acceptodds
Under review as a conference paper at ICLR 2027

Rank-1 Factored Gradient Descent without Restricted Isometry: Application to Quantum Pure State Tomography on Multi-Qubits

Abstract

Factored gradient descent (FGD) is a scalable nonconvex method for low-rank positive semidefinite matrix recovery, in which the rank and positivity constraints are satisfied automatically by the factorization. Most of the existing theory that certifies benignity of its loss landscape, however, relies on a restricted isometry property (RIP) of the measurement map. In quantum state tomography, one of its applications, the measurements that can be acquired with stable statistics are limited by detection efficiency and drift to Pauli observables of low weight; we first show that under such partial-weight Pauli measurements RIP fails in the worst possible way. We then bundle the mutually commuting observables setting by setting, pushing the state forward to the probability distribution of each setting, and introduce a max-sliced total-variation loss. On the stratum where the setting attaining the maximum is unique there is no spurious local minimum, and provided only that the measurement map is injective on generic pure states, the global minimum is the truth up to a phase. Finally we verify these claims numerically on four qubits and compare against FGD with the ordinary loss.

Primary area:
optimization
Keywords:
factored gradient descent, quantum state tomography, restricted isometry property, Pauli measurement, non-convex optimization, non-smooth analysis, sliced Wasserstein distance

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_FMLmOpenReview ↗
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
  1. 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.
  2. 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.
  3. 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
  1. 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?
  2. 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?
  3. Could the authors provide a simple worked example connecting the measurement data, predicted distributions, loss calculation, and parameter updates?
4 comments

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\ell^2 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\ell^2 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\ell^2. 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\ell^2 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 nn, and the relevant comparison is on the acquisition side. A full-weight setting needs a joint readout of all nn sites, with rate ηn\eta^{n}. A setting of Assumption 2.1 needs n/2n/2 sites, with rate ηn/2\eta^{n/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(2\sqrt3)^n. Computationally there are K=(nn/2)3n/2K=\binom{n}{n/2}3^{n/2} settings with 2n/22^{n/2} cells each (K=54K=54 for n=4n=4), one slice costs O(n2n)O(n2^n) and one iteration O(Kn2n)O(Kn2^n). The ℓ2\ell^2 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)\tau=(S,b) with S={1,3}S=\{1,3\} and b=(x,z)b=(x,z), each readout returns two bits, so the data are counts in four cells and their relative frequencies form qτq^\tau. For a candidate uu the Born rule gives pkτ(u)=u†Πkτup^\tau_k(u)=u^\dagger\Pi^\tau_k u on the same cells. The slice loss is ℓτ(u)=∑k∣pkτ(u)−qkτ∣\ell_\tau(u)=\sum_k|p^\tau_k(u)-q^\tau_k| and L(u)L(u) is the largest of these over the 54 settings. The update takes the maximizing setting τ⋆\tau_\star, forms g=2∑ksign⁡(pkτ⋆(u)−qkτ⋆)Πkτ⋆ug=2\sum_k\operatorname{sign}\big(p^{\tau_\star}_k(u)-q^{\tau_\star}_k\big)\Pi^{\tau_\star}_k u, moves uu against gg 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.

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:

  1. 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?
  2. 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 uu with positive loss, the rank-two curve Xt=[1−t u, t ψ]X_t = [\sqrt{1-t}\,u,\ \sqrt{t}\,\psi] gives ρ(Xt)=(1−t) uu∗+t ψψ∗\rho(X_t) = (1-t)\,uu^* + t\,\psi\psi^*. Every outcome probability is linear in ρ\rho, so every residual shrinks by the factor 1−t1-t, whether the setting is tied or not. An escape direction that does not use ψ\psi can be computed explicitly from the active settings. We expect that the landscape theorem extends to rank-rr 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 BB of the 54 settings is counted as B/54B/54 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.

MethodSuccess from one startCost, medianCost, mean
Least squares, L-BFGS, restart until success43/60 per start5579
Least squares, first-order (Polyak step)60/60176255
Max-sliced TV, all 54 settings (Polyak step ×1.95\times 1.95)60/6016162904
Max-sliced TV, 2 random settings per step (Polyak step ×1.8\times 1.8)60/60103113

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.

