from __future__ import annotations

import numpy as np


def mean_squared_error(reference: np.ndarray, test: np.ndarray) -> float:
    diff = np.asarray(reference, dtype=np.float32) - np.asarray(test, dtype=np.float32)
    return float(np.mean(diff**2))


def peak_signal_to_noise_ratio(reference: np.ndarray, test: np.ndarray, data_range: float) -> float:
    mse = mean_squared_error(reference, test)
    if mse <= 1e-12:
        return float("inf")
    data_range = max(float(data_range), 1e-6)
    return float(20.0 * np.log10(data_range) - 10.0 * np.log10(mse))


def gray_levels_used(level_indices: np.ndarray) -> int:
    return int(np.unique(np.asarray(level_indices)).size)


def shannon_entropy(level_indices: np.ndarray) -> float:
    values, counts = np.unique(np.asarray(level_indices), return_counts=True)
    if values.size == 0:
        return 0.0
    probabilities = counts.astype(np.float64) / counts.sum()
    return float(-(probabilities * np.log2(probabilities + 1e-12)).sum())


def gray_levels_used_from_image(image: np.ndarray) -> int:
    return int(np.unique(np.asarray(image)).size)


def shannon_entropy_from_image(image: np.ndarray) -> float:
    values, counts = np.unique(np.asarray(image), return_counts=True)
    if values.size == 0:
        return 0.0
    probabilities = counts.astype(np.float64) / counts.sum()
    return float(-(probabilities * np.log2(probabilities + 1e-12)).sum())


def clipped_pixel_ratios(image: np.ndarray, low: float, high: float) -> tuple[float, float]:
    image = np.asarray(image, dtype=np.float32)
    total = float(image.size)
    low_ratio = float(np.count_nonzero(image < low) / total)
    high_ratio = float(np.count_nonzero(image > high) / total)
    return low_ratio, high_ratio


def format_metric(value: float, digits: int = 3) -> str:
    if np.isinf(value):
        return "inf"
    return f"{value:.{digits}f}"