Chapter map

1. 1. Cell-coordinate measures
2. 2. Gaussian convolution
3. 3. Grid sampling
4. 4. Mixture components
5. 5. Source and target images
6. 6. Affine matrix
7. 7. Homogeneous coordinates
8. 8. Velocity field
9. 9. Flow integration
10. 10. Diffeomorphic map
11. 11. Matching functional
12. 12. Regularization energy
13. 13. Landmark term
14. 14. Gradient descent
15. 15. Inverse mapping
16. 16. Governance and learning

Visit the How Mathematics Works hub for the wider series and its connections to education, reasoning, science, computing, design and careers.

---

1. Cell-coordinate measures

This step deserves a careful pause. Cell positions become weighted point masses in continuous space.

A concrete situation

A section has 80,000 observed cell coordinates.

What the mathematics does not license

Segmentation and density bias alter the measure. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Audit counts and sampling density. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

2. Gaussian convolution

The useful question is not only how to calculate, but what the calculation assumes. A spatial kernel spreads each point mass smoothly.

A concrete situation

Each cell contributes a bell-shaped density.

What the mathematics does not license

Kernel width erases or preserves detail. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Sweep raster width. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

3. Grid sampling

A small example makes the mechanism visible. The continuous density is evaluated on pixels.

A concrete situation

A 30-micrometre grid produces a finite image.

What the mathematics does not license

Coarse pixels hide small structures. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Compare several resolutions. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

4. Mixture components

This is where modelling choices begin to affect evidence. Signal, background and artifacts can receive different image-model terms.

A concrete situation

Sparse isolated cells are downweighted as artifacts.

What the mathematics does not license

Real rare regions may look like artifacts. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Inspect component assignments. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

5. Source and target images

The central mathematical idea is concrete. Rasterized sections are compared in one objective.

A concrete situation

Two density images have partly overlapping tissue.

What the mathematics does not license

Intensity scaling changes matching. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Normalize transparently. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

6. Affine matrix

Here the geometry becomes operational. Linear and translation terms handle global pose.

A concrete situation

Rotation, scale and shear provide a coarse match.

What the mathematics does not license

Affine flexibility can distort anatomy. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Report matrix and determinant. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

7. Homogeneous coordinates

This step deserves a careful pause. Translation joins the matrix calculation by adding one coordinate.

A concrete situation

A 2D point becomes a three-entry vector.

What the mathematics does not license

Units remain essential. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Calculate one transformed point. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

8. Velocity field

The useful question is not only how to calculate, but what the calculation assumes. A vector is assigned to every location and time.

A concrete situation

Local tissue regions flow in different directions.

What the mathematics does not license

Sparse evidence underdetermines the field. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Plot vector magnitude. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

9. Flow integration

A small example makes the mechanism visible. An ordinary differential equation carries points through time.

A concrete situation

Three numerical timesteps approximate the final map.

What the mathematics does not license

Few timesteps create integration error. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Check step-size convergence. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

10. Diffeomorphic map

This is where modelling choices begin to affect evidence. The integrated smooth flow produces an invertible warp.

A concrete situation

A curved boundary bends without folding.

What the mathematics does not license

Poor optimization can still approach singular behavior. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Inspect Jacobian determinants. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

11. Matching functional

The central mathematical idea is concrete. Warped source and target images receive a dissimilarity score.

A concrete situation

Density overlap improves after transformation.

What the mathematics does not license

Density agreement may align wrong structures. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Use molecular holdouts. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

12. Regularization energy

Here the geometry becomes operational. The squared velocity norm penalizes complex deformation.

A concrete situation

A rough warp costs more than a smooth one.

What the mathematics does not license

Regularization strength controls bias. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Publish a trade-off curve. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

13. Landmark term

This step deserves a careful pause. Optional paired points add supervised geometric evidence.

A concrete situation

Three anatomical landmarks guide initialization.

What the mathematics does not license

Wrong landmarks dominate the fit. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Leave landmarks out in turn. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

14. Gradient descent

The useful question is not only how to calculate, but what the calculation assumes. Affine and velocity parameters update iteratively.

A concrete situation

Five thousand steps reduce the objective.

