The central-slice (Fourier-slice) theorem

The Fourier transform of a projection equals a central slice through the object's spectrum, linking tomographic tilt angles to coverage of frequency space

The central-slice theorem, also called the Fourier-slice theorem, relates a projection of an object to its Fourier transform. It states that the one-dimensional Fourier transform of a parallel projection of a two-dimensional object equals a line through the origin of the object’s two-dimensional transform, oriented perpendicular to the projection direction. The three-dimensional version replaces the line with a plane: the two-dimensional transform of a projection image is a central slice through the three-dimensional spectrum of the object. The theorem equates a real-space summation with sampling along a line or plane in frequency space, which gives tomographic acquisition and reconstruction a single geometric description.

Why does this deserve its own theorem? Because it turns a seemingly hopeless inverse problem into a solvable one. Recovering a three-dimensional interior from a stack of projections has no obvious handle in real space; but once you accept that each projection corresponds to one definite slice of the spectrum, reconstruction becomes a jigsaw: place each angle’s slice into the same 3D spectrum, and once it is filled, a single inverse transform recovers the object. The difficulty simply moves — no longer “how do we invert this,” but “do we have enough slices, and are they placed accurately.”

Sampled slicesMissing wedge (half-angle 30°)

Maximum tilt: ±60°Projections: 15

Each projection samples the 2D spectrum along one line through the origin (the central-slice theorem). For a tilt range of ±60°, two wedges of half-angle 30° about the vertical axis stay unmeasured; the wedges close as the maximum tilt approaches 90°.

The slice contributed by every tilt angle passes through the origin of the spectrum, so low frequencies are sampled repeatedly by all projections while high-frequency detail is recorded only along each slice’s own orientation. This has two immediate consequences: low frequencies (overall shape, large-scale contrast) are always redundant and robust, almost independent of how many angles you collect; high frequencies (edges, small particles, atomic-scale features) are far more fragile — only the part that happens to fall along some slice’s direction is measured, and rotating the same feature by 90° can give you entirely different information.

Formally, let $p_\theta(s)$ be the projection of $f(\mathbf{x})$ along direction $\theta$ . Then

\widehat{p_\theta}(k) = F\big(k\cos\theta,\; k\sin\theta\big),

where $f(\mathbf{x})$ is the object (density) we want, $p_\theta(s)$ is the one-dimensional projection obtained by summing the object along a line at tilt $\theta$ , and $s$ is the coordinate along that projection. On the left, $\widehat{p_\theta}(k)$ is the one-dimensional Fourier transform of that projection, with $k$ the spatial frequency along the projection direction. On the right, $F$ is the object’s two-dimensional Fourier transform, and $\big(k\cos\theta,\;k\sin\theta\big)$ is a point on a line through the origin oriented at angle $\theta$ . In words: transforming the 1D projection gives exactly the values of the object’s 2D spectrum along one central line, with no information lost in either direction. Each projection therefore samples the object’s spectrum along one central line (in 2D) or plane (in 3D). Collecting projections over many angles fills frequency space; an exact inverse transform then recovers the object. This is the principle behind tomographic reconstruction from a tilt series.

Intuition

A projection sums the object along the beam, which discards all variation in that direction. In frequency terms, summing along an axis keeps only the slice of the spectrum that lies orthogonal to it. Each tilt angle therefore contributes one new slice; the more angles, the more of frequency space is filled.

One way to remember it: summing keeps only the zero-frequency bin. Summing along a direction zeroes out every nonzero frequency along that direction, leaving only the frequency components perpendicular to it — and those survivors are precisely the slice through the origin orthogonal to the summation direction.

Deeper

Why the theorem holds takes one line. Project along $y$ : $p(x)=\int f(x,y)\,dy$ . Take its one-dimensional Fourier transform,

\widehat{p}(k_x)=\int\!\!\int f(x,y)\,e^{-2\pi i k_x x}\,dy\,dx = \int\!\!\int f(x,y)\,e^{-2\pi i (k_x x + 0\cdot y)}\,dx\,dy = F(k_x, 0).

The key step equates the inner sum over $y$ with setting $k_y=0$ in the 2D transform: because $e^{-2\pi i\cdot 0\cdot y}=1$ , integrating over $y$ is exactly the 2D transform evaluated on $k_y=0$ . So projecting along $y$ corresponds to the line $k_y=0$ in the spectrum — the central slice perpendicular to the projection direction. Rotating the coordinates by $\theta$ turns that line into $\big(k\cos\theta,\;k\sin\theta\big)$ , giving the general form above. The rotation property of the Fourier transform (rotating the object rotates its spectrum identically) makes this step legitimate.

Several quantitative facts follow. First, the angular spacing between slices widens with frequency: at radius $k$ , the arc between adjacent tilts $\Delta\theta$ is about $k\,\Delta\theta$ , so gaps between slices are larger at high frequency and sampling is sparser there — which is why packing in more angles still struggles to recover the finest detail. Second, discrete acquisition is usually reconstructed by filtered back-projection: the polar-coordinate Jacobian in the inverse transform contributes a factor $|k|$ , equivalent to multiplying each projection by a ramp filter (gain proportional to $|k|$ ) before back-projecting, which compensates for the over-sampled low frequencies and undoes their repeated counting. Third, phase matters more than amplitude: edge positions are set by the alignment of phases across frequencies, and in slice-sampling errors a phase mismatch is more damaging than an amplitude attenuation.

The theorem has a direct and unavoidable consequence for Cryo-ET. A specimen can only be tilted over a limited range, typically about $\pm 60^\circ$ to $\pm 70^\circ$ , before its effective thickness and the grid geometry make higher tilts unusable. The slices contributed by the available angles leave a wedge-shaped region of frequency space unsampled — the missing wedge. For a concrete feel of the size: with a maximum tilt of $\pm 60^\circ$ , the plane perpendicular to the beam has $30^\circ$ on each side — about $60^\circ$ in total — that no slice ever reaches, so roughly a third of that plane’s frequency content is never measured. Information in that wedge is never measured, producing anisotropic resolution: elongation along the beam axis and characteristic distortion of features (spheres stretched into ellipsoids, edges parallel to the gap washed out). Because the gap is defined in frequency space, it cannot be filled by any operation in real space alone — you can interpolate and reweight, but you cannot conjure frequencies that were never recorded. Addressing it is a central problem that methods such as CryoGEN confront: rather than “filling the hole,” they use a learned prior to infer which complete structure lies beyond the missing frequencies while staying consistent with every measured slice.

← Signal Processing