differential geometry

Differential geometry studies smooth shapes using calculus, linear algebra, and topology. It begins with curvature and torsion of curves, develops metrics and curvature on surfaces, and extends to smooth manifolds whose geometry is defined intrinsically rather than by an ambient Euclidean space. Tangent spaces linearize local motion, connections compare vectors at nearby points, geodesics generalize straight lines, and curvature measures the failure of Euclidean behavior. The subject links local derivatives to global topology and supports mechanics, relativity, optimization, robotics, graphics, and data analysis. ```svg ``` **A regular curve is a smooth map with nonzero velocity.** A parametrized curve $\gamma:I\to\mathbb R^n$ is regular when $\gamma'(t)\ne0$. Its image is the geometric path, while the parameter determines speed and orientation. Two regular reparametrizations can describe the same oriented curve. A vanishing derivative may be a bad parameter or a genuine singularity. Arc length is $L=\int_a^b\|\gamma'(t)\|dt$ and is invariant under orientation-preserving regular reparametrization. The arc-length parameter $s$ gives unit speed, so the tangent $T=d\gamma/ds$ has norm one. Differentiating $T\cdot T=1$ shows $T'$ is perpendicular to $T$. **Curvature measures tangent rotation per unit arc length.** For unit-speed curves, $\kappa=\|dT/ds\|$. For a general regular plane curve, $\kappa=|x'y''-y'x''|/(x'^2+y'^2)^{3/2}$. Signed curvature additionally records turning orientation. A straight line has zero curvature, while a circle of radius $R$ has constant curvature $1/R$. Where curvature is nonzero, the principal normal points along $dT/ds$ and the binormal is $B=T\times N$ in three dimensions. The Frenet–Serret equations describe how $T,N,B$ rotate. They depend on nonzero curvature; alternative moving frames remain well-defined through inflection points. Torsion measures how a space curve leaves its osculating plane. A planar curve has zero torsion where defined, while a circular helix has constant curvature and torsion. Curvature and torsion determine a regular space curve up to rigid motion under appropriate positivity and regularity conditions. The osculating circle matches position, tangent, and curvature locally. Its radius is $1/\kappa$ and its center lies along the principal normal. At zero curvature the radius is infinite and the construction degenerates. Evolutes trace curvature centers and can develop cusps even when the original curve is smooth. Total curvature accumulates $\int\kappa ds$. For a simple closed plane curve, signed total turning is $2\pi$ up to orientation. Fenchel's theorem gives total curvature at least $2\pi$ for closed space curves, and knotting raises further constraints. These results connect local bending with global closure and topology. Variations of curves perturb an entire path and differentiate length or energy as functionals. Critical points of length under fixed endpoints are straight lines in Euclidean space and geodesics on manifolds. Energy is often easier to differentiate; constant-speed energy-critical curves also minimize or stationarize length locally under suitable conditions. ```svg ``` **A regular surface patch has two independent tangent directions.** A map $X(u,v)$ parametrizes a surface where $X_u$ and $X_v$ are linearly independent. Their span is the tangent plane, and their cross product determines a normal direction and area scale. Coordinate singularities can occur even when the surface itself is smooth. Implicit surfaces $F(x,y,z)=c$ are regular where $\nabla F\ne0$. The gradient is normal to the level surface because directional derivatives vanish along tangent curves. The implicit function theorem supplies local graph coordinates. Critical level points may create cones, crossings, or topology changes. **The first fundamental form is the metric induced on a surface.** In coordinates it has coefficients $E=X_u\cdot X_u$, $F=X_u\cdot X_v$, and $G=X_v\cdot X_v$. It computes lengths, angles, areas, and intrinsic distances without referring to a particular drawing. Positive definiteness follows from patch regularity. The surface area element is $dA=\|X_u\times X_v\|dudv=\sqrt{EG-F^2}dudv$. Changing coordinates introduces a Jacobian that exactly compensates so geometric area is invariant. Orientation affects the signed normal but not ordinary area magnitude. The Gauss map assigns the unit normal $N(p)$ to each oriented surface point. Its derivative is tangent-valued and, up to sign convention, defines the shape operator. Eigenvectors of the shape operator are principal directions and eigenvalues are principal curvatures. Umbilic points have equal principal curvatures and no distinguished principal direction. **Gaussian curvature is the product of principal curvatures.** Positive curvature is locally sphere-like, negative curvature saddle-like, and zero curvature developable in at least one principal direction. Mean curvature is their average under a common convention and controls first variation of area. Sign conventions for the shape operator can reverse mean curvature but not Gaussian curvature. Normal curvature in a tangent direction is the second fundamental form divided by the first. Euler's formula interpolates between principal curvatures by direction. Meusnier's theorem relates curvature of a surface curve to its normal component. A curve can bend strongly in ambient space while having zero geodesic curvature on the surface. The second fundamental form measures extrinsic bending relative to a chosen normal. Plane and cylinder have different second fundamental forms though both have zero Gaussian curvature. This distinction foreshadows Gauss's theorem that Gaussian curvature itself can be computed intrinsically from the metric. **Gauss's theorema egregium makes Gaussian curvature intrinsic.