Choosing K¶
The fit never selects the number of terms K; you do. Every check below reads fits at one
temperature, selected at K = 1 on your table (the recipe). mean_slack_ is
comparable across K, and loss_ is not. The package has no held-out criterion for K.
Methods¶
- Mean slack against K. Fit
n_terms= 1, 2, ... and read the drop inmean_slack_, the mean shortfall of the designs below the fitted reference surface, per added term. A break, one large drop then small ones, shows where added terms stop paying. The mean slack falls with every added term by construction, so it cannot choose K. A table may show no break. - Thinnest term.
term_share_.min()is the share of the near-frontier designs in the thinnest term's cell. The band is the designs within 1/tau of the surface in log V; a cell is the designs for which the term is largest. Read the share against 1/K, what each term holds if the cells are equal. It says a term is thin, not which K is right. - Movement under resampling. Refit on resamples of the rows and measure how far each term's exponent vector sits from the nearest term of the refit (summed absolute differences; exponents sum to one, so at most 2). It is noisy: draws of one procedure differ by several tenths on the cubesat case, so it offers no threshold.
- Planted check. Fit
make_frontier_data(n_terms=K0)at your table's shape to see how recovery looks. The planted K0 is known there; your table's is not. - The certificate.
termination_.statussays the optimizer reached a local minimizer;termination_.radiusis the stable radius, the size of a change of the terms that leaves every design's cell unchanged. Every fit at K = 2 and above is certified, so it does not say K is right.
The temperature¶
For your table M of conditioned metrics, the fit temperature is selected at K = 1 and held for
every K (Fitting; the path against tau):
from tradespace import (
make_frontier_data,
select_temperature,
)
M, _ = make_frontier_data(400, 4, n_terms=2, random_state=0)
tau = select_temperature(M, 0.10, bounds=(3, 1000))
print(round(tau, 1))
0.10 is the share of designs allowed above V = 1. The planted table returns 50.0, used
below for all three (26.5 and 38.1 on the other two planted tables); real data may need bounds=(3, 1e4).
Helpers¶
import numpy as np
from tradespace import HyperbolicTradeSpace
def fit_k(M, K, tau, seed=0):
return HyperbolicTradeSpace(
n_terms=K, temperature=tau, random_state=seed
).fit(M)
def sweep(M, tau, Ks, planted=None):
"""One line per K at one temperature; returns the mean slacks."""
slack = []
for K in Ks:
fit = fit_k(M, K, tau)
t = fit.termination_
line = (
f"K={K} slack={fit.mean_slack_:.3f}"
f" thinnest={fit.term_share_.min():.2f}"
f" {t.status} radius={t.radius:.0e}"
)
if planted is not None:
d = np.abs(fit.exponents_[:, None] - planted[None]).max(
axis=2
)
line += (
f" found={(d.min(axis=0) < 0.05).sum()}"
f" unmatched={(d.min(axis=1) >= 0.05).sum()}"
)
print(line)
slack.append(fit.mean_slack_)
drops = -np.diff(slack)
print("drops", drops.round(3))
return np.array(slack)
def movement(M, K, tau, draws=2, seed=0):
"""Largest distance to the nearest term of a refit on a resample."""
ref = fit_k(M, K, tau)
rng = np.random.default_rng(seed)
moved = np.zeros(K)
for s in range(draws):
W = fit_k(M[rng.choice(len(M), len(M))], K, tau, s).exponents_
d = np.abs(W[:, None] - ref.exponents_).sum(axis=2)
moved = np.maximum(moved, d.min(axis=0))
return moved.max()
A planted table¶
found counts planted terms matched by a fitted term (largest exponent
difference under 0.05); unmatched counts fitted terms matching none.
for K0, seed in ((2, 0), (3, 0), (3, 1)):
X, info = make_frontier_data(400, 4, n_terms=K0, random_state=seed)
print(f"planted K0 = {K0}, table {seed}")
sweep(X, 50.0, range(1, 6), np.asarray(info["exponents"]))
With two planted terms the mean slack drops by 0.108 at K = 2 and by 0.005 or less after; both terms are found at K = 2. With three, table 0 drops by 0.163 and 0.029, then 0.005 or less, and all three are found at K = 3. Table 1 drops by 0.181, then 0.006 or less, and finds two of the three at every K from 2. The break reads the terms the fit recovers, not the number planted.
X, _ = make_frontier_data(400, 4, n_terms=3, random_state=0)
for K in (3, 5):
print(
K, np.round([movement(X, K, 50.0, seed=s) for s in range(2)], 2)
)
On table 0 the movement is 0.04 and 0.02 at K = 3, and 0.23 and 0.24 at K = 5.
The same checks on the cubesat case are on the Examples page.
Large K and the elbow in PCA¶
The elbow heuristic for the number of components in principal component analysis, or PCA, reads a break the same way. The limit differs: variance explained ends at d components, while here K has no ceiling and the mean slack no floor.