The Luna Bound Propagator for Formal Analysis of Neural Networks
Henry LeCates, Haoze Wu · 2026
Read The Luna Bound Propagator for Formal Analysis of Neural Networks by Henry LeCates, Haoze Wu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-23878.txt.
Introduction
LeCates and Wu introduce Luna, a C++ tool for formal analysis of neural networks. Luna propagates bounds through a network to check its behavior. It supports interval propagation, the DeepPoly and CROWN analyses, and the parameterized alpha-CROWN method. Earlier alpha-CROWN code was Python only, which made integration into other verifiers harder. Luna works over a general computational graph. On VNN-COMP 2025 benchmarks, they report tighter bounds and faster runs than the leading alpha-CROWN implementation. The code is publicly available.
Catalog record
| CREATOR | Henry LeCates, Haoze Wu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 26429 |
| LICENSE | http://creativecommons.org/licenses/by-sa/4.0/ |
| RIGHTS | CC BY SA |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.23878
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
The Luna Bound Propagator for Formal Analysis of Neural Networks Henry LeCates, Haoze Wu Abstract The parameterized CROWN analysis, a.k.a., α \alpha -CROWN, has emerged as a practically successful bound propagation method for neural network verification. However, existing implementations of α \alpha -CROWN are limited to Python, which complicates integration into existing DNN verifiers and long-term production-level systems. We introduce Luna , a new bound propagator implemented in C++. Luna supports Interval Bound Propagation, the CROWN analysis, and the α \alpha -CROWN analysis over a general computational graph. We describe the architecture of Luna and show that it is competitive with the state-of-the-art α \alpha -CROWN implementation in terms of both bound tightness and computational efficiency on benchmarks from VNN-COMP 2025. 1 Introduction Bound propagation is an important technique in the formal analysis of deep neural networks (DNNs). Modern DNN reasoners, such as \alpha - \beta -CROWN [ 24 , 20 , 22 ] , ERAN [ 14 , 15 , 16 ] , Marabou [ 8 , 21 ] , NeuralSAT [ 4 ] , NNV [ 18 , 10 ] , nnenum [ 2 ] , VeriNet [ 6 , 7 ] , etc, all employ some version of bound propagation, either using it as stand-alone, incomplete DNN verification procedures, or as in-processing passes in a branch-and-bound/case-analysis-based search shell. Over the past few years, numerous bound propagation methods (e.g., [ 19 , 15 , 25 , 12 , 23 ] , inter alia) have been proposed. They differ in the choice of the abstraction domain and abstraction refinement strategies, and span over the spectrum of the efficiency and precision trade-off, from exact bounding based on MILP [ 17 ] , to fast bounding based on interval arithmetic [ 5 ] . In this paper, we focus on the parameterized CROWN (a.k.a. \alpha -CROWN) analysis [ 25 ] , which empirically strikes a good balance between precision and efficiency, and underlies the top two performers ( \alpha - \beta -CROWN and NeuralSAT) in the past two VNN-COMPs [ 3 , 9 ] . To date, practical implementations of \alpha -CROWN are only available in Python. The most widely used implementation is auto_LiRPA [ 22 ] , which serves as the bound-propagation backend for both \alpha - \beta -CROWN and NeuralSAT. While auto_LiRPA is in itself a high-performance tool, invoking the Python-based \alpha -CROWN from systems written in other programming languages, e.g., C++ or MATLAB, can incurs nontrivial integration and start-up overhead. To reduce this barrier and make \alpha -CROWN accessible as a more easily integrable component for a wide range of verification tools (including future production-level verifiers), we develop Luna , the first C++ implementation of \alpha -CROWN with a stable interface suitable for foreign-function integration. Luna is designed for the common setting in which a verifier must obtain sound output bounds for a given input region, while incurring minimal per-invocation overhead and still leveraging parallelization when available. Luna currently includes the following features: • support for interval bound propagation (IBP), CROWN, and \alpha -CROWN analyses; • parallelization for the core tensor and bound computations leveraging the Torch Deep Learning Library; • multiple integration modes, including a command-line interface, a native C++ API, and Python bindings. On benchmarks from VNN-COMP 2025, Luna produces output bounds that are comparable with auto_LiRPA , the state-of-the-art \alpha -CROWN implementation, with overall lower runtime and start-up overhead. We expect that Luna will help to lower the engineering effort required to integrate strong bound-propagation techniques into existing tools and to promote research work that focus on other algorithmic bottlenecks in neural network verification. 