# Dimension-aware power-mean fusion

- ID: 118
- Canonical URL: https://synthcore.org/idea/118/dimension-aware-power-mean-fusion
- API JSON: https://synthcore.org/api/idea/118.json
- API Markdown: https://synthcore.org/api/idea/118.md
- Verification status: unverified
- Source: [arXiv:2608.23963](https://arxiv.org/abs/2608.23963)
- Category: architecture
- Solves: accuracy, stability
- ML areas: mlp, cnn, moe-routing
- Math tags: inequalities, convex-analysis, measure-theory
- Ratings: usefulness 5/10; difficulty 3/10; novelty 5/10

## Idea description

Use the paper's dimension-dependent exponent transformation to fuse nonnegative outputs from several branches. Instead of selecting an arbitrary generalized-mean exponent, choose the output exponent q=Q_d(p) induced by an input exponent p, making the fusion rule explicitly sensitive to the dimension of the barycentric variables.

## Mathematical statement

For positive weights lambda_i summing to one and nonnegative branch values a_i, define the weighted power mean M_q^lambda(a)=(sum_i lambda_i a_i^q)^(1/q), with the geometric mean at q=0 and the minimum at q=-infinity. The paper maps an input exponent p to Q_d(p)=p/(1-dp), where d is the dimension of the variables combined by the barycenter and p<=1/d. Define the neural fusion operator F_{p,d}(a)=M_{Q_d(p)}^lambda(a). At p=0, Q_d(0)=0 and fusion is geometric. As p approaches 1/d from below, Q_d(p) approaches positive infinity and fusion approaches a weighted maximum. Negative p emphasizes weak branches and approaches a minimum as p approaches negative infinity. The rigorous BBL guarantee applies when a_i arise from ratios of integrable functions evaluated at barycentric inputs; using the operator directly on features is a testable architectural heuristic rather than an automatic theorem.

## Key formulas

- $$Q_d(p)=\frac{p}{1-dp},\qquad p\le\frac{1}{d}.$$
- $$F_{p,d}(a_1,\ldots,a_m)=M_{Q_d(p)}^{\lambda}(a_1,\ldots,a_m)=\left(\sum_{i=1}^{m}\lambda_i a_i^{Q_d(p)}\right)^{1/Q_d(p)}.$$
- $$M_0^{\lambda}(a_1,\ldots,a_m)=\prod_{i=1}^{m}a_i^{\lambda_i},\qquad M_{-\infty}^{\lambda}(a_1,\ldots,a_m)=\min_i a_i.$$

## Implementation notes

Modify a multi-branch fusion point such as multi-sensor feature fusion, confidence aggregation, or an MoE gate when every branch emits a nonnegative scalar or channel-wise activation. Set a_i=softplus(h_i)+epsilon, use fixed convex weights lambda_i or obtain them from a softmax router, and compute y=F_{p,d}(a_1,...,a_m). For a chosen dimension d and exponent p<1/d, calculate q=p/(1-dp), then compute y=(sum_i lambda_i a_i^q)^(1/q). Use y=exp(sum_i lambda_i log(a_i)) when absolute q is below 1e-4. For large positive q, evaluate in the log domain using log-sum-exp; for negative q, clamp a_i to epsilon. A learnable version can parameterize p=(1/d-epsilon)tanh(s), where s is a scalar or per-layer parameter, and add a weak penalty encouraging s near zero. The paper supplies the exponent map and endpoint behavior; the experiment must determine whether the effective dimension should be one, the spatial dimension, or the feature dimension. First compare arithmetic mean, geometric mean, fixed quadratic mean, minimum, and F_{p,d} in a two-sensor CIFAR-10 classifier and a small segmentation network with one randomly corrupted sensor. Sweep p in {-1,-0.25,0,0.1,0.25} for d=1 and d=2. Record clean accuracy, corrupted-sensor accuracy, gradient variance across branches, and sensitivity to branch dropout. Success is higher robustness or lower gradient imbalance at matched parameter count and FLOPs.

## Disclaimer

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