Migration to 1.0.0¶

This guide covers migration from the last published PyPI version, 0.0.3, to the first stable release, 1.0.0.

Python Requirement¶

Version 1.0.0 requires Python 3.10 or higher. Version 0.0.3 supported Python 3.8 and higher; upgrade the interpreter before installing 1.0.0.

Package Surfaces¶

online-fdr exposes p-value and e-value methods through separate package surfaces:

• online_fdr.p_values for p-value procedures
• online_fdr.e_values for e-value procedures

The root package no longer re-exports every method.

Import Changes¶

# Before:
# from online_fdr.investing.addis.addis import Addis
# from online_fdr.batching.bh import BatchBH
# from online_fdr.utils.generation import DataGenerator

# Now
from online_fdr.p_values import Addis
from online_fdr.p_values import BatchBH
from online_fdr.core.utils import DataGenerator

For common p-value methods, prefer package-level imports:

from online_fdr.p_values import Addis, BatchBH, Saffron

For e-value methods:

from online_fdr.e_values import EBH, ELond

Validation Changes¶

Constructors reject invalid parameter tuples that older versions accepted. LordPlusPlus requires reward == alpha. LordDependent also requires wealth <= reward, and LordDiscard requires wealth <= tau * alpha. LORD rewards must be finite and positive. SAFFRON gamma sequences require finite nonnegative c and finite gamma_exp > 1; normalization of a custom sequence remains the caller's responsibility. Batch p-values receive shared strict validation. An empty batch is a no-op, and exceeding a finite spending horizon raises a user-facing ValueError.

Weighted GAI++ requires finite positive numeric prior_weight and penalty_weight values (excluding booleans) and a positive integer gamma_horizon. It processes exactly that many tests; the next call raises before changing wealth, counters, or history. The prior-weight cap now enforces the paper's nonnegative reward bounds. Penalty weights that cannot support a nonnegative reward raise ValueError before changing state. error_rate now reports "weighted FDR" or "weighted mem-FDR" according to the configured decay. These describe penalty-weighted error ratios; ordinary FDR requires unit penalty weights and no decay, under the documented independence assumptions. Invalid p-values passed to an asynchronous method's synchronous test_one convenience call leave the testing state unchanged.

Finite spending horizons and Gaussian e-value generator dimensions/batch sizes must be positive integers; booleans and fractional values are rejected. An invalid requested batch size leaves the generator's cursor unchanged. NumPy integer scalars are accepted for these counts and horizons and stored as Python integers. Finite configuration checks also accept NumPy real scalars; accepted configuration values are normalized to Python numbers.

TOAD arrivals advance from the current stage, including after advance_to; finalized decisions remain final. Arrivals also finalize expired decisions; those appear in final_decisions but are not returned again by a later advance_to. Deadlines and stages must be integers (NumPy integers are accepted, 3.0 is not), and streaming weights must lie in (0, 1]. Its identity-shape guarantee requires conditional positive dependence and appropriately prespecified or predictable weights, as detailed in the guarantee matrix.

BatchBHOfficial retains one gamma profile for the whole stream. Automatic selection uses the first nonempty batch. Later changes in batch size do not recalculate historical allocations. Pass gamma_profile="poly" or gamma_profile="half" to choose explicitly; the read-only gamma_profile property reports the selected profile, or None before automatic selection.

Likelihood-ratio mixtures own a fixed prior. The weights property returns a copy; construct a new process to choose a different prior. Products containing both zero and infinity raise ValueError instead of depending on input order. Likelihood-ratio updates also reject undefined cumulative log capital without committing a partial component update. This protects the process's internal capital; caller-supplied functions remain responsible for their own side effects.

State Names¶

New code should use:

• target_level for the configured error-control level
• error_rate for the controlled error-rate family (FDR, mFDR, FWER, weighted FDR, or weighted mem-FDR)
• last_test_level for the last method-specific test level
• last_rejection_threshold for the last input-scale rejection boundary
• num_hypotheses for processed hypotheses
• num_batches for processed non-empty batches where applicable

Use test_one_detail(...) and test_batch_detail(...) when downstream audit logs need thresholds, indices, or metadata in addition to boolean decisions.

Detailed calls execute the same validation and state updates as boolean calls. Requesting a record does not make numeric strings valid input. E-value methods continue to accept e_value= and e_values= keyword arguments.

Persisting Method State¶

Use snapshot() to obtain an opaque JSON-compatible payload, then restore it with the same method's from_snapshot(...) class method. Snapshot state is detached from the running method and supports NumPy scalar values produced by array-based streams. Existing schema-1 e-LOND and weighted GAI++ snapshots remain restorable, including snapshots containing a prefetched next test level. Weighted GAI++ restores historical levels and weights as recorded, while future calls use the corrected prior cap and updated error-target metadata. Restoring a snapshot cannot repair invalid past statistical decisions.

Legacy BatchBHOfficial snapshots infer their fixed profile from the first recorded allocation. Snapshots with negative or nonfinite past levels, or accounting that exceeds the inferred fixed gamma budget, are rejected. Older snapshots do not store historical batch sizes, so restoration cannot establish the validity of past mixed-profile allocations; it preserves those decisions and fixes the profile for future batches.

The built-in gamma sequences and spending functions have explicit snapshot codecs. Custom helpers, including subclasses of those built-ins, raise TypeError when taking a snapshot; they are never silently restored as a different helper implementation.

Memory-Decay Evaluation¶

MemoryDecayFDR uses running discounted counts, so a scalar score_one call takes constant time. Requesting cumulative output copies the full score history. delta and offset are fixed after construction; labels and ground_truth return detached lists. Construct a new tracker to change its configuration. The writable cumulative flag can still switch between scalar and historical outputs at any point. Configuration requires finite delta in [0, 1] and finite nonnegative offset. Floating-point summation order can differ from the previous convolution implementation by roundoff.