Random Orthogonal Initializations

• Random orthogonal initializations are procedures that generate weight matrices uniformly from orthogonal groups using the Haar measure, ensuring stability in neural network training.
• Algorithms such as QR decomposition and Householder reflections efficiently produce Haar-uniform matrices with O(n^3) computation and O(n^2) memory requirements.
• Empirical studies show that using orthogonal initializations improves gradient flow and overall accuracy in deep architectures by preventing signal vanishing or explosion.

Random orthogonal initializations refer to procedures for generating weight matrices (or transformation matrices) that are uniformly distributed in orthogonal groups—typically with respect to the Haar measure—and are central in numerous fields including machine learning, probability, theoretical physics, and computational geometry. Orthogonal initializations ensure stability of signal propagation and gradient flow in deep architectures, and provide symmetry properties critical to applications requiring conservation laws or metric preservation.

1. Mathematical Foundations of Orthogonal Groups and Haar Measure

The classical real orthogonal group $$O(n) = \{A \in \mathbb{R}^{n \times n} : A^T A = I_n\}$$ is comprised of all distance-preserving linear maps on $\mathbb{R}^n$ (Saraeb, 2024). The Haar measure $\mu$ on $O(n)$ is the unique probability measure invariant under left and right multiplication by fixed orthogonal matrices, ensuring uniform sampling from $O(n)$. For generalized forms, the set of matrices $A$ satisfying $A^T S A = S$ (with $S$ a fixed invertible symmetric or skew-symmetric matrix) preserves the associated bilinear form. Special cases yield groups such as symplectic, Lorentz, and indefinite orthogonal groups, underpinning applications in theoretical physics, computational geometry, and number theory.

2. Algorithms for Generating Haar-Uniform Orthogonal Matrices

Efficient generation of Haar-distributed orthogonal matrices can be achieved via two primary algorithms:

• QR decomposition of Gaussian matrices: For $Z$ with i.i.d. $N(0,1)$ entries, perform economy-size QR decomposition ($\mathbb{R}^n$0). Construct $\mathbb{R}^n$1 so that $\mathbb{R}^n$2 is Haar-uniform (Saraeb, 2024).
• Householder reflections: Iteratively apply random Householder reflections to the identity ($\mathbb{R}^n$3), where each reflection is defined by a random Gaussian vector in progressively lower-dimensional blocks, yielding a Haar-uniform orthogonal $\mathbb{R}^n$4.

Both methods require $\mathbb{R}^n$5 operations and $\mathbb{R}^n$6 memory; standard QR algorithms are backward-stable. For very high dimensions ($\mathbb{R}^n$7), structured or block-orthogonal initializations may be preferable.

3. Information-Geometric and Kernel-Theoretic Properties

Mean-field theory in deep networks shows that orthogonal weights ensure both the forward-propagated activations and backpropagated gradients remain near $\mathbb{R}^n$8 isometries, preventing vanishing or exploding signals (Sokol et al., 2018). Concretely, for networks initialized with orthogonal $\mathbb{R}^n$9, the spectral radius $\mu$0 of the input-output Jacobian satisfies $\mu$1 as depth grows, in contrast to Gaussian-initialized networks where $\mu$2 for large depth.

The Fisher information matrix (FIM) $\mu$3 curvature is controlled by the maximal singular value of $\mu$4; near-isometric initialization allows larger learning rates. Manifold-based optimization (e.g. Stiefel manifold) can maintain exact orthogonality during training, stabilizing Fisher curvature but not necessarily guaranteeing improved optimization speed.

For kernel approximations, single-layer neural networks initialized with Haar-distributed (possibly rescaled) orthogonal matrices converge, as width increases, to the same deterministic kernel as their Gaussian-initialized counterparts. This equivalence holds for activation functions with bounded derivatives, and the finite-width convergence rate matches the Gaussian case (Martens, 2021).

