Theory Digest — Apr 9, 2026

1. Motivation: Financial time series generation requires capturing both marginal distributions and temporal dynamics, including drift and stochastic volatility. Existing methods fail: Schrödinger Bridge fixes volatility, Bass martingale transport ignores drift. This limits realistic financial modeling applications like stress testing and data augmentation.

2. Mathematical setup: Let $\Omega = C([0,T], \mathbb{R}^d)$ be the canonical space of continuous paths. Consider probability measures $P \in \mathcal{P}$ where the canonical process $X$ has diffusion decomposition:

   \[X_t = X_0 + \int_0^t \alpha_s ds + \int_0^t \sigma_s dW_s\]

   Given joint distribution $\mu \in \mathcal{P}((\mathbb{R}^d)^{n+1})$ for time series at dates $0 = t_0 < \cdots < t_n = T$, define cost functional:

   \[J(P) = \mathbb{E}_P\left[\int_0^T \|\alpha_t\|^2 + \beta\|\sigma_t - I_d\|^2 dt\right]\]

   Assumptions:

   1. Each conditional $\mu_{i+1

      0:i}(\cdot

      x_{0:i})$ has finite second moment

   2. Absolutely continuous w.r.t. Gaussian with positive continuous density

   3. Finite relative entropy: $KL(\mu_{i+1

      0:i}

      N_{t_{i+1}-t_i}) < \infty$

   4. Constraint: $\beta\Delta t_i > 1$ for all $i$

3. Toy example: When $d=2$, $n=2$ (three time points), and $\beta$ large, this reduces to sequential Schrödinger bridges between consecutive marginals, but with learned volatility structure capturing correlations.

4. Formal objective: Find

   \[\text{SBBTS}(\mu) := \inf_{P \in \mathcal{P}(\mu)} J(P)\]

   where $\mathcal{P}(\mu) = {P : P \circ (X_{t_0}, \ldots, X_{t_n})^{-1} = \mu}$.

The method decomposes SBBTS into sequential conditional transport problems via Theorem 3.2:

\[\text{SBBTS}(\mu) = \sum_{i=0}^{n-1} \mathbb{E}_{\mu_i}[V_i(X_{t_0:t_i})]\]

where

\[V_i(x_{0:i}) = \text{SBB}(\delta_{x_i}, \mu_{i+1|0:i}(\cdot|x_{0:i}))\]

Algorithm steps:

1. Initialize transport map $Y^0 = \text{Id}$ and score network $s_\theta^0 \equiv 0$

2. For each iteration $k$: sample mini-batch, compute transformed variables via

   \[Y_{t_i}^b = X_{t_i}^b - \frac{1}{\beta}s_\theta^k(t_i, X_{t_i}^b, \Phi_\theta^k(X_{t_0:t_i}^b))\]

3. Sample Brownian bridges $Y_t \sim W

   Y_{t_i}, Y_{t_{i+1}}$ using

   \[y_t = \frac{t_{i+1}-t}{\Delta t_i}y_{t_i} + \frac{t-t_i}{\Delta t_i}y_{t_{i+1}} + \sigma_t Z\] \[\sigma_t^2 = \frac{(t-t_i)(t_{i+1}-t)}{\Delta t_i}\]

4. Update $\theta$ by minimizing loss:

   \[L(\theta) = \mathbb{E}\left[\left\|s_\theta(t, Y_t, \Phi_\theta(Y_{t_0:t_i})) - \frac{Y_{t_{i+1}} - Y_t}{t_{i+1} - t}\right\|_2^2\right]\]

   Toy example application: For $d=2$, $n=2$, the method learns score functions $s_\theta^{(1)}, s_\theta^{(2)}$ for intervals $[t_0,t_1]$ and $[t_1,t_2]$, where each depends on path history via transformer encoder $\Phi_\theta$.

