TG-CIR: Retrieval, Finance & Quantum

• TG-CIR is a multi-context framework combining target-guided composed image retrieval, time-changed affine models, and integrable quantum many-body regimes.
• In composed image retrieval, TG-CIR leverages a dual teacher–student architecture with CLIP-based feature extraction to achieve state-of-the-art performance on benchmarks.
• For finance and quantum systems, TG-CIR refines classical models via deterministic jumps and finite-range corrections, enhancing predictive power and experimental validation.

TG-CIR refers to distinct—yet precisely defined within their respective fields—models and regimes: (1) Target-Guided Composed Image Retrieval for multimodal information retrieval (Wen et al., 2023), (2) a time-changed Cox–Ingersoll–Ross process with deterministic jump times for affine interest rate modeling (Fontana et al., 19 Sep 2025), and (3) the Tonks–Girardeau regime across a confinement-induced resonance in quantum gases, featuring integrable many-body physics with finite-range corrections (Qi et al., 2012). The following aims to encapsulate the technical underpinnings and theoretical advancements associated with TG-CIR in these contexts.

1. TG-CIR in Composed Image Retrieval: Model Architecture and Principles

Target-Guided Composed Image Retrieval extends the composed image retrieval (CIR) paradigm by addressing two limitations in previous approaches: adaptive modeling of conflict relationships between reference image and modification text, and metric learning driven by adaptive, graded matching rather than binary positive–negative targets.

The architecture centers on three components:

• Attribute Feature Extraction: Unified global and local attribute features are extracted from images and text using CLIP as the backbone. For each modality, the CLIP output is decomposed by learnable attribute masks (for global features) and by attention-based pooling of local tokens (for local features). The result is a $K$-by-$D$ matrix $E_r$, $E_m$, $E_t$ (for reference, modification, and target, respectively) forming a cross-modal aligned representation.
• Target-Query Relationship-Guided Composition: The model operates with two parallel branches. The student branch (used at inference) forms query compositions via learned “keep” and “replace” masks over attributes:

$$\hat m_k = \sigma(\mathrm{MLP}_s([E_r, E_m])), \quad \hat m_r = 1 - \hat m_k, \quad \hat E_c = \hat m_k \odot E_r + \hat m_r \odot E_m$$

The teacher branch provides supervision by incorporating the target attributes:

$$\tilde m_k = \sigma(\mathrm{MLP}_{t1}([E_t, E_r])), \quad \tilde m_r = \sigma(\mathrm{MLP}_{t2}([E_t, E_m]))$$

Consistency and knowledge distillation losses are imposed to align student and teacher branch predictions.

• Metric Learning and Orthogonal Regularization: The model uses batch-based cross-entropy losses to promote correct ranking in the retrieval task, augmented by a “matching degree” regularization term aligning the matching probability distribution with actual inter-image similarity distribution. Orthogonal regularization is applied to attribute vectors to promote disentanglement and decorrelation:

$$\mathcal{L}_{\mathrm{ortho}} = \sum_{X \in \{E_r,E_m,E_t\}} \| X X^\top - I_K \|_F^2$$

The aggregation of these modules culminates in a composite loss with tunable weighting for each regularizer (Wen et al., 2023).

2. Theoretical Foundations and Affine Framework of TG-CIR in Financial Mathematics

In mathematical finance, the TG-CIR (“time-changed and jumped CIR”) model generalizes the classical CIR process by introducing stochastic discontinuities (jumps at deterministic times, typically central bank dates) with jump sizes that may depend on the pre-jump state:

$$dr_t = \kappa(\theta - r_t)\,dt + \sigma\sqrt{r_t}\,dW_t + \sum_{n \ge 1} g_n(r_{s_n-})\,\delta_{t=s_n},\quad r_0 = x \ge 0$$

Here, $(s_n)$ is a deterministic sequence of strictly increasing jump times, and $D$0 are functions satisfying $D$1 to guarantee $D$2 and thus nonnegativity. The process remains affine if, for each $D$3 and all real $D$4,

$D$5

with analytic functions $D$6 on $D$7 (Fontana et al., 19 Sep 2025).

Affine property allows the conditional transform

$D$8

where $D$9 are updated across continuous intervals by CIR Riccati ODEs and, at each jump $E_r$0, via

$E_r$1

A deterministic time-change construction forms a TG-CIR process by $E_r$2 with $E_r$3 a standard CIR, $E_r$4 a strictly increasing, càdlàg function capturing the jump schedule (Fontana et al., 19 Sep 2025). Infinite divisibility holds if the jump transforms admit Lévy–Khintchine representations.

