"""ENOB sweep analysis versus number of calibration bits.
Sweeps through bit counts to evaluate how calibration quality improves with more bits.
"""
import numpy as np
import matplotlib.pyplot as plt
from adctoolbox.calibration import calibrate_weight_sine
from adctoolbox.spectrum import analyze_spectrum
[文档]
def analyze_enob_sweep(
bits: np.ndarray,
freq: float | None = None,
harmonic_order: int = 1,
osr: int = 1,
win_type: str = 'hamming',
calibration_mode: str = "prefix_of_full_calibration",
frequency_policy: str = "python",
create_plot: bool = True,
ax=None,
title: str | None = None,
verbose: bool = False
) -> tuple[np.ndarray, np.ndarray]:
"""
Sweep ENOB vs number of bits used for calibration.
Incrementally adds bits (MSB to LSB) and measures ENOB after calibration
to understand diminishing returns and optimal bit count.
Parameters
----------
bits : np.ndarray
Binary matrix (N samples x M bits, MSB to LSB order)
freq : float, optional
Normalized frequency (0-0.5). If None, auto-detect from data
harmonic_order : int, default=1
Harmonic order for calibrate_weight_sine
osr : int, default=1
Oversampling ratio for spectrum analysis
win_type : str, default='hamming'
Window function: 'boxcar', 'hann', 'hamming'
calibration_mode : {'prefix_of_full_calibration',
'recalibrate_each_subset'}, default='prefix_of_full_calibration'
ENOB sweep calibration policy. ``'prefix_of_full_calibration'``
calibrates all bits once and sweeps prefixes of the full-weight
solution. ``'recalibrate_each_subset'`` estimates the frequency once
when needed, then recalibrates each bit-prefix subset independently.
frequency_policy : {'python', 'matlab'}, default='python'
Coarse frequency estimator passed to ``calibrate_weight_sine`` when
automatic frequency search is requested.
create_plot : bool, default=True
If True, plot ENOB sweep curve
ax : plt.Axes, optional
Axes to plot on. If None, uses current axes (plt.gca())
title : str, optional
Title for the plot. If None, uses default title
verbose : bool, default=False
If True, print progress messages
Returns
-------
tuple[np.ndarray, np.ndarray]
- enob_sweep: ENOB for each bit count (length M)
- n_bits_vec: Bit counts from 1 to M
Notes
-----
The default ``'prefix_of_full_calibration'`` mode answers the calibration
ablation question: "after a full-bit calibration, how much performance
remains if lower-bit terms are removed?" Use
``'recalibrate_each_subset'`` only when intentionally asking: "if only the
first n bits are available, how well can that n-bit subsystem be
calibrated?"
What to look for in the plot:
- Increasing trend: More bits improve resolution
- Plateau: Additional bits don't help (noise/distortion limited)
- Decrease: Extra bits add noise/calibration errors
"""
bits = np.asarray(bits)
_, m_bits = bits.shape
valid_modes = {"recalibrate_each_subset", "prefix_of_full_calibration"}
if calibration_mode not in valid_modes:
raise ValueError(
f"Unknown calibration_mode {calibration_mode!r}. "
f"Expected one of {sorted(valid_modes)}."