** Although defined initially through embedding and normal change, Gaussian curvature depends only on the first fundamental form and its derivatives. Isometries preserve it. A flat sheet cannot be smoothly stretched onto a sphere without metric distortion, while it can roll into a cylinder. Developable surfaces such as planes, cylinders, cones away from the apex, and tangent developables have zero Gaussian curvature. They can locally unfold into the plane without stretching. Creases and singularities fall outside smooth curvature theory and may concentrate bending in generalized formulations. Minimal surfaces have zero mean curvature and are critical points of area under compactly supported variations. Soap films motivate them, but physical films also face boundaries, pressure, gravity, and instability. Planes, catenoids, and helicoids are classical examples. Zero mean curvature does not mean zero Gaussian curvature. Constant-mean-curvature surfaces model interfaces with constant pressure jump under surface tension. Spheres are compact embedded examples. Mean-curvature flow evolves a surface in its normal direction proportional to mean curvature and tends to smooth while possibly developing singularities. ```svg ``` **Gauss–Bonnet links integrated curvature with Euler characteristic.** For a closed oriented surface, $\int_M KdA=2\pi\chi(M)$. Boundary versions add geodesic curvature and corner-angle terms. The theorem explains why total curvature cannot be adjusted independently of topology and is a model for local-to-global geometry. The Euler characteristic of a triangulated closed surface is vertices minus edges plus faces and is independent of triangulation. A sphere has $\chi=2$, a torus $\chi=0$, and an oriented genus-$g$ surface has $\chi=2-2g$. Gauss–Bonnet converts these discrete counts into a metric integral. Geodesic curvature measures the tangential part of a surface curve's acceleration. It vanishes for a geodesic under unit-speed parametrization, even though ambient curvature may remain. On a sphere, great circles are geodesics while smaller latitude circles are not. Boundary orientation determines the sign in Gauss–Bonnet. Parallel transport moves tangent vectors along a curve while keeping covariant derivative zero. On a curved surface, transporting around a loop can rotate a vector; the holonomy encodes integrated curvature for small loops. This provides an operational view of curvature without embedding. Intrinsic distance is the infimum of lengths of surface curves joining two points. A length-minimizing curve is geodesic in its interior, but a geodesic need only be locally minimizing. Multiple geodesics can join points, and beyond cut loci an initially minimizing geodesic can cease to minimize. ```svg ``` **A smooth manifold is locally Euclidean but can be globally different.** Each point has a neighborhood mapped homeomorphically to an open subset of $\mathbb R^n$, and overlapping charts have smooth transition maps. Hausdorff and countability assumptions prevent pathological global behavior. No single chart need cover the manifold. An atlas is a compatible family of charts; a smooth structure is a maximal compatibility class. Coordinates are not geometric observables by themselves. A formula is intrinsic when its transformed versions agree on overlaps. Coordinate singularities such as longitude at a pole do not imply geometric singularity. Smooth maps between manifolds are defined by smooth coordinate representations. A diffeomorphism is a smooth bijection with smooth inverse and identifies smooth structures. Homeomorphic manifolds can carry inequivalent smooth structures in some dimensions, showing that smooth geometry contains information beyond topology. **The tangent space is the intrinsic linearization of a manifold at one point.** It can be defined through velocities of curves, derivations on smooth functions, or coordinate vectors modulo transformation. Each chart supplies a basis $\partial/\partial x^i$, but the vector is independent of that basis. The cotangent space is the dual space of covectors. The derivative or pushforward $dF_p:T_pM\to T_{F(p)}N$ maps tangent vectors through a smooth map. Pullback sends covectors and differential forms in the opposite direction. These directions are essential: vectors push forward naturally, while covariant objects pull back naturally. Immersions have injective derivative and model locally embedded submanifolds, while submersions have surjective derivative and regular level sets as fibers. An embedding is an immersion that is also a homeomorphism onto its image. An immersed curve can cross itself; an embedded one cannot at distinct parameter points. The regular value theorem says $F^{-1}(q)$ is a submanifold when $dF$ is surjective along the level set. Its tangent space is the kernel of $dF$. This generalizes the gradient criterion for implicit surfaces and supports constrained optimization and configuration manifolds. Vector fields assign tangent vectors smoothly. Their integral curves solve ODEs on the manifold, producing a local flow. Complete vector fields generate flows for all time. The Lie bracket $[X,Y]$ measures failure of flows to commute and detects whether distributions of tangent subspaces are integrable. The Frobenius theorem characterizes when a smooth distribution is tangent to a foliation by submanifolds: closure under Lie brackets is the central local condition. Constraints defined by