1. Introduction
The Pauli problem [
1] questions the possibility of deducing the theoretical quantum state (wave function) from the observed statistics of quantum measurements. The measurements are assumed to be ideal (i.e., infinitely accurate), and a reconstruction of the state, if at all possible, requires measuring several non-commuting operators (see for example [
2,
3,
4,
5,
6,
7] and Refs. therein).
Below we will consider the problem in a somewhat broader context. Quantum mechanics predicts probabilities of the outcomes of series of consecutive measurements by defining probability amplitudes for virtual histories (Feynman paths), followed by the system. The path amplitudes are then added, as appropriate, and the absolute square of the sum gives the probability of a particular scenario to occur [
8]. The measurements usually considered in connection with the Pauli problem are, in fact two-step sequences, consisting of preparing the system in an initial state and measuring the chosen variable later. The case of a preparation followed by two or more measurements made on the system is richer, since it allows for a different type of interference not available to the two-step histories.
In this paper we consider a different “Pauli problem”, namely the possibility of recovering the system’s path amplitudes from the results of intermediate fuzzy measurements, and discuss how it can be done in practice. The rest of the paper is organised as follows. In 
Section 2 we revisit the basic rules for constructing probabilities with the help of virtual (Feynman) paths. In 
Section 3 we apply the recipe of 
Section 2 to a composite 
system + 
probe, and explain why the two-step histories are insufficient for our purpose. In 
Section 4 we extend the approach to three-step histories where a system is “pre- and post-selected” in known initial and final states. In 
Section 5 and 
Section 6 we apply the method to the simplest case of a two-level system (a qubit), and in 
Section 7 provide a numerical simulation. 
Section 8 discusses the “strong” and “weak” limiting cases where the approach of 
Section 4, 
Section 5, 
Section 6 and 
Section 7 fails, and analyses the reason for the failure. In 
Section 9 we evaluate the “strong” and “weak” averages of the pointer’s readings, and briefly comment on the popular subject of the so-called “weak measurements” [
9]. Usefulness of the paths amplitudes for predicting the outcomes of future measurements, and making statements about the system’s past is analysed in 
Section 10. 
Section 11 contains our conclusions.
  2. From Amplitudes to Probabilities
We start by recalling the rules for evaluating the probabilities of three consecutive measurements, given in [
8], and recently revisited in [
10,
11,
12]. Consider such measurements performed on a quantum system (
s), with which one associates a 
N-dimensional Hilbert space, 
. Three quantities, represented by operators 
, 
, are measured at 
, 
, and 
, respectively. Each operator has 
 possibly degenerate eigenvalues, 
, 
, and can be written as
      
      where 
, 
 form a suitable orthonormal basis, and 
 for 
, and 0 otherwise.
To be able to define a statistical ensemble, one needs the first measurement to yield a non-degenerate eigenvalue 
, which prepares the system in the corresponding state 
. The next step consists of evaluating the 
probability amplitudes for all virtual (Feynman) paths, 
, starting at 
 and passing through all possible states at 
 and 
 (see 
Figure 1),
      
      where 
 is the system’s evolution operator. An amplitude for obtaining at 
 a value 
 is found by summing (
2) over all states 
, consistent with 
, 
,
      
Finally the 
probability for having a sequence of observed outcomes 
 is found by summing absolute squares of the amplitudes (
3) over the degeneracies of the last eigenvalue 
,
      
Two relevant observations can be made here. Firstly, the scheme is explicitly causal in the sense that future observations cannot affect the statistics of the ones already made. In particular, summing (
4) over all outcomes 
 restores the probabilities for the experiment in which only 
 and 
 are measured,
      
