The Missing Wedge

Why are Cryo-ET reconstructions always stretched along the vertical axis?

Intuition

To reconstruct a 3D structure you tilt the specimen inside the microscope and record a 2D projection at each angle. But the stage cannot reach ±90° — as the tilt $\theta$ grows, the beam path through the ice slab lengthens as $1/\cos\theta$ (already doubled at $60^\circ$ ), so signal is progressively attenuated — not cut off sharply at any one angle — until electrons can no longer usefully penetrate. In practice you stop around ±60°. So there is a whole range of angles you never imaged.

The missing wedge is not a parameter you can tune away; it is a hard constraint set by geometry. The Fourier coefficients inside it never entered a single projection, so everything the reconstruction “says” about them is inferred, not measured. That is the deepest difference between Cryo-ET and ordinary microscopy.

Those un-imaged angles, shown in Fourier space; the slider sets the tilt range:

SampledMissing wedge

Max tilt angle: ±60°

Tilting to ±60° leaves about 33% of Fourier-space angles never sampled. That gap is the "missing wedge" — it stretches and blurs the reconstruction along the vertical axis.

Purple is the sampled region; the two dark wedges are the missing wedge.

Why it’s a wedge in Fourier space

Depth

The key is the central-slice theorem: the 2D Fourier transform of a projection equals a 2D slice through the origin of the object’s 3D Fourier transform, oriented perpendicular to the projection direction.

\mathcal{F}_{2D}\big[ P_\theta(x,y) \big] \;=\; \widehat{f}\big(\text{plane} \perp \theta\big)

Read it term by term: on the left, $P_\theta(x,y)$ is the 2D projection recorded at tilt $\theta$ , and $\mathcal{F}_{2D}$ takes its 2D Fourier transform; on the right, $\widehat{f}$ is the Fourier transform of the 3D object $f$ , and $\text{plane}\perp\theta$ is the plane through the origin whose normal points along the projection direction $\theta$ . The equality says: recording one projection is the same as copying out one slice through the origin of the 3D spectrum — and about every frequency off that slice, the image tells you nothing.

Each tilt angle $\theta$ contributes one slice through the origin. Sweeping $\theta$ from $-\theta_{\max}$ to $+\theta_{\max}$ sweeps out a double cone, leaving two wedge-shaped voids around the beam axis. The half-angle of the missing wedge is

\alpha = 90^\circ - \theta_{\max}.

Here $\theta_{\max}$ is the largest reachable tilt and $\alpha$ is the opening angle of one wedge measured from the equatorial plane of frequency space. At $\theta_{\max}=60^\circ$ , $\alpha=30^\circ$ — roughly one third of all angles are never sampled. Even pushing the stage to $\theta_{\max}=70^\circ$ only shrinks $\alpha$ to $20^\circ$ ; the wedge is still there. Removing it entirely would need $\theta_{\max}=90^\circ$ , which a flat slab of ice does not allow.

Intuition

Why a wedge and not some other shape? Because what’s missing isn’t a scatter of isolated frequency points but a whole range of directions — every frequency component pointing nearly vertical (along the beam axis) goes unsampled. The direction of a frequency corresponds to the orientation of structure in real space: vertical frequencies encode how things vary up-and-down across horizontal interfaces, like a membrane lying flat. Lose that band of directions and every “top-to-bottom” boundary goes soft. So the missing wedge doesn’t lower resolution uniformly — it lowers it only along certain orientations, which is exactly why its artifacts are so recognizable.

Consequences

Missing vertical information in Fourier space is equivalent to a point-spread function elongated along Z in real space. One way to see it: the reconstruction is the true object convolved with a blur kernel, and that kernel is the inverse Fourier transform of the missing wedge — smeared into a tail along Z. The visible effects:

membranes and filaments blur and distort along the vertical axis; horizontal membranes suffer most, because the frequencies that encode them fall almost entirely inside the wedge;
spherical particles look squashed into ellipsoids, often with “ray”-like vertical streaks;
resolution along Z is markedly worse than in the XY plane — the anisotropy can reach roughly 1.4× depending on $\theta_{\max}$ .

A common mitigation is dual-axis tomography: collect one tilt series, rotate the specimen $90^\circ$ in plane, then collect a second. The two series have orthogonal missing wedges, and combining them shrinks the void to a missing pyramid — artifacts are reduced but not removed, at the cost of doubling the dose.

None of this is noise; it is deterministic loss of information: the frequencies inside the wedge were never measured, and no linear filter can manufacture them. This is exactly what methods like CryoGEN try to “fill in” using a learned prior and self-supervised learning — treating the missing wedge as an inverse problem to be completed, inferring the never-recorded frequencies from a statistical prior over real biological structure. In the pipeline, the missing wedge corrupts the observed volume $y$ produced at the reconstruction stage, and the geometric ceiling of tilt-series acquisition is its root cause.

← Electron Tomography