allowable velocities may be holonomic when integrable or nonholonomic when not. This distinction matters in mechanics and robotics. Lie groups combine smooth manifolds with group operations. Their tangent space at the identity is a Lie algebra whose bracket captures infinitesimal commutators. Matrix groups such as rotations provide concrete examples. The exponential map relates algebra elements to one-parameter subgroups but need not be globally one-to-one or onto. Group actions encode symmetry. Orbits are reachable sets under the group, stabilizers fix points, and quotient spaces remove redundant coordinates when regularity holds. Singular actions can produce orbifolds or stratified spaces rather than manifolds. Symmetry reduction can simplify dynamics while changing topology. Partitions of unity blend local constructions into global ones using smooth nonnegative functions subordinate to an open cover. They build global metrics, extend local functions, and define integration. Their existence relies on paracompactness, commonly ensured by standard manifold assumptions. ```svg ``` **Differential forms are alternating covariant tensors designed for integration.** A $k$-form consumes $k$ tangent vectors and changes sign when two are swapped. Functions are zero-forms, covector fields are one-forms, and top-degree forms act as oriented volume densities. Alternation makes determinants and orientation changes natural. The wedge product combines a $k$-form and an $l$-form into a $(k+l)$-form with graded commutativity. Repeated one-form factors vanish. Coordinate expressions use antisymmetric coefficient arrays, but the form itself is invariant. Exterior algebra packages oriented area and volume elements without choosing a metric. **The exterior derivative generalizes gradient, curl, and divergence structure.** It maps $k$-forms to $(k+1)$-forms, obeys a graded product rule, commutes with pullback, and satisfies $d^2=0$. The last identity encodes curl of a gradient and divergence of a curl under Euclidean identifications. Closed forms satisfy $d\omega=0$, while exact forms satisfy $\omega=d\eta$ and are automatically closed. The converse holds locally on star-shaped or contractible regions by the Poincaré lemma but can fail globally around holes. De Rham cohomology measures this obstruction and links differential equations to topology. Integration of a differential form pulls it back to coordinate domains and uses compatible orientation. A change of coordinates is built into the pullback rather than added as an afterthought. Nonorientable manifolds need densities or orientation covers for global integration of ordinary top forms. **Generalized Stokes says integration of a derivative equals integration over the boundary.** For a compact oriented manifold with boundary, $\int_Md\omega=\int_{\partial M}\omega$ under compatible orientation. The fundamental theorem of calculus, Green's theorem, classical Stokes, and divergence theorem are instances. Boundary orientation is part of the theorem. Orientation is a consistent choice of handedness across tangent spaces. An atlas with positive transition determinants defines one. The Möbius strip is nonorientable, while its boundary is orientable. A Riemannian metric supplies volume density but not automatically a global orientation. The interior product inserts a vector field into the first slot of a form. Lie derivative describes change along a flow and satisfies Cartan's formula $\mathcal L_X\omega=d(\iota_X\omega)+\iota_Xd\omega$. This identity connects symmetry, conservation, transport, and exterior calculus. On a Riemannian oriented manifold, the Hodge star maps $k$-forms to complementary-degree forms using metric and orientation. It defines the codifferential and Hodge Laplacian. Hodge theory represents cohomology classes by harmonic forms under compactness and boundary conditions, linking topology to elliptic analysis. Symplectic geometry uses a closed nondegenerate two-form rather than a distance metric. Hamiltonian vector fields satisfy insertion into the symplectic form equal to a differential of the Hamiltonian up to convention. Their flows preserve symplectic structure and phase volume. Darboux's theorem gives standard local coordinates despite global differences. Contact geometry is the odd-dimensional counterpart describing maximally nonintegrable hyperplane distributions. It appears in optics, thermodynamics, and constrained dynamics. A contact form is not unique; multiplication by a positive function preserves the distribution while changing its Reeb dynamics. ```svg ``` **A Riemannian metric assigns an inner product to every tangent space smoothly.** It defines vector lengths, angles, curve length, volume, gradient, and distance. In coordinates the metric is a positive-definite matrix field $g_{ij}$. Coordinate components change between charts while scalar geometric measurements remain invariant. The gradient of $f$ is defined intrinsically by $g(\operatorname{grad}f,X)=df(X)$ for every vector field $X$. Thus the metric converts the covector $df$ into a vector. In coordinates this uses the inverse metric $g^{ij}$, not merely a list of partial derivatives. The divergence measures infinitesimal volume expansion of a vector field relative to the Riemannian volume. The Laplace–Beltrami operator is divergence of gradient and generalizes the Euclidean Laplacian. Sign conventions differ. On compact manifolds its spectrum encodes both geometry and topology. **A connection defines directional differentiation of vector fields.