Lesson overview | Lesson overview | Next part

Optimization on Manifolds: Part 1: Intuition

1. Intuition

Intuition develops the part of optimization on manifolds specified by the approved Chapter 25 table of contents. The treatment is geometry-first and AI-facing.

1.1 Constrained optimization as unconstrained manifold optimization

Constrained optimization as unconstrained manifold optimization belongs to the canonical scope of Optimization on Manifolds. The goal is to make curved-space reasoning concrete enough for ML practice without turning the section into a pure topology course.

Working scope for this subsection: Riemannian gradient descent, retractions, vector transport, Riemannian Hessian preview, first-order optimality, matrix manifolds, and ML optimization examples. The recurring pattern is localize, linearize, measure, move, and return to the manifold.

$$\mathbf{x}_{k+1}=R_{\mathbf{x}_k}\left(-\eta_k\operatorname{grad} f(\mathbf{x}_k)\right).$$

Operational definition.

Manifold optimization updates in a tangent space and maps the step back to the manifold with an exponential map or retraction.

Worked reading.

On the sphere, take a tangent gradient step and normalize. Normalization is a simple retraction because it returns to the sphere and agrees with the tangent direction to first order.

Three examples of constrained optimization as unconstrained manifold optimization:

1. PCA on Grassmann manifolds.
2. Orthogonal weights on Stiefel manifolds.
3. Covariance learning on SPD manifolds.

Two non-examples clarify the boundary:

1. Euclidean gradient descent followed by arbitrary clipping.
2. A projection step that destroys the first-order update direction.

Proof or verification habit for constrained optimization as unconstrained manifold optimization:

Check tangent feasibility, descent direction under the metric, and retraction properties.

global object      -> curved manifold or constraint set
local object       -> chart, tangent space, or coordinate patch
linear operation   -> derivative, gradient, velocity, Hessian approximation
geometric measure  -> metric, length, distance, curvature
algorithmic move   -> tangent step followed by geodesic or retraction

In AI systems, constrained optimization as unconstrained manifold optimization matters because learned representations and constrained parameter spaces are rarely globally flat. A local linear approximation may be useful, but it must be attached to the point where it is valid.

This turns constraints such as orthogonality, low rank, and positive definiteness into native geometry instead of penalties.

Mini derivation lens.

1. Choose a point $p$ on the manifold $M$ and name the local representation used near $p$.
2. Move the question into a chart, tangent space, or embedded constraint where first-order calculus is available.
3. Compute the local object: derivative, tangent projection, metric-weighted gradient, path velocity, or retraction step.
4. Translate the result back into coordinate-free language so the answer is not tied to one chart by accident.
5. Check the invariant: the point remains on $M$, the direction remains in $T_pM$, or the distance/gradient uses the stated metric.

Implementation lens.

A practical ML implementation should store both the ambient array representation and the geometric contract attached to it. For example, a normalized embedding is not just a vector; it is a point on a sphere. An orthogonal weight matrix is not just a matrix; it is a point on a Stiefel-type constraint. A covariance matrix is not just a symmetric array; it must stay positive definite.

The clean computational pattern is: encode the state, compute an ambient derivative if needed, convert it into a tangent or metric-aware object, take a small local step, and then return to the manifold with a geodesic formula or retraction. This is the same pattern used in the companion notebooks, just scaled down to visible two- and three-dimensional examples.

The important warning is that coordinate code can pass shape checks while still violating geometry. Differential geometry adds checks that are semantic: tangentness, smooth compatibility, metric choice, path validity, and constraint preservation.

Practical checklist:

• State the manifold and whether it is abstract, embedded, or quotient-like.
• State the local coordinates or tangent representation being used.
• Separate ambient vectors from tangent vectors.
• Name the metric before computing distances, angles, or gradients.
• Use geodesics or retractions when moving on the manifold.
• For ML claims, identify whether geometry is data geometry, parameter geometry, or statistical geometry.

Local diagnostic: Write gradient, tangent projection, retraction, and stopping criterion.

The companion notebook uses low-dimensional synthetic examples: circles, spheres, tangent projections, spherical interpolation, SPD matrices, and orthogonality constraints. These examples keep geometry visible while preserving the same update logic used in higher-dimensional ML systems.

A useful learning move is to compute everything first on a sphere. The sphere has visible curvature, simple tangent spaces, closed-form geodesics, and practical retractions. Once those are clear, Stiefel, Grassmann, SPD, and information-geometric examples become less mysterious.

For implementation, the main discipline is to avoid leaving the manifold silently. If a gradient step violates a constraint, either project the gradient into the tangent space before stepping or use a method whose update is intrinsic by design.