For someone interested in the statistics of only the first two measurements it does not, therefore, matter what would be measured at , or if anything would be measured in the future at all.
Secondly, the scheme treats the “past” (any 
) and the “present” (at 
) differently. If the measured eigenvalue is degenerate, in the “past” one sums the amplitudes, as in Equation (
3). At “present”, the probabilities 
 are added, as in Equation (
4). The latter rule [
8] can be traced to the need to ensure causality in the case an operator commuting with 
 and having 
N distinct eigenvalues is measured in the immediate future at 
, 
 [
10].
The rules, although formulated for systems in a finite-dimensional Hilbert space are readily generalised to the case where the measured operators have continuous spectra. Note also that they can be used to obtain more compact expressions for the probabilities (
5) in terms of the projectors onto the eigen-subspaces of the chosen operators (see, e.g., Section 2.2 of [
12]).
  3. Von Neumann Measurements and the Two-Step Histories
The standard approach to quantum measurements involves coupling the system of interest to another degree of freedom (a probe), and deducing the system’s properties from the probe’s statistics. One choice of a probe is a von Neumann pointer (
p) [
13], a one dimensional free particle of a mass 
M with a coordinate 
f and a momentum 
. The pointer is briefly coupled to the system (
s) at some 
, so that the full Hamiltonian is given by
      
      where 
 refers to the system, the operator
      
      represents the system’s variable to be measured, and 
 is the Dirac delta. We will start by assuming that all eigenvalues of 
, 
, 
 are different, and return to degenerate 
’s at the end of the next Section.
A possible experiment consists of preparing the composite 
{system+
pointer} in a known initial state measuring 
, applying the coupling (
6), and then making a measurement on the pointer, 
. The purpose of the experiment is to learn something about the system in the absence of the probe. We, therefore, have a two-step history, which can be treated by the method of 
Section 2. At 
 the system and the pointer are prepared in states 
 and 
, respectively. In particular, we suppose that the outcome of 
 is 
, so
      
      with 
 and 
 acts only on the pointer.
Thus, obtaining an outcome 
 one prepares the composite in an initial state
      
In what follows, we will assume 
 to be a real symmetric Gaussian,
      
      where the width 
 determines the uncertainty in the pointer’s 
initial position and, therefore, affects the accuracy of the measurement.
To describe the pointer at  we will use a continuous basis , , and the measured operator, , with a discrete spectrum , . Thus, after observing an outcome  one knows that the (yet undefined) variable  has a value in an interval . The newly introduced parameter  determines the accuracy with which the pointer is read. Note that an eigenvalue  is highly degenerate, since , for any , and .
The evolution operator for the composite {
system + 
pointer} is a product
      
      where 
, and 
.
Following the recipe of 
Section 2 we write down the probability amplitudes for all virtual (Feynman) paths 
 in the composite’s Hilbert space, 
,
      
      where
      
      is the systems transition amplitude [
8] between the states 
 and 
, defined in the absence of the pointer, and the factor 
 can be written as
      
      since 
.
With only two measurements, all paths 
 lead to distinguishable (orthogonal) final states, and the probability to have a pointer reading 
 is found by adding absolute squares of the amplitudes (
13)
      
Therefore, regardless of how accurately the meter was prepared and read, all information about the phases of the amplitudes 
 is lost. This is because, according to the rules of 
Section 2 none of the virtual paths are allowed to interfere. For someone still wishing to determine the system’s initial state 
, the standard way to proceed is to choose a different operator 
, 
 repeat the measurement, and use the obtained data [
2,
3,
4,
5,
6,
7]. We will, however, consider a different problem, in order to exploit the interference associated with the measurements made in the “past”.
  4. From Probabilities to Amplitudes. Three-Step Histories
Suppose next that an additional measurement is made on the system at a . Now the measurement made on the pointer at  belongs to the past, and a different rule will apply.
The new experiment is as follows. At 
 the system and the pointer are prepared in a state 
, and coupled according to (
6) just before 
. At 
, a measurement made on the pointer yields an outcome 
. At 
 the outcome 
 is recorded, but only if a measurement 
 made on the system at 
 yields a particular outcome 
. The three steps are repeated enough times to evaluate the probabilities of having an outcome 
, given a later outcome 
. The purpose of the experiment is to recover the values of the system’s amplitudes, defined in the absence of the pointer.
This is a three-step measurement, for which we have
      
