TSQP: Quantum, Optimization & Security

Updated 3 July 2026

TSQP is a multi-context term encompassing quantum process tomography, sparse quadratic programming, iterative quantum state preparation, trust-region optimization, and cryptographic protocols.
In quantum applications, TSQP protocols reconstruct process matrices using SIC-POVMs to minimize estimation error and employ the Q-Tucker method for efficient shallow circuit synthesis.
In optimization and security, TSQP addresses NP-hard sparse quadratic challenges with conic relaxations and identifies vulnerabilities in TEE-shielded ML inference, prompting future research.

TSQP refers to several distinct and technically advanced constructs across quantum information processing, mathematical optimization, and security protocols. Below is a comprehensive treatment of the principal meanings of TSQP found in the arXiv literature, focusing on quantum state/process tomography, sparse quadratic programming, trust-region stochastic optimization, and TEE-shielded cryptographic inference.

1. Standard Quantum Process Tomography (SQPT/TSQP)

Standard Quantum Process Tomography (SQPT)—occasionally abbreviated as TSQP in vectorization-based treatments—provides a protocol to characterize unknown quantum channels using only linear algebraic operations and no ancillary resources. For a system with Hilbert space dimension $d$ , SQPT reconstructs the process matrix $\chi$ that encodes all physical transformations as a system of $d^4$ linear equations (Xiaohua, 2012).

The formalism adopts a vectorization map: for any $A\in\mathcal B(\mathbb{C}^d)$ ,

$|A\rangle\!\rangle = (A \otimes I)\,|S_+\rangle = \sum_{i,j=1}^d A_{ij}\,|i\rangle\otimes|j\rangle,$

enabling a transparent linear representation of superoperators. The essential linear system is

$\mathfrak{B}\,|\chi\rangle\!\rangle = |\Lambda\rangle\!\rangle,$

where $\mathfrak{B}$ is formed by the choices of input states and measurement bases; $\chi$ is uniquely obtainable via Cramer's rule.

The optimal choice for both input states and measurements is a rank-one symmetric informationally complete POVM (SIC-POVM), or their tensor product for the full channel tomography, which minimizes the mean-square Hilbert–Schmidt estimation error. This guarantees the protocol achieves the lowest possible average error among all informationally complete schemes for SQPT (Xiaohua, 2012).

2. Sparse Standard Quadratic Programming (TSQP)

In continuous optimization, TSQP often denotes the sparse standard quadratic optimization problem: $\ell_\rho(Q) = \min_{x\in \mathbb{R}^n} \left\{ x^T Q x : x \in F_n, \|x\|_0 \le \rho \right\},\quad F_n = \{x\in\mathbb{R}^n_+ : 1^Tx = 1\}$ where $Q\!\in\!\mathcal{S}^n$ and $\chi$ 0 is the sparsity budget (Bomze et al., 2024). The combinatorial $\chi$ 1-constraint renders TSQP NP-hard, even for convex $\chi$ 2.

The state-of-the-art relaxation strategies convert the MIQP to conic programs using completely positive (CP) and doubly nonnegative (DNN) cones, with tractable SDP relaxations D1B (stronger) and D2B (dimension-reduced). The reduced SDP in D1B dramatically cuts computational cost, solving practical $\chi$ 3 problems in seconds. D2B remains useful for even higher $\chi$ 4 due to further reduction in variable and constraint counts. Instance generation algorithms guarantee nontriviality and can enforce (non-)exactness of the DNN bounds. For convex $\chi$ 5, and depending on the structure of $\chi$ 6, D1B may even provide exact bounds, but for more pathological copositive cases, DNN relaxations exhibit small optimality gaps but remain more scalable than MIQP solvers (Bomze et al., 2024).

3. Tucker Iterative Quantum State Preparation (Q-Tucker / TSQP)

TSQP also denotes the Tucker-based iterative quantum state preparation algorithm (“Q-Tucker”), which synthesizes shallow quantum circuits for efficient amplitude encoding of classical data into quantum registers (Blank et al., 10 Feb 2026). Given an $\chi$ 7-qubit state $\chi$ 8, TSQP represents it as a higher-order tensor and applies the Tucker decomposition: $\chi$ 9 with $d^4$ 0. This factorizes the state into a “core” tensor with low entanglement and mode matrices corresponding to small qubit blocks, each of which is synthesized into local unitaries or isometries.