What the mathematics does not license

Local minima depend on starts and rates. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Repeat initializations. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

15. Inverse mapping

A small example makes the mechanism visible. The smooth transform maps original cells and can be inverted.

A concrete situation

Target coordinates return near their source positions.

What the mathematics does not license

Numerical inversion has residual error. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Measure round-trip distance. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

16. Governance and learning

This is where modelling choices begin to affect evidence. Version raster, mixtures, affine start, kernels, penalties and solver.

A concrete situation

A student rasterizes five points and integrates a constant field.

What the mathematics does not license

A smooth warp is conditional evidence. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Write a registration ledger. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

A seven-day practice lab

1. Day 1: define the decision, observation unit, X variables and Y or reference condition.
2. Day 2: reproduce the smallest worked example and check every dimension, unit and sign.
3. Day 3: build a transparent baseline and write what it assumes.
4. Day 4: fit the main method inside a training-only pipeline; record every parameter.
5. Day 5: challenge one assumption with a deliberately difficult dataset or operating regime.
6. Day 6: compare validation metrics with the practical workload created by errors.
7. Day 7: present one supported conclusion, one unresolved uncertainty and one safe next experiment.

Questions for a student, parent or mentor

• Which part of the model is a mathematical definition, and which part is a choice?
• What small calculation would catch a sign, scale or dimension error?
• What evidence comes from training, validation and the untouched test?
• What benign condition could look like a fault or prediction failure?
• How would the answer change for another response, asset, school project or time period?
• What should a human review before acting on the score?

A method-specific audit before deployment

STalign turns large sets of cell coordinates into smooth images before registration. Each cell contributes mass to a spatial measure, Gaussian convolution spreads that mass, and sampling on a regular grid produces a rasterized density. This replaces an expensive all-pairs cell comparison with image calculations while retaining tissue shape at a chosen spatial scale.

Registration combines an affine map with a diffeomorphism generated by integrating a time-varying velocity field. The objective trades image similarity against a regularization energy for the flow, with optional landmark terms. Smoothness makes the transformation continuous and invertible, so mapped points can be carried forward and backward without tearing the mathematical space.

Raster width, artifact and background mixture components, affine initialization, smoothness scale, regularization, timesteps and optimizer settings all influence the answer. A visually pleasing warp is not proof of anatomical correspondence. Check inverse consistency, Jacobian determinants, held-out landmarks, gene correspondence and sensitivity to partial tissue coverage.

A useful small example begins with five coordinates, places one Gaussian kernel on each, and compares two candidate affine maps. Students can then integrate a constant velocity field, verify the forward and inverse coordinates, and see how a stronger smoothness penalty changes the preferred solution. The arithmetic shows why image match, deformation energy and landmark error must be reported together.

---

Frequently asked questions

What is STalign?

STalign is a spatial-transcriptomics registration method that rasterizes cell positions and uses large-deformation diffeomorphic metric mapping to estimate smooth invertible transformations.

Why rasterize cells?

Direct all-pairs cell matching is expensive; Gaussian smoothing and grid sampling create images that can be compared efficiently.

What is a diffeomorphism?

It is a smooth, one-to-one transformation with a smooth inverse, preserving topology while allowing large local deformation.

What does the velocity field do?

A time-varying velocity field is integrated to generate the non-linear flow from source to target.

Why include regularization?

It penalizes rough or energetically extreme flows so improved image matching does not justify arbitrary distortion.

Can a diffeomorphic alignment prove two cells correspond?

No. It maps coordinates under density and smoothness assumptions; molecular, anatomical and landmark evidence must test the result.

---

Primary source and fact-check note

Andrew E. Clifton, Yuxuan Liu, Alexander J. Leon et al., “STalign: Alignment of spatial transcriptomics data using diffeomorphic metric mapping,” Nature Communications 14, 8123 (2023), DOI 10.1038/s41467-023-43915-7. Definitions and core claims in this article were checked against this primary publication. Calculations shown here were independently recomputed. Examples are educational illustrations; they are not reported industrial performance results.

Useful next reading

• SANTO coarse-to-fine alignment
• ST-GEARS elastic recovery
• GPSA probabilistic warps
• How Mathematics Works

Keep the options open. Mathematics education supports clearer questions, transferable problem-solving skills and informed choices across science, engineering, computing, AI, statistics, design and many other careers. It does not guarantee intelligence, income, admission or employment; those outcomes depend on far more than one subject or model.

