Hybrid Projection Decomposition (HPD)

• Hybrid Projection Decomposition (HPD) is a methodology that combines analytical, numerical, and data-driven projection techniques to break down complex models into interpretable components.
• HPD is applied across diverse domains such as dual-energy CT, large-scale inverse problems, and matrix product factorization to enhance computational efficiency and robustness.
• Hybrid strategies in HPD enable automatic parameter tuning and noise reduction in hardware-robust systems, providing scalable and interpretable solutions for ill-posed problems.

Hybrid Projection Decomposition (HPD) refers to a class of methodologies that combine distinct projection techniques—typically blending analytical, numerical, or data-driven algorithms—to decompose complex models or measurements into interpretable or basis components. HPD is deployed in domains such as dual-energy computed tomography (CT), large-scale Bayesian and deterministic inverse problems, matrix product factorization, hardware-robust deep learning, and high-dimensional data visualization. The unifying concept is the hybrid orchestration of projections—often in conjunction with optimization, filtering, evidence aggregation, or regularization—resulting in robust, scalable, and typically interpretable solutions for otherwise ill-posed or highly structured mathematical problems.

1. Physical Modeling and Analytic–Numerical Hybridization

In dual-energy CT, HPD addresses the nonlinear decomposition of polychromatic X-ray projections into physical basis components. The core physical model expresses line-integral measurements at energies $i=1,2$ as: $$I_i = \int_{E_{\min}}^{E_{\max}} S_i(E)\,\exp\Bigl[ -\int_{\ell}\mu(E,\mathbf{r})\,d\ell \Bigr]\,dE$$ Here, $\mu(E,\mathbf{r})$ is well-modeled as a linear combination of photoelectric and Compton-scattering terms: $$\mu(E,\mathbf{r}) = p(E)\,f_{\mathrm{ph}}(\mathbf{r}) + c(E)\,f_{\mathrm{co}}(\mathbf{r})$$ HPD proceeds by analytically approximating the response of one spectrum to the Compton basis via a Taylor expansion, reducing the two-dimensional nonlinear inverse to a single-variable quartic polynomial in a deviation parameter $\Delta x$. The subsequent stage constrains the solution using the second energy measurement, minimizing a residual function $F(y)$ in the photoelectric projection. This fusion of semi-analytic quartic solution with a one-dimensional line search over $y$ represents a hybridization that sharply accelerates and stabilizes projection decomposition relative to brute-force discretization or joint-optimization schemes (Cong et al., 2018).

2. Algorithmic Structures in Large-scale Inverse Problems

In solution decomposition for large-scale inverse/Bayesian problems, HPD encompasses methods that simultaneously project onto flexible Krylov subspaces and optimize with composite regularizations—typically mixing smooth (quadratic, Gaussian) and sparse (Laplacian, $\ell_1$) priors. The forward model

$A s = d$

is decomposed with $s = s_1 + s_2$, where $$I_i = \int_{E_{\min}}^{E_{\max}} S_i(E)\,\exp\Bigl[ -\int_{\ell}\mu(E,\mathbf{r})\,d\ell \Bigr]\,dE$$0 is regularized by a Gaussian prior and $$I_i = \int_{E_{\min}}^{E_{\max}} S_i(E)\,\exp\Bigl[ -\int_{\ell}\mu(E,\mathbf{r})\,d\ell \Bigr]\,dE$$1 by an $$I_i = \int_{E_{\min}}^{E_{\max}} S_i(E)\,\exp\Bigl[ -\int_{\ell}\mu(E,\mathbf{r})\,d\ell \Bigr]\,dE$$2-type prior. Iterative HPD algorithms use combined generalized/flexible Golub–Kahan bidiagonalization to build joint subspaces accommodating both priors, employing majorization-minimization for convex surrogate optimization, and adaptively tuning regularization parameters at each iteration (Chung et al., 2022, Chung et al., 2020, Chung et al., 2021). This procedure avoids costly nested optimization and realizes highly scalable, robust, and interpretable decompositions suitable for, for example, anomaly detection in spatiotemporal environmental inversion.

