RingBuffer.Resample repeats source elements unevenly because of banker's rounding, and its docs promise linear interpolation it doesn't do

[matt-edmondson]

What's wrong

The XML doc on RingBuffer<T>.Resample (Containers/RingBuffer.cs:237) says it "preserves the data pattern by applying simple linear interpolation". It also says that when upsampling, new values are created between existing ones.

The implementation (RingBuffer.cs:274-275) does not interpolate. It picks the nearest existing element:

double oldIndex = (double)i * (oldCount - 1) / Math.Max(length - 1, 1);
int index = (int)Math.Round(oldIndex);

Nearest-neighbour is unavoidable for a generic T. The problem is Math.Round, which defaults to MidpointRounding.ToEven. For common ratios the mapping often lands exactly on .5, and every such tie goes to the even index. Some source elements then appear three times while their neighbours appear once, which distorts the resampled shape.

Reproduction (verified)

new RingBuffer<float>([0, 10], 2).Resample(3);        // [0, 0, 10]
new RingBuffer<float>([0, 1, 2, 3, 4], 5).Resample(9); // [0,0,1,2,2,2,3,4,4]
new RingBuffer<float>([0, 1, 2], 3).Resample(5);       // [0,0,1,2,2]

In the 5→9 case, element 2 appears three times while elements 1 and 3 each appear once.

The existing test ResampleWithSpecificValues_CorrectlyInterpolates (Containers.Test/RingBufferTests.cs:221-235) has a comment saying it expects 10, 15, 20, 25, 30. The actual output is 10, 10, 20, 30, 30. The test passes only because it asserts indices 0, 2 and 4.

Why it matters

Resample is offered for resampling time series and changing sample rates. A caller who relies on the documented interpolation, or at least on even duplication of samples, gets a distorted signal with no error.

Suggested fix / acceptance criteria

• Make the mapping evenly distributed. Either use Math.Round(oldIndex, MidpointRounding.AwayFromZero), or sample from the centre: index = (int)((i + 0.5) * oldCount / length), clamped to [0, oldCount - 1].
• Correct the XML doc to describe nearest-neighbour selection, since T is generic.
• Fix the misleading test comment. Add tests that assert the full output for 5→9 (no source element appears more than ⌈9/5⌉ = 2 times) and for 2→3.
• Keep the large-buffer overflow behaviour from RingBuffer.Resample throws IndexOutOfRangeException and loses the buffer's contents for ~46k+ elements (int overflow) #73 / Stop RingBuffer.Resample's index mapping overflowing for large buffers #76.

[matt-edmondson]

Triage

• Category: Bug (behaviour and docs)
• Priority: Medium. Resample quietly distorts the signal shape at common ratios, its XML doc promises interpolation it doesn't do, and an existing test hides the problem by asserting only indices 0, 2 and 4.
• Area / suggested assignment: Containers/RingBuffer.cs Resample and Containers.Test/RingBufferTests.cs. @matt-edmondson
• Duplicates: None open. It builds on the overflow fix from RingBuffer.Resample throws IndexOutOfRangeException and loses the buffer's contents for ~46k+ elements (int overflow) #73 and Stop RingBuffer.Resample's index mapping overflowing for large buffers #76, which has to be preserved.
• In progress: No open Containers PR.

Notes: Centre sampling, (int)((i + 0.5) * oldCount / length), gives the evenest spread and avoids floating-point ties altogether. Assert the full output arrays in the new tests.

---

Generated by Claude Code