Extended laboratory: topic-specific validation and responsible use

Raster audit

Sweep grid resolution and Gaussian width.

Mixture audit

Inspect signal, background and artifact components.

Affine audit

Report matrix, units and determinant.

Flow audit

Plot velocity magnitude and smoothness.

Topology audit

Check Jacobian determinants and foldings.

Inverse audit

Measure round-trip coordinate error.

Landmark audit

Reserve points and molecular features.

Final claim

Smooth registration remains model-based.

Mechanism workshop: calculate, challenge and transfer

Workshop 1: Cell-coordinate measures

Begin with this mechanism: Cell positions become weighted point masses in continuous space. Work through the concrete case: A section has 80,000 observed cell coordinates. Make the limitation visible: Segmentation and density bias alter the measure. Complete the evidence step: Audit counts and sampling density. Write the coordinates, graphs, kernels, images, matrices, losses, transformations, vector fields, attention weights, distances or tree edges with named dimensions and units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve full sections, donors, stages, modalities, genes or landmarks outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 2: Gaussian convolution

Begin with this mechanism: A spatial kernel spreads each point mass smoothly. Work through the concrete case: Each cell contributes a bell-shaped density. Make the limitation visible: Kernel width erases or preserves detail. Complete the evidence step: Sweep raster width. Write the coordinates, graphs, kernels, images, matrices, losses, transformations, vector fields, attention weights, distances or tree edges with named dimensions and units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve full sections, donors, stages, modalities, genes or landmarks outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 3: Grid sampling

Begin with this mechanism: The continuous density is evaluated on pixels. Work through the concrete case: A 30-micrometre grid produces a finite image. Make the limitation visible: Coarse pixels hide small structures. Complete the evidence step: Compare several resolutions. Write the coordinates, graphs, kernels, images, matrices, losses, transformations, vector fields, attention weights, distances or tree edges with named dimensions and units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve full sections, donors, stages, modalities, genes or landmarks outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 4: Mixture components

Begin with this mechanism: Signal, background and artifacts can receive different image-model terms. Work through the concrete case: Sparse isolated cells are downweighted as artifacts. Make the limitation visible: Real rare regions may look like artifacts. Complete the evidence step: Inspect component assignments. Write the coordinates, graphs, kernels, images, matrices, losses, transformations, vector fields, attention weights, distances or tree edges with named dimensions and units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve full sections, donors, stages, modalities, genes or landmarks outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 5: Source and target images

Begin with this mechanism: Rasterized sections are compared in one objective. Work through the concrete case: Two density images have partly overlapping tissue. Make the limitation visible: Intensity scaling changes matching. Complete the evidence step: Normalize transparently. Write the coordinates, graphs, kernels, images, matrices, losses, transformations, vector fields, attention weights, distances or tree edges with named dimensions and units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve full sections, donors, stages, modalities, genes or landmarks outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 6: Affine matrix

Begin with this mechanism: Linear and translation terms handle global pose. Work through the concrete case: Rotation, scale and shear provide a coarse match. Make the limitation visible: Affine flexibility can distort anatomy. Complete the evidence step: Report matrix and determinant. Write the coordinates, graphs, kernels, images, matrices, losses, transformations, vector fields, attention weights, distances or tree edges with named dimensions and units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve full sections, donors, stages, modalities, genes or landmarks outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 7: Homogeneous coordinates

Begin with this mechanism: Translation joins the matrix calculation by adding one coordinate. Work through the concrete case: A 2D point becomes a three-entry vector. Make the limitation visible: Units remain essential. Complete the evidence step: Calculate one transformed point. Write the coordinates, graphs, kernels, images, matrices, losses, transformations, vector fields, attention weights, distances or tree edges with named dimensions and units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve full sections, donors, stages, modalities, genes or landmarks outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 8: Velocity field