Reviewer_GPqDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper suggests a nonconvex gradient optimization over the manifold of rank 1 matrices, to support more efficient quantum state tomography. This is a quantum version of low rank matrix recovery. The particular difficulties that arise from having quantum states and Pauli measurements rather than general ones, are shown to be circumvented via a different error metric that is used as an objective. The overall running time is still at least 2^d where d is the number of qubits, since that is the representation that is extracted. Further, shadow tomography and improper learning via Frank Wolfe are not discussed.

Critical strengths

The main strength is overcoming the difficult of Pauli measurements and failures of the RIP condition which is usually used, with cleverly changing the objective. It is a solid theoretical result.

Critical weaknesses

The context is severely lacking: Shaddow tomorgraphy by Aaronson needs to be discussed as a viable and AFAIK superior alternative. Low rank matrix prediction is also studied via the Frank Wolfe algorithm and improper learning, as per: Near-Optimal Algorithms for Online Matrix Prediction, Elad Hazan, Satyen Kale, Shai Shalev-Shwartz

Questions for authors

Can you relate to shadow tomorograph, and explain why it is not superior? (at least there is no exponential in qubits dependence)

2 comments

Response to Reviewer GPqD

Thank you for the review and for the positive assessment of the theoretical result. We agree that the context was incomplete and answer your question directly.

1. Relation to shadow tomography

Shadow tomography (Aaronson) and classical shadows (Huang, Kueng and Preskill) address a different task from ours, so we would describe them as complementary and not as superior or inferior.

Those methods predict the expectation values of a given collection of observables without producing a description of the state. Their sample complexity is logarithmic in the number of observables for exactly this reason. Our task is full reconstruction of a pure state, that is, the output is the vector u∈Cdu\in\mathbb{C}^d with d=2nd=2^n itself. Any method with this output has cost at least 2n2^n, since the answer has that many amplitudes. This exponential dependence is therefore a property of the task and we do not claim to avoid it. If one needs only a prescribed set of expectation values, shadow methods are the right tool and our method is not.

There is a real point in common, and we should have stated it. Both lines of work rest on the same observation, that only local, low-weight readouts are reliably available, and both work with estimates of local quantities. In that sense the two fit together. Classical shadows provide estimates of low-weight Pauli expectations, which is precisely the data of Assumption 2.1, and our question starts where that leaves off, namely whether the global pure state can be recovered from such local data by a factored method when RIP fails. We will add this discussion to the related work.

2. Improper learning and Frank–Wolfe type methods

Thank you for the reference to Hazan, Kale and Shalev-Shwartz. As we understand it, that line of work concerns prediction of matrix entries with regret guarantees and allows improper hypotheses, that is, predictors outside the low-rank class. In our problem properness is essential. The target is a pure state, and as noted in Remark 1.4 the partial measurements determine the state only among pure states. The set of density matrices consistent with the data is in general not a single point, so a method that returns an element of the convex set, or a predictor outside the rank-one class, does not identify the state. This is the reason we optimize on the rank-one manifold directly. We will cite this work and explain the distinction.

3. Computational cost and limitations

We agree that cost should be stated explicitly. Any full reconstruction is exponential in nn, so the relevant comparison is on the acquisition side, and there partial-weight measurement scales considerably better. A full-weight setting needs a joint readout of all nn sites, with rate ηn\eta^{n} for per-site detection efficiency η\eta. A setting of Assumption 2.1 needs n/2n/2 sites, with rate ηn/2\eta^{n/2}, so the exponential suppression is reduced to its square root.

The number of settings is not reduced. For n=2Nn=2N qubits there are K=(nn/2) 3n/2K=\binom{n}{n/2}\,3^{n/2} settings with 2n/22^{n/2} cells each (K=54K=54 for n=4n=4). Asymptotically KK grows like (23)n(2\sqrt3)^n, which exceeds the 3n3^n settings of full product-basis tomography for n≥10n\ge 10. We use all subsets of size n/2n/2 and have not studied how far this family can be reduced. Computationally, one slice costs O(n2n)O(n2^n) and one iteration O(Kn2n)O(Kn2^n). The memory is 2n2^n amplitudes for the variable and K2n/2K2^{n/2} frequencies for the data, and no d×dd\times d density matrix is formed.

