PP-FormulaNet: Unified Formula Recognition

• PP-FormulaNet is a unified framework that integrates advanced methodologies for formula recognition, Petri net analysis, and gravitational computations.
• It employs sophisticated deep learning architectures, including transformer encoders and autoregressive decoders, for accurate LaTeX transcription from images.
• The system leverages formal analysis, dynamic data augmentation, and symbolic computation to enhance efficiency and precision across diverse computational domains.

PP-FormulaNet is a unified designation for advanced frameworks and models at the intersection of mathematical formula recognition and formal symbolic computation, as exemplified in the domains of document intelligence, Markovian Petri-net analysis, and parametrized post-Newtonian formalism. Spanning deep learning systems for LaTeX transcription from images, product-form analysis in Markov processes, and computer algebra components for gravitational theory, the PP-FormulaNet landscape integrates state-of-the-art technical apparatus for extraction, normalization, and formal inference in mathematical structures.

1. Formula Recognition: Architecture and Algorithmic Innovations

PP-FormulaNet denotes a family of end-to-end formula recognition models, optimized for transcribing mathematical expressions from images into LaTeX markup. The architecture consists of:

• Visual encoder: For PP-FormulaNet-L (high-accuracy), a Vary-VIT-B (GOT 2.0) transformer is used: 12 layers, 1024 hidden units, 16 heads, accepting 768×768 input patches. For PP-FormulaNet-S (high-efficiency), a PP-HGNetV2-B4 convolutional backbone is distilled from Vary-VIT-B, reducing feature dimension to P=384 and model size to 15.6M parameters.
• Autoregressive decoder: Both variants implement an MBart-style Transformer decoder. PP-FormulaNet-L deploys 8 layers (512 dimension), while S uses 6 layers (384 dimension).
• Special modules:
  • Weight interpolation via nearest-neighbor resizing, facilitating direct import of pre-trained Vary-VIT-B weights into downsampled layers.
  • Knowledge distillation, where the S variant is supervised at the feature level using L₂ loss against the frozen teacher backbone.
  • Multi-token parallel prediction (only S): enables prediction of k>1 tokens per inference iteration by modifying the causal attention mask.

All variants employ cross-entropy loss for sequence generation, formalized as: $$\mathcal L_{\text{CE}} = -\frac{1}{B}\sum_{i=1}^B \sum_{t=1}^{N_i} \log p(y_{t}^{(i)} \mid y_{<t}^{(i)}, F^{(i)})$$ Feature-level supervision for distillation utilizes a learned projection φ to match student and teacher representations.

2. Mathematical Formalism, Training Routines, and Model Efficiency

Architectural efficiency is achieved through:

• Weight interpolation: Linear weights $F^{D}_{\text{Linear}}$ and normalization parameters are resized using nearest-neighbor interpolation. Empirical results show CPE-BLEU recovery from 0.7970 (no interpolation, Donut-Swin@512) up to 0.9148 (Vary-VIT-B@512 with interpolation).
• Multi-token prediction: A custom causal mask allows block-wise autoregressive decoding,

$$M_{ij} = \begin{cases} 0, & \lfloor i/k \rfloor \geq \lfloor j/k \rfloor \ -\infty, & \text{otherwise} \end{cases}$$

reducing decoding iterations from $N$ to $\lceil N/k \rceil$, at a controlled BLEU loss versus speedup. For instance, increasing $k$ from 1 to 5 decreases batch-1 latency from 2779.8 ms to 600.1 ms, at a cost of 10.4 pp in CPE-BLEU.

• Knowledge distillation: Distilling Vary-VIT-B into PP-HGNetV2-B4 backbone yields CPE-BLEU improvements of 5.17 pp on UniMERNet-1M.

The two models support accuracy–efficiency trade-offs:

Model	Backbone	Avg-BLEU	Time/batch=1 (ms)	Time/batch=15 (ms)
PP-FormulaNet-L	Vary-VIT-B	0.9213	1976.5	332.1
PP-FormulaNet-S	PP-HGNetV2-B4	0.8712	202.3	32.9
UniMERNet	Donut-Swin@1024	0.8613	2267.0	536.8
Pix2tex	[various]	0.7163	1244.6	147.4

PP-FormulaNet-L leads in accuracy by ∼6 pp over UniMERNet, and PP-FormulaNet-S offers a ∼16× speedup in high-throughput scenarios (Liu et al., 24 Mar 2025).