Prior work:

1. “Schrödinger Bridge Methods for Time Series” (Chen et al., 2021) - introduced SBTS but fixed volatility structure
2. “Bass Martingale Transport” (Backhoff-Veraguas et al., 2020) - calibrated volatility but ignored drift

3. “Schrödinger-Bridge-Bass Problem” (Loeper et al., 2023) - unified SB and Bass for two-marginal case only

   Delta: This paper extends SBB from two-marginal to full multi-step time series via decomposition theorem. Key contributions:

4. Theorem 3.2 showing SBBTS decomposes into conditional transport problems
5. Neural architecture with path-dependent encoding

6. Large-β approximation for practical implementation.

   Theory-specific assessment:

   • Main theorem (3.2) is non-obvious extension requiring careful conditional probability analysis
   • Proof technique combines measure-theoretic argument with dynamic programming structure
   • Decomposition result is surprising and enables tractable computation
   • No known lower bounds for this specific multi-step setting

   The decomposition $\text{SBBTS}(\mu) = \sum_{i=0}^{n-1} \mathbb{E}_{\mu_i}[V_i(X_{t_0:t_i})]$ is genuinely novel and non-trivial, requiring sophisticated pasting arguments for stochastic processes.

   Verdict: SIGNIFICANT — Extends important optimal transport framework to multi-step time series with non-obvious decomposition enabling practical algorithms.

The key proof of Theorem 3.2 uses measure-theoretic pasting arguments:

1. Step 1 - Conditional expectation decomposition: For $\nu \in \mathcal{P}((\mathbb{R}^d)^{i+2})$, shows

   \[\tilde{V}_i(\mu_{i+1}) = \mathbb{E}_{\mu_i}[V_i(X_{t_0:t_i})]\]

   using regular conditional probabilities $P_{x_{0:i}}$ given $X_{t_0:t_i} = x_{0:i}$.

2. Key inequality: For any $P \in \mathcal{P}(\mu_{i+1})$:

   \[\mathbb{E}_P\left[\int_{t_i}^{t_{i+1}} L(\alpha_t, \sigma_t) dt\right] \geq \int V_i(x_{0:i}) \mu_i(dx_{0:i})\]

3. Measurable selection: Constructs universally measurable family $x_{0:i} \mapsto P_x_{0:i}^\varepsilon$ such that

   \[\mathbb{E}_{P_x_{0:i}^\varepsilon}\left[\int_{t_i}^{t_{i+1}} L(\alpha_t, \sigma_t) dt\right] \leq V_i(x_{0:i}) + \varepsilon\]

4. Step 2 - Dynamic programming recursion: Defines

   \[\bar{V}_i := \inf_{P \in \mathcal{P}(\mu_i)} \mathbb{E}_P\left[\int_0^{t_i} L(\alpha_t, \sigma_t) dt\right]\]

   and proves

   \[\bar{V}_{i+1} = \bar{V}_i + \mathbb{E}_{\mu_i}[V_i(X_{t_0:t_i})]\]

5. Stochastic process pasting: Uses concatenation of prefix law $P^{1,\varepsilon}_{x_{0:i}}$ on $[0,t_i]$ with continuation law $P^\varepsilon_{x_{0:i}}$ on $[t_i, t_{i+1}]$ to construct global measure $Q^\varepsilon \in \mathcal{P}(\mu_{i+1})$.

   The proof technique is genuinely sophisticated, requiring careful handling of conditional measures and process concatenation beyond standard optimal transport methods.

Experiments on:

1. Heston stochastic volatility model: 5000 trajectories of length 252, parameters $(κ, θ, ξ, ρ, r)$ sampled from prescribed ranges. SBBTS accurately recovers “vol of vol” parameter $ξ$ and correlation $ρ$ that prior SBTS methods failed to capture (concentrated around center rather than full range).

2. S&P 500 data augmentation: Daily returns 2010-2021, 433 stocks, features with 252-day lookback. TabICL transformer for binary return direction prediction.

   Key numbers:

   • SBBTS augmentation: Accuracy 0.532 vs 0.521 (real only), ROC AUC 0.521 vs 0.497, Sharpe ratio 2.113 vs 1.613
   • Used 200× more synthetic paths than real data
   • White noise augmentation fails to improve consistently
   • Performance scales with synthetic data amount up to saturation

   Dataset: S&P 500 from WRDS, train/val/test split 2010-2018/2019-mid2020/mid2020-2021. PCA + clustering for 433-dimensional reduction.