)
enob_sweep = np.zeros(m_bits)
n_bits_vec = np.arange(1, m_bits + 1)
if calibration_mode == "recalibrate_each_subset":
enob_sweep = _sweep_recalibrating_each_subset(
bits=bits,
freq=freq,
harmonic_order=harmonic_order,
osr=osr,
win_type=win_type,
frequency_policy=frequency_policy,
verbose=verbose,
)
else:
enob_sweep = _sweep_prefix_of_full_calibration(
bits=bits,
freq=freq,
harmonic_order=harmonic_order,
osr=osr,
win_type=win_type,
frequency_policy=frequency_policy,
verbose=verbose,
)
# Plotting
if create_plot:
if ax is None:
ax = plt.gca()
ax.plot(n_bits_vec, enob_sweep, 'o-k', linewidth=2, markersize=8, markerfacecolor='k')
ax.grid(True)
ax.set_xlabel('Number of Bits Used for Calibration')
ax.set_ylabel('ENOB (bits)')
# Set title if provided
if title is not None:
ax.set_title(title)
else:
ax.set_title('ENOB vs Number of Bits Used for Calibration')
ax.set_xlim([0.5, m_bits + 0.5])
ax.set_xticks(n_bits_vec)
valid_enob = enob_sweep[~np.isnan(enob_sweep)]
if len(valid_enob) > 0:
ax.set_ylim([np.min(valid_enob) - 0.5, np.max(valid_enob) + 2])
# Annotate with delta ENOB
delta_enob = np.concatenate([[enob_sweep[0]], np.diff(enob_sweep)])
if len(valid_enob) > 0:
y_offset = (np.max(valid_enob) - np.min(valid_enob)) * 0.06
else:
y_offset = 0.1
for i in range(m_bits):
if not np.isnan(enob_sweep[i]) and not np.isnan(delta_enob[i]):
if i == 0:
annotation_text = f'{delta_enob[i]:.2f}'
text_color = [0, 0, 0]
else:
annotation_text = f'+{delta_enob[i]:.2f}'
normalized_delta = max(0, min(1, delta_enob[i]))
text_color = [1 - normalized_delta, 0, 0]
ax.text(n_bits_vec[i], enob_sweep[i] + y_offset, annotation_text,
ha='center', va='bottom', fontsize=10, fontweight='bold',
color=text_color)
return enob_sweep, n_bits_vec
def _sweep_recalibrating_each_subset(
bits: np.ndarray,
freq: float | None,
harmonic_order: int,
osr: int,
win_type: str,
frequency_policy: str,
verbose: bool,
) -> np.ndarray:
"""Diagnostic mode: recalibrate every bit-prefix subset independently."""
_, m_bits = bits.shape
enob_sweep = np.zeros(m_bits)
auto_frequency_requested = freq is None or np.all(np.asarray(freq) == 0)
if auto_frequency_requested:
full_result = calibrate_weight_sine(
bits,
freq=freq,
harmonic_order=harmonic_order,
frequency_policy=frequency_policy,
)
subset_freq = full_result["refined_frequency"]
else:
subset_freq = freq
for n_bits in range(1, m_bits + 1):
bits_subset = bits[:, :n_bits]
try:
result = calibrate_weight_sine(
bits_subset,
freq=subset_freq,
force_search=False,
harmonic_order=harmonic_order,
frequency_policy=frequency_policy,
)
spectrum_result = analyze_spectrum(
result["calibrated_signal"],
osr=osr,
win_type=win_type,
create_plot=False,
)
enob_sweep[n_bits - 1] = spectrum_result["enob"]
except Exception as exc:
enob_sweep[n_bits - 1] = np.nan
if verbose:
print(f"[{n_bits:2d} bits] FAILED: {exc}")
continue
if verbose:
print(f"[{n_bits:2d} bits] ENOB = {enob_sweep[n_bits - 1]:.2f}")
return enob_sweep
def _sweep_prefix_of_full_calibration(
bits: np.ndarray,
freq: float | None,
harmonic_order: int,
osr: int,
win_type: str,
frequency_policy: str,
verbose: bool,
) -> np.ndarray:
"""Historical Python mode: sweep prefixes of one full-bit calibration."""
_, m_bits = bits.shape
enob_sweep = np.zeros(m_bits)
result = calibrate_weight_sine(
bits,
freq=freq,
harmonic_order=harmonic_order,
frequency_policy=frequency_policy,
)
weights_all = result["weight"]
for n_bits in range(1, m_bits + 1):
bits_subset = bits[:, :n_bits]
weights_subset = weights_all[:n_bits]
calibrated_signal = bits_subset @ weights_subset
try:
spectrum_result = analyze_spectrum(
calibrated_signal,
osr=osr,
win_type=win_type,
create_plot=False,
)
enob_sweep[n_bits - 1] = spectrum_result["enob"]
except Exception as exc:
enob_sweep[n_bits - 1] = np.nan
if verbose:
print(f"[{n_bits:2d} bits] FAILED: {exc}")
continue
if verbose:
print(f"[{n_bits:2d} bits] ENOB = {enob_sweep[n_bits - 1]:.2f}")
return enob_sweep
