How Frequency Resolution Really Works Beyond Δf
Most people learn frequency resolution as
“Δf = sampling rate / N”
But this is only part of the story.
True frequency resolution depends on more than just Δf.
The Basic Formula (Δf)
Meaning
- Fs → sampling frequency
- N → number of samples
Intuition
“Spacing between FFT bins”
Why Δf Is NOT the Whole Story
Even if Δf is small, you may still NOT separate two close frequencies.
Example
- Signal contains 50 Hz and 52 Hz
- Δf = 1 Hz
Still might look like one spectral peak!
Why? → Spectral leakage & window effects
Practical Frequency Resolution = Ability to Separate Peaks
Real resolution means
“Can you distinguish two nearby frequencies?”
Key Practical Insight
- Resolution ≠ bin spacing
- Resolution = peak separability
Role of Signal Length (Time Duration)
Practical frequency resolution improves with
Longer observation time
Relationship
Where
- T = total signal duration
Intuition
“Longer listening → better frequency clarity”
Windowing Effect (Critical!)
Windowing controls
Key Idea
- Rectangular window → sharp but noisy
- Hanning/Hamming window → wider but cleaner
Trade-off
- Narrow peak → good resolution
- Low leakage → clean spectrum
Spectral Leakage (Why Peaks Spread)
If signal is NOT perfectly periodic, energy spreads across frequency bins.
Result
- Spectral peaks become wider
- Harder to distinguish frequencies if sevel frequencies exist
When the truncation matches an integer number of periods (coherent sampling),
the DTFT exhibits a sinc-shaped spectrum due to the windowing effect,
while the FFT places energy only at discrete frequency bins corresponding to the main-lobe, with no leakage
When truncation does not align with an integer number of periods (non-coherent sampling),
the DTFT remains sinc-shaped spectrum due to the windowing effect,
while the FFT distributes energy beyond the main-lobe into neighboring frequency bins, resulting in spectral leakage
Zero Padding Myth
Important misconception
- Zero padding does NOT increase resolution
What it actually does
- Adds more points between bins
- Makes plot denser
- Zero padding provides a denser sampling of the DTFT, making the FFT appear closer to the continuous spectrum
Key Insight
“Interpolation, not resolution improvement”
FFT with zero-padding vs FFT without zero-padding, just plot denser (refer to Samples/zero-padding effect.mmj)
What REALLY Improves Resolution?
Increase signal duration
- More data → better separation, It'll describe MALMIJAL example below
Proper window selection
- Balance of leakage vs sharpness
Higher SNR
- Less noise → clearer peaks
Averaging (PSD)
MALMIJAL Workflow
Can't distinguish adjacent frequency
- Fs = 1000Hz, T = 0.1sec, Sine 50Hz and Sine 52Hz
- # of FFT = 100, Δf = Fs/NFFT = 10Hz
- Apply FFT
- Can't distinguish 50Hz from 52Hz by spectral leakage & smearing
Can't distingush 50Hz from 52Hz by smearing (lack of freqeuncy resolution)
Improving Frequency Resolution
- Fs = 1000Hz, T = 1sec, Sine 50Hz and Sine 52Hz
- # of FFT = 1000, Δf = Fs/NFFT = 1Hz
- Apply FFT
- Distinguish 50Hz from 52Hz by improving frequency resolution
Distingush 50Hz from 52Hz by improving frequency resolution (long record length)
Can't distinguish adjacent frequency by windowing
- Fs = 1000Hz, T = 1sec, Sine 50Hz and Sine 52Hz
- # of FFT = 1000, Δf = Fs/NFFT = 1Hz
- Apply the Power Spectrum with a Blackman window, whose main-lobe width is approximately 6Δf
- Even with the above frequency resolution, the two frequencies can't be distinguishable due to windowing effects
Can't distingush 50Hz from 52Hz by smearing (windowing effect)
Key Takeaways
- Δf = bin spacing, NOT true resolution
- Practical resolution = ability to separate frequencies
- Controlled by time length, window, leakage
- Zero padding ≠ real resolution improvement
Conclusions
Frequency resolution is often misunderstood as simply Δf = fs / N, but practical resolution goes beyond bin spacing.
- Δf only represents FFT bin spacing, not the actual ability to distinguish close frequencies.
- True frequency resolution depends on the ability to separate nearby spectral peaks, which is influenced by signal duration, windowing, and spectral leakage.
- Longer observation time improves resolution, while window choice and noise level affect how clearly peaks can be identified.
- Zero padding only interpolates the spectrum and does not improve real resolution.
In summary,
practical frequency resolution is determined by peak separability, not just Δf, and requires careful control of signal length, windowing, and noise conditions.
Suggested Further Reading
You may also find these topics helpful:
How Frequency Resolution Really Works Beyond Δf
Most people learn frequency resolution as
“Δf = sampling rate / N”
But this is only part of the story.
True frequency resolution depends on more than just Δf.
The Basic Formula (Δf)
Meaning
Intuition
“Spacing between FFT bins”
Why Δf Is NOT the Whole Story
Even if Δf is small, you may still NOT separate two close frequencies.
Example
Still might look like one spectral peak!
Why? → Spectral leakage & window effects
Practical Frequency Resolution = Ability to Separate Peaks
Real resolution means
“Can you distinguish two nearby frequencies?”
Key Practical Insight
Role of Signal Length (Time Duration)
Practical frequency resolution improves with
Longer observation time
Relationship
Where
Intuition
“Longer listening → better frequency clarity”
Windowing Effect (Critical!)
Windowing controls
Key Idea
Trade-off
Spectral Leakage (Why Peaks Spread)
If signal is NOT perfectly periodic, energy spreads across frequency bins.
Result
When the truncation matches an integer number of periods (coherent sampling),
the DTFT exhibits a sinc-shaped spectrum due to the windowing effect,
while the FFT places energy only at discrete frequency bins corresponding to the main-lobe, with no leakage
When truncation does not align with an integer number of periods (non-coherent sampling),
the DTFT remains sinc-shaped spectrum due to the windowing effect,
while the FFT distributes energy beyond the main-lobe into neighboring frequency bins, resulting in spectral leakage
Zero Padding Myth
Important misconception
What it actually does
Key Insight
“Interpolation, not resolution improvement”
FFT with zero-padding vs FFT without zero-padding, just plot denser (refer to Samples/zero-padding effect.mmj)
What REALLY Improves Resolution?
Increase signal duration
Proper window selection
Higher SNR
Averaging (PSD)
MALMIJAL Workflow
Can't distinguish adjacent frequency
Can't distingush 50Hz from 52Hz by smearing (lack of freqeuncy resolution)
Improving Frequency Resolution
Distingush 50Hz from 52Hz by improving frequency resolution (long record length)
Can't distinguish adjacent frequency by windowing
Can't distingush 50Hz from 52Hz by smearing (windowing effect)
Key Takeaways
Conclusions
Frequency resolution is often misunderstood as simply Δf = fs / N, but practical resolution goes beyond bin spacing.
In summary,
practical frequency resolution is determined by peak separability, not just Δf, and requires careful control of signal length, windowing, and noise conditions.
Suggested Further Reading
You may also find these topics helpful: