1. Motivation (2–3 sentences): Matrix scaling via the Sinkhorn algorithm is fundamental in optimal transport and contingency table analysis. When the input matrix is random rather than deterministic, understanding the behavior of the rescaled matrix becomes crucial for applications in machine learning, statistics, and computational biology.

2. Mathematical setup: Let $X \in \mathbb{R}^{m \times n}_{\geq 0}$ be a random matrix with independent entries and positive mean $\Lambda = \mathbb{E}[X] \in \mathbb{R}^{m \times n}_{> 0}$. Given target margins $(r,c)$ with $r \in \mathbb{R}^m_{> 0}$, $c \in \mathbb{R}^n_{> 0}$, the Sinkhorn-rescaled random matrix is defined as:

   \[X^{r,c} = D_1(X) X D_2(X)\]

   where $D_1(X), D_2(X)$ are random diagonal scaling matrices such that $X^{r,c}$ has exact row sums $r$ and column sums $c$. The scaling factors are obtained by solving:

   \[X^{r,c} = \arg\min_{Z \in \mathcal{T}(r,c)} D_{KL}(Z \| X)\]

   where $\mathcal{T}(r,c)$ is the transportation polytope.

   Assumptions:

   1. $X$ has independent nonnegative entries with subexponential tails: $\mathbb{E}[

      X_{ij} - \Lambda_{ij}

      ^q] \leq \frac{q!}{2}\sigma^2 R^{q-2}$ for all $q \geq 2$

   2. The margin $(r,c)$ is $\delta$-smooth: $\frac{\delta N}{m} \leq r(i) \leq \frac{N}{\delta m}$ and $\frac{\delta N}{n} \leq c(j) \leq \frac{N}{\delta n}$

   3. Bounded cost condition: $

      \log\frac{r(i)c(j)|\Lambda|_1}{\Lambda_{ij} N^2}

      \leq K$

3. Toy example: When $m = n = 2$, $r = c = (1/2, 1/2)$, and $\Lambda = \begin{pmatrix} 0.3 & 0.2 \ 0.1 & 0.4 \end{pmatrix}$, the deterministic Sinkhorn scaling yields potentials $\alpha, \beta$ such that $\Lambda^{r,c} = \text{diag}(e^\alpha) \Lambda \text{diag}(e^\beta)$ has exact margins $(1/2, 1/2)$. The core difficulty is that when $X$ is random, the scaling factors $D_1(X), D_2(X)$ are nonlinear functions of $X$ without explicit form.

4. Formal objective: The main quantities to control are:

   \[\|(\alpha_X, \beta_X) - (\alpha, \beta)\|_\infty \quad \text{and} \quad d_H(\phi_{X^{r,c}}, W^{r,c;\Lambda})\]

   where $(\alpha_X, \beta_X)$ are random Schrödinger potentials, $(\alpha, \beta)$ are deterministic ones, and $W^{r,c;\Lambda}$ is the limiting static Schrödinger bridge density.

The method proceeds through a three-layer approximation scheme: $X^{r,c} \approx \hat{X}^{r,c} \approx \Lambda^{r,c}$ where $\hat{X}^{r,c} = D_1(\Lambda) X D_2(\Lambda)$ is a comparison model using deterministic scaling factors.

Key steps:

1. Establish quantitative stability bounds for the static Schrödinger bridge (SSB) under perturbations of reference measure and margins
2. Apply these bounds to show concentration of random Schrödinger potentials around deterministic ones
3. Use concentration to prove scaling limit convergence to continuous SSB

4. Develop fluctuation theory around the limit

   The core stability results are:

   • Kernel stability: For fixed margins,
   \[d_H(\pi_{\mu_0,\mu_1;R}, \pi_{\mu_0,\mu_1;R'}) \leq \exp\left(\frac{3}{2}\|\kappa_+\|_\infty \vee \|(\kappa')_+\|_\infty\right) d_H(R,R')\]
   • Margin stability: For fixed reference measure,
   \[d_H(\pi_{\mu_0,\mu_1;R}, \pi_{\mu'_0,\mu'_1;R})^2 \leq 4(\|\kappa\|_\infty \vee \|\kappa'\|_\infty)(\|\rho_r - \rho'_r\|_{L^1} + \|\rho_c - \rho'_c\|_{L^1})\]
   • Discrete potential stability: Under margin perturbation $\epsilon$,
   \[\|(\alpha_A, \beta_A)\|_\infty \leq 4C_A \epsilon\]

   where $C_A = 1 + \frac{9\max_i r_i \max_j c_j}{\rho_A mn}$ and $\rho_A$ measures row-alignment.