** Ordinary derivatives of coordinate components are not tensorial because bases change from point to point. A covariant derivative $\nabla_XY$ corrects for this change. Connections may have torsion or fail to preserve a metric; these are additional properties rather than automatic facts. The Levi–Civita connection is uniquely torsion-free and metric-compatible. Christoffel symbols represent it in coordinates and are computed from first derivatives of the metric. They do not transform as tensor components and can vanish at one chosen point in normal coordinates even when curvature is nonzero there. Covariant differentiation extends to covectors and tensors by product and contraction rules. Along a curve it defines acceleration and parallel transport. A tensor equation remains meaningful across coordinate changes; a bare partial-derivative component equation usually does not without connection terms. **Geodesics have zero covariant acceleration.** In coordinates they satisfy $\ddot x^k+\Gamma^k_{ij}\dot x^i\dot x^j=0$. Initial position and velocity determine a local geodesic by ODE theory. Affine reparametrizations preserve this equation; arbitrary reparametrizations preserve the path but add a tangential acceleration term. The exponential map sends a tangent vector $v$ at $p$ to the point reached at unit time by the geodesic with initial velocity $v$. Near zero it gives normal coordinates. It can fail to be injective at the cut locus or singular at conjugate points. Completeness determines whether it is defined on the entire tangent space. Hopf–Rinow connects metric completeness, geodesic completeness, compactness of closed bounded sets, and existence of minimizing geodesics in connected finite-dimensional Riemannian manifolds. The equivalences are special to this setting; general metric spaces and indefinite metrics behave differently. Jacobi fields describe first-order separation of nearby geodesics and satisfy a linear second-order equation involving curvature. Zeros correspond to conjugate points and loss of local minimizing behavior. Positive curvature tends to focus geodesics, while negative curvature tends to separate them, with precise comparison theorems requiring bounds. The Riemann curvature tensor is $R(X,Y)Z=\nabla_X\nabla_YZ-\nabla_Y\nabla_XZ-\nabla_{[X,Y]}Z$ under a common sign convention. It measures noncommutation of covariant derivatives and infinitesimal holonomy. Symmetries reduce its independent components. The opposite overall sign convention is also common. Sectional curvature assigns a scalar to each tangent two-plane and determines the full Riemann tensor. In two dimensions it equals Gaussian curvature. Constant positive, zero, and negative sectional curvature model spherical, Euclidean, and hyperbolic geometries locally after scaling. Ricci curvature traces sectional curvature over directions and governs volume distortion, geodesic focusing, and the Einstein field equation. Scalar curvature traces Ricci again. These contractions discard information in dimensions above three, so equal Ricci tensors need not imply equal full curvature. The Bianchi identities constrain derivatives and cyclic sums of curvature. Their contracted form yields divergence-free Einstein tensor and supports conservation structure in general relativity. Component verification is sensitive to index positions and sign convention; invariant definitions should anchor calculations. Curvature comparison theorems turn upper or lower curvature bounds into estimates on distances, volumes, triangles, and topology. Rauch compares Jacobi fields, Bishop–Gromov compares volume growth under Ricci bounds, and Toponogov compares triangles under sectional bounds. Each uses a specific curvature notion and completeness hypothesis. Positive curvature can force compactness or topology restrictions under global assumptions. Myers's theorem uses a positive Ricci lower bound to bound diameter and fundamental group. Cartan–Hadamard says a complete simply connected manifold of nonpositive sectional curvature has globally well-behaved exponential map and unique geodesics between points. Isometries preserve the metric and therefore distances, Levi–Civita connection, geodesics, and curvature. Infinitesimal isometries are Killing vector fields satisfying a metric Lie-derivative equation. Along geodesics they generate conserved quantities through Noether-type reasoning. Conformal changes preserve angles but rescale lengths by a positive function. They alter curvature through derivatives of the scale. In two dimensions every metric is locally conformally flat, but global conformal structure remains rich. Conformal maps are not generally isometries. Pseudo-Riemannian metrics are nondegenerate but not positive definite. Lorentzian geometry uses signature with one time direction, dividing vectors into timelike, null, and spacelike classes. Distance and completeness intuition from Riemannian geometry requires revision; null curves can have zero proper length without being stationary. General relativity models spacetime with a Lorentzian metric, free-fall trajectories as geodesics, and gravity as curvature sourced through Einstein's equation. Coordinate effects can mimic forces, while curvature invariants reveal geometric effects. Singular coordinates and genuine curvature singularities must be distinguished. Fiber bundles organize spaces that locally look like a base times a fiber but can twist globally. The tangent bundle collects tangent spaces, frame bundles collect bases, and principal bundles encode gauge symmetry. Connections can be described as horizontal subspaces or connection forms, and curvature measures their nonintegrability. Characteristic classes extract global topological invariants from bundle curvature. Chern