The final question for this subsection is whether a Euclidean formula is being used as an approximation, a coordinate expression, or a mistaken replacement for geometry. Differential geometry is the habit of telling those cases apart.

1.2 Why projection alone is not always geometry-aware

Why projection alone is not always geometry-aware belongs to the canonical scope of Optimization on Manifolds. The goal is to make curved-space reasoning concrete enough for ML practice without turning the section into a pure topology course.

Working scope for this subsection: Riemannian gradient descent, retractions, vector transport, Riemannian Hessian preview, first-order optimality, matrix manifolds, and ML optimization examples. The recurring pattern is localize, linearize, measure, move, and return to the manifold.

$$R_p(\mathbf{0}_p)=p,\qquad dR_p(\mathbf{0}_p)=\operatorname{id}_{T_pM}.$$

Operational definition.

Why projection alone is not always geometry-aware belongs to the canonical scope of Optimization on Manifolds: Riemannian gradient descent, retractions, vector transport, Riemannian Hessian preview, first-order optimality, matrix manifolds, and ML optimization examples.

Worked reading.

Start from a concrete embedded example, compute the local tangent or metric object, then translate back to intrinsic notation.

Three examples of why projection alone is not always geometry-aware:

1. Sphere geometry.
2. Embedding-space local coordinates.
3. Matrix-manifold parameter constraints.

Two non-examples clarify the boundary:

1. A flat Euclidean approximation used globally.
2. A geometric claim made without metric or tangent space.

Proof or verification habit for why projection alone is not always geometry-aware:

The proof habit is to compute locally and verify coordinate-independent meaning.

global object      -> curved manifold or constraint set
local object       -> chart, tangent space, or coordinate patch
linear operation   -> derivative, gradient, velocity, Hessian approximation
geometric measure  -> metric, length, distance, curvature
algorithmic move   -> tangent step followed by geodesic or retraction

In AI systems, why projection alone is not always geometry-aware matters because learned representations and constrained parameter spaces are rarely globally flat. A local linear approximation may be useful, but it must be attached to the point where it is valid.

The AI relevance is that model spaces are often curved even when implemented as arrays.

Mini derivation lens.

1. Choose a point $p$ on the manifold $M$ and name the local representation used near $p$.
2. Move the question into a chart, tangent space, or embedded constraint where first-order calculus is available.
3. Compute the local object: derivative, tangent projection, metric-weighted gradient, path velocity, or retraction step.
4. Translate the result back into coordinate-free language so the answer is not tied to one chart by accident.
5. Check the invariant: the point remains on $M$, the direction remains in $T_pM$, or the distance/gradient uses the stated metric.

Implementation lens.

A practical ML implementation should store both the ambient array representation and the geometric contract attached to it. For example, a normalized embedding is not just a vector; it is a point on a sphere. An orthogonal weight matrix is not just a matrix; it is a point on a Stiefel-type constraint. A covariance matrix is not just a symmetric array; it must stay positive definite.

The clean computational pattern is: encode the state, compute an ambient derivative if needed, convert it into a tangent or metric-aware object, take a small local step, and then return to the manifold with a geodesic formula or retraction. This is the same pattern used in the companion notebooks, just scaled down to visible two- and three-dimensional examples.

The important warning is that coordinate code can pass shape checks while still violating geometry. Differential geometry adds checks that are semantic: tangentness, smooth compatibility, metric choice, path validity, and constraint preservation.

Practical checklist:

• State the manifold and whether it is abstract, embedded, or quotient-like.
• State the local coordinates or tangent representation being used.
• Separate ambient vectors from tangent vectors.
• Name the metric before computing distances, angles, or gradients.
• Use geodesics or retractions when moving on the manifold.
• For ML claims, identify whether geometry is data geometry, parameter geometry, or statistical geometry.

Local diagnostic: Name the manifold, tangent space, metric, and map being used.

The companion notebook uses low-dimensional synthetic examples: circles, spheres, tangent projections, spherical interpolation, SPD matrices, and orthogonality constraints. These examples keep geometry visible while preserving the same update logic used in higher-dimensional ML systems.

A useful learning move is to compute everything first on a sphere. The sphere has visible curvature, simple tangent spaces, closed-form geodesics, and practical retractions. Once those are clear, Stiefel, Grassmann, SPD, and information-geometric examples become less mysterious.

For implementation, the main discipline is to avoid leaving the manifold silently. If a gradient step violates a constraint, either project the gradient into the tangent space before stepping or use a method whose update is intrinsic by design.

