelihood ratio test while being simple and based on a single-mode measurement, like the tests in Eqs. (18) and (21).On the downside, Eq. (23) is not separation-independent.To obviate to this problem, let us replace $x$ on the right-hand side of Eq. (23) by some $x_{\min }$ and then test for the modified hypotheses: (H0) there is only one source or (H1) there are two sources with separation $x\geqslant x_{\min }$. In this case, we will still obtain $\lim _{N\to \infty }P_{\textrm {e}}=0$ (see Supplement 1). Furthermore, for fixed $N$ and $x_{\min }$, $P_{\textrm {e}}$ is decreasing with growing $x$, meaning that the probability of error of the algorithm is upper bounded by $P_{\textrm {e}}$ calculated with $x=x_{\min }$. Therefore, the test in Eq. (23) is semi-separation-independent, in the sense that for a fixed $x_{\min }$, despite being not optimal for every value of $x$, it allows for an easy access to a maximal probability of error independent of any a priori knowledge of the separation. This, in particular, avoids any possible underestimation of the actual error probability $P_{\textrm {e}}$. All of this provides a reliable method of planning experiments: it is sufficient to set the minimal separation $x_{\min }$ into Eq. (23) to determine the number of photons $N$ to detect to be sure not to exceed a pre-established maximal tolerable probability of error. The minimal separation $x_{\min }$ can be dictated by multiple factors, such as physical constraints (e.g., atoms in molecules cannot be packed too tightly) or technical limitations of the experimental setups. Figure 3(b) shows a comparison between the real probability of error for this test and its upper bound. Additionally, we note that the semi-separation-independent test can outperform ideal direct imaging even for some separations $x\geq x_{\min }$. This can be seen in Fig. 3(c) where we present a comparison of the upper bound of the probability of error for semi-separation-independent test and probability of error for their ideal direct imaging estimated using the approximate Chernoff exponent in Eq. (10).
5. Conclusions
We have scrutinized the effectiveness of realistic SPADE-based discrimination between one and two closely separated light sources. We analytically showed that the presence of cross talk heavily affects the probability of successful discrimination, causing it to scale suboptimally with the source separation $d$ even for relatively large values of $d$. Similarly, any cross talk renders separation-independent hypothesis testing non-viable in practice. To remedy this, we proposed a simple semi-separation-independent algorithm based on the likelihood-ratio test, which, even for imperfect demultiplexers, gives access to the maximal probability of error without requiring separation estimation.
Our results suggest that it is mandatory to include the role of cross talk, and other experimental imperfection, e.g., electronic noise [21,25], in SPADE-based source discrimination. In particular, it would be interesting to see how significant experimental imperfections are in multiple-hypotheses testing [16], and when considering potentially unequal brightnesses of the sources [26]. Finally, our findings could be easily extended to the problem of pulse discrimination in the time-frequency domain [31,32]. In that context, demultiplexing can be performed using nonlinear optical devices [33], for which the finite width of the phase-matching function induces unavoidable cross talk.
Funding
Narodowe Centrum Nauki (UMO-2019/32/Z/ST2/00017); Agence Nationale de la Recherche; QuantERA (ApresSF); Horizon 2020 Framework Programme (731473); ERCIM (‘Alain Bensoussan’ Fellowship Programme).
Acknowledgments
Project ApresSF is supported by the National Science Centre (No. 2019/32/Z/ST2/00017), Poland; ANR under QuantERA, which has received funding from the European Union’s Horizon 2020 research and innovation programme under Grant Agreement No. 731473. This work was carried out during the tenure of an ERCIM ‘Alain Bensoussan’ Fellowship Programme.
Disclosures
The authors declare no conflicts of interest.
Data availability
No data were generated or analyzed in the presented research.
Supplemental document
See Supplement 1 for supporting content.
References
1. J. L. Harris, “Resolving power and decision theory,” J. Opt. Soc. Am. 54(5), 606–611 (1964). [CrossRef]
2. C. O. Acuna and J. Horowitz, “A statistical approach to the resolution of point sources,” J. App. Stat. 24(4), 421–436 (1997). [CrossRef]
3. M. Shahram and P. Milanfar, “Statistical and information-theoretic analysis of resolution in imaging,” IEEE Trans. Inform. Theory 52(8), 3411–3437 (2006). [CrossRef]
4. A. Labeyrie, S. G. Lipson, and P. Nisenson, An Introduction to Optical Stellar Interferometry (Cambridge University Press, 2006).
5. Z. Huang and C. Lupo, “Quantum hypothesis testing for exoplanet detection,” Phys. Rev. Lett. 127(13), 130502 (2021). [CrossRef]
6. U. Zanforlin, C. Lupo, P. W. R. Connolly, P. Kok, G. S. Buller, and Z. Huang, “Optical quantum super-resolution imaging and hypothesis testing,” Nat. Commun. 13(1), 5373 (2022). [CrossRef]
7. S.-H. Lee, J. Y. Shin, A. Lee, and C. Bustamante, “Counting single photoactivatable fluorescent molecules by photoactivated localization microscopy (PALM),” Proc. Natl. Acad. Sci. 109(43), 17436–17441 (2012). [CrossRef]
8. X. Nan, E. A. Collisson, S. Lewis, J. Huang, T. M. Tamgüney, J. T. Liphardt, F. McCormick, J. W. Gray, and S. Chu, “Single-molecule superresolution imaging allows quantitative analysis of RAF multimer formation and signaling,” Proc. Natl. Acad. Sci. 110(46), 18519–18524 (2013). [CrossRef]
9. D. Baddeley and J. Bewersdorf, “Biological insight from super-resolution microscopy: what we can learn from localization-based images,” Annu. Rev. Biochem. 87(1), 965–989 (2018). [CrossRef]