     We will assume the eigenvalues 
 to be non-degenerate provided 
 is acting in the Hilbert space 
 of the system. They are, however, highly degenerate, if 
 acts the 
, since 
, for any 
.
Next we apply the rules of 
Section 2. Evaluating the amplitudes for all virtual paths in 
, connecting 
 with 
, and 
 with 
 [cf. Equation (
2)], we find
      
      where
      
      is the amplitude for the system to follow a path 
 in 
.
Summing the amplitudes over the degeneracies of the operator 
 acting in the 
 [cf. Equation (
3)], yields
      
Finally, summing over the degeneracies of 
 in the 
 [cf. Equation (
4)], we have
      
      where 
 is the overlap matrix of the pointer’s states,
      
      and 
 is the probability density of the pointer’s readings, obtained for a system ending up in 
 at 
 (see 
Appendix A).
The measured system contributes to 
 with the path amplitudes given in (
21), whose values can, in principle, be determined by rewriting Equation (
23) in an equivalent form,
      
      where
      
     The system of linear equations (
25) can be solved if the probabilities 
 have been measured for 
 different intervals 
, 
. Solving Equation (
26) for 
 and 
, 
, one will be able to determine the values of all amplitudes 
 up to an unimportant overall phase.
To conclude the Section, we note that the measured operator 
 may have 
 degenerate eigenvalues,
      
      where 
 projects onto its 
j-th eigen-subspace. In this case the analysis remains the same, except that 
N is replaced by 
J, and the 
J amplitudes to be determined,
      
      result from the interference between the virtual paths (
21) not distinguished by a measurement of 
.
Next we see how the scheme would work in the simplest case of a two-level system, .
  5. An Inverse Measurement Problem
We can write Equation (
23) as
      
      where 
 is a complex “vector” with the components 
, 
, and 
 is an “operator” with matrix elements 
, 
, and the subindex 
 accounts for the specific subset of intervals used. The problem now takes a more familiar form. One needs to find the components of a (fictitious) state 
, given the expectation values of the hermitian operators 
, 
.
Equationation (
29) is particularly useful in the case 
, where 
 can be seen as an unnormalised state of a fictitious “spin”, and 
 can be expanded in terms of the Pauli matrices (
, 
, 
, and 
),
      
      with four intervals 
 corresponding to the pointer’s readings 
, chosen at one’s convenience. The resulting four equations (
29) determine the “spin”’s projections 
 onto the three spatial axes, as well as the state’s norm, 
,
      
      The polar angles 
 and 
 of the axis along which the “spin” is polarised,
      
      determine the “spin”’s state 
 up to an arbitrary overall phase and, returning to our original notations, we have the desired result,
      
      Next we see how the scheme will work in practice.
  6. Double-Slit Interference
It is natural to start with the simplest case, where one measures the final position of a massive pointer, 
, and
      
      which moves only when it interacts with the system, and whose state remains the same once this interaction is over [
13]. If so, the matrices 
 are real,
      
      and the coefficient multiplying 
 in Equation (
30) vanishes, 
. One can still solve any three of Equation (
31) for 
, 
, and 
 but would be unable to decide between 
 and 
, as illustrated in 
Figure 2. (Note that the problem is exacerbated for 
 where 
 signs would remain indeterminate when calculating the relative phases.)
Measuring instead the final pointer’s momentum 
, 
, one encounters a similar difficulty. In this case we have
      
      so that 
, and having solved the three remaining equations one will not be able to decide between 
 and 
 (see 
Figure 2).
However, provided both 
 and 
 have been measured independently in two different experiments, one can combine the results to obtain the four Equation (
31). For example, choosing any three equations employing 
 and 
, and one using 
 and 
, will determine the two amplitudes (
33) unambiguously (up to a global phase).
Finally we note that the case of a two-level system is conceptually similar to a Young’s double-slit experiment. Here the two states  and  play the roles of the two holes, and the target state , together with its orthogonal companion, ,  are the “points on the screen”. Unperturbed probabilities  and  correspond to having an “interference pattern on the screen”. Thus, if a von Neumann measurement perturbs the interference pattern (the probability to be detected in  is ), yet does not destroy it completely, and the values of the path amplitudes  can be deduced from the measurement’s statistics.
  7. A Simple Example
To see the efficiency of the proposed scheme, we first change it a little. It is possible to avoid mixing the results of measuring the pointer’s position and momentum, if the condition (
34) is relaxed, and the pointer’s initial state 
 is allowed to spread in the coordinate space. Choosing 
, from (
15) we have
      