The final question for this subsection is whether a Euclidean formula is being used as an approximation, a coordinate expression, or a mistaken replacement for geometry. Differential geometry is the habit of telling those cases apart.

1.3 Tangent-space updates and retractions

Tangent-space updates and retractions belongs to the canonical scope of Optimization on Manifolds. The goal is to make curved-space reasoning concrete enough for ML practice without turning the section into a pure topology course.

Working scope for this subsection: Riemannian gradient descent, retractions, vector transport, Riemannian Hessian preview, first-order optimality, matrix manifolds, and ML optimization examples. The recurring pattern is localize, linearize, measure, move, and return to the manifold.

$$\operatorname{grad} f(p)=\Pi_{T_pM}\left(\nabla \bar{f}(p)\right)\quad\text{for embedded submanifolds with Euclidean metric}.$$

Operational definition.

A tangent space is the vector space of allowable first-order velocities through a point on a manifold.

Worked reading.

For the unit sphere, tangent vectors at $\mathbf{x}$ are exactly vectors $\mathbf{v}$ satisfying $\mathbf{x}^\top\mathbf{v}=0$.

Three examples of tangent-space updates and retractions:

1. Velocity of a curve on a sphere.
2. Jacobian pushing embedding perturbations forward.
3. A vector field assigning one tangent direction per point.

Two non-examples clarify the boundary:

1. An arbitrary ambient vector not tangent to the constraint.
2. A finite difference step that leaves the manifold without retraction.

Proof or verification habit for tangent-space updates and retractions:

For embedded manifolds, differentiate the constraint; for abstract manifolds, use curves or derivations.

global object      -> curved manifold or constraint set
local object       -> chart, tangent space, or coordinate patch
linear operation   -> derivative, gradient, velocity, Hessian approximation
geometric measure  -> metric, length, distance, curvature
algorithmic move   -> tangent step followed by geodesic or retraction

In AI systems, tangent-space updates and retractions matters because learned representations and constrained parameter spaces are rarely globally flat. A local linear approximation may be useful, but it must be attached to the point where it is valid.

Tangent spaces are where local sensitivity, Jacobians, and first-order optimization live.

Mini derivation lens.

1. Choose a point $p$ on the manifold $M$ and name the local representation used near $p$.
2. Move the question into a chart, tangent space, or embedded constraint where first-order calculus is available.
3. Compute the local object: derivative, tangent projection, metric-weighted gradient, path velocity, or retraction step.
4. Translate the result back into coordinate-free language so the answer is not tied to one chart by accident.
5. Check the invariant: the point remains on $M$, the direction remains in $T_pM$, or the distance/gradient uses the stated metric.

Implementation lens.

A practical ML implementation should store both the ambient array representation and the geometric contract attached to it. For example, a normalized embedding is not just a vector; it is a point on a sphere. An orthogonal weight matrix is not just a matrix; it is a point on a Stiefel-type constraint. A covariance matrix is not just a symmetric array; it must stay positive definite.

The clean computational pattern is: encode the state, compute an ambient derivative if needed, convert it into a tangent or metric-aware object, take a small local step, and then return to the manifold with a geodesic formula or retraction. This is the same pattern used in the companion notebooks, just scaled down to visible two- and three-dimensional examples.

The important warning is that coordinate code can pass shape checks while still violating geometry. Differential geometry adds checks that are semantic: tangentness, smooth compatibility, metric choice, path validity, and constraint preservation.

Practical checklist:

• State the manifold and whether it is abstract, embedded, or quotient-like.
• State the local coordinates or tangent representation being used.
• Separate ambient vectors from tangent vectors.
• Name the metric before computing distances, angles, or gradients.
• Use geodesics or retractions when moving on the manifold.
• For ML claims, identify whether geometry is data geometry, parameter geometry, or statistical geometry.

Local diagnostic: Verify the proposed direction satisfies the tangent constraint.

The companion notebook uses low-dimensional synthetic examples: circles, spheres, tangent projections, spherical interpolation, SPD matrices, and orthogonality constraints. These examples keep geometry visible while preserving the same update logic used in higher-dimensional ML systems.

A useful learning move is to compute everything first on a sphere. The sphere has visible curvature, simple tangent spaces, closed-form geodesics, and practical retractions. Once those are clear, Stiefel, Grassmann, SPD, and information-geometric examples become less mysterious.

For implementation, the main discipline is to avoid leaving the manifold silently. If a gradient step violates a constraint, either project the gradient into the tangent space before stepping or use a method whose update is intrinsic by design.

