What the Lens Loses

On the difference between a single-axis read and a reconstruction. Most of what gets called measurement is the first kind, and its seams do not show.

I. The Lab

Focused-Beam Imaging

An electron microscope takes pictures with electrons instead of light. The wavelengths involved are a hundred thousand times shorter than visible light: a few picometers, fine enough to image atoms. And yet sub-angstrom resolution remained out of reach for most of the twentieth century.

The wavelength was not the limit. The lens was.

Focused-beam imaging (the shape that has defined microscopy for a century) has the same structure across light, electrons, and most ordinary microscopes. A probe is focused to a point inside the sample. Intensity is measured at the focal point or along a focal plane. The image is the intensity map. Resolution is bounded by how sharply the probe can be focused.

Better images come from better focusing. That clause is the field's conceptual frame for most of its history.

With visible light, the focal point cannot be made finer than the wavelength itself; the wavelength sets the floor. Electrons remove the wavelength as the constraint. The lens becomes the constraint instead.

Scherzer's Theorem

In 1936 Otto Scherzer proved that the lens limit is structural, not engineering. Any electromagnetic lens that is rotationally symmetric, static, and free of internal charge must have positive spherical aberration.

The proof is short and devastating. Maxwell's equations plus rotational symmetry plus the absence of free charge produce a constrained class of magnetic field configurations, and every member of that class focuses electrons more strongly the further they pass from the optical axis. There is no symmetric configuration that produces an under-focused beam.

To compensate for spherical aberration you would need a divergent component somewhere in the optics, and Scherzer's theorem says no such component exists within the rotationally symmetric, static, charge-free regime. The ceiling is enforced by the geometry of allowed electromagnetic fields, not by lens manufacturing.

For sixty years the response was to engineer better within Scherzer's regime. Stabler power supplies, cleaner vacuums, more carefully wound coils. Resolution crept upward. The structural ceiling held.

Then it broke. Through the 1990s, Harald Rose, Maximilian Haider, and Knut Urban built aberration correctors that worked by breaking rotational symmetry: multipole magnets (hexapoles, octupoles) that introduced compensating divergent terms outside Scherzer's symmetric regime. The theorem was satisfied but routed around. Sub-angstrom resolution arrived. Individual atoms in a crystal lattice became routinely visible.

This is real. Worth seeing clearly: aberration correction is the maturation of focused-beam imaging, not an escape from it. The shape of the measurement is unchanged. A probe is focused (more sharply now, with the symmetric and broken-symmetric components working together) and intensity is measured at the focal point. The ceiling has moved, but the kind of measurement is the same kind it was in 1931.

Phase

There is a ceiling underneath the one Scherzer named, and aberration correction cannot reach it either. It is structural to the shape of focused-beam measurement itself.

A focused-beam measurement detects intensity at a focal point. Intensity is the squared magnitude of the electron wave's amplitude — it throws away phase. The wave that emerges from the sample carries two kinds of information: amplitude (how much wave came through) and phase (where each scattered component is in its cycle, relative to the others). Phase encodes the interference structure of how different parts of the sample scattered the probe. Detecting intensity at a focal point integrates over that structure and loses most of it.

Aberration correction does not address this loss; it cannot. Sharper focus produces a cleaner intensity map. The phase information is gone before the focusing happens.

Ptychography

Ptychography is a different shape of measurement.

The probe is no longer focused to extract one read at one point. It is scanned across the sample in a dense grid of known overlapping positions. Each illumination spot covers territory the previous spot already touched. At every probe position, the detector records not an image and not a focal-point intensity, but the full diffraction pattern: intensity across every angle at which the probe scattered into the detector.

The diffraction patterns are not images of the sample. They are records of how the probe and the sample interfered, position by position.

There is no image yet. There is a four-dimensional dataset (two probe-position axes, two diffraction-angle axes) recording how the sample's response varied as the probe's known position moved across it.

The image is reconstructed.

An iterative algorithm works backward from the diffraction patterns and the known overlap structure of the probe positions to recover, simultaneously, the complex electron wave that emerged from the sample and the probe itself. Phase is recovered because the redundancy from overlapping illumination at known positions provides enough constraint for the algorithm to converge on a consistent solution. The phase information that focused-beam imaging discards is preserved in the diffraction patterns and recoverable from the structure of how those patterns relate across known perturbations of the probe.

In 2021, a group at Cornell led by David Muller used electron ptychography to map atomic positions with precision well below the Bohr radius. The remaining limit was thermal vibration. Atoms in the sample do not stand still.

The General Move

The structural difference is sharper than the resolution number suggests. Focused-beam imaging — including its aberration-corrected form — detects intensity at a focal point and accepts the loss of phase as the cost of the measurement. Ptychography records the response of the sample to a known probe across many structured perturbations and reconstructs the full wave from that response structure.

These are not different qualities of the same measurement. They are different shapes of measurement, with different floors.

Naming the move requires care, because the move is general. The probe-and-focus shape is what physics, biology, and most analytical work call "taking a measurement." The ptychographic shape is something else:

Probe with known structured perturbations. Record the response across the perturbation structure. Reconstruct the underlying object from the response structure.

