Conditional Normalizing Flows (CNFs)

• Conditional Normalizing Flows (CNFs) are deep generative models that map data to latent spaces through invertible transformations, enabling tractable likelihood estimation of complex conditional distributions.
• They employ diverse architectures—including affine coupling layers, Neural ODEs, and graph neural networks—to integrate contextual information and manage non-Gaussian, multimodal outputs.
• CNFs are trained using maximum-likelihood objectives via gradient-based optimization, ensuring sample efficiency and calibrated uncertainty quantification across various real-world applications.

Conditional Normalizing Flows (CNFs) are deep generative models that define families of invertible, flexible maps between simple latent distributions and complex high-dimensional conditional target distributions. By directly parameterizing the change-of-variables between observed variables and a tractable base density, CNFs enable efficient likelihood-based modeling of conditional distributions $p(x|c)$, where $x$ is the target variable and $c$ is a context or conditioning variable. CNFs have become prominent in applications that require calibrated conditional uncertainty quantification, sample efficiency, and the ability to handle highly non-Gaussian or multi-modal posteriors.

1. Mathematical Foundations of Conditional Normalizing Flows

Let $x \in \mathcal{X}$ denote the target variable, and $c \in \mathcal{C}$ the conditioning variable. A conditional normalizing flow defines a smooth, invertible mapping

$z = f_\theta(x; c)$

from $x$ to a latent variable $z$ with a simple, tractable conditional base distribution $p_z(z|c)$. The density $p(x|c)$ is given by the change-of-variables formula: $x$0 The context $x$1 can represent class labels, temporal histories, raw detector readouts, or arbitrary high-dimensional side information.

For Euclidean targets, $x$2 is typically a standard Gaussian; for manifold-valued variables (e.g., directions on the sphere $x$3), $x$4 may be a uniform or Fisher–von Mises distribution, with the flow parameterized to preserve manifold structure (Glüsenkamp, 2023).

Maximum-likelihood training minimizes the expected negative log-likelihood over a dataset $x$5: $x$6 Gradient-based optimization and mini-batch training are standard.

2. Architectural Variants and Conditioning Mechanisms

CNFs’ expressivity derives from the architecture of the invertible map and the treatment of conditioning:

• Affine Coupling and Gaussianization Flows: Typical 1D and low-dimensional flows are constructed by stacking invertible affine-coupling or specialized Gaussianization blocks, with scale/shift parameters produced by conditioning networks (Glüsenkamp, 2023).
• Continuous-Time CNFs: In high-dimensional or continuous settings, the flow is realized via a Neural ODE whose dynamics are parameterized as $x$7, with conditioning $x$8 injected via, e.g., small neural networks (Voleti et al., 2021).
• Graph Neural Network Conditioners: When context has a non-trivial geometric or relational structure (e.g., IceCube detector modules), graph neural networks process $x$9 and emit layerwise flow parameters (Glüsenkamp, 2023).
• Hierarchical/Residual Structures: For robustness and capacity, multi-resolution CNFs decompose the modeling task into hierarchical scales, factorizing the target as products of conditional flows between coarse and fine information (Voleti et al., 2021).
• Mixture/Factorization Methods: In settings with extremely high-dimensional $c$0, hierarchical or soft-gated mixture-of-experts parameterizations are employed to prevent overfitting and promote statistical efficiency (Ausset et al., 2021).

Table: CNF Conditioning Mechanisms (selected settings)

Context Structure	Conditioning Architecture
Tabular, vectors	MLP, feature concatenation
Spatial/temporal grids	CNN/Transformer/State-space model
Graphs	GNN-based per-layer parameterization
Survival/covariates	Softmax-gated vector fields (ODE flows)

3. Training Methodologies and Likelihood Objectives

Across applications, CNFs are trained by exact or approximate maximization of the conditional log-likelihood