The final question for this subsection is whether a Euclidean formula is being used as an approximation, a coordinate expression, or a mistaken replacement for geometry. Differential geometry is the habit of telling those cases apart.

1.4 Matrix manifolds in ML

Matrix manifolds in ML belongs to the canonical scope of Optimization on Manifolds. The goal is to make curved-space reasoning concrete enough for ML practice without turning the section into a pure topology course.

Working scope for this subsection: Riemannian gradient descent, retractions, vector transport, Riemannian Hessian preview, first-order optimality, matrix manifolds, and ML optimization examples. The recurring pattern is localize, linearize, measure, move, and return to the manifold.

$$\operatorname{grad} f(p)=0\quad\text{is the first-order criticality condition on }M.$$

Operational definition.

Matrix manifolds in ML belongs to the canonical scope of Optimization on Manifolds: Riemannian gradient descent, retractions, vector transport, Riemannian Hessian preview, first-order optimality, matrix manifolds, and ML optimization examples.

Worked reading.

Start from a concrete embedded example, compute the local tangent or metric object, then translate back to intrinsic notation.

Three examples of matrix manifolds in ml:

1. Sphere geometry.
2. Embedding-space local coordinates.
3. Matrix-manifold parameter constraints.

Two non-examples clarify the boundary:

1. A flat Euclidean approximation used globally.
2. A geometric claim made without metric or tangent space.

Proof or verification habit for matrix manifolds in ml:

The proof habit is to compute locally and verify coordinate-independent meaning.

global object      -> curved manifold or constraint set
local object       -> chart, tangent space, or coordinate patch
linear operation   -> derivative, gradient, velocity, Hessian approximation
geometric measure  -> metric, length, distance, curvature
algorithmic move   -> tangent step followed by geodesic or retraction

In AI systems, matrix manifolds in ml matters because learned representations and constrained parameter spaces are rarely globally flat. A local linear approximation may be useful, but it must be attached to the point where it is valid.

The AI relevance is that model spaces are often curved even when implemented as arrays.

Mini derivation lens.

1. Choose a point $p$ on the manifold $M$ and name the local representation used near $p$.
2. Move the question into a chart, tangent space, or embedded constraint where first-order calculus is available.
3. Compute the local object: derivative, tangent projection, metric-weighted gradient, path velocity, or retraction step.
4. Translate the result back into coordinate-free language so the answer is not tied to one chart by accident.
5. Check the invariant: the point remains on $M$, the direction remains in $T_pM$, or the distance/gradient uses the stated metric.

Implementation lens.

A practical ML implementation should store both the ambient array representation and the geometric contract attached to it. For example, a normalized embedding is not just a vector; it is a point on a sphere. An orthogonal weight matrix is not just a matrix; it is a point on a Stiefel-type constraint. A covariance matrix is not just a symmetric array; it must stay positive definite.

The clean computational pattern is: encode the state, compute an ambient derivative if needed, convert it into a tangent or metric-aware object, take a small local step, and then return to the manifold with a geodesic formula or retraction. This is the same pattern used in the companion notebooks, just scaled down to visible two- and three-dimensional examples.

The important warning is that coordinate code can pass shape checks while still violating geometry. Differential geometry adds checks that are semantic: tangentness, smooth compatibility, metric choice, path validity, and constraint preservation.

Practical checklist:

• State the manifold and whether it is abstract, embedded, or quotient-like.
• State the local coordinates or tangent representation being used.
• Separate ambient vectors from tangent vectors.
• Name the metric before computing distances, angles, or gradients.
• Use geodesics or retractions when moving on the manifold.
• For ML claims, identify whether geometry is data geometry, parameter geometry, or statistical geometry.

Local diagnostic: Name the manifold, tangent space, metric, and map being used.

The companion notebook uses low-dimensional synthetic examples: circles, spheres, tangent projections, spherical interpolation, SPD matrices, and orthogonality constraints. These examples keep geometry visible while preserving the same update logic used in higher-dimensional ML systems.

A useful learning move is to compute everything first on a sphere. The sphere has visible curvature, simple tangent spaces, closed-form geodesics, and practical retractions. Once those are clear, Stiefel, Grassmann, SPD, and information-geometric examples become less mysterious.

For implementation, the main discipline is to avoid leaving the manifold silently. If a gradient step violates a constraint, either project the gradient into the tangent space before stepping or use a method whose update is intrinsic by design.

The final question for this subsection is whether a Euclidean formula is being used as an approximation, a coordinate expression, or a mistaken replacement for geometry. Differential geometry is the habit of telling those cases apart.

