Overview

The meeting reviewed statistical library design updates and focused on how users should control result types for statistical functions.

Design Decisions

• Clarified standard deviation as the square root of the absolute variance and corrected range terminology for parallel implementations.
• Tightened statistical definitions to prevent division by zero.
• Preferred named structures, including statistical_set-style types and named mean/standard-deviation results, over Boolean flags or positional values.
• Special values will not currently be checked; the treatment of silent NaNs and other constraints remains unresolved.

Type Promotion

• Automatic promotion, user-side data transformation, and initialization parameters were viewed as poor approaches because they can be surprising, costly, or mathematically unsuitable.
• The preferred design is an explicit return-type template argument; the group suggested requiring it for every call rather than providing an overload with an implicit default.

Open Questions

• Should the API require users to specify the internal and return type on every call?
• How will the design behave in real codebases, and what usage patterns will emerge from repository-based examples?

Action Items

• Update the Croydon-related values in the paper to reflect the committee's vote.
• Change the statistical functions to require an explicitly specified return type rather than supporting the overload that infers it.
• Search GitHub repositories for programs that compute statistics and identify candidate codebases to update with ranges and the proposed statistical API, then use them to evaluate the interface.

Outline

Standards Updates

• Standard deviation uses the square root of the absolute variance.
• Parallel implementations use forward ranges.
• Statistical definitions were revised to avoid division by zero.

API Naming

• Statistical sample and population options are represented as structures rather than Boolean parameters.
• Named structures are favored for paired results such as mean and standard deviation.

Special Values

• The current direction is not to check for special values.
• Silent NaNs and additional constraints requiring branching remain under consideration.

Return-Type Promotion

• Discussed four approaches: no transformation, automatic promotion, an initialization parameter, and an explicit return-type template argument.
• The explicit return-type approach was favored, with discussion about requiring it on every call.

Implementation Evidence

• The discussion emphasized the need to examine real programs that compute statistics and use ranges, rather than relying only on library publication examples.