cornetto — compact reference for LLMs¶

This page is a dense, self-contained reference intended for AI assistants helping users with cornetto.

Package identity¶

PyPI: pip install cornetto
Version: 0.2.0
License: MIT
GitHub: https://github.com/nicoborghi/cornetto
Backend: KDExpress FFT-KDE

Data format¶

data = {
    "param_name": array,   # shape (N,) single chain  OR  (N_chains, N) multi-chain
    ...
}

All arrays in the same dict must share the same N (samples per chain).
1-D arrays are broadcast to all chains when mixed with 2-D arrays.
Hard cap: MAX_CHAINS = 10.

Main functions¶

from cornetto import corner, quick_corner, marginal, trace, trace_marginal, Cornetto

# Full corner plot (FFT-KDE, smooth contours)
fig, axes = corner(data, **kwargs)

# Fast histogram-only corner (~7× faster on 8-param chains)
fig, axes = quick_corner(data, **kwargs)

# 1-D marginal grid
fig, axes = marginal(data, **kwargs)

# Trace plot (sample index vs. value)
fig, axes = trace(data, **kwargs)

# 1-D KDE + trace side-by-side
fig, axes = trace_marginal(data, **kwargs)

# Stateful class — caches KDEs between plot() calls
c = Cornetto(data, **analysis_kwargs)
fig, axes = c.plot(**plot_kwargs)
c.info()       # print setup
c.summary()    # SummaryTable (HTML in Jupyter, text in terminal)
c.latex()      # AASTeX LaTeX table (requires astropy)
a.pairplot(b)  # lower triangle = a, upper triangle = b

Key parameters (corner / Cornetto)¶

Parameter	Type	Default	Notes
`data`	`dict[str, array]`	required	See data format above
`params`	`list[str]`	`None`	Subset to draw; others are ignored entirely
`labels`	`dict[str, str]`	`{}`	Display labels; LaTeX via raw strings works
`truths`	`dict[str, scalar\\|array]`	`None`	Scalar = shared; array of length `N_chains` = per-chain
`chain_labels`	`list[str]`	`None`	Legend labels, one per chain
`weights`	array or list	`None`	Importance-sampling weights; normalised internally
`subsample`	`int`	`None`	Thin chain before KDE
`stat`	`str\\|callable`	`"median"`	`"median"`, `"median_mad"`, `"median_hdi"`, `"mean"`
`sigmas`	`tuple[float]`	`(1, 2)`	n-sigma contour levels
`color`	`str\\|list`	`"cornetto"`	Named: `"indigo"`, `"coral"`, `"teal"`, `"gold"`, `"ink"`
`dark`	`bool`	`False`	Dark theme
`smooth`	`bool`	`True`	KDE (`True`) or histogram (`False`) on diagonal
`fill_contours`	`bool`	`True`	Filled 2-D contour bands
`show_titles`	`bool`	`True`	`value⁺ᵃ₋ᵦ` heading (single-chain only)
`ax_lims`	`dict[str, (lo,hi)]`	`None`	Explicit axis ranges (parameter: `ax_lims`, NOT `range`)
`kwargs_truths`	`dict`	`None`	Truth line style: `color`, `lw`, `ls`, `alpha`, `marker`

Gotchas¶

ax_lims, not range — parameter is ax_lims to avoid shadowing Python's built-in range().
KDExpress fft_kde2d always needs weights — cornetto passes uniform weights internally when the user doesn't supply any; do not pass weights=None directly to KDExpress.
Transpose — fft_kde2d returns shape (Mx, My); cornetto transposes to (My, Mx) for matplotlib's convention.
JAX CPU warning — suppressed automatically at import; no user action needed.
LaTeX labels — use raw strings (r"$m_1$"); usetex=False by default for portability.
Small N — fewer than ~50 samples per parameter disables KDE for that panel and falls back to a histogram with a RuntimeWarning.
N_chains limit — default max_chains=10; override per call.

Custom stat function signature¶

def my_stat(samples: np.ndarray, weights=None) -> dict:
    return {"center": float, "lo": float, "hi": float, "label": str}

corner(data, stat=my_stat)

Minimal working example¶

import numpy as np
from cornetto import corner

rng = np.random.default_rng(0)
data = {"m1": rng.normal(30, 4, 10_000), "m2": rng.normal(25, 3, 10_000)}
fig, axes = corner(data, labels={"m1": r"$m_1$", "m2": r"$m_2$"},
                   truths={"m1": 30.0}, chain_labels=["posterior"])
fig.savefig("corner.pdf", bbox_inches="tight")