We also state the limitations plainly. The experiments are limited to n=4n=4 on synthetic data, and detection efficiency and drift are not simulated (Section 4). Publicly accessible quantum processors do not at present operate at a precision where the distinction we draw would be visible, so we could not provide an experimental demonstration. The target family of Section 4 is a pathological case constructed to trap the ℓ2\ell^2 loss. For generic targets such as Haar-random states, experiments made after submission show that the ℓ2\ell^2 loss with a suitable optimizer is sufficient, and we see no advantage of the proposed loss. On the theory side, the landscape result clears only the stratum where the maximizing slice is unique. The remaining difficulty is concentrated on the tie set, where we have no classification.

Reviewer_GPqDOpenReview ↗

thank you for your comments

they are very helpful, I have no further questions.

Reviewer_HVFZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Within the setting of Factored gradient descent (FGD), the authors show that under partial weight Pauli measurement matrices, RIP fails. Moreover, they introduce a total variation loss for which they show that under injectivity of the measurements map on pure states, the global minimum is the ground truth.

Critical strengths

There is novelty with respect to the measurement matrices.

Critical weaknesses

The problem is not new. You have to be careful (look at Boumal et al, Waldspurger, etc…). Perhaps Boumal et al is unconstrained. The set of measurement matrices considered is rather narrow.

Abstract:

  • The sentence “We then bundle the mutually commuting the observable …” is not clear
  • The term “positivity constraint” is usually reserved for matrices with non negative entries. I would replace by “positive semidefinite constraints

Introduction

  • line 077 : “The random Pauli measurements for which RIP is known to hold require building the measurement pool uniformly at random under the premise that all n qubits can be read out simultaneously” “the success rate of such a joint readout decays exponentially in n”—> does that appear in Liu ? if not a reference would be welcome.
  • lines 87 - 88 : “For the global solution to be confined to the truth” —> do you mean for the global minimum to be equal to the ground truth ?
  • since the search space of rank-1FGD is the manifold of pure states —> I would recommend tagging your equations using the names you use to refer to them. E.g. why not label (3) as QST and (4) as FGD ?
  • 86 - 87: Why do you say that the search space is the manifold of pure states ? To me the manifold of pure states is the rank one manifold while the search space from
  • lines 98 - 103, in your counterexample 1.1. I would not use the word “Hence” as it implies a conclusion. What you want to say is that the measurements are independent of \varphi so if you want to keep the “Hence” I would say something like “For any measurements of the form … Since the expectation is … \varpsi_0 and \varpsi_p cannot be distinguished”
  • In your counterexample what I don’t understand is that if the measurement setting lacks injectivity, why would one be interested in recovering the solution? Is it because as your work on the rank one manifold, you can get rid of the ambiguity associated to the rank two matrix X_\varphi ?
  • lines 97 - 106: In counterexample 1.1. : I think you have to be careful about the fact that you first introduce \varpsi_p as a tensor or order n then use it as a vector.
  • line 125 - 128 : When you say that the \ell_2 loss demands as a regularity condition on the landscape information about high order coherences ? could you give a reference on this ?
  • line 130-131, why do you require the sum \sum_{k} \pi_k^\tau = I ? What is the intuition behind this ?
  • line 142 : “This is an analogue of the max-sliced Wasserstein metric, except that the variables selecting the slice” —> I would remove this. This does not clarify anything.
  • line 142 : does || ||_1 mean total variation? Because in CS || ||_1 is usually used for the ell_1 norm. I would use _TV for total variation. Your use of distributions is not very clear either. One usually use the total variation when optimization is performed on measures.. but I don’t really understand where the probability measure appears here.. Why does it make sense to view the difference p^\tau - q^\tau as a measure ?
Questions for authors

see weaknesses

1 comment

Response to Reviewer HVFZ

Thank you for the careful reading.

1. Why partial-weight Pauli measurements (line 077)

This does not come from Liu (2011). It is an experimental constraint. The expectation of a Pauli string of weight ww is obtained from the joint outcome of ww detectors. With per-site detection efficiency η∈(0,1)\eta\in(0,1) the acquisition rate scales as ηw\eta^{w}, as stated in Figure 1 and Section 2.1. The RIP result for random Pauli measurements draws strings uniformly from all 4n4^n Paulis, so a typical string has weight about 3n/43n/4 and its statistics are suppressed exponentially in nn. Assumption 2.1 formalizes the regime that remains accessible, namely strings of weight at most NN on n=2Nn=2N qubits.

