# Technical Document Extraction: Comparative Analysis of MLP, KAN, and RAN Models
This document provides a comprehensive extraction of data and visual information from the provided technical graphic, which compares three neural network architectures: **MLP** (Multi-Layer Perceptron), **KAN** (Kolmogorov-Arnold Network), and **RAN** (Rational Activation Network - the proposed method).
---
## 1. Header Section: Error Heatmaps
The top row consists of three 2D heatmaps representing the absolute error $|f - \hat{f}|$ on a log scale.
### Common Scale (Right-most Legend)
* **Metric:** Absolute Error $|f - \hat{f}|$ (Log Scale)
* **Range:** $10^{-8}$ (Dark Purple/Black) to $10^{-2}$ (White/Yellow)
### Heatmap 1: MLP (~5.3k Params)
* **Subtitle:** Fail at Heavy Tails
* **Visual Description:** Shows a large, diffused orange/yellow circular region.
* **Error Analysis:** High error concentrated in the center and spreading outward, indicating a failure to capture the sharp peak and the specific decay of the function.
### Heatmap 2: KAN (~5.2k Params)
* **Subtitle:** Spline Oscillations
* **Visual Description:** Displays a distinct "grid-like" or "checkerboard" pattern of orange/yellow lines.
* **Error Analysis:** The error is structured, revealing artifacts caused by the spline-based nature of the KAN architecture.
### Heatmap 3: RAN (Ours, 72 Params)
* **Subtitle:** Structural Resonance
* **Visual Description:** Almost entirely dark purple/black with uniform, low-intensity noise.
* **Error Analysis:** Significantly lower error across the entire spatial domain compared to MLP and KAN, despite having orders of magnitude fewer parameters.
---
## 2. Main Chart (a): Cross-Section Analysis
**Title:** (a) Cross-Section Analysis: Capturing Global Topology
### Axis Definitions
* **Y-Axis:** Potential $f(x)$ (Range: 0.0 to 1.2)
* **X-Axis:** Spatial Coordinate $x$ ($y=0$) (Range: -3 to 3)
### Legend and Data Series (Spatial Grounding [x=0.8, y=0.8] of plot area)
1. **Ground Truth (Solid Thick Light Grey Line):** Represents the target function. It shows a smooth, bell-shaped curve peaking at $x=0, f(x)=1.0$.
2. **MLP (Tail Bias) (Red Dashed Line):**
* *Trend:* Slopes upward toward the peak but stays consistently above the ground truth at the tails ($x < -1$ and $x > 1$).
* *Observation:* Fails to capture the "heavy tails" of the distribution.
3. **KAN (Ripple) (Blue Dotted Line):**
* *Trend:* Follows the general shape but exhibits high-frequency oscillations (ripples) around the ground truth.
* *Observation:* Visible instability near the peak.
4. **RAN (Ours) (Solid Green Line):**
* *Trend:* Overlaps almost perfectly with the Ground Truth.
* *Observation:* High fidelity across both the peak and the tails.
### Component: Zoom: Peak Fidelity
* **Description:** An inset box magnifying the region $x \in [-0.5, 0.5]$.
* **Detail:** Clearly shows the Blue Dotted line (KAN) oscillating significantly above and below the peak, while the Green line (RAN) remains smooth and centered.
---
## 3. Main Chart (b): The Efficiency Gap
**Title:** (b) The Efficiency Gap: Params vs. Precision
### Axis Definitions
* **Primary Y-Axis (Left):** Parameter Count (Bars) - Log Scale ($10^1$ to $10^4$).
* **Secondary Y-Axis (Right):** MSE Loss (Line, Lower is Better) - Log Scale ($10^{-7}$ to $10^{-3}$).
* **X-Axis:** Model Categories (MLP, KAN, RAN).
### Data Table Reconstruction
| Model | Parameter Count (Bar Height) | MSE Loss (Data Point) | Visual Trend |
| :--- | :--- | :--- | :--- |
| **MLP** | 5300 | $\approx 10^{-5}$ | High params, moderate error. |
| **KAN** | 5200 | $\approx 10^{-4}$ | High params, highest error. |
| **RAN** | 72 | $\approx 10^{-6}$ | Lowest params, lowest error. |
### Key Annotations and Logic Checks
* **Trend Verification (Line):** The black dashed line representing MSE Loss rises from MLP to KAN, then drops sharply to its lowest point at RAN.
* **Trend Verification (Bars):** The bars for MLP (Red) and KAN (Blue) are nearly equal in height, while the RAN bar (Green) is significantly shorter.
* **Callout Box:** A green text box points to the RAN data point stating: **"72 params beat 5k+ params (100x lower MSE vs. KAN)"**.
* **Numerical Labels:**
* Above MLP bar: **5300**
* Above KAN bar: **5200**
* Above RAN bar: **72**
---
## Summary of Findings
The document illustrates that the **RAN** architecture achieves superior precision (lower MSE and better topological fit) while utilizing approximately **1.3%** of the parameters required by MLP or KAN architectures. It specifically eliminates the "Tail Bias" seen in MLPs and the "Spline Oscillations" seen in KANs.