Limitations:

1. Parameter β selection lacks systematic criterion - TECHNICAL (practical guidelines exist: avoid small values, use β∆ti > 1, but optimal choice unclear)

2. Large-β approximation restricts regime - TECHNICAL (could adapt more general iterative scheme from Chen et al.)

3. No formal convergence guarantees for Algorithm 1 - TECHNICAL (empirically converges in K=5 iterations but theory missing)

4. Constraint β∆ti > 1 limits time resolution - NATURAL (typical financial time scales make this reasonable)

5. Assumes continuous paths - RESTRICTIVE (many financial assets exhibit jumps, though framework could potentially extend)

6. Requires absolute continuity w.r.t. Gaussian - NATURAL (standard assumption in optimal transport, satisfied by most continuous distributions)

   Open problems:

7. Develop principled β calibration methods: Currently relies on heuristics, could use cross-validation or information-theoretic criteria

8. Extend to jump-diffusion dynamics: Incorporate Lévy processes for realistic modeling of financial extremes and discontinuous price movements

Motivation: Optimal switching with discrete intervention dates on high-dimensional state spaces is computationally challenging due to the curse of dimensionality. Classical PDE/QVI methods fail in high dimensions, while existing neural approaches typically lack computable upper bounds or rigorous approximation guarantees.

Mathematical setup: Fix probability space $(\Omega, \mathcal{F}, \mathbb{F}, P)$ with filtration $\mathbb{F} = (\mathcal{F}_t)_{t \in [0,T]}$ and discrete intervention grid $\pi: t_n = nT/N$, $n \in {0,\ldots,N}$. State process $X$ follows Itô diffusion:

\[dX_t = \mu(t,X_t)dt + \sigma(t,X_t)dW_t\]

Given $J$ regimes with running payoffs $f^i$, terminal payoffs $\Phi^i$, and switching costs $l_{ij}(t)$ satisfying:

1. Integrability: $l_{ij}(t) \in L^1(\mathcal{F}_t; \mathbb{R})$

2. Strict triangular condition: $l_{ii} \equiv 0$, $l_{ij} + l_{jk} > l_{ik}$ for $i \neq j \neq k$

   Toy example: When $J=2$, $d=2$, $T=1$, $N=2$, starting in regime $i=1$ at $t_0=0$, the controller can switch to regime $j=2$ at $t_1=0.5$ paying cost $l_{12}(t_1)$, then collect running reward $f^2$ over $[0.5,1]$ plus terminal payoff $\Phi^2(X_1)$.

   Formal objective: Find the value process:

   \[Y^i_{t_n} = \text{ess sup}_{\alpha \in \mathcal{A}^i_n} J^i_n(\alpha)\]

   where $J^i_n(\alpha)$ is the expected discounted reward from admissible switching strategy $\alpha$.

The method develops a martingale duality framework extending from optimal stopping to optimal switching.

Key steps:

1. Regime-decision reformulation: Transform switching control to regime-decision sequences $d = (d_m)_{m=n}^N$ where $d_m \in \mathcal{J}$ is $\mathcal{F}_{t_m}$-measurable. Show equivalence:

   \[Y^i_{t_n} = \text{ess sup}_{d \in \mathcal{D}^i_n} L^i_n(d)\]

2. Dual upper bound construction: For martingale family $M = (M^j)_{j \in \mathcal{J}}$, define:

   \[\tilde{U}^i_n(M) = \max_{j \in \mathcal{J}_n} \left[ \sum_{m=n}^{N-1} \left( \int_{t_m}^{t_{m+1}} f^{j_m}(s)ds - l_{j_{m-1}j_m}(t_m) - \Delta M^{j_m}_{t_m} \right) + \Phi^{j_N} \right]\]

3. Doob characterization: Prove the Doob martingales $\bar{M} = (\bar{M}^i)_{i \in \mathcal{J}}$ from the Doob decomposition are surely optimal:

   \[Y^i_{t_n} = \tilde{U}^i_n(\bar{M}) = \mathbb{E}_{t_n}[\tilde{U}^i_n(\bar{M})]\]

4. DeepMartingale approximation: Parameterize martingale increments using neural networks:

   \[\xi^{i,\theta^i_n;K}_n = \sum_{k=0}^{K-1} z^{i,\theta^i_n}_n(t^n_k, X_{t^n_k}) \cdot \Delta W_{t^n_k}\]

   Applied to toy example: For $J=2$, $d=2$, $N=2$, the dual upper bound becomes:

   \[\tilde{U}^1_0(M) = \max \left\{ \int_0^{0.5} f^1(s)ds - \Delta M^1_{t_1} + \tilde{U}^1_1(M), \int_0^{0.5} f^2(s)ds - l_{12}(0) - \Delta M^2_{t_1} + \tilde{U}^2_1(M) \right\}\]

Prior work:

1. Lin and Ludkovski (2014) - developed dual methods for switching but upper bounds depend on unknown value functions, not fully computable
2. Ye and Wong (2023, DeepMartingale) - martingale duality for optimal stopping with neural approximation

3. Nellis et al. (2021, DeepOSJ) - deep learning for switching via reflected BSDEs, but lacks computable upper bounds

   Delta: This paper provides the first fully computable martingale dual representation for finite-horizon optimal switching with rigorous neural approximation theory.