     The appearance of a “complex width” 
 allows one to determine all coefficients 
 in Equation (
30) from the statistics of the final pointer’s positions. Explicitly, from (
24) we have
      
     In an actual experiment set-up to evaluate the amplitudes 
 a successful post-selection of the system in the final state, 
, will occur 
K times out of the total number of trials, 
. It is convenient to divide the full range of 
f into four regions, 
, 
, each containing a quarter of all cases, 
 (see 
Figure 3). To simulate the measurements, we use a random number generator, obtain four numbers 
, 
, 
, and 
, 
, replace the probabilities in the r.h.s. of Equation (
31) by their estimates,
      
      and solve Equation (
39) for different values of 
.
The results of three simulations for
      
      arbitrarily chosen (unnormalised) initial and final states,
      
      and
      
      are shown in 
Figure 4. It takes approximately 
 successful post-selections in order to recover the amplitudes of 
, given the values in Equation (
42). Next we discuss the two limiting cases, in which the method of this Section will fail.
  8. Accurate (Strong) and Inaccurate (Weak) Limits
The method of 
Section 8 fails, or at least becomes impractical as 
 or 
. The uncertainty in the initial pointer’s position determines the perturbation suffered by the measured system, as shown in 
Figure 5a, where the probability of detecting the system in the final state 
,
      
      is seen to vary from 
 to 
, as 
 increases from 0 to ∞. In both limits the matrix in the l.h.s. of Equation (
39) becomes singular, as shown in 
Figure 5b.
In particular, if the initial position of the pointer is highly uncertain, 
, from (
23) we have
      
     The pointer decouples from the system, and the interference between the two virtual paths is preserved,
      
     Note that the same effect can be achieved by leaving the width 
 constant, and multiplying the coupling term in Equation (
6), 
, by a constant 
. Indeed, scaling the pointer’s coordinate by choosing 
, would result in
      
If, on the other hand, the initial position of the pointer is known accurately, 
, Equation (
23) yields
      
      since by (
15)
      
      Thus, an accurately set pointer strongly perturbs the system by completely destroying interference between the virtual paths, even when the spreading of its initial state is taken into account.
  9. Averages and the “Weak Measurements”
Another possibility to explore the limits 
 and 
 is to evaluate the moments of the distribution of the pointer’s readings (
23),
      
These are, of course, also expressed in terms of the system’s path amplitudes 
, and we will look at the 
’s first moments in the “strong” and the “weak” limits. For simplicity, we will restore the condition (
34) or, what is the same, assume that the times 
 and 
 are so short that 
.
Expansions around 
 are not particularly interesting. Bearing in mind that 
, in the two-level case of 
Section 6, for 
 we have
      
      where 
 is the average, obtained in a highly accurate measurement, and the factors 
 and 
, which only depend on the parameters of the pointer and the eigenvalues 
, rapidly decrease for 
 (See 
Appendix B).
Similarly, since 
, for the mean pointer’s momentum, 
 we obtain
      
      where 
 is given in the Appendix B. Thus, some information about the relative phase of the two path amplitudes can be obtained from accurate yet not too accurate measurements. However, the expressions (
50) and (
51) are cumbersome and, as we already said, not particularly interesting.
The opposite limit 
 is involved in the controversy surrounding the so-called “weak measurements”. Returning to the notations of 
Section 3 and noting that 
, for the pointer’s mean position we have
      
      Similarly, for the mean pointer’s momentum we find
      
      where 
 is the variance of the distribution of the momenta in the initial pointer’s state 
,
      
      which vanishes when 
 becomes very broad in the coordinate space, and 
.
Equationations (
50)–(
53), although different, illustrate the same point. Any average, evaluated for a pointer coupled, as in Equation (
6), to a system making a transition between initial and final states will have to be expressed in terms of certain combinations of the system’s transition amplitudes, defined in the absence of the pointer. Transition amplitudes are the basic elements of the description of quantum motion [
8], and this is really all that can be said about this matter.
The above analysis relates to the so called “weak values” WV(for a recent review see [
9]). For short enough 
 and 
, 
, the quantity in the square brackets in Equations (
52) and (
53) reduces to a ratio of matrix elements
      