   Applied to toy example: For the $2 \times 2$ case, if $X$ has entries with variance $\sigma^2$, the random potentials concentrate around deterministic ones with error $O(\sigma\sqrt{\log n}/|\Lambda|_1)$.

Step 1 — Prior work:

• Eckstein-Nutz (2022): Established Hölder-$p/(p+1)$ stability of SSB under cost perturbations and Wasserstein bounds for margin perturbations
• Landa (2022): Proved concentration of scaling factors for strictly positive bounded random matrices with constants depending on min/max ratio
• Carlier-Laborde (2020): Local Lipschitz continuity of Schrödinger potentials via implicit function theorem

Step 2 — Delta: This paper provides:

• Lipschitz (not Hölder) kernel stability in Hellinger distance with explicit exponential constants
• Linear cost dependence for margin stability (vs exponential in kernel case)
• Discrete potential stability with row-alignment parameter $\rho_A$ instead of min/max ratio, allowing zero entries
• Complete scaling limit theory connecting discrete bridges to continuous SSB

Step 3 — Theory-specific assessment:

• The main theorems are predictable extensions of known stability techniques to the random matrix setting
• Proof techniques are routine assemblies of concentration inequalities, fixed-point arguments, and existing SSB theory
• The Lipschitz improvement over Hölder bounds is modest - achieved via standard operator-theoretic arguments on square-root densities
• No lower bounds are established; tightness of stability constants is unclear except for one example

The discrete-to-continuous scaling limit is expected given the stability framework. The bulk rigidity result follows standard random matrix theory via local laws for inhomogeneous Gram matrices.

Verdict: INCREMENTAL — solid technical execution extending known stability results to random matrices, but the main conclusions follow predictably from the deterministic theory.

The proofs use several key technical tools:

1. Fixed-point stability for discrete potentials: The Sinkhorn map is shown to be a contraction in a neighborhood of the correct scaling via the row-alignment parameter:

   \[C_A = 1 + \frac{9\max_i r_i \max_j c_j}{\rho_A mn}\]

   where $\rho_A := \min_{i_1,i_2} \sum_j A_{i_1 j} A_{i_2 j}$ measures overlap between rows.

2. Operator-theoretic stability for continuous SSB: The key insight is that the square-root density map

   \[\Gamma: \sqrt{\frac{dR}{d(\mu_0 \otimes \mu_1)}} \mapsto \sqrt{\frac{d\pi_{\mu_0,\mu_1;R}}{d(\mu_0 \otimes \mu_1)}}\]

   has bounded derivative as a linear operator, yielding Lipschitz stability:

   \[d_H(\pi_R, \pi_{R'}) \leq C e^{(3/2)\|\kappa_+\|_\infty} d_H(R, R')\]

3. Bernstein concentration with non-uniform scaling: The concentration bound uses:

   \[\mathbb{P}\left(\left\|\sum_{ij} a_{ij}(X_{ij} - \Lambda_{ij})\right\| \geq t\right) \leq 2\exp\left(-\frac{t^2/2}{\sigma^2 \sum_{ij} a_{ij}^2 + Rt/3}\right)\]

   Applied with scaling-dependent weights $a_{ij} = e^{\alpha(i)+\beta(j)}$.

4. Local law for inhomogeneous Gram matrices: For the fluctuation matrix $A$ with variance profile $S_{ij} = \frac{e^{2\alpha(i)+2\beta(j)}\text{Var}(X_{ij})}{(m+n)s_{\max}}$, the eigenvalue rigidity follows from the Dyson equation:

   \[-\frac{1}{m_i(z)} = z - \sum_{k=1}^n \frac{S_{ik}}{1 + \sum_{j=1}^m S_{jk} m_j(z)}\]

5. Delta method for CLT: The covariance of empirical Schrödinger potentials is:

   \[L^{\dagger} H \Sigma H^T (L^{\dagger})^T\]

   where $L$ is the Jacobian of margin constraints and $H$ maps matrix perturbations to margin changes.

Purely theoretical. The paper includes numerical illustrations (Figure 1) showing empirical singular value distributions of rescaled fluctuation matrices for different entry distributions (Poisson, Bernoulli, Exponential, Uniform) at dimensions $m = n = 5000$, comparing against theoretical predictions from the Dyson equation.

Empirical validation would require:

• Verification of concentration rates for Schrödinger potentials across different matrix dimensions and noise levels
• Testing scaling limit convergence for various margin sequences and reference measures
• Comparison of bulk rigidity predictions with sample covariance eigenvalue locations
• Validation of CLT predictions for empirical potentials with finite sample sizes

Limitations:

1. δ-smooth margins assumption - RESTRICTIVE: requires all margin entries to be bounded away from zero and infinity, significantly limiting applicability to sparse or highly unbalanced margins
2. Subexponential entry assumption - NATURAL: standard in random matrix theory and satisfied by most common distributions
3. Mass growth condition $|\Lambda|_1 \gg (m \vee n)^{3/2}$ - TECHNICAL: needed for concentration but likely removable with more refined analysis
4. Bounded cost assumption - RESTRICTIVE: limits the allowed mismatch between mean matrix and target margins

5. Independence of entries - TECHNICAL: standard assumption that could potentially be relaxed to martingale conditions

   Open problems:

6. Optimal concentration rates: Are the $O((m \vee n)^{3/2}\sqrt{\log(m \vee n)}/|\Lambda|_1)$ rates for potential concentration tight, or can they be improved to the information-theoretic rate $O(\sqrt{mn}/|\Lambda|_1)$?
7. Edge universality: The bulk rigidity result only covers eigenvalues in the interior of the support - can this be extended to establish Tracy-Widom fluctuations at the spectral edges for the rescaled covariance matrices?

Motivation: Environmental health research faces a critical challenge in evaluating causal effects of multivariate continuous exposures like air pollution mixtures. Traditional single-pollutant analyses fail to capture confounding and interaction effects, while standard deterministic interventions suffer from frequent positivity violations that render effects unidentified.

Mathematical setup:

• Observed data: $n$ i.i.d. samples ${Z_i}_{i=1}^n = {(X_i, W_i, Y_i)}_{i=1}^n$
• $X_i \in \mathcal{X} \subseteq \mathbb{R}^p$: covariates
• $W_i \in \mathcal{W} \subseteq \mathbb{R}^q$: $q$-dimensional exposure vector (environmental mixture)
• $Y_i \in \mathbb{R}$: scalar health outcome

• Conditional exposure density: $f(w

  x)$

Assumptions:

1. SUTVA: No interference and well-defined treatments

2. No unmeasured confounding: $Y(w) \perp W

   X$ for all $w \in \mathcal{W}$

3. Positivity violations occur for many exposure combinations

   Toy example: When $q=2$ with exposures $(W_1, W_2)$ representing PM2.5 components, certain combinations like high $W_1$ with low $W_2$ may never occur naturally due to emission source correlations, violating positivity for deterministic interventions setting $W = (w_1^*, w_2^*)$.

   Formal objective: Estimate the incremental causal effect under exponentially tilted interventions:

   \[\psi(\delta) = \mathbb{E}[Y^{g_\delta}]\]

   where $g_\delta(w

   x) = \frac{\exp(\delta^\top w)f(w

   x)}{\int_{\mathcal{W}} \exp(\delta^\top v)f(v

   x) dv}$

Core Method: Exponentially tilted stochastic interventions for multivariate continuous treatments.

Key steps:

1. Intervention definition: For intervention vector $\delta \in \mathbb{R}^q$, define tilted density:

   \[g_\delta(w|x) = \frac{\exp(\delta^\top w)f(w|x)}{\int_{\mathcal{W}} \exp(\delta^\top v)f(v|x) dv}\]

2. Causal estimand:

   \[\psi(\delta) = \int_{\mathcal{X}} \int_{\mathcal{W}} \mathbb{E}[Y|W=w, X=x] g_\delta(w|x) dw dP(x)\]

3. Fair comparison constraint: Use Gelbrich distance $G(\delta) = c^2$ to ensure interventions have same “size”

4. One-step estimation: Using efficient influence function:

   \[\hat{\psi}_{\text{onestep}}(\delta) = \frac{1}{n}\sum_{i=1}^n \hat{r}(W_i, X_i)[Y_i - \mathbb{E}_{\hat{g}_\delta}[\hat{\mu}|X_i]] + \frac{1}{n}\sum_{i=1}^n \mathbb{E}_{\hat{g}_\delta}[\hat{\mu}|X_i]\]

5. Optimization: Riemannian BFGS on manifold ${δ: G(δ) = c^2}$ to find optimal policy:

   \[\delta^*_c = \arg\min_{\delta \in \mathcal{A}_c} \psi(\delta)\]

   Toy example application: For $q=2$, $X=(1,0)$, and normal exposures with $\Sigma = \begin{pmatrix} 1 & 0.5 \ 0.5 & 1 \end{pmatrix}$, the method computes $\hat{\psi}(\delta)$ for intervention $\delta = (0.1, -0.2)$ by estimating the density ratio $\hat{r}(w,x) = \exp(0.1w_1 - 0.2w_2)/\hat{\nu}(x)$ and integrating the outcome regression under the tilted distribution.

