Normalization Factor Explorer
Run the Normalization Factor Explorer MicroSim Fullscreen
You can include this MicroSim on your website using the following iframe:
1 2 | |
About This MicroSim
Run the same signal through two different FFT libraries and you can get answers that differ by a factor of 512. Neither library is wrong. They picked different scaling conventions, and there is no universal standard.
Three conventions are in common use:
| Convention | Forward | Inverse | Round trip |
|---|---|---|---|
| No forward scaling | 1 | 1/N | exact |
| Unitary | 1/√N | 1/√N | exact |
| Forward-normalized | 1/N | 1 | exact |
Every one of them reconstructs the original signal exactly on a forward-then-inverse round trip. The scaling has to land somewhere; the only question is which side of the transform pair carries it.
Why This Matters for Benchmarking
This is a trap when comparing implementations. If you benchmark library A against library B and their magnitudes differ by a factor of N, your first instinct will be that one of them has a bug. Usually neither does — you are comparing a sum against an average.
Before you compare any two FFT implementations numerically, find out what each one does with the scaling. It is often buried in a single sentence of the documentation, and sometimes it is not documented at all and you have to determine it by feeding in a known signal.
How to Use
- At N = 512 with the default convention, note the peak reads 8.00.
- Switch to 1/N forward. The same peak now reads 0.016 — the bars have collapsed to nothing against the fixed axis. Nothing about the signal changed.
- Switch to unitary. The peak reads 0.354, between the two extremes.
- Now change N to 8 and cycle through the conventions again. Note how much less the conventions differ when N is small — and why that makes the problem easy to miss during small-scale testing.
- Read the round-trip figure at every setting. It is always exactly 1.000.
The Axis Does Not Auto-Scale
The y-axis is deliberately pinned to the unscaled peak of 8.0. An auto-scaling chart would make all three conventions look identical, which would hide the entire point. The numeric label above each bar stays readable even when the bar itself has vanished.
Lesson Plan
Grade Level
Undergraduate (college junior/senior)
Duration
10 minutes
Prerequisites
- The DFT produces a magnitude per bin
- Forward and inverse transforms as a pair
Learning Objective
Students will be able to calculate how a chosen scaling convention changes displayed magnitude values, and demonstrate that the choice does not affect correctness.
Activities
- Compute the factors (4 min): For N = 1024, students compute all three forward factors by hand and verify against the readout.
- Predict the peak (3 min): Before switching conventions, students predict the displayed peak value under each.
- Design a test (3 min): Students describe how they would determine an undocumented library's convention using a single known input.
Assessment
Ask: "Library A reports a peak of 256 for a signal; library B reports 0.5 for the same signal at N = 512. Are they inconsistent? What is each one's likely convention?" (No — A uses no forward scaling, B uses 1/N.)
Related Resources
References
- DFT normalization conventions — the standard formulations and where the factor is placed.
- Parseval's theorem — why the unitary convention is the energy-preserving one.