2 Preliminaries on Bound Propagation Let f:\mathbb{R}^{n}\to\mathbb{R}^{m} be a feedforward neural network. Given vectors \ell,u\in\mathbb{R}^{n} , define the input box \mathcal{X}:=\{x\in\mathbb{R}^{n}\mid\ell\leq x\leq u\} . The goal of bound propagation is to compute an output box \mathcal{Y}:=\{y\in\mathbb{R}^{m}\mid a\leq y\leq b\} , such that f(x)\in\mathcal{Y} for all x\in\mathcal{X} . CROWN The CROWN analysis [ 25 ] (or equivalent the DeepPoly analysis [ 15 ] ) operates by constructing affine relaxations of each nonlinear layer. For a nonlinear layer z^{\prime}=\sigma(z) where z\in[\ell,u] , it builds affine functions h_{L}(z) and h_{U}(z) such that h_{L}(z)\leq\sigma(z)\leq h_{U}(z)\text{ for all }z\in[\ell,u] . By composing these affine bounds layer by layer, one obtains global affine lower and upper bounds on the network output. The concrete output bounds are then obtained by concretizing the global affine bounds using the input box \mathcal{X} . Parameterized CROWN ( \alpha -CROWN) Many nonlinearities admit multiple valid affine relaxations. For example, \mathrm{ReLU}(z)=\max(z,0)\geq\alpha\cdot z\text{ for all }\alpha\in[0,1] . Instead of fixing the relaxation, \alpha -CROWN [ 22 ] considers a family of linear relaxations parameterized by a variable \alpha . This induces parameterized global bounds L(x,\alpha)\leq f(x)\leq U(x,\alpha) where L(x,\alpha) and U(x,\alpha) are affine in x for fixed \alpha . The choice of \alpha is then optimized (once for lower bound and once for upper bound) by solving \max_{\alpha\in\mathcal{A}}\;\min_{x\in\mathcal{X}}L(x,\alpha),\qquad\min_{\alpha\in\mathcal{A}}\;\max_{x\in\mathcal{X}}U(x,\alpha), where \mathcal{A} represents the valid ranges of \alpha (e.g., [0,1] for ReLU lower bound). In practice, this nonconvex optimization problem is solved using projected gradient descent [ 11 ] over \alpha . 3 The Luna Bound Propagator Figure 1 : High-level overview of Luna ’s system architecture. This section describes the architecture design of Luna . A high-level overview is presented in Figure 1 . Luna takes as input a neural network in the ONNX format [ 1 ] and a specification in the VNN-LIB format, and computes sound bounds with respect to the input region defined in the specification. 3.1 Bounded Model Luna parses the inputs and convert them into (i) a bound-aware computational graph on top of the Torch Deep Learning Library [ 13 ] , (ii) an input domain with element wise lower and upper bounds, and (iii) a set of linear output constraints in the form \mathbf{C}\mathbf{y}\leq\mathbf{t} . These three components constitute the Bounded Model, which is an internal representation of the verification problem shared by all analysis routines. App. 0.A.1 provides implementation details on parsing. The computational graph is a directed acyclic graph whose nodes correspond to layer operations. For each node, three types of bounds are recorded: concrete bounds from interval arithmetic (IBP), linear relaxations from CROWN, and optimized bounds from \alpha -CROWN. The specification matrix C represents a collection of linear constraints on the neural network output, which can be viewed as an additional layer following the output layer. During bound propagation, Luna treats this additional layer as the actual output layer and seeks to bound its output. For example, for an output constraint y_{1}\leq y_{2} , Luna seeks to bound the value of an auxiliary variable a=y_{1}-y_{2} . Bounded Node Modules Each node in the computational graph is associated with a Bounded Node module, which implements operator-specific bound computation. Two types of propagation are implemented: 1) Interval Bound Prorogation (IBP), which is executed before the CROWN and \alpha -CROWN analysis to obtain initial bounds on the neurons and detect stable neurons; 2) the CROWN backward propagation, which takes as input affine bounds of the output node in terms of the subsequent node, propagates them backward through the node, and outputs affine bounds of the output node in terms of the current node. Non-linear optimizable activation nodes inherit from