Prior work:

• Kennedy (2019): “Nonparametric causal effects based on incremental propensity score interventions” - introduced exponential tilting for binary treatments
• Díaz & Hejazi (2020): “Causal mediation analysis for stochastic interventions” - extended to univariate continuous treatments
• Schindl et al. (2024): “Incremental causal effects for continuous exposures” - developed efficient estimation for univariate case

Delta: This paper extends exponential tilting to multivariate continuous treatments, introducing:

1. Fair comparison framework via Gelbrich distance constraints
2. Riemannian optimization for finding optimal intervention directions

3. Multivariate semiparametric efficiency theory with minimax rates.

   Theory-specific assessment:

   • Main theorem: The minimax bound $\text{RMSE}(\hat{\theta}) \gtrsim \sqrt{\delta^\top \Gamma \delta / n}$ is predictable from univariate theory but the multivariate geometry via matrix $\Gamma$ is new
   • Proof technique: Extends standard semiparametric efficiency analysis but requires novel treatment of multivariate tilting geometry and Riemannian manifold constraints
   • Tightness: No known lower bounds for multivariate incremental effects, so tightness unclear

   Verdict: INCREMENTAL — Solid extension of existing exponential tilting methods to multivariate setting with reasonable theoretical analysis, but represents expected generalization rather than fundamental breakthrough.

Main proof strategy:

1. Efficient influence function derivation: Uses standard semiparametric theory to show:

   \[\phi(Z;\delta) = \frac{g_\delta(W|X)}{f(W|X)}(Y - \mu(X,W)) + \frac{g_\delta(W|X)}{f(W|X)}(\mu(X,W) - \mathbb{E}_{g_\delta}[\mu(X,W)|X]) + \mathbb{E}_{g_\delta}[\mu(X,W)|X] - \psi\]

2. Minimax lower bound: Key steps involve: - Decomposition: $h(Z) = \tilde{W}[Y - \mathbb{E}[\mu|X]] - \mathbb{E}[\mathbb{E}[\mu\tilde{W}|X]]$ where $\tilde{W} = W - \mathbb{E}[W|X]$ - Covariance structure:

   \[H = \Sigma_{\tilde{W}\tilde{W}^\top, \varepsilon} + \Sigma_{\mu,\text{full}}\]

   - **Le Cam's method**: Constructs hardest two-point testing problem to establish
   

   \[\inf_{\hat{\theta}} \sup_P \mathbb{E}_P[(\hat{\theta} - \theta_P(\delta))^2] \geq C \cdot \frac{\delta^\top \Gamma \delta}{n}\]

3. Asymptotic normality: - Lipschitz continuity: Shows $(μ, f) \mapsto r_\delta$ and $(μ, f) \mapsto m_\delta$ are Lipschitz:

   \[\|\hat{r}_\delta - r_\delta\|_2 \leq C_1(\Delta)\|\hat{f} - f\|_2\]

   - **Product condition**: Requires $\|\hat{r}\_\delta - r\_\delta\|\_2 \|\hat{m}\_\delta - m\_\delta\|\_2 = o\_P(n^{-1/2})$
   - **Rate requirement**: Sufficient if both $\hat{\mu}$ and $\hat{f}$ converge at $n^{-1/4}$ rate
   

4. Riemannian optimization: - Tangent space projection: $\text{grad}\,\psi(\delta) = P_\delta \nabla\psi(\delta)$ where $P_\delta = I - n(\delta)n(\delta)^\top$ - Retraction: Uses projection retraction $R_\delta(\xi) = \Pi(\delta + \xi)$ onto constraint manifold - Convergence: Invokes Huang et al. (2018) global convergence theory for RBFGS with weak Wolfe conditions

Datasets:

• Simulation studies with $n=5000$, $p=10$ covariates, $q=6$ exposures across three scenarios:
  1. Gaussian exposures
  2. Skew-normal exposures
  3. Truncated contaminated normal exposures
  • Real data application to nationwide environmental health dataset examining PM2.5 chemical mixture effects on health outcomes

  Baselines: Compares plug-in estimator vs. one-step estimator vs. regression-based approach that avoids density estimation

  Key numbers: In simulations, one-step estimator achieves coverage rates 92-96% across scenarios. Real data analysis identifies optimal intervention reducing adverse health outcomes by estimated 0.15 units (specific health outcome not clearly defined in abstract).

