Deep-QPP Model Overview

• Deep-QPP models are deep learning frameworks that unroll optimization algorithms to efficiently solve convex quadratic programs.
• They employ learned gradient steps, adaptive step-size modulation, and conic projections to significantly reduce iteration counts and warm-start traditional solvers.
• The framework extends to query performance prediction and scientific signal detection, demonstrating versatile applications across optimization and information retrieval.

Deep-QPP refers to several distinct but thematically linked families of models that use deep neural architectures to either (i) solve quadratic programs (QPs) more efficiently by unrolling operator splitting or first-order optimization schemes as neural networks, or (ii) address the unrelated challenge of query performance prediction (QPP) in information retrieval by learning latent structures associated with query/document interactions. Across these lines of work, Deep-QPP approaches combine the principles of algorithm unrolling, problem-structure injection, and deep learning to accelerate, predict, or interpret solutions for complex optimization or prediction tasks.

1. Deep-QPP for Quadratic Programming: Problem Setting

Deep-QPP models for optimization take as input convex QPs, typically in the general form: $\min_{x} \; \frac{1}{2}x^T P x + c^T x$ subject to affine and/or conic constraints: $A x = b, \quad Gx \leq h, \quad \ell \leq x \leq u$ or

$$A x + s = b, \quad s \in \mathcal{K}, \quad \mathcal{K} = \mathbb{R}_+^m$$

Such QPs arise in finance, engineering, control, and energy systems. Instances may reach sizes with $n$ (variables) and $m$ (constraints) in the range $10^2$–$10^4$ or more. The focus is on computational efficiency, warm-start capabilities, and scalability for both synthetic and real-world datasets (Xiong et al., 16 Aug 2025).

2. Unrolled Operator-Splitting Neural Solvers

The representative Deep-QPP (DR-GD Net) approach (Xiong et al., 16 Aug 2025) is built by unrolling a variant of the Douglas–Rachford (DR) splitting algorithm. The classical DR algorithm requires explicit linear system solves at each iteration: $\tilde u^{k+1} = (I+M)^{-1}(w^k - q)$ which becomes computationally expensive at scale. The DR-GD variant replaces this with a single gradient step per iteration: $\tilde u^{k+1} = \tilde u^k - \eta^k g^k$ where $g^k$ is the gradient of the quadratic least-squares residual,

$g^k = (I+M)^\top((I+M)\tilde u^k - (w^k - q))$

followed by conic projections and state updates. This iterative procedure is then unrolled into a finite-depth ($L$-layer; typically $L = 4$) feed-forward neural architecture, where each layer corresponds to a DR-GD update, with untied parameters across layers.

Within each layer, channel expansion, step-size modulation, and mixing matrices are learned to adaptively rescale and combine the updates. The output is mapped back to a primal-dual vector for the QP. The network is trained in supervised mode to minimize the Euclidean error on true QP solutions $(x^*, y^*)$ as computed by a generic convex solver (SCS) (Xiong et al., 16 Aug 2025).

3. Architecture Details and Theoretical Guarantees

The DR-GD Net implements the following core mechanisms per layer:

• Channel expansion of inputs by learned matrices to increase embedding dimension
• Gradient computation and adaptive step-size modulation via learned parameters and elementwise nonlinearities (sigmoid)
• Projection onto the constraint set (nonnegativity via ReLU for conic constraints)
• Mixing and update steps using learned linear combinations

Layer-specific step sizes ($\eta^\ell$) and projection/mixing matrices are untied and trained per layer.

Under standard monotonicity and Lipschitz conditions for $M$ and appropriate selection or learning of $\eta^\ell$, convergence to the unique QP solution $(u^*, w^*)$ is guaranteed. Specifically, as $k\to\infty$, $u^k\to u^*$, $w^k\to w^*$, and $\|w^{k+1} - w^k\|\to 0$ [(Xiong et al., 16 Aug 2025), Proposition 4.1].