The iteration proceeds by partitioning qubits using a correlation-graph heuristic to minimize inter-block entanglement, extracting mode matrices via SVD, aligning “monotone gauge” for guaranteed fidelity increase, and assembling the corresponding parallel circuit layers. Convergence is monotonic in fidelity and is guaranteed to reach arbitrary precision in finite steps as $d^4$ 1 (block-size) increases. Circuit complexity scales as $d^4$ 2 for each iteration for small $d^4$ 3, with overall circuit depth and width favorable compared to MPS and recursive bipartition methods. Practical benchmarks demonstrate significant reductions in circuit depth for realistic datasets, with trade-offs between block-size and gate synthesis complexity (Blank et al., 10 Feb 2026).

4. Trust-Region Sequential Quadratic Programming for Stochastic Optimization (TR-SQP)

A separate strand refers to Trust-Region Sequential Quadratic Programming algorithms for stochastic optimization with random models under deterministic equality constraints (Fang et al., 2024). Here, TSQP denotes algorithms that, at each iteration, build a local quadratic model of the stochastic objective (with adaptively accurate random sampling) and solve a subproblem subject to a trust-region and linearized constraints. In addition to the standard gradient step, the inclusion of an “eigen step” effectively explores negative curvature directions, facilitating escape from strict saddle points and improving global convergence.

The core algorithm includes:

Stochastic accuracy adaptation of function, gradient, and Hessian estimates according to the trust-region size,
Parameter-free decomposition of the trial-step into feasibility and optimization directions,
Second-order correction (SOC) to overcome the Maratos effect,
Almost sure convergence to first- and second-order stationary points given bounded stochastic model error.

Extensive numerical results on CUTEst and regression datasets indicate superior robustness and convergence relative to line-search-based methods under strong noise (Fang et al., 2024).

5. TEE-Shielded QNN Partition: Cryptographic TSQP for LLM Integrity

In secure machine learning inference, TSQP (TEE-Shielded QNN Partition) is a cryptographic protocol for neural network inference on untrusted accelerators (Saini et al., 11 Feb 2026). Here, TSQP partitions the model into “sensitive” layers inside a TEE (with secret scales and masking “noise” bases) and the bulk of the model on the GPU, with integrity-checked exchanges of masked activations. The primary integrity mechanism relies on precomputed noise vectors stored in the TEE and their linear combinations as fingerprints, with output verification via linearity.

A critical flaw arises from the reuse of a static corner-basis for mask generation. All challenge fingerprints fall in a fixed subspace; an adversary can reconstruct this subspace by collecting outputs for a small number of fingerprint challenges. This enables bypassing integrity checks for all future queries by distinguishing fingerprinted activations and returning forged outputs that nonetheless pass the TEE check. Experimental attacks against TSQP and similar Soter systems achieve subspace recovery and complete integrity bypass in seconds to minutes, breaking the intended security model unless fresh randomness or more complex cryptographic primitives are incorporated per inference (Saini et al., 11 Feb 2026).

6. Open Questions and Future Directions

For SQPT, efficient construction of higher-dimensional SIC-POVMs, as well as extensions to quantum process tomography for noisy or non-trace-preserving channels, remain technically open (Xiaohua, 2012).
In sparse TSQP, closing the gap between DNN bounds and true optima for pathological copositive cases is a current research target (Bomze et al., 2024).
For Q-Tucker/TSQP, incorporation of efficient multi-qubit isometry compilers, extensions to mixed-state preparation, and adaptive block-size selection are active areas (Blank et al., 10 Feb 2026).
For cryptographic TSQP, provably secure, low-latency protocols that avoid fixed subspace keying and support fresh, TEE-efficient masking are required (Saini et al., 11 Feb 2026).
For trust-region stochastic SQP, theoretical analysis of convergence under weaker smoothness or stochasticity assumptions and hybridization with other second-order optimization strategies are ongoing efforts (Fang et al., 2024).

7. Summary Table of TSQP Contexts

Domain	Meaning of TSQP	Core Reference
Quantum Process Tomography	Standard Quantum Process Tomography	(Xiaohua, 2012)
Sparse Quadratic Optimization	Sparse Standard Quadratic Program	(Bomze et al., 2024)
Quantum State Preparation	Tucker-based iterative state preparation	(Blank et al., 10 Feb 2026)
Trust-Region Optimization	Trust-Region Sequential Quadratic Programming	(Fang et al., 2024)
Secure ML Inference	TEE-Shielded QNN Partition (cryptographic)	(Saini et al., 11 Feb 2026)

Each variant of TSQP captures a distinct, highly technical construct: either a fundamental protocol in quantum information science, a hard combinatorial programming model with advanced relaxations, a tensor-network–based quantum algorithm, a stochastic trust-region optimization method, or a cryptographic protocol with concrete attack surfaces. The technical depth and context-specific meanings warrant precise usage and attention to domain when encountering “TSQP” in the literature.