Filtering: ramp, low-pass, high-pass

Reweighting the spectrum to suppress noise, sharpen detail, or compensate the geometry of backprojection

Filtering is “picking signal by frequency”. Any signal splits into a sum of sinusoids at different frequencies, and each one is a single number in the spectrum; a filter does just one thing — it assigns a gain to each frequency and multiplies. The gain curve $H(k)$ is the frequency response: $H=1$ at some frequency passes it through untouched, $H=0$ discards it entirely, and values in between scale it proportionally. “Low-pass”, “high-pass”, and “ramp” are nothing more than different shapes of this single curve $H(k)$ .

A linear filter reshapes a signal by multiplying its spectrum by a chosen frequency response $H(k)$ and transforming back. Because this is multiplication in frequency space, it is equivalent to a convolution in real space, and the same operation can be described either way. Filters are classified by which frequencies they pass.

Here $H(k)$ is the gain at frequency $k$ , and larger $k$ means faster variation, finer detail. That “multiplication in frequency equals convolution in real space” is the convolution theorem at work: in real space, filtering is a sliding weighted average of the signal with a kernel $h(x)$ , and $h(x)$ is exactly the inverse Fourier transform of $H(k)$ . Each view earns its keep — to design (“which band do I want to suppress?”) the frequency picture is clearest, while to diagnose artifacts (“how far did a single point get smeared?”) real space is more telling.

OriginalFiltered

Cutoff kc: 5

Low-pass keeps low frequencies and attenuates high ones: slow structure survives while broadband noise is suppressed — the core mechanism of denoising.

The synthetic trace above is a sum of sinusoids at known frequencies plus a broadband noise field. Because every component frequency is known, reconstruction requires no Fourier transform: scaling each component by the selected response $H(k)$ reads off the filtered result directly. A low-pass keeps the slow components and suppresses high-frequency noise, while the ramp grows linearly with frequency — the two responses correspond to the distinct weighting goals of denoising and of tomographic reconstruction.

Put differently, what you move in the demo is not the signal but that gain curve: the low-pass pushes the right end (high frequency) down, so the jittery noise is smoothed away and the slow large waves survive; the ramp does the reverse, lifting the right end so detail is emphasized and low frequencies are relatively cut. Note the two go opposite ways — denoising pushes high frequency down, tomographic reconstruction lifts it up — two opposite uses of the same frequency-domain mixing board.

A low-pass filter keeps low frequencies and attenuates high ones, smoothing the signal and suppressing noise at the cost of fine detail. A high-pass filter does the opposite, retaining rapid variation and edges while removing slow background; combined with a low-pass it forms a band-pass that isolates a chosen range of scales. An abrupt cutoff in frequency creates ringing artifacts in real space (the Gibbs phenomenon), so practical filters use smooth roll-offs such as Gaussian, Butterworth, or cosine-tapered edges.

A concrete number: if voxels are sampled at 1 nm and you want to keep structure down to 2 nm, the cutoff is about $1/(2\,\text{nm}) = 0.5\,\text{nm}^{-1}$ ; the band above that mostly holds noise, and the low-pass attenuates it. Why can’t you just “cut sharply” — drop the gain from 1 to 0 at the cutoff? Because a box (a sharp boxcar window) in frequency space inverse-transforms to a $\mathrm{sinc}$ function in real space, which has infinitely trailing oscillations to either side. Every edge then sprouts rings of light and dark fringes — this is ringing / the Gibbs phenomenon. Replacing the box with a curve that descends smoothly at both ends (Gaussian, Butterworth, cosine taper) tames the tail of its real-space kernel and the ringing disappears, at the cost of a wider transition band and a less “crisp” cutoff.

Intuition

Noise tends to spread across all frequencies, while large-scale structure concentrates at low frequencies. A low-pass filter throws away the band where noise dominates and structure is weak, trading resolution for clarity. The cutoff is chosen near the frequency where the signal-to-noise ratio falls below one.

Think of the spectrum as a curve of “diminishing value” from low to high frequency: at low frequency the signal is strong and noise weak, so every band is worth keeping; toward high frequency the signal fades and the noise stays flat, until past some point a band is almost pure noise and keeping it only roughens the image. The low-pass cuts at that frequency where signal has drowned in noise. The high-pass does the reverse — when a slowly varying background (uneven illumination, an ice-thickness gradient) swamps the detail you care about, removing the lowest few bands makes edges and particles stand out.

