Retractions, vector transports and exponential maps

A smooth retraction is defined on a neighborhood of the zero section of the tangent bundle. At each x, its local map Rₓ sends zero to x and has derivative at zero equal to the identity on TₓM. Thus the curve t↦Rₓ(tη) starts at x with velocity η. These conditions let objective pullbacks reproduce first derivatives correctly. They do not assert that the curve is a geodesic, that a finite displacement has unchanged length, or that the map is globally defined for arbitrarily large tangent vectors.

The exponential map also satisfies the retraction conditions, but is defined by solving a geodesic initial-value problem. On a complete manifold it is defined for every tangent vector, although it need not be globally injective. A second-order retraction additionally has zero intrinsic acceleration at time zero along every radial tangent curve. In smooth local coordinates it agrees with the exponential through second order, giving an order-three local difference. A generic first-order retraction need only have an order-two difference. The adjective second-order refers to covariant acceleration, not merely to normalizing a point twice.

On the unit sphere with its induced metric, normalizing x+η gives a smooth retraction for every tangent η, since x and η are orthogonal and their sum cannot vanish. Its radial curve has initial ambient acceleration −∥η∥²x, which is normal, hence zero intrinsic acceleration. It is therefore second order. The exact exponential uses sine and cosine instead. Their Taylor expansions agree through degree two, but normalization uses an angular displacement arctan∥η∥ rather than ∥η∥ for a finite step.

A retraction can be first order without being second order. On a circle at (1,0), the curve with angle t+t² has the correct initial tangent velocity but a nonzero tangential acceleration at zero. This difference affects a nonstationary objective’s second-order pullback model. Choosing a retraction should consequently be part of the algorithm specification, especially for Newton or trust-region methods. First-order methods can often use either order, provided their required smoothness and step-size bounds hold on the actual retraction domain.

A vector transport associated with a retraction maps an old tangent vector into the tangent space at the new point, is linear in that vector, and agrees with the identity at a zero step, with suitable smoothness. Projection onto the new tangent space supplies such a local construction for embedded manifolds. It can shrink a vector or even lose invertibility at a large step. Parallel transport along a specified curve instead preserves the metric inner product through a covariant differential equation. The two constructions should not be conflated.

Conjugate-gradient recurrences, quasi-Newton updates and momentum methods compare tangent data across iterates. Their analyses may require bounded transport norms, metric compatibility, an inverse transport, or explicit scaling to avoid undesirable expansion. These properties must be checked for the selected transport and step range. A valid retraction by itself does not imply them. Conversely, insisting on exact geodesics and parallel transport can impose unnecessary cost when a proof only needs a controlled local approximation. The correct choice balances geometric accuracy with the assumptions used by the algorithm.