4. Empirical Results and Performance

When used to warm-start off-the-shelf solvers (e.g., SCS), DR-GD Net achieves substantial reductions in both iteration count and wall-clock time on diverse QP testbeds:

• QP(RHS), $n = 1000$: iteration count reduced from 24,268 to 11,266 ($-53.6\%$); solve time reduced from $85.89$ s to $40.16$ s ($-53.5\%$)
• General QP, $n = 1000$: $-14.1\%$ iterations, $-13.0\%$ time
• QPLIB instance 4270: $-57.8\%$ iterations, $-55.7\%$ time
• Portfolio optimization, $k=200$: $-45\%$ iterations, $-40.8\%$ time

All gains are realized with a lightweight ($L=4$) unrolled network and negligible inference overhead ($\sim10$ ms) (Xiong et al., 16 Aug 2025).

5. Comparative Context: Related Deep QP Solvers

The Deep-QPP (DR-GD Net) framework is part of a broader trend in learning-to-optimize, in which optimization algorithms are viewed as computation graphs for neural parametrization. Notable comparisons include:

• PDQP-Net, which unrolls Primal–Dual Hybrid Gradient (PDHG) steps with unsupervised KKT-based loss, and achieves up to $45\%$ acceleration as a warm-start for PDQP (Yang et al., 2024)
• DeepQP and DeepDistributedQP, which unfold OSQP or ADMM-style distributed QP solvers, learning penalty and relaxation schedules; these models show order-of-magnitude wall-clock time speedups and are certified by PAC-Bayes bounds for generalization (Saravanos et al., 2024)
• Partially-supervised neural architectures (PSNNs) for multiparametric QP, in which active-set structure and KKT-derived weights are analytically injected into the first layer to enforce piecewise-affine solution structure, yielding near-exact feasibility and optimality across training and "extreme" out-of-distribution inputs (Beylunioglu et al., 5 Jun 2025)

6. Deep-QPP in Query Performance Prediction (IR/QPP)

The Deep-QPP identifier is also used in a distinct set of works in information retrieval:

• As unsupervised coherence-based predictors exploiting dense retrieval model embeddings (e.g., AC-embs, WAND-embs, A-pairRatio), trained to predict effectiveness of query rankings from ANCE or TCT-ColBERT (Vlachou et al., 2023)
• As a supervised pairwise interaction-based deep model for QPP, utilizing 2D convolutional processing of term-term cosine interaction histograms between queries and top/bottom retrieved documents to predict query difficulty and effectiveness (Datta et al., 2022)

These models typically operate within the QPP paradigm, evaluating post-retrieval statistical or learned features to estimate ranking quality in the absence of relevance annotations. They have demonstrated significant accuracy advances (up to $188\%$ relative improvement in Kendall $\tau$) over prior sparse or bag-of-words predictors (Vlachou et al., 2023). The key methodological difference is early modeling of fine-grained query$\times$document interaction structure, either via dense embedding similarity matrices (unsupervised) or pairwise convolutional architectures (supervised) (Datta et al., 2022).

7. Open Variants: Deep-QPP in Scientific Signal Detection

Deep-QPP has also been used (unrelated to QP or IR) to denote CNN-based classifiers for the detection of quasi-periodic pulsations (QPP) in solar and stellar flare lightcurves:

• A two-channel fully convolutional network (FCN) trained on 90,000 synthetic flare lightcurves (with and without exponentially decaying QPP) identifies QPP events in large-scale astronomical datasets (e.g., Kepler flares), achieving $87.2\%$ test accuracy on synthetics and a $7\%$ high-confidence detection rate in real white-light flare samples (Belov et al., 2024)

This usage highlights the breadth of the Deep-QPP label but is methodologically unrelated to the QP optimization and IR/QPP prediction families above.

---

References:

• "Solving Quadratic Programs via Deep Unrolled Douglas-Rachford Splitting" (Xiong et al., 16 Aug 2025)
• "An Efficient Unsupervised Framework for Convex Quadratic Programs via Deep Unrolling" (Yang et al., 2024)
• "Deep Distributed Optimization for Large-Scale Quadratic Programming" (Saravanos et al., 2024)
• "Partially-Supervised Neural Network Model For Quadratic Multiparametric Programming" (Beylunioglu et al., 5 Jun 2025)
• "On Coherence-based Predictors for Dense Query Performance Prediction" (Vlachou et al., 2023)
• "Deep-QPP: A Pairwise Interaction-based Deep Learning Model for Supervised Query Performance Prediction" (Datta et al., 2022)
• "Detecting quasi-periodic pulsations in solar and stellar flares with a neural network" (Belov et al., 2024)