   Theory-specific assessment:

   • Main theorem (strong duality) is non-obvious: extending stopping duality to switching requires handling coupled continuation values across regimes and joint optimization over intervention times and post-intervention regimes
   • Proof technique is genuinely new: the regime-decision reformulation and iterative expansion using Doob martingales is novel for switching problems
   • Bounds are not compared against known lower bounds - no lower bounds mentioned for this problem class
   • Expressivity result achieving $O(d^q \varepsilon^{-r})$ network size avoiding curse of dimensionality is significant under stated assumptions

   Verdict: SIGNIFICANT - First computable dual upper bounds for high-dimensional optimal switching with rigorous approximation theory represents a clear advance that switching researchers should read.

Main proof strategy:

1. Regime-decision equivalence: Use dynamic programming to show switching and regime-decision formulations yield identical values. Key insight: every switching strategy induces a regime-decision sequence and vice versa.

2. Doob martingale optimality: Apply iterated stopping duality to each regime separately, then use triangular cost condition to prove consecutive switches are suboptimal with probability zero:

   \[P(\{τ^{\iota^i_n}_n = n\} \cap \{\iota^{\iota^i_n}_n \neq i\}) = 0\]

3. Pathwise expansion: Construct optimal regime-decision candidates $j^{i,n}$ recursively and prove surely optimal property:

   \[Y^i_{t_n} = \tilde{U}^{i,j^{i,n}}_n(\bar{M}) \text{ P-a.s.}\]

4. Strong duality: Use dual dynamic programming principle:

   \[\tilde{U}^i_n(M) = \max_{j \in \mathcal{J}} \left[ \int_{t_n}^{t_{n+1}} f^j(s)ds - l_{ij}(t_n) - \Delta M^j_{t_n} + \tilde{U}^j_{n+1}(M) \right]\]

5. Neural approximation: Apply universal approximation for ReLU networks to approximate Doob martingale integrands. Key inequality from expressivity analysis:

   \[E\left[ \max_{i \in \mathcal{J}} |\tilde{U}^{i;d}_n(\tilde{M}^d_\varepsilon) - Y^{i;d}_{t_n}|^2 \right]^{1/2} \leq J(N-n)\varepsilon\]

   Technical insight: The triangular cost condition eliminates consecutive switching, enabling the pathwise construction of optimal candidates that can be expanded along the Doob martingales.

Datasets:

1. Geometric Brownian Motion with $J=3$ regimes, dimensions $d \in {2,10,20,30,50,100}$

2. Exponential OU with Brownian-Poisson jumps, dimensions $d \in {2,10,30,50}$

   Baselines: DeepOSJ (reflected BSDE approach), Least-Squares regression method

   Key results:

   • DeepPD maintains duality gaps ≈ 0.1 across all dimensions up to $d=100$
   • DeepOSJ runs out of memory for $d \geq 20$
   • DeepPD shows superior CVaR hedging performance (95% CVaR: 2.7 vs 4.3 for DeepOSJ at $d=10$)
   • Training: 1000+20d epochs for GBM, 300+3d for jump diffusion
   • Networks: 3 layers, batch size 4096, ReLU activation
   • Final evaluation on 1.6M samples

Limitations:

1. Discrete intervention assumption - TECHNICAL (continuous-time switching would require different mathematical framework)

2. Markovian dynamics in expressivity analysis - NATURAL (standard assumption for high-dimensional problems)

3. Structural assumptions (4.6, 4.9, 4.10) for curse-free approximation - RESTRICTIVE (polynomial growth, specific Lipschitz conditions limit applicability)

4. Reference regime training heuristic - TECHNICAL (trains dual solver for only one regime, though empirically works well)

5. Strong triangular cost condition - NATURAL (prevents arbitrage-like consecutive switching)

   Open problems:

6. Extend to continuous-time intervention with impulse control formulation
7. Remove or weaken structural assumptions while preserving curse-free approximation guarantees

1. Motivation: Bayesian inference in high-dimensional discrete-input additive noise models faces the curse of dimensionality, as the posterior distribution’s support grows exponentially with the number of variables. This challenge is fundamental in communication systems, particularly for MIMO detection and channel decoding where exact inference becomes computationally prohibitive.