This shape has been running for decades in other fields under other names. Imaging happened to hit a ceiling sharp enough that the difference became inescapable. The ceiling exists in other domains too. It is usually less obvious because nothing as concrete as a Bohr radius is sitting just past it.

The shape generalizes regardless.

II. Already Running Elsewhere

It's likely you or someone you know has been in an MRI scanner.

Ask how the machine works and you'll usually get a story about powerful magnets being used to scan the tissue in your body.

A science YouTuber would explain further: the magnets align the hydrogen nuclei in your tissue, radio pulses cause them to emit signals at frequencies that depend on their local environment, and the computer assembles those signals into pictures.

Generalize past the specific physics, and you get something like this: the scanner probes the body with known structured perturbations of its magnetic field, records how every proton responds (magnitude and phase), and computes an image from the structure of those responses.

Gradients

Inside the scanner, a strong static magnetic field aligns the hydrogen nuclei in the body's tissue. An RF pulse tips them out of alignment. As they relax back, they emit RF signals at a frequency determined by their local magnetic field strength.

If the field were perfectly uniform, every nucleus in the body would emit at the same frequency. The receiver would record one integrated tone. No spatial information.

The gradient coils break the uniformity. They impose magnetic field gradients along chosen axes: fields whose strength varies linearly across position. Under a gradient, nuclei at different positions experience different total magnetic fields and precess at different frequencies.

Position is encoded into frequency. And the encoding is known, because the scanner controls the gradient precisely.

A second technique, called phase encoding, works differently. The scanner applies brief gradient pulses before the signal acquisition. During the pulse, nuclei at different positions accumulate different phases: where they are in their precession cycle. The pulse ends, but the phase difference doesn't. It persists into the signal that follows.

A typical scan combines both. Frequency encoding during acquisition, phase encoding before, repeated at many gradient strengths. The body's response is sampled across many known perturbations of the magnetic field.

The Signal

The receiver coils record what the nuclei emit as they relax. At any instant, the signal is the integrated response of every proton in the body, under whatever gradient is currently active. Different gradients produce different signals.

The receiver records both magnitude and phase.

Each gradient configuration writes data into a structured set called k-space. K-space is the spatial Fourier transform of the image: same content, expressed in coordinates that match the gradient encoding. The scanner fills in k-space gradient by gradient, until there's enough to reconstruct.

Reconstruction

When enough of k-space has been filled, the scanner computes the image. An inverse Fourier transform converts the gradient-encoded data back to spatial coordinates. The body's geometry emerges from the structured set of responses, not from any single measurement.

Same move, different mathematics. Ptychography reconstructs phase iteratively; MRI inverts a Fourier transform.

The recognition extends.

CT scans do it with X-rays: known paths through the body, absorption recorded along each, image computed from the response.

Synthetic aperture radar does it with pulses from a moving platform: the same shape, in a different probe.

Compressed sensing asks how little is enough: how few structured perturbations does the move need, if the underlying object is sparse?

These are not analogies. They are the same shape of measurement, recognized in different fields under different names.

III. The Calibration

Most of what gets called measurement is a single-axis read.

You can build a revenue dashboard without ever hitting a Bohr radius. You can run a customer survey without phase encoding. You can read a model output without iterative reconstruction. Single-probe mode keeps working because the work doesn't push hard enough to reveal where it stops. The walls that forced ptychography and MRI and synthetic aperture radar into being are walls that most analytical work never approaches, which is why most analytical work has never had to develop the move.

It has the cost anyway. Single-axis reads inherit the axis they were taken along; they report what got measured, not what's there. In imaging, with a Bohr radius staring back at the field, the cost was visible. In a quarterly report, it isn't.

Take a chart of quarterly revenue trending upward. What it actually shows is one number that got chosen to be reported, integrated over a complex distribution of customers, products, pricing decisions, and seasonal effects. The chart is a single-axis read along the axis of what got measured and how. A different choice of axis would have produced a different chart, possibly trending in the opposite direction, drawn from the same underlying business.

The chart is not a measurement of the business. It's a measurement of what got reported.

The same shape runs in any artifact treated as data. A model output is a single-axis read along the axis of one prompt, one temperature, one moment in the model's training distribution; the response is internally coherent because single-axis reads always are, and the coherence is what makes it feel like a measurement of the topic rather than a single position on it. An interview captures answers to specific questions in a specific moment, with a specific person in a specific mood; the transcript reads as testimony about the subject when it's testimony about that crossing. A dashboard aggregates whatever the system was wired to aggregate; the aggregation is treated as a measurement of the system's state when it's a measurement of what the wiring decided to count.

Single-axis reads don't show their seams. In ptychography, the seams are the diffraction pattern — visible structure in the data that proves the reading is a probe position rather than a complete picture. In a chart or a dashboard, no comparable structure is visible. The read is internally smooth, which is what makes the category error invisible. The reader looking at a clean chart cannot see, from the chart alone, what the chart is leaving out.

Stop trusting them as reconstructions.

What changes when you do is not what the artifact says. The chart still says what it says. What changes is the trust you extend to it — calibrated to what the artifact actually is, which is a single-axis read along the axis of how the data got produced, rather than what it's been treated as, which is a measurement of the underlying thing.

What am I treating as a measurement that's actually a single-axis read?

What would reconstructing it look like?