  Empirical validation needed: More extensive real data applications across different environmental mixtures and health outcomes would strengthen the practical utility claims.

Limitations:

1. Multivariate density estimation requirement - TECHNICAL: One-step estimator requires estimating $f(w

   x)$ which is challenging for moderate-to-large $q$, though authors provide regression-based alternative

2. Normality assumptions for analytical results - TECHNICAL: Closed-form expressions for efficiency and Wasserstein distance only hold under multivariate normal exposures; extends heuristically to general distributions

3. Local intervention assumption - NATURAL: Theory requires $|\delta| \leq \Delta$ for fixed finite $\Delta$, excluding extreme interventions

4. No unmeasured confounding - RESTRICTIVE: Standard assumption but particularly strong for environmental mixtures where many confounders may be unmeasured (though authors provide sensitivity analysis)

5. Gelbrich distance as proxy - TECHNICAL: Uses lower bound approximation to Wasserstein distance rather than true distance for computational tractability

   Open problems:

6. Optimal rate characterization: Determine whether $n^{-1/4}$ nuisance function rate requirement is tight or can be relaxed
7. High-dimensional extensions: Develop methods for $q$ growing with $n$, particularly relevant for modern environmental mixture analysis with many pollutants

1. Motivation: Quantum channel tomography is fundamental for characterizing and validating quantum hardware. The question of optimal query complexity for learning an unknown quantum channel’s full classical description remains poorly understood.

2. Mathematical setup: Consider an unknown quantum channel

   \[E: L(\mathbb{C}^{d_1}) \to L(\mathbb{C}^{d_2})\]

   with Kraus rank at most $r$. The channel has Choi operator

   \[C_E = (E \otimes I)(|I\rangle\rangle\langle\langle I|) \in L(\mathbb{C}^{d_2} \otimes \mathbb{C}^{d_1})\]

   Key parameter: dilation rate $\tau = rd_2/d_1 \geq 1$.

   Assumptions:

   1. CPTP map: $E$ is completely positive and trace preserving
   2. Bounded rank: Kraus rank of $E$ is at most $r$
   3. Black-box access: Can query $E$ but don’t know its explicit form

3. Toy example: When $d_1 = d_2 = 2$ and $r = 1$, we have $\tau = 1$ (boundary regime). The channel is a unitary $E(\rho) = U\rho U^\dagger$ for some $2 \times 2$ unitary $U$. This reduces to learning a 2-dimensional unitary matrix up to error $\varepsilon$.

4. Formal objective: Learn channel $\hat{E}$ such that

   \[\|E - \hat{E}\|_{\diamond} \leq \varepsilon\]

   or equivalently for Choi trace norm error

   \[\frac{1}{d_1}\|C_E - C_{\hat{E}}\|_1 \leq \varepsilon\]

The main algorithmic contribution is a local test theorem enabling reduction from general channel tomography to isometry tomography.

Key steps:

1. Local Test Construction: For any parallel tester making $n$ queries to dilation $V \in \text{Dilation}_r(E)$, construct equivalent tester making $n$ queries to $E$ itself with expected performance:

   \[\tilde{P}_i(E) = \mathbb{E}_{V \sim \text{Dilation}_r(E)}[P_i(V)]\]

2. Reduction Algorithm: Apply existing isometry tomography algorithms (Yang-Renner-Chiribella for boundary regime, Haah-Kothari-O’Donnell-Tang for away-from-boundary regime) to dilations, then use local test to simulate with direct channel access.

3. Specific Bounds: - Boundary regime ($\tau = 1$): Use unitary tomography achieving $O(rd_1d_2/\varepsilon)$ queries - Away-from-boundary ($\tau \geq 1 + c$): Use general isometry tomography achieving $O(rd_1d_2/\varepsilon^2)$ queries

   Applied to toy example: For $d_1 = d_2 = 2, r = 1$ (unitary channel), the algorithm reduces to applying HKOT unitary tomography requiring $O(4/\varepsilon) = O(1/\varepsilon)$ queries, achieving Heisenberg scaling in the boundary regime.

