- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper studies pure quantum state reconstruction from partial Pauli measurements. The authors group the outcomes of each measurement setting into a probability distribution and propose a max-sliced total-variation loss, which minimizes the largest distribution discrepancy across settings instead of the usual least-squares objective. They prove that the proposed loss has no spurious local minima at points where the maximizing measurement setting is unique, while the case of multiple maximizing settings remains unresolved. Numerical experiments on four-qubit systems show improved reconstruction success on the tested target states. Under finite sampling, combining the proposed method with subsequent least-squares refinement yields better recovery performance than either method alone in the reported comparisons.
Critical strengths
The paper addresses a practically relevant problem. It studies pure quantum state reconstruction from limited local measurements, with a clear motivation grounded in measurement constraints. The approach has both theoretical and empirical support. The authors provide a conditional theoretical analysis of the proposed loss and compare different objectives using the same data and initializations, offering preliminary evidence of its effectiveness.
Critical weaknesses
- The paper builds on an existing pure-state parameterization and iterative optimization framework. The main change is to compare measurement outcome distributions and minimize the largest discrepancy across settings. The theoretical analysis provides partial support for this design, but its conclusions require a unique maximizing measurement setting, and no convergence guarantee is provided for the algorithm used in practice. It therefore remains unclear under what conditions the improvement can be expected consistently. A clearer explanation of the connection between the theoretical results and practical recovery performance would strengthen the paper.
- All experiments are limited to four-qubit systems, with target states primarily drawn from perturbations around a particular constructed state. Although the results demonstrate benefits on these examples, they provide limited evidence of generalization to other state families or larger systems. Experiments covering more diverse states and larger systems, together with runtime and computational cost comparisons, would make the evaluation more convincing.
- The paper introduces concepts from quantum measurement, matrix recovery, and nonsmooth optimization in quick succession, making the overall procedure difficult to follow. A simple worked example showing how measurement data are converted into distributions, how the loss is evaluated, and how the candidate state is updated would help. The advantages over the standard least-squares objective could also be explained more intuitively.
Questions for authors
- Under what conditions should the proposed loss be expected to improve recovery over least squares? How do the theoretical results explain the behavior of the algorithm, particularly when multiple measurement settings attain the maximum?
- Do the observed improvements extend to other target-state families and systems larger than four qubits? How do the runtime and computational costs compare with the baselines?
- Could the authors provide a simple worked example connecting the measurement data, predicted distributions, loss calculation, and parameter updates?
4 comments
Thank you for the response, and I apologize for the delayed response. I agree that studying the optimization landscape of pure-state reconstruction from partial measurements is worthwhile. Given the paper’s theoretical focus, I placed too much weight in my initial review on the four-qubit experiments and the lack of a full convergence proof. Now I think neither limitation is decisive on its own.
I would still like to see a stronger answer to at least one of two questions:
- When several measurement settings tie for the largest error, is there still a way to reduce the overall loss, or can the method get stuck at a wrong solution?
- On the same challenging target states, does the proposed method retain its advantage over least squares with a restarted quasi-Newton optimizer, using the same data and a comparable total computational budget that includes restarts?
A convincing answer to either question would make me to support acceptance of the paper as a useful first step.
Response to the two questions
Thank you for the follow-up. We answer both questions and try to be explicit about what does not work in our favor. All numbers are for the four-qubit hard target of the paper with exact data, 54 settings, and the same 60 random starts.
1. Ties
Tied local minima with positive loss do exist in the rank-one landscape of this target. A descent method started in their basins converges to them (loss between 0.04 and 0.06, with most of the 54 settings attaining the maximum simultaneously).
There is always a way to reduce the loss at such a point, by lifting the rank. For exact data and a unit vector with positive loss, the rank-two curve gives . Every outcome probability is linear in , so every residual shrinks by the factor , whether the setting is tied or not. An escape direction that does not use can be computed explicitly from the active settings. We expect that the landscape theorem extends to rank- factors by the same argument, although we have not written out the proof. If so, one can raise the rank each time the iteration is trapped, and at full rank the problem becomes a convex one. This route converges under a slightly stronger uniqueness assumption, namely that the target is the unique compatible state among all density matrices. The paper assumes uniqueness only among pure states.
We do not claim this as an advantage. It is not clearly better than solving the convex problem without a rank constraint, and the same lifting mechanism exists for least squares. The strength of the proposed loss lies only in the global landscape under the hard rank-one constraint.
We also examined a temporary lift: raise the rank by one, take steps at rank two, and then project back to a normalized rank-one vector. Starting from 24 tied local minima, this recovered the target in 24 of 24 cases with 3000 rank-two steps and in 18 of 24 with 1000 steps. This is empirical. Guaranteeing that the projection does not return the iterate to the original trap is a separate difficulty that we have not resolved.
2. Comparison with least squares under a comparable budget
Definition of cost. We measure cost by the number of evaluations, not by running time. One evaluation means computing the model predictions for all data terms together with the gradient. For least squares this is all 66 Pauli expectations. For the max-sliced TV loss this is the outcome probabilities of all 54 settings. A minibatch step that uses only of the 54 settings is counted as of an evaluation. The full evaluations used periodically for the stopping test are added to the count. For L-BFGS with restarts, the evaluations of all failed starts are included.
| Method | Success from one start | Cost, median | Cost, mean |
|---|---|---|---|
| Least squares, L-BFGS, restart until success | 43/60 per start | 55 | 79 |
| Least squares, first-order (Polyak step) | 60/60 | 176 | 255 |
| Max-sliced TV, all 54 settings (Polyak step ) | 60/60 | 1616 | 2904 |
| Max-sliced TV, 2 random settings per step (Polyak step ) | 60/60 | 103 | 113 |
L-BFGS is a descent method and is trapped in 17 of 60 starts, but its total cost including restarts is the lowest. The proposed method does not beat it. We have not tested how this comparison changes with the number of qubits, where full-batch evaluations and the number of restarts may weigh more.