10. A. J. den Dekker and A. van den Bos, “Resolution: a survey,” J. Opt. Soc. Am. A 14(3), 547–557 (1997). [CrossRef]
11. J. W. Goodman, Introduction to Fourier optics (Roberts and Company Publishers, 2005).
12. C. Helstrom, “Resolution of point sources of light as analyzed by quantum detection theory,” IEEE Trans. Inform. Theory 19(4), 389–398 (1973). [CrossRef]
13. M. Tsang, R. Nair, and X.-M. Lu, “Quantum theory of superresolution for two incoherent optical point sources,” Phys. Rev. X 6, 031033 (2016). [CrossRef]
14. M. Tsang, “Resolving starlight: a quantum perspective,” Contemp. Phys. 60(4), 279–298 (2019). [CrossRef]
15. X.-M. Lu, H. Krovi, R. Nair, S. Guha, and J. H. Shapiro, “Quantum-optimal detection of one-versus-two incoherent optical sources with arbitrary separation,” npj Quantum Inf. 4(1), 64 (2018). [CrossRef]
16. M. R. Grace and S. Guha, “Identifying objects at the quantum limit for superresolution imaging,” Phys. Rev. Lett. 129(18), 180502 (2022). [CrossRef]
17. R. Nichols, T. R. Bromley, L. A. Correa, and G. Adesso, “Practical quantum metrology in noisy environments,” Phys. Rev. A 94(4), 042101 (2016). [CrossRef]
18. C. Oh, S. Zhou, Y. Wong, and L. Jiang, “Quantum limits of superresolution in a noisy environment,” Phys. Rev. Lett. 126(12), 120502 (2021). [CrossRef]
19. G. Sorelli, M. Gessner, M. Walschaers, and N. Treps, “Optimal observables and estimators for practical superresolution imaging,” Phys. Rev. Lett. 127(12), 123604 (2021). [CrossRef]
20. C. Lupo, “Subwavelength quantum imaging with noisy detectors,” Phys. Rev. A 101(2), 022323 (2020). [CrossRef]
21. Y. L. Len, C. Datta, M. Parniak, and K. Banaszek, “Resolution limits of spatial mode demultiplexing with noisy detection,” Int. J. Quantum Inform. 18(01), 1941015 (2020). [CrossRef]
22. J. O. de Almeida, J. Kołodyński, C. Hirche, M. Lewenstein, and M. Skotiniotis, “Discrimination and estimation of incoherent sources under misalignment,” Phys. Rev. A 103(2), 022406 (2021). [CrossRef]
23. M. R. Grace, Z. Dutton, A. Ashok, and S. Guha, “Approaching quantum-limited imaging resolution without prior knowledge of the object location,” J. Opt. Soc. Am. A 37(8), 1288–1299 (2020). [CrossRef]
24. M. Gessner, C. Fabre, and N. Treps, “Superresolution limits from measurement crosstalk,” Phys. Rev. Lett. 125(10), 100501 (2020). [CrossRef]
25. G. Sorelli, M. Gessner, M. Walschaers, and N. Treps, “Moment-based superresolution: Formalism and applications,” Phys. Rev. A 104(3), 033515 (2021). [CrossRef]
26. T. Linowski, K. Schlichtholz, G. Sorelli, M. Gessner, M. Walschaers, N. Treps, and Ł. Rudnicki, “Application range of crosstalk-affected spatial demultiplexing for resolving separations between unbalanced sources,” New J. Phys. 25(10), 103050 (2023). [CrossRef]
27. J. W. Goodman, Statistical Optics (Wiley, 1985).
28. H. Chernoff, “A measure of asymptotic efficiency for tests of a hypothesis based on the sum of observations,” Ann. Math. Statist. 23(4), 493–507 (1952). [CrossRef]
29. T. Ogawa and M. Hayashi, “On error exponents in quantum hypothesis testing,” IEEE Trans. Inform. Theory 50(6), 1368–1372 (2004). [CrossRef]
30. H. L. Van Trees, Detection, Estimation, and Modulation Theory, Part I. (John Wiley and Sons, Inc., 2001).
31. J. M. Donohue, V. Ansari, J. Řeháček, Z. Hradil, B. Stoklasa, M. Paúr, L. L. Sánchez-Soto, and C. Silberhorn, “Quantum-limited time-frequency estimation through mode-selective photon measurement,” Phys. Rev. Lett. 121(9), 090501 (2018). [CrossRef]
32. M. Mazelanik, A. Leszczyński, and M. Parniak, “Optical-domain spectral super-resolution via a quantum-memory-based time-frequency processor,” Nat. Commun. 13(1), 691 (2022). [CrossRef]
33. L. Serino, J. Gil-Lopez, M. Stefszky, R. Ricken, C. Eigner, B. Brecht, and C. Silberhorn, “Realization of a multi-output quantum pulse gate for decoding high-dimensional temporal modes of single-photon states,” PRX Quantum 4(2), 020306 (2023). [CrossRef]