2. Mathematical setup: Consider the real-valued observation model:

   \[y = f(x) + n\]

   where $x \in A^{NT}$ with discrete alphabet $A \subset \mathbb{R}$ of cardinality $

   A

   = L$, and $n \in \mathbb{R}^{NR}$ is additive white noise with independent components from an exponential family. The posterior distribution is:

   \[P(x|y) \propto p(y|x)P(x) \propto \prod_{j=1}^{NR} e^{\ell(y_j|x)} \prod_{i=1}^{NT} P(x_i)\]

   In logarithmic domain:

   \[\log P(x|y) = C + \sum_{j=1}^{NR} \ell(y_j|x) + \sum_{i=1}^{NT} \log P(x_i) =: C + \Lambda(x)\]

   Assumptions:

   1. Noise components $n_1, \ldots, n_{NR}$ are i.i.d. from exponential family
   2. Transmit symbols $x_i$ are i.i.d. with known prior $P(x_i)$
   3. Discrete alphabet $A$ is finite

3. Toy example: For $NT = 2$, $L = 2$, $A = {0,1}$, the joint posterior is a $2 \times 2$ table. With uniform priors and AWGN, this reduces to evaluating $4$ likelihood values. The tensor-train representation exploits structure in the log-APP to avoid explicit storage.

4. Formal objective: Compute symbol-wise marginal posteriors:

   \[P(x_i = a|y) = \sum_{\mathbf{a} \in A^{NT}, a_i = a} P(\mathbf{x} = \mathbf{a}|y)\]

The method represents the log-posterior $\Lambda(x)$ exactly in tensor-train (TT) format, then approximates the exponential for marginalization.

Construction steps:

1. Build log-prior $\sum_{i=1}^{NT} \log P(x_i)$ as rank-2 TT with cores:

   \[G_1 = \begin{pmatrix} \mathbf{1}_L \\ \mathbf{v} \end{pmatrix} \in \mathbb{R}^{1 \times L \times 2}\] \[G_{NT} = \begin{pmatrix} \mathbf{v}^T \\ \mathbf{1}_L^T \end{pmatrix} \in \mathbb{R}^{2 \times L \times 1}\]

   where $\mathbf{v} = (\log P(x_i = a_1), \ldots, \log P(x_i = a_L))^T$.

2. Construct each log-likelihood term $\ell(y_j

   x)$ in TT format (problem-specific).

3. Sum all terms using TT arithmetic to obtain $\Lambda(x)$ in TT format.

4. Approximate $\exp(\Lambda(x))$ using TT-cross algorithm initialized with truncated Taylor series:

   \[\exp(\Lambda) \approx \sum_{k=0}^p \frac{\Lambda^k}{k!}\]

5. Compute marginals via tensor-matrix multiplication:

   \[P(x_i = a|y) \propto \exp(\Lambda) \times_j \mathbf{1}_{n_j}^T\]

   for all $j \neq i$.

   Toy example application: For $NT = 2$, the method builds a $2 \times 2$ TT representation of the log-posterior, exponentiates it approximately, then marginalizes to obtain $P(x_1

   y)$ and $P(x_2

   y)$.

Step 1 — Prior work:

1. “Factor graphs and the sum-product algorithm” (Kschischang & Loeliger, 2001): Established belief propagation for graphical models with local message passing.
2. “Expectation propagation detection for high-order high-dimensional MIMO systems” (Céspedes et al., 2014): Applied EP with Gaussian approximations to MIMO detection.

3. Classical tensor networks literature (Oseledets, 2011; Hackbusch, 2019): Developed tensor-train format and basic arithmetic operations.

   Step 2 — Delta: This paper applies tensor-train representations specifically to the inference domain rather than the signal domain. Key contributions: (i) exact low-rank TT constructions for log-posterior in discrete-input additive noise models, (ii) TT-cross approximation of matrix exponential for marginalization, (iii) explicit constructions for MIMO detection and channel decoding.