1.5 Optimization pipeline: gradient step retract transport

Optimization pipeline: gradient step retract transport belongs to the canonical scope of Optimization on Manifolds. The goal is to make curved-space reasoning concrete enough for ML practice without turning the section into a pure topology course.

Working scope for this subsection: Riemannian gradient descent, retractions, vector transport, Riemannian Hessian preview, first-order optimality, matrix manifolds, and ML optimization examples. The recurring pattern is localize, linearize, measure, move, and return to the manifold.

$$\mathbf{x}_{k+1}=R_{\mathbf{x}_k}\left(-\eta_k\operatorname{grad} f(\mathbf{x}_k)\right).$$

Operational definition.

The Riemannian gradient is the tangent vector whose inner product with any direction equals the directional derivative.

Worked reading.

In coordinates with metric matrix $G$, the Riemannian gradient is $G^{-1} abla f$, not usually the raw Euclidean gradient.

Three examples of optimization pipeline: gradient step retract transport:

1. Natural gradient using Fisher information.
2. Projected gradient on the sphere.
3. Geometry-aware update for SPD covariance matrices.

Two non-examples clarify the boundary:

1. Raw parameter gradient treated as invariant under reparameterization.
2. A direction off the tangent space called a manifold gradient.

Proof or verification habit for optimization pipeline: gradient step retract transport:

Use the defining identity $g_p(\operatorname{grad} f,\mathbf{v})=df_p[\mathbf{v}]$ for all tangent directions.

global object      -> curved manifold or constraint set
local object       -> chart, tangent space, or coordinate patch
linear operation   -> derivative, gradient, velocity, Hessian approximation
geometric measure  -> metric, length, distance, curvature
algorithmic move   -> tangent step followed by geodesic or retraction

In AI systems, optimization pipeline: gradient step retract transport matters because learned representations and constrained parameter spaces are rarely globally flat. A local linear approximation may be useful, but it must be attached to the point where it is valid.

Natural gradient and second-order preconditioning are geometry choices, not only optimizer tricks.

Mini derivation lens.

1. Choose a point $p$ on the manifold $M$ and name the local representation used near $p$.
2. Move the question into a chart, tangent space, or embedded constraint where first-order calculus is available.
3. Compute the local object: derivative, tangent projection, metric-weighted gradient, path velocity, or retraction step.
4. Translate the result back into coordinate-free language so the answer is not tied to one chart by accident.
5. Check the invariant: the point remains on $M$, the direction remains in $T_pM$, or the distance/gradient uses the stated metric.

Implementation lens.

A practical ML implementation should store both the ambient array representation and the geometric contract attached to it. For example, a normalized embedding is not just a vector; it is a point on a sphere. An orthogonal weight matrix is not just a matrix; it is a point on a Stiefel-type constraint. A covariance matrix is not just a symmetric array; it must stay positive definite.

The clean computational pattern is: encode the state, compute an ambient derivative if needed, convert it into a tangent or metric-aware object, take a small local step, and then return to the manifold with a geodesic formula or retraction. This is the same pattern used in the companion notebooks, just scaled down to visible two- and three-dimensional examples.

The important warning is that coordinate code can pass shape checks while still violating geometry. Differential geometry adds checks that are semantic: tangentness, smooth compatibility, metric choice, path validity, and constraint preservation.

Practical checklist:

• State the manifold and whether it is abstract, embedded, or quotient-like.
• State the local coordinates or tangent representation being used.
• Separate ambient vectors from tangent vectors.
• Name the metric before computing distances, angles, or gradients.
• Use geodesics or retractions when moving on the manifold.
• For ML claims, identify whether geometry is data geometry, parameter geometry, or statistical geometry.

Local diagnostic: Ask which metric converts covectors into update vectors.

The companion notebook uses low-dimensional synthetic examples: circles, spheres, tangent projections, spherical interpolation, SPD matrices, and orthogonality constraints. These examples keep geometry visible while preserving the same update logic used in higher-dimensional ML systems.

A useful learning move is to compute everything first on a sphere. The sphere has visible curvature, simple tangent spaces, closed-form geodesics, and practical retractions. Once those are clear, Stiefel, Grassmann, SPD, and information-geometric examples become less mysterious.

For implementation, the main discipline is to avoid leaving the manifold silently. If a gradient step violates a constraint, either project the gradient into the tangent space before stepping or use a method whose update is intrinsic by design.

The final question for this subsection is whether a Euclidean formula is being used as an approximation, a coordinate expression, or a mistaken replacement for geometry. Differential geometry is the habit of telling those cases apart.