Under the same first-order Polyak rule with all data evaluated at every step, the proposed method also loses to least squares, by a factor of about ten. One reason for this cost is that the maximum is computed over the losses of all settings at every step. The iteration should instead be minibatched over settings. With two random settings per step, the cost falls by a factor of about 25 without losing any run. This is below first-order least squares and about 1.4 times the cost of L-BFGS with restarts.
These results concern one target with exact data, and the step scaling and batch size were chosen after seeing the results. In our unoptimized Python implementation at four qubits, running time is dominated by per-iteration overhead, and by that measure the minibatched method is not faster than the full-batch one and both are slower than least squares.
Thank you for running the additional experiments and reporting the results candidly. I still think the paper addresses a worthwhile problem.
On the original challenging target state, the proposed method did not outperform the stronger least-squares baseline under the reported cost measure. The new experiments also clarify what can happen when several measurement settings tie for the largest error, but the behavior under the pure-state constraint remains open theoretically. For these reasons, I will keep my original score.
Response to Reviewer FMLm
Thank you for the accurate summary and the constructive questions.
1. Theory versus practice (Weakness 1, Question 1)
Our aim was theoretical. We looked for a formulation of rank-one factored recovery with a landscape guarantee that does not rely on RIP, since RIP fails for the partial-weight measurements that experiments can afford (Counterexample 1.1). For the measurement map of Assumption 2.1 we construct targets on which the ordinary ℓ2 loss is trapped at spurious points (25 of 100 trials in Table 1). For the max-sliced loss, Theorem 3.5 and Corollary 3.6 show that spurious local minima cannot occur where the maximizing slice is unique. They are confined to the tie set, which for generic data has Lebesgue measure zero.
This concentrates the difficulty and does not remove it. The tie set is small in measure, but descent trajectories are attracted to it. What we gain is an explicit form of the remaining obstruction. By Clarke regularity (Lemma 3.2), a point of the tie set is stationary exactly when zero lies in the convex hull of the subdifferentials of the active slices, which can be tested numerically by a small convex program. For the ℓ2 loss we know of no comparable description of the spurious points. We did not carry out this test in the submission, and a classification on the tie set is open.
In practice an exact tie is never attained in floating-point arithmetic. Near the tie set the active slice switches from step to step and the iterates oscillate instead of stopping (mentioned only in Section 5 for lack of space). The implementation suppresses the oscillation by running the Clarke subgradient method with momentum, which raises the success rate from 72 to 100 out of 100. There is no convergence guarantee for the method we use, and the theorem does not say how often the tie set is approached.
2. Other targets, larger systems and cost (Weakness 2, Question 2)
The target family of Section 4 was chosen because it creates traps for ℓ2. It is a pathological case in that sense, but not an artificial one. It is a neighbourhood of the ring cluster state, and such highly symmetric entangled states are among the physically most important targets, as resources for measurement-based quantum computation and error correction. The benefit of the max-sliced loss is visible there. For generic targets such as Haar-random states, experiments made after submission show that the ℓ2 loss with a suitable optimizer (quasi-Newton with restarts) is sufficient, and we see no advantage of the proposed loss. We have no evidence of an advantage at larger sizes.
On scalability, the cost of any full reconstruction is exponential in n, and the relevant comparison is on the acquisition side. A full-weight setting needs a joint readout of all n sites, with rate ηn. A setting of Assumption 2.1 needs n/2 sites, with rate ηn/2, so the exponential suppression is reduced to its square root. This is the sense in which partial-weight measurement scales considerably better. The number of settings is not reduced. It grows like (23)n. Computationally there are K=(n/2n)3n/2 settings with 2n/2 cells each (K=54 for n=4), one slice costs O(n2n) and one iteration O(Kn2n). The ℓ2 baseline uses the same data and has cost of the same order per iteration.
3. A worked example (Weakness 3, Question 3)
We agree that the presentation is too dense. The idea is to recast Pauli expectations as histograms of counts. A setting reads out a block of qubits in a fixed product basis, each readout returns a bit string, and the counts over bit strings form a histogram. By (7) and (8) this histogram carries every Pauli expectation that is diagonal in that basis, so nothing is lost, and the comparison of expectations becomes a comparison of two probability vectors.
As an example (Figure 2(a)), for τ=(S,b) with S={1,3} and b=(x,z), each readout returns two bits, so the data are counts in four cells and their relative frequencies form qτ. For a candidate u the Born rule gives pkτ(u)=u†Πkτu on the same cells. The slice loss is ℓτ(u)=∑k∣pkτ(u)−qkτ∣ and L(u) is the largest of these over the 54 settings. The update takes the maximizing setting τ⋆, forms g=2∑ksign(pkτ⋆(u)−qkτ⋆)Πkτ⋆u, moves u against g with the Polyak step and momentum, and renormalizes.
Intuitively, a sum can be stationary while individual settings still disagree with the data, because their gradients cancel. A maximum cannot hide a single disagreeing setting. Cancellation remains possible only among settings tied at the maximum.
The contribution is a landscape statement that holds without RIP, with an explicit description of where it stops. Its distance from practical performance is still large.