numpy.fft计算exp(-|x|)的解析与数值傅里叶变换差异原因探究

Great question! Let's break down why you're seeing this cosine modulation, and how to verify this from numpy's documentation.

Core Reason: FFT's Periodicity Assumption #

The key issue here is that numpy's fft implements the Discrete Fourier Transform (DFT), which inherently treats your input signal as a periodic, infinitely repeating sequence.

Your signal sig = np.exp(-np.abs(x)) is truncated to the interval [-10, 10] (2001 points). The DFT doesn't see this as a finite signal—it assumes this segment repeats infinitely in both directions. But the infinite exp(-|x|) is not periodic, so this periodic extension creates subtle discontinuities at the boundaries of your segment (e.g., the value at x=10 is exp(-10), while the next "period" starts at x=10+dx with value exp(-(10+dx)), creating a small jump).

Convolution Theorem in Action #

Truncating your infinite signal to [-10,10] is equivalent to multiplying it by a rectangular window function (1 inside the interval, 0 outside). By the Fourier transform convolution theorem:

The FT of a product of two functions is the convolution of their individual FTs.

So:
FT(truncated_sig) = FT(infinite_sig) * FT(rectangular_window) (where * denotes convolution)

The rectangular window's FT is a sinc function: FT(window) = 2 * sin(10k)/k (since the window spans from -10 to 10). For low frequencies (where your plot focuses, -15 < k <15), this sinc function can be approximated by a cosine-like modulation—which is why multiplying your numeric FT by that cosine term brings it close to the analytic result.

How to Confirm This from numpy's Documentation #

You can find all this context directly in numpy's official docs:

• numpy.fft.fft: The docstring explicitly states the DFT definition:

  X[k] = sum_{n=0}^{N-1} x[n] * exp(-2πi k n / N)
  

  This formula assumes x[n] is a periodic sequence with period N, which is the root of the periodicity assumption we discussed.
• numpy.fft.fftfreq: This function's docs explain how it calculates frequency bins, which is critical to correctly mapping your numeric FT results to the analytic k axis. Make sure you understand how d=dx scales the frequency output to match physical units.
• General FFT Notes: The numpy.fft module docs also emphasize that FFT is just an efficient implementation of DFT, so all DFT's properties (like periodicity) apply directly.

Quick Fix for Your Test Code #

If you want the numeric FT to match the analytic result more closely, you can use a smoother window (like a Hann window) instead of a sharp rectangular truncation to reduce boundary discontinuities. For example:

window = np.hanning(len(x))
sig_windowed = sig * window
ft_numeric = fft.fftshift(fft.fft(sig_windowed))

This will reduce the sinc-like modulation, though it will broaden the FT peaks slightly (a standard tradeoff between frequency resolution and spectral leakage).

内容的提问来源于stack exchange，提问作者markmuetz