4. Generalized Random Orthogonal Initializations

Sampling $\mu$5 so that $\mu$6 (for invertible symmetric or skew-symmetric $\mu$7) is generalized as follows (Saraeb, 2024):

1. Decompose $\mu$8: Apply real Schur or spectral decomposition $\mu$9, where $O(n)$0 is block-diagonal.
2. Blockwise sampling: Draw block-diagonal $O(n)$1 satisfying $O(n)$2; each block is sampled from $O(n)$3 or $O(n)$4 as appropriate.
3. Form $O(n)$5: Set $O(n)$6; for indefinite orthogonal ($O(n)$7) $O(n)$8 and $O(n)$9 with $O(n)$0. For symplectic ($O(n)$1) $O(n)$2, reduction uses $O(n)$3 draws.

Applications include Hamiltonian neural networks (canonical 2-form preservation), Lorentzian/hyperbolic embeddings (Minkowski metric preservation), and metric learning with indefinite inner products.

5. Cayley Transform Parametrization and Statistical Approximations

The Cayley transform provides a practical parametrization for generating random orthogonal matrices on the Stiefel ($O(n)$4) and Grassmann ($O(n)$5) manifolds (Jauch et al., 2018). For $O(n)$6,

$O(n)$7

Stiefel points are obtained by constraining $O(n)$8 to block-skew forms; Grassmann points by further simplification. The induced density under change-of-variables is given by the Jacobian determinant $O(n)$9 for Euclidean parameters $A$0.

Asymptotic theory shows that, for large $A$1, the components of $A$2 behave nearly independently and normally,

$A$3

with total error $A$4. For weight initialization, a Gaussian-approximation sampler provides nearly Haar-uniform orthogonality and is computationally preferable to exact MCMC on manifold coordinates.

6. Empirical Results, Biological Plausibility, and Practical Guidance

Empirically, in recurrent and deep feedforward architectures, random orthogonal initialization yields substantial improvements over random Gaussian weights:

• In synthetic RNN tasks, maximum sequence length solved increased substantially under separate pre-training or penalty-enforced orthogonality (Manchev et al., 2022).
• In deep feedforward MNIST networks, test accuracy exceeded 97% with orthogonal initialization vs. baseline 11.35% with random initialization.

Biologically plausible schemes are presented:

• Layer-wise pre-training: Each weight matrix $A$5 is optimized locally using $A$6 until nearly orthogonal.
• Penalty enforcement during training: Adds a term $A$7 to the loss.

Convergence of such pre-training is theoretically ensured: for large dimensions $A$8, loss minimization reliably drives $A$9 toward orthogonality. Also, local plasticity and global homeostatic constraints provide plausible neurobiological analogs to orthogonal weight evolution.

Implementation tips: Standard linear algebra libraries (NumPy, MATLAB) suffice for QR-based and Householder generation. For very high dimension or structured applications, block-orthogonal or sparse Householder layer products may be necessary.

7. Summary of Key Results and Limitations

• Uniform (Haar) random orthogonal initializations can be efficiently generated and provide stable signal and gradient dynamics.
• In both mean-field and kernel-theoretic perspectives, random orthogonal and Gaussian-initialized networks converge to identical kernels in the infinite-width limit, given rescaling.
• Generalized orthogonal initializations enable structure-preserving initial weights for specialized applications.
• Exact maintenance of orthogonality through manifold optimization stabilizes curvature but is not sufficient for optimal learning rates; the trajectory of Fisher curvature and NTK eigenvalues is critical.
• Gaussian approximation via the Cayley transform produces high-fidelity orthogonal matrices for initialization in high dimensions.
• Biologically-motivated approaches demonstrate empirical benefits and offer plausible mechanisms for orthogonal matrix formation in neural architectures.

This collective body of work delineates the theory, algorithms, and practical utility of random orthogonal initializations and provides rigorous foundations for their continued application and generalization in research and practice.