3. Formula Mining System: Data Pipeline and Augmentation

Efficient large-scale formula recognition hinges on a robust data pipeline. PP-FormulaNet introduces a five-stage extraction and normalization process:

1. Source indexing: LaTeX documents are ordered to ensure macro definitions precede their use.
2. Formula extraction: Nonstandard math environments (e.g., align, multline) are parsed; interference from figures/tables is suppressed.
3. Macro recovery: Regular-expression and stack-based parsing expands user-defined macros into canonical LaTeX.
4. Syntax normalization: Formulas are re-rendered via KaTeX and canonicalized (e.g., matrix vs. array environments).
5. Rendering: Each formula is embedded in a template, compiled (PDF via pdflatex), isolated with Fitz+OpenCV, and paired (image, LaTeX).

Augmentations for visual diversity include dilation/erosion, noise, affine transforms, and contrast jitter, consistent with UniMERNet practices.

4. Product-Form Petri Nets and Symbolic Methods

In formal model analysis, PP-FormulaNet encompasses frameworks for the synthesis and steady-state analysis of product-form Markovian Petri nets:

• Product-form Petri net: For a net $(P, T, W^-, W^+, m_0, (\mu_t)_{t\in T})$, a product-form invariant measure exists if for all choices $\{\mu_t\}$ there is $\nu(m) = \prod_{p\in P} u_p^{m(p)}$, $u_p>0$. Steady-state probabilities are then $\pi(m) = Z^{-1}\nu(m)$ with $Z = \sum_{m\in \mathcal{R}(m_0)}\nu(m)$ (Haddad et al., 2011).
• Synthesis rules:
  • S-add (disjoint state-machine insertion)
  • C-update (complex update by modifying input/output bags)
  • P-delete (remove isolated place)
  • These rules generate exactly the WR, deficiency-0 net class, ensuring product-form invariants.
• Complexity: Reachability and liveness for safe product-form nets are PSPACE-complete; unbounded coverability is EXPSPACE-complete.
• Polynomially normalizable subclass: Ordered $\Pi^r$-nets (n-level) admit polynomial-time computation of $Z$ for fixed $n$. The reachability set and normalization constants are systematically constructed via dynamic programming recurrences on marking invariants.

5. Computer Algebra: Parametrized Post-Newtonian Infrastructure

The term PP-FormulaNet, in the context of symbolic computation for gravitational theory, references the design patterns established by xPPN, a Mathematica/xAct package for implementing the PPN formalism (Hohmann, 2020):

• 3+1 decomposition: Space-time tensors are split as $X^\mu = X^0 \partial_t + X^a \partial_a$, with time/space indices LI[0], T3a,... as slots for split components.
• Velocity expansion: VelocityOrder extracts $\mathcal{O}(n)$ velocity-order terms from composite tensor expressions, using a recursive rule set for products and derivatives.
• PPN constants and potentials: Ten standard parameters ($\gamma,\beta,\xi,\alpha_1,\ldots,\zeta_4$) enter field equations with predefined gauge structure, enabling solution workflows across Riemannian, teleparallel, and symmetric teleparallel geometries.
• Rule-based computation: Replacement and expansion rules (OrderSet, ApplyPPNRules) and built-in routines for Euler and potential substitutions facilitate systematic derivation of PPN coefficients for broad families of metric theories.
• Performance and extensibility: Modular index splitting and aggressive pruning of vanishing terms enable scalability. The infrastructure supports arbitrary addition of new background geometries, potential types, and matter couplings using consistent symbolic patterns.

6. Domain Integration and Application Scenarios

PP-FormulaNet frameworks are designed for integration into:

• Document understanding platforms requiring high-accuracy transcription of mathematical content, as needed by publishing, search, and knowledge extraction in research domains (Liu et al., 24 Mar 2025).
• Markovian performance modeling, queuing theory, and analysis of concurrency systems via tractable steady-state analysis of Petri nets (Haddad et al., 2011).
• Automated derivation and symbolic solution of gravitational field equations in metric and non-metric settings, streamlining the calculation of post-Newtonian parameters for extended classes of gravity theories (Hohmann, 2020).

The unification under the PP-FormulaNet nomenclature emphasizes the convergence of neural, algebraic, and formal methods for mathematical pattern recognition, structural analysis, and knowledge base construction across computational disciplines.