      Presented in this manner in [
14], the r.h.s of Equation (
55) was dubbed “the weak value of 
 for a system pre- and post-selected in its initial and final states”, which be can obtained in a particular kind of “weak quantum measurements”. Various weak values have been measured experimentally, yet their place and status within conventional quantum mechanics remain unclear. We have long advocated the interpretation of the “weak measurements” in terms of Feynman’s transition amplitudes, and refer the reader to [
15,
16,
17,
18] for an analysis of the role of the Uncertainty Principle and the significance of “anomalous weak values”. Here we further support this view by placing the “weak measurements” within a broader context of measuring the transition amplitudes, absent in the simple two-step histories of 
Section 2.
  10. Prediction and Retrodiction
Having recovered the transition amplitudes 
, it is reasonable to question the usefulness of what has been found. If the amplitudes are known for all 
, 
, they can be used to predict the results of other measurements made on the same system staring in the same 
, and ending in the same 
, provided the new operator 
 commutes with 
, 
. Indeed, their values is all that required to compute the probabilities 
 in Equation (
23), for any choice of 
, 
, 
 and 
, even if the Hamiltonian of the system, 
, is not known. The task is not entirely trivial for 
, where 
 can have degenerate eigenvalues, and the corresponding amplitudes must be added, as described in 
Section 2.
On the other hand, little can be learnt about an intermediate measurement of a 
 which does not commute with the 
. This is seen already from the 
 example, discussed in the previous Section. Suppose one replaces 
 with a 
, so that now
      
      and
      
     Of the four quantities in the r.h.s. of Equation (
57) needed to evaluate 
, only the first two are known from measuring the 
, and this is clearly not enough.
Another use of the path amplitudes (
21) and (
28) is retrodictive reconstruction of the system’s past. Classically, the knowledge of a system’s current position, velocity, and its Lagrangian is sufficient for predicting its position in the past. Quantally, one may wish to determine the system’s initial state, 
, from the values of the path amplitudes. The state is fully determined by the coefficients 
 of its expansion in some known basis 
, 
. If the system’s Hamiltonian 
 is known, the operator 
 has non-degenerate eigenvalues [cf. Equation (
7)], and the values of 
 have been measured, the problem is easily solved. Indeed, using (
21) we obtain
      
      and with 
 thus determined, the results of other possible measurements can be predicted.
However, if some of the 
’s eigenvalues are degenerate [cf. Equation (
27)], full reconstruction of the initial state is not possible, since important information is lost to interference. From (
28) we have
      
      so that the values of 
 cannot be recovered from the known values of 
.
  11. Conclusions
In summary, we have shown that the values of system’s transition amplitudes can be deduced from the statistics of an intermediate measurement. The deduction is possible provided the measurement is “fuzzy”, and does not destroy interference between the system’s virtual paths. What distinguishes our method from usual approach to the “Pauli problem” [
2,
3,
4,
5,
6,
7] is its reliance on a type of interference, absent in two-step histories, consisting only of preparation and the actual measurement. With the post-selection step added, the situation is conceptually similar to a double-(multiple-) slit experiment [
8], in which a probe, designed to determine the path taken by the system, does its job imperfectly, so a vestige of the interference pattern is retained on the screen. There is a two-way relationship between a result of observation and what can be considered a computational tool, although it is typically more difficult to deduce amplitudes from the probabilities than to construct the probabilities from the known amplitudes.
Finally, it is worth noticing that neither reconstructing the amplitudes as in 
Section 7, nor evaluating their combinations by measuring the “weak values” of 
Section 9, would serve to provide a deeper insight into quantum mechanical formalism. If asked “what has been evaluated?” one can only answer “amplitudes”. Moreover, if asked further “what are these amplitudes?” one can only reply “something quantum theory uses to predict the observable probabilities”.