Step 1 — Prior work:

1. Haah-Kothari-O’Donnell-Tang (2023): Established $\Theta(d^2/\varepsilon)$ for unitary channel tomography
2. Yoshida-Miyazaki-Murao (2025): Lower bound $\tilde{\Omega}((d_2-d_1)d_1/\varepsilon^2)$ for isometry tomography

3. Oufkir (2023): $\tilde{\Theta}(d_1^3d_2^3/\varepsilon^2)$ for full-rank channel tomography with non-adaptive queries

   Step 2 — Delta: This paper identifies dilation rate $\tau = rd_2/d_1$ as key parameter governing phase transition between Heisenberg scaling (1/ε) and classical scaling (1/ε²). Proves tight bounds across three regimes: boundary ($\tau = 1$), near-boundary ($1 < \tau < 4/3$), away-from-boundary ($\tau \geq 1 + \Omega(1)$).

   Step 3 — Theory assessment:

   • Main theorem is surprising: the sharp phase transition at $\tau = 1$ was not predicted by prior work
   • Proof techniques are genuinely new: local test construction using quantum combs + Schur-Weyl duality, and packing net constructions with orthogonality constraints
   • Bounds are tight in away-from-boundary regime; gaps remain in boundary regime for diamond norm

   Verdict: SIGNIFICANT — establishes fundamental phase transition in quantum channel tomography that changes how the field should think about this problem.

The proof combines several sophisticated techniques:

1. Local Test via Quantum Combs: For parallel tester ${T_i}$ accessing dilation $V$, construct local tester ${\tilde{T}_i}$ accessing channel $E$ directly. Uses Schur-Weyl duality on bipartite systems:

   \[\bigotimes_{i=1}^n H_{A,i} \otimes H_{B,i} \cong \bigoplus_{\lambda \vdash d_1 n, \mu \vdash d_2 n} P_\lambda \otimes P_\mu \otimes Q_{d_1}^\lambda \otimes Q_{d_2}^\mu\]

2. Packing Net Construction: Build nets of channels with form $E_k(\cdot) = \text{tr}[(V_0 + \varepsilon\Delta_k)(\cdot)(V_0 + \varepsilon\Delta_k)^\dagger]$ where:

   \[V_0 \text{ is center map}, \quad \Delta_k \text{ are random perturbations}\]

   Key insight: orthogonality between center and perturbation images enables classical scaling.

3. Hardness via Discrimination: For type II instances (away-from-boundary), decompose $

   V\rangle\rangle^{\otimes n}$ by number of perturbations appearing:

   \[|V\rangle\rangle^{\otimes n} = \sum_{j=0}^n P_j\]

   where $P_j$ involves exactly $j$ perturbation components. Averaging over Haar measure yields universal upper bound $\lambda \cdot \Gamma$ where $\Gamma$ is an $n$-comb.

4. Success Probability Bound: For discrimination among $

   N

   $ channels:

   \[P[\text{success}] \leq \frac{1}{|N|} \sum_{V \in N} T_V \star (λ \cdot Γ) = \frac{λ}{|N|}\]

   Setting $P[\text{success}] \geq 2/3$ gives $λ \geq 2

   N

   /3$, yielding query lower bounds.

Purely theoretical. Empirical validation would require:

1. Implementing the local test construction on quantum hardware to verify the reduction from dilation access to direct channel access maintains performance
2. Testing the phase transition behavior on synthetic channels with varying dilation rates τ
3. Comparing against existing tomography algorithms (Choi state tomography, direct process tomography) across different parameter regimes
4. Benchmarking on realistic quantum channels from NISQ devices to validate practical relevance of theoretical bounds

Limitations:

1. Gap between upper and lower bounds in boundary regime for diamond norm error - TECHNICAL (proof techniques could potentially be tightened)
2. Results proven only for parallel testers, not sequential adaptive algorithms - TECHNICAL (local test theorem may extend to sequential case)
3. Near-boundary regime bounds not tight - NATURAL (inherent difficulty of precise characterization in transition region)

4. Assumes exact black-box access to channels - NATURAL (standard assumption in quantum tomography)

   Open Problems:

5. Determine optimal query complexity in near-boundary regime $1 < \tau < 1 + o(1)$ with tight constants
6. Extend local test theorem to sequential testers to potentially close remaining gaps via reduction to bootstrapped algorithms