2. What Counterexample 1.1 says (lines 97–106 and 125–128)

In this regime the regularity condition behind FGD fails in the worst possible way, and this is so even for rank-one FGD on the sphere Sd\mathbb{S}_d. The states ψ0\psi_0 and ψφ\psi_\varphi have identical data under every partial measurement, so the rank-two matrix Xφ=ψ0ψ0†−ψφψφ†X_\varphi=\psi_0\psi_0^\dagger-\psi_\varphi\psi_\varphi^\dagger lies in the kernel of A\mathcal{A} and δ2≥1\delta_2\ge 1. The only string that separates them has weight nn. This is what we mean by saying that the ℓ2\ell^2 landscape guarantee demands information about high-order coherences. It is a consequence of the counterexample and not a statement from the literature.

Recovery is still meaningful because the failure occurs on an exceptional family. By Proposition 2.2 almost every pure state is uniquely determined among pure states (UDP). What fails is the uniform condition, not identifiability of a generic target.

3. Relation to landscape results for optimization on the sphere (Boumal et al., Waldspurger)

We agree that this literature must be discussed and will add it. Our search space is of the kind studied there. The variable is u∈Sdu\in\mathbb{S}_d, and modulo the global phase this is the manifold of rank-one positive semidefinite matrices of unit trace.

The difference is the role of the rank. In Burer–Monteiro theory the factor rank is a computational device, the target is a solution of the convex problem, and benignity is obtained when the rank is large relative to the number of constraints. Here rank one is the model itself. The partial measurements guarantee uniqueness only among pure states (UDP), not among all states, as noted in Remark 1.4. Recovering the solution set of the convex problem is therefore not sufficient, and in Counterexample 1.1 every mixture of the two states is also feasible. The mechanism that a rank-deficient second-order critical point is globally optimal gives nothing here, since a unit vector cannot be rank deficient.

Since the loss L(u)=max⁡τℓτ(u)L(u)=\max_\tau \ell_\tau(u) is nonsmooth, we work with Clarke stationarity on the sphere (Definition 3.1, Lemma 3.2). Theorem 3.5 classifies the stationary points on the stratum where the maximizing slice is unique. They are exactly the points satisfying the support condition, and each of them is a saddle with an explicit descent direction. Corollary 3.6 then confines spurious local minima to the tie set.

4. Total variation

For each setting τ\tau the projectors satisfy ∑kΠkτ=I\sum_k \Pi^\tau_k = I, which is completeness of a projective measurement. By the Born rule pτ(u)=(u†Πkτu)kp^\tau(u)=(u^\dagger \Pi^\tau_k u)_k is a probability vector over the cells, and ℓτ(u)=∥pτ(u)−qτ∥1\ell_\tau(u)=\|p^\tau(u)-q^\tau\|_1 is twice the total variation distance between two distributions on a finite set. We will change the notation to avoid confusion with the entrywise ℓ1\ell_1 norm.

5. A clarification on uniqueness (lines 83–85)

The results of Linden et al. and of Jones and Linden cited in lines 83–85 concern uniqueness among all states (UDA). The uniqueness that our analysis uses is the weaker one among pure states (UDP), given by Proposition 2.2 for the measurements of Assumption 2.1. The current text does not separate the two clearly, and we will do so.

Limitations

First, the paper motivates its setting by real experiments but validates it only on synthetic data with n=4n=4. Detection efficiency and drift are not simulated, as stated in Section 4. Publicly accessible quantum processors do not at present operate at a precision where the distinction we draw would be visible, so we could not provide an experimental demonstration.

Second, we have not removed the difficulty of the landscape. We have relocated it. Theorem 3.5 clears the stratum where the maximizer is unique, and what remains is concentrated on the tie set, where the loss is nonsmooth and where we have no classification. In the finite-shot experiment at R=103R=10^3 the max-sliced loss alone does not reach the threshold, and the two-stage scheme relies on ℓ2\ell^2 polishing.

We accept the remaining suggestions on wording and notation and thank you for them.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.