3. TG-CIR in One-Dimensional Quantum Gases: Tonks–Girardeau Regime across CIR

The term TG-CIR also denotes the many-body regime of a quasi-1D Bose gas at a Tonks–Girardeau (TG) limit tuned across a narrow confinement-induced resonance (CIR). The effective interparticle interaction becomes momentum dependent due to a finite-range parameter $E_r$5:

$E_r$6

The many-body Hamiltonian involves a polynomial of derivative $E_r$7-function interactions; the Bethe ansatz solution leads to integral equations for the root density and ground-state energy (Qi et al., 2012).

Low-energy properties map to a Tomonaga–Luttinger liquid (TLL):

• Sound velocity: $E_r$8, with $E_r$9, $E_m$0
• Luttinger parameter: $E_m$1

Finite $E_m$2 introduces nontrivial corrections to phase shifts and exclusion statistics, such that the Haldane statistical parameter $E_m$3 interpolates between free-Fermi, fractional, and “super-TG” limits.

Experimentally, the breathing-mode frequency in a shallow trap responds to $E_m$4 and provides a direct probe of these quantum many-body effects.

4. Representative Results and Empirical Performance

In Composed Image Retrieval (Wen et al., 2023):

TG-CIR achieves new benchmarks across FashionIQ, Shoes, and CIRR datasets:

• FashionIQ: R@10 increases from 46.84% (prior best) to 51.32%
• Shoes: Average R rises from 52.05% to 58.05%
• CIRR: R@1 improves from 38.53% to 45.25%, with overall average from 69.09% to 75.57%

Ablation studies substantiate the necessity of both global and local attributes, orthogonal regularization, teacher guidance, composition distillation, and similarity distribution regularization, with each removal resulting in measurable performance drops.

In Affine Processes (Fontana et al., 19 Sep 2025):

The TG-CIR model allows upward and downward jumps while preserving nonnegativity and the affine property under explicit analytic and growth constraints on jump transforms. Explicit constructions (e.g., Gamma-distributed jumps or time-changed CIR) yield shifted non-central chi-square laws for jumps, with closed-form Laplace transforms and infinite divisibility if the involved functions have Lévy–Khintchine form.

In Quantum Many-Body Systems (Qi et al., 2012):

A nonzero $E_m$5 modifies the universal scaling relations for the equation of state and TLL coefficients. The quantum-critical scaling near $E_m$6 is governed by $E_m$7, dynamic exponent $E_m$8, and correlation-length exponent $E_m$9, with scaling variables rescaled by $E_t$0. The predicted shift in the breathing-mode frequency is directly tied to $E_t$1 and is proposed as an experimental hallmark of TG-CIR physics.

5. Limitations and Extensions

For TG-CIR in vision:

• The dual-branch (teacher–student) design increases training complexity, though inference is efficient using only the student branch.
• Extension to multi-turn interactive retrieval, such as dialog-based systems, remains an open direction (Wen et al., 2023).

For affine jump-diffusion models:

• Existence, uniqueness, and nonnegativity require strict adherence to jump admissibility and non-accumulation of jump times.
• Preservation of the affine property is not universal; explicit characterizations of jump transforms are required, and practical calibration may depend on tractability for specific jump forms (Fontana et al., 19 Sep 2025).

In quantum gases:

• The regime is experimentally accessible only where transverse trapping and Feshbach resonance structure permit a significant finite-range correction.
• The polynomial-derivative structure complicates analytic progress for some observables but yields to exact Bethe ansatz treatment in the thermodynamic limit (Qi et al., 2012).

6. Comparative Table of TG-CIR Realizations

Context	Definition/Mechanism	Mathematical Structure
Vision (CIR)	Target-guided teacher–student fusion, CLIP, loss	Attribute matrix, cross-modal fusion
Affine Process	Time-changed CIR with deterministic jumps	SDE with jump kernel, affine Riccati
1D Bose Gas	TG regime across narrow CIR (finite $E_t$2)	Bethe ansatz, TLL, derivative $E_t$3

Each instance of TG-CIR signifies an advancement over its predecessor—whether by refined compositional modeling and metric learning (CIR), explicit affine jump handling (finance), or unveiling fractional-statistics many-body regimes (quantum gases).

7. Key Insights and Theoretical Significance

TG-CIR exemplifies the modern trend of extending classical models via explicit structural innovations—be it compositional fusion in multimodal retrieval, deterministic time-changes with affine jumps in finance, or momentum-dependent interactions in integrable quantum systems. In each realization, rigorous mathematical criteria (loss function engineering, affine transform preservation, Bethe solvability) are paramount, and empirical or computational gains are directly tied to the theoretical underpinnings meticulously laid out in the cited research.