# Dimension-calibrated contractive branch attractor

- ID: 2419
- Canonical URL: https://synthcore.org/idea/2419/dimension-calibrated-contractive-branch-attractor
- API JSON: https://synthcore.org/api/idea/2419.json
- API Markdown: https://synthcore.org/api/idea/2419.md
- Verification status: unverified
- Source: [arXiv:2608.16817](https://arxiv.org/abs/2608.16817)
- Category: architecture
- Solves: stability, generalization, sample-efficiency
- ML areas: rnn, world-model, vae, architecture
- Math tags: dynamical-systems, geometry, measure-theory, approximation-theory
- Ratings: usefulness 6/10; difficulty 5/10; novelty 7/10

## Idea description

Construct a recurrent or generative network from finitely many contractive branches whose hidden-state attractor has a prescribed similarity dimension. The branch contraction ratios determine the target complexity through the equation sum_i r_i^s = 1, while a separation penalty approximates the open set condition and prevents branch collapse or excessive overlap.

## Mathematical statement

Let F = {f_i} be a finite iterated-function system on a complete geometrically doubling state space (X,d). Each branch is a similitude satisfying d(f_i(x), f_i(y)) = r_i d(x,y), with 0 < r_i < 1. Its attractor A is the unique compact set satisfying A = union_i f_i(A). The similarity dimension s is the unique solution of sum_i r_i^s = 1. The open set condition requires an open set U such that f_i(U) is contained in U and f_i(U) and f_j(U) are disjoint for i not equal to j. Under the paper's regularity assumptions, the open set condition implies 0 < H^s(A) < infinity, while positivity of H^s(A) implies the open set condition. In a neural implementation, exact similitudes are approximated by branch maps with controlled Jacobian singular values, and covering-number slopes estimate the attractor dimension.

## Key formulas

- d(f_i(x), f_i(y)) = r_i d(x,y), with 0 < r_i < 1
- A = union_i f_i(A), and sum_i r_i^s = 1
- f_i(U) subset U and f_i(U) intersection f_j(U) = empty set for i != j implies 0 < H^s(A) < infinity
- N_A(epsilon) proportional to epsilon^(-s), so the estimated dimension is s_hat(epsilon) = - d log N_A(epsilon) / d log epsilon

## Implementation notes

1. Exact integration point: use this mechanism as the transition module in a branching RNN, recurrent world model, or generative decoder. Represent hidden state h in R^D and branch i by f_i(h) = c_i + r_i Q_i h, where c_i is learned, Q_i is orthogonal or spectrally normalized, and r_i = sigmoid(alpha_i) is constrained below 1 minus a safety margin. A gate selects or samples a branch at every transition. Choose a target dimension s0 and regularize the contraction ratios so that sum_i r_i raised to s0 is close to 1. 2. Pseudocode: for every minibatch of states h_b, compute z_bi = c_i + r_i Q_i h_b; compute L_sep = sum over b and i less than j of exp(-||z_bi - z_bj|| squared divided by tau squared); compute L_dim = (log(sum_i r_i raised to s0)) squared; optimize L_task + lambda_sep L_sep + lambda_dim L_dim. Estimate each Q_i spectral norm and rescale it if necessary. 3. Taken from the paper: the contraction requirement, similarity-dimension equation, and implication from open-set separation to positive finite Hausdorff measure. Estimated empirically: Jacobian contraction factors, overlap rates, covering numbers, and dimension slopes. 4. First cheap experiment: train a four-branch model on sequential MNIST or a two-dimensional mixture-of-Gaussians generator. Compare unconstrained branches, contraction-only branches, and contraction-plus-separation branches. Roll out 10,000 hidden states after burn-in, estimate covering numbers over logarithmically spaced epsilon values, and fit the log-log slope. The predicted signature is that the separated model approaches target slope s0, while removing separation produces a lower slope and more duplicate states. Perturbation norms should decay at least approximately as (max_i r_i)^t times the initial perturbation.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.