   Step 3 — Theory-specific assessment:

   • The main insight that log-posteriors admit low-rank TT representations is somewhat predictable given the additive structure in (4), but the explicit constructions are non-trivial.
   • The proof technique is largely constructive, assembling known TT arithmetic operations rather than developing fundamentally new theory.
   • No theoretical analysis of approximation error bounds or rank growth is provided. The approach relies on empirical validation of low ranks.
   • The connection between problem structure and TT ranks is not rigorously established.

   The paper demonstrates that known TT techniques can be effectively applied to a new domain (Bayesian inference), but the theoretical understanding remains limited.

   Verdict: INCREMENTAL — solid engineering application of existing tensor methods to communication problems, but lacks deep theoretical insights or guarantees.

The paper is primarily constructive rather than proof-based. The main technical contributions are explicit tensor-train constructions:

1. Log-prior construction: Exact rank-2 representation using the recursive structure:

   \[\sum_{i=1}^{NT} \log P(x_i) = \log P(x_1) + \sum_{i=2}^{NT} \log P(x_i)\]

   Each core encodes the contribution $\log P(x_i)$ through matrix slices.

2. MIMO log-likelihood construction: For linear model $y = Hx + n$, builds:

   \[\ell(y_j|x) = -\frac{(y_j - \mathbf{h}_j^T \mathbf{x})^2}{2\sigma^2}\]

   Key insight: $\mathbf{h}_j^T \mathbf{x} = \sum_{i=1}^{NT} h_{ji} x_i$ admits rank-3 TT representation through careful core design.

3. Channel decoding construction: For binary linear codes $c = Gu$, expands:

   \[\Lambda(u) = -\frac{1}{N_0} \sum_{j=1}^n \left(y_j - (-1)^{(Gu)_j}\right)^2\]

   The key technical step is representing:

   \[\prod_{i=1}^k G_{ji}=1 (-1)^{u_i}\]

   as rank-1 TT, enabling the full sum to have rank at most $n$.

4. Approximation via TT-cross: Uses truncated Taylor series:

   \[\exp(A) \approx \sum_{k=0}^p \frac{A^k}{k!}\]

   followed by cross-interpolation algorithms (multifuncrs, funcrs) to refine the approximation within controlled TT ranks.

   The analysis lacks rigorous error bounds or convergence guarantees for the cross approximation step.

Datasets: Synthetic Monte Carlo simulations for:

1. MIMO detection: 4-QAM and 16-QAM with $\tilde{N} \times \tilde{N}$ systems ($\tilde{N} = 8, 16, 32$), Rayleigh fading channels

2. Channel decoding: BCH codes (15,7), (31,16), (63,30) over BI-AWGN channels

   Baselines:

   • MIMO: LMMSE detector, expectation propagation (10 iterations), sphere decoder (optimal MLSE)
   • Coding: Belief propagation with sum-product algorithm, ordered statistics decoding (OSD-3)

   Key results:

   • MIMO: Near-optimal performance matching sphere decoder for 4-QAM, 0.7 dB gain over EP at SER = 10^{-2}
   • Required TT ranks: modest (r_max = 10 for 4-QAM, r_max = 20-40 for 16-QAM)
   • Memory savings: 10^{13}-10^{14} reduction vs explicit storage for large systems
   • Channel decoding: 2 dB gain over BP at BER = 10^{-4}, performance comparable to OSD-3

   Analysis: Empirical validation shows the low-rank nature emerges in practice, with rank requirements increasing at high SNR where approximation becomes more challenging.

Limitations:

1. TECHNICAL: TT-cross approximation quality depends on hyperparameter tuning (maximum rank r_max, truncation tolerance). No theoretical guarantees on approximation error.
2. RESTRICTIVE: Framework limited to discrete-input additive noise models. Extension to continuous alphabets unclear.
3. TECHNICAL: Performance degrades in high-SNR regime where TT-cross fails to approximate exponential accurately, suggesting initialization strategy is insufficient.
4. NATURAL: Requires known channel state information and noise statistics.
5. TECHNICAL: No analysis of computational complexity scaling with problem dimensions, only empirical rank observations.

6. RESTRICTIVE: Adaptive rank strategy for decoding relies on specific code structure (minimum distance) and may not generalize.

   Open problems:

7. Develop rigorous error analysis for TT-cross approximation of matrix exponential in inference context, potentially leading to certified approximation bounds.
8. Investigate tensor-train representations for continuous-alphabet systems and non-Gaussian noise models, extending beyond exponential family assumptions.