Begin with this mechanism: A vector is assigned to every location and time. Work through the concrete case: Local tissue regions flow in different directions. Make the limitation visible: Sparse evidence underdetermines the field. Complete the evidence step: Plot vector magnitude. Write the coordinates, graphs, kernels, images, matrices, losses, transformations, vector fields, attention weights, distances or tree edges with named dimensions and units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve full sections, donors, stages, modalities, genes or landmarks outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 9: Flow integration

Begin with this mechanism: An ordinary differential equation carries points through time. Work through the concrete case: Three numerical timesteps approximate the final map. Make the limitation visible: Few timesteps create integration error. Complete the evidence step: Check step-size convergence. Write the coordinates, graphs, kernels, images, matrices, losses, transformations, vector fields, attention weights, distances or tree edges with named dimensions and units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve full sections, donors, stages, modalities, genes or landmarks outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 10: Diffeomorphic map

Begin with this mechanism: The integrated smooth flow produces an invertible warp. Work through the concrete case: A curved boundary bends without folding. Make the limitation visible: Poor optimization can still approach singular behavior. Complete the evidence step: Inspect Jacobian determinants. Write the coordinates, graphs, kernels, images, matrices, losses, transformations, vector fields, attention weights, distances or tree edges with named dimensions and units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve full sections, donors, stages, modalities, genes or landmarks outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 11: Matching functional

Begin with this mechanism: Warped source and target images receive a dissimilarity score. Work through the concrete case: Density overlap improves after transformation. Make the limitation visible: Density agreement may align wrong structures. Complete the evidence step: Use molecular holdouts. Write the coordinates, graphs, kernels, images, matrices, losses, transformations, vector fields, attention weights, distances or tree edges with named dimensions and units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve full sections, donors, stages, modalities, genes or landmarks outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 12: Regularization energy

Begin with this mechanism: The squared velocity norm penalizes complex deformation. Work through the concrete case: A rough warp costs more than a smooth one. Make the limitation visible: Regularization strength controls bias. Complete the evidence step: Publish a trade-off curve. Write the coordinates, graphs, kernels, images, matrices, losses, transformations, vector fields, attention weights, distances or tree edges with named dimensions and units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve full sections, donors, stages, modalities, genes or landmarks outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 13: Landmark term

Begin with this mechanism: Optional paired points add supervised geometric evidence. Work through the concrete case: Three anatomical landmarks guide initialization. Make the limitation visible: Wrong landmarks dominate the fit. Complete the evidence step: Leave landmarks out in turn. Write the coordinates, graphs, kernels, images, matrices, losses, transformations, vector fields, attention weights, distances or tree edges with named dimensions and units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve full sections, donors, stages, modalities, genes or landmarks outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 14: Gradient descent

Begin with this mechanism: Affine and velocity parameters update iteratively. Work through the concrete case: Five thousand steps reduce the objective. Make the limitation visible: Local minima depend on starts and rates. Complete the evidence step: Repeat initializations. Write the coordinates, graphs, kernels, images, matrices, losses, transformations, vector fields, attention weights, distances or tree edges with named dimensions and units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve full sections, donors, stages, modalities, genes or landmarks outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 15: Inverse mapping

Begin with this mechanism: The smooth transform maps original cells and can be inverted. Work through the concrete case: Target coordinates return near their source positions. Make the limitation visible: Numerical inversion has residual error. Complete the evidence step: Measure round-trip distance. Write the coordinates, graphs, kernels, images, matrices, losses, transformations, vector fields, attention weights, distances or tree edges with named dimensions and units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve full sections, donors, stages, modalities, genes or landmarks outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.

Workshop 16: Governance and learning

Begin with this mechanism: Version raster, mixtures, affine start, kernels, penalties and solver. Work through the concrete case: A student rasterizes five points and integrates a constant field. Make the limitation visible: A smooth warp is conditional evidence. Complete the evidence step: Write a registration ledger. Write the coordinates, graphs, kernels, images, matrices, losses, transformations, vector fields, attention weights, distances or tree edges with named dimensions and units. Calculate one small example by hand, then vary one assumption and explain which output moves and why. Reserve full sections, donors, stages, modalities, genes or landmarks outside tuning when testing transfer. Keep measurement, fitted representation, statistical evidence, biological interpretation and decision on separate notebook lines so a student, parent or reviewer can audit the mathematics.