The ramp filter is specific to tomography. Filtered backprojection smears each projection back across the volume and sums the results, but the central-slice theorem shows that central slices through frequency space are sampled more densely near the origin, so simple backprojection overweights low frequencies and blurs the result. The ramp filter $H(k) = |k|$ exactly compensates this radial density, restoring a faithful reconstruction.

Here $H(k)=|k|$ means a component at frequency $k$ is multiplied by its distance to the origin. The origin ( $k=0$ , the overall gray level) gets gain zero, and the gain grows linearly outward. Why linear? Because the central slices of all projections pass through the origin and fan out like the spokes of a wheel: near the hub the spokes crowd together and the slices overlap heavily, far out they thin. How many slices “count” a given frequency is proportional to the spoke density there, and that density falls off inversely with the distance $|k|$ from the hub — so naive summation implicitly multiplies each frequency by $1/|k|$ , and multiplying back by $|k|$ cancels it exactly.

Depth

In polar coordinates the inverse 2D Fourier transform carries a Jacobian factor $|k|$ , the area element of frequency space. Backprojection alone omits it, leaving each frequency weighted by $1/|k|$ ; multiplying every central slice by $|k|$ before summing cancels the factor. The unbounded growth of $|k|$ amplifies high-frequency noise, so the pure ramp is windowed by a low-pass (the Shepp–Logan, Hann, or similar apodization) to limit it. This filtering step is the defining ingredient of weighted backprojection used in Cryo-ET reconstruction.

In more detail: the 2D inverse transform is $f(\mathbf{x}) = \iint F(\mathbf{k})\,e^{\,2\pi i\,\mathbf{k}\cdot\mathbf{x}}\,d^2\mathbf{k}$ , and switching to polar coordinates $\mathbf{k}=(k,\theta)$ makes the area element $d^2\mathbf{k} = |k|\,dk\,d\theta$ — that $|k|$ is the Jacobian. The central-slice theorem says the frequency-space slice along angle $\theta$ is precisely the 1D Fourier transform of the projection at that angle. Backprojecting each angle’s projection and integrating directly performs $\int dk\,d\theta$ (missing $|k|$ ), so each frequency is short one factor of $|k|$ ; multiplying every slice by $|k|$ before integrating restores the correct area element. Why must it be windowed? $H(k)=|k|$ grows without bound, and the high-frequency band is exactly where the signal-to-noise ratio is lowest — the pure ramp amplifies noise there the hardest. In practice the ramp is multiplied by a window that decays at high frequency: $H(k)=|k|\cdot W(k)$ , with $W$ a Hann, Shepp–Logan, or cosine apodization, a fusion of “ramp to fix geometry plus low-pass to suppress noise”. In Cryo-ET each projection also carries the oscillations of its contrast transfer function and the gaps left by the missing wedge, so the real weighting is more than $|k|$ alone — it chains the ramp, CTF correction, and SNR weighting (in a Wiener-like form) together. This is the frequency-domain ledger that the CryoGEN family of methods cannot avoid when reconstructing structure under the joint pressure of the missing wedge and noise.

Three filters at a glance

Filter	Shape of $H(k)$	Effect	Where it’s used
Low-pass	1 at low, →0 at high	Smooth, denoise, lose detail	Denoising, visualization, anti-aliasing before downsampling
High-pass	→0 at low, 1 at high	Keep edges, remove slow background	Removing background shading, emphasizing particles
Band-pass	Keep only a middle band	Isolate a range of scales	Scale selection before template matching
Ramp	$\\|k\\|$ , rising linearly (needs a window)	Compensate backprojection’s radial density	Filtered / weighted backprojection

Where it sits in Cryo-ET

The three things on this page each have a job in the reconstruction pipeline: the ramp is the geometric correction at the heart of filtered backprojection, without which the reconstruction blurs; the low-pass recurs in denoising, in visualization, and in anti-aliasing before data is downsampled; the high-pass subtracts the slow background from ice thickness and illumination. They are all the same action — assigning a gain per frequency — differing only in the shape of that $H(k)$ curve. Stacked with CTF correction and missing-wedge handling, this frequency-domain weighting makes up the full verdict that Cryo-ET reconstruction renders on every frequency — how much to trust it, how much to amplify it — and the starting point on which the CryoGEN family learns a prior to fill in the missing information.

← Signal Processing