3. Hybrid Projection Schemes in Matrix Product Constraints

HPD approaches extend to matrix factorization with product constraints and structural restrictions. Given $$I_i = \int_{E_{\min}}^{E_{\max}} S_i(E)\,\exp\Bigl[ -\int_{\ell}\mu(E,\mathbf{r})\,d\ell \Bigr]\,dE$$3, the task is to find $$I_i = \int_{E_{\min}}^{E_{\max}} S_i(E)\,\exp\Bigl[ -\int_{\ell}\mu(E,\mathbf{r})\,d\ell \Bigr]\,dE$$4 minimizing

$$I_i = \int_{E_{\min}}^{E_{\max}} S_i(E)\,\exp\Bigl[ -\int_{\ell}\mu(E,\mathbf{r})\,d\ell \Bigr]\,dE$$5

subject to $$I_i = \int_{E_{\min}}^{E_{\max}} S_i(E)\,\exp\Bigl[ -\int_{\ell}\mu(E,\mathbf{r})\,d\ell \Bigr]\,dE$$6 and further structural constraints (e.g., nonnegativity, low rank). Hybrid projections iterate between: (a) “quasiprojection” enforcing the exact product constraint (fixing one operand and updating the other via minimum-norm), and (b) tangent-space projections accounting for differentiable constraints via Lagrange multiplier and Sylvester equation. These steps are coupled with fixed-point iterations, often using relaxed-reflect-reflect (RRR) schemes. Structural projections (e.g., elementwise bounds, rank restrictions) are interleaved to realize hybrid algorithms converging efficiently even in nonconvex scenarios (Elser, 2016).

4. Hybridization in Data-driven and Learning-based Frameworks

HPD is also realized by combining analytical inversions with deep networks, achieving scalable and high-fidelity solutions for ill-posed multi-material decomposition in dual-energy CT. In the rFast-MMDNet pipeline, an initial stage (SinoNet) maps input sinograms to basis-material projections in the transmission domain using a learned encoder–decoder. These are analytically inverted by Filtered Backprojection (FBP), followed by a dedicated denoising network (DenoiseNet) to remove artifacts. Analytical and learned components are trained separately, enabling independent convergence and control. This hybrid architecture outperforms purely image-domain, algebraic, or single-stage learned baselines, both in quantitative error (per-material RMSE, SSIM) and in interpretability and runtime (Xu et al., 2023).

Hybrid strategies are critical for hardware-robust inference on analog compute-in-memory (CIM) platforms. For state-space models (SSMs), HPD decomposes the most noise-sensitive weight (the output projection) by singular value decomposition: $$I_i = \int_{E_{\min}}^{E_{\max}} S_i(E)\,\exp\Bigl[ -\int_{\ell}\mu(E,\mathbf{r})\,d\ell \Bigr]\,dE$$7. Only $$I_i = \int_{E_{\min}}^{E_{\max}} S_i(E)\,\exp\Bigl[ -\int_{\ell}\mu(E,\mathbf{r})\,d\ell \Bigr]\,dE$$8 is implemented on the noisy analog hardware, while $$I_i = \int_{E_{\min}}^{E_{\max}} S_i(E)\,\exp\Bigl[ -\int_{\ell}\mu(E,\mathbf{r})\,d\ell \Bigr]\,dE$$9 is updated in digital logic, yielding a composite inference pathway that significantly reduces noise-induced perplexity or accuracy degradation in SSMs such as Mamba/Mamba2, with minimal hardware changes and almost complete restoration of baseline robustness (Feng et al., 16 Aug 2025).

5. Interpretability, Evidence Aggregation, and Information Preservation

HPD also addresses the interpretability/accuracy tradeoff in high-dimensional visualization. Given linear projections (e.g., PCA, LPP), HPD decomposes each such projection into sparse sets of axis-aligned subspaces (pairs of original axes). This is achieved by greedy selection of subspaces minimizing the residual structural distortion in local neighborhood relations, encoded on the Grassmannian. When multiple linear projections are considered, Dempster–Shafer theory combines evidence to quantify the cumulative explanatory power of each axis-aligned subspace. Experimental datasets demonstrate that HPD's axis-aligned decompositions retain $\mu(E,\mathbf{r})$0 of local neighborhood structure while providing immediate interpretability (Thiagarajan et al., 2017).

6. Performance, Numerical Evidence, and Practical Considerations

Extensive quantitative and qualitative benchmarks, especially in biomedical imaging and large-scale inverse problems, highlight HPD’s computational and statistical advantages:

Method	Memory/Compute (abbr.)	Error (RMSE/%)	Key Limitations
HPD—dual-energy CT (Cong et al., 2018)	10 min, negligible RAM	Rel. error < 1.5%	Taylor-expansion error in extremes; $\mu(E,\mathbf{r})$1 initialization
Lookup-table (LUT)	62 min, 10 GB	Noisy photoelectric, unstable
Direct optimization	60 min+, unstable/noisy	Local minima, slow	Initialization sensitivity
rFast-MMDNet (Xu et al., 2023)	~78 ms, 124M params	RMSE $\mu(E,\mathbf{r})$20.004, SSIM $\mu(E,\mathbf{r})$30.002	FBP artifacts before denoising
Standard HyBR (Chung et al., 2020)	Storage $\mu(E,\mathbf{r})$4, $\mu(E,\mathbf{r})$5	See main text	Memory growth in large $\mu(E,\mathbf{r})$6
HPD with recycling	Fixed memory budget	Error near full HyBR	Compression selection
SSM HPD (Feng et al., 16 Aug 2025)	$\mu(E,\mathbf{r})$71 cycle latency, digital overhead	PPL up to $\mu(E,\mathbf{r})$8% robust	Storage for $\mu(E,\mathbf{r})$9, $$\mu(E,\mathbf{r}) = p(E)\,f_{\mathrm{ph}}(\mathbf{r}) + c(E)\,f_{\mathrm{co}}(\mathbf{r})$$0

In inverse problems, HPD enables automatic parameter selection (GCV, L-curve, discrepancy principle) within the projected small-scale subspaces, scales to $$\mu(E,\mathbf{r}) = p(E)\,f_{\mathrm{ph}}(\mathbf{r}) + c(E)\,f_{\mathrm{co}}(\mathbf{r})$$1 with only $$\mu(E,\mathbf{r}) = p(E)\,f_{\mathrm{ph}}(\mathbf{r}) + c(E)\,f_{\mathrm{co}}(\mathbf{r})$$2 storage, and, in its recycling variant, permits fixed storage at the cost of negligible error increase.

Hardware-robust HPD in SSMs leads to perplexity reductions of up to 99.57% under moderate analog noise, with accuracy recovery up to 96.67% on commonsense benchmarks, by protecting the most noise-sensitive projection via hybrid analog-digital factorization (Feng et al., 16 Aug 2025).

7. Concluding Synthesis and Open Frontiers

HPD frameworks unify a broad class of methods that interleave projection, decomposition, and constraint enforcement—often across distinct algorithmic or domain boundaries. In essence, HPD leverages hybridization (analytic–numerical, data-driven–analytic, hardware–software, structure–statistical) to achieve computational and estimation efficiencies not attainable by a single-category method.

Limitations are context-dependent and include accuracy loss from approximation (e.g., Taylor expansion order in CT), reliance on suitable initialization, memory or storage for hybrid factors, or complexity in combining evidence from diverse projections. Open directions include fully adaptive rank or parameter selection in hardware-robust HPD, extension to nonlinearly coupled or hierarchical mixed priors in large-scale inversion, and new intersectional applications at the analytic–learning boundary.

HPD continues to expand as a framework for interpretable, robust, and high-performance decomposition in inverse problems, imaging, machine learning, and beyond (Cong et al., 2018, Chung et al., 2022, Elser, 2016, Feng et al., 16 Aug 2025, Xu et al., 2023, Thiagarajan et al., 2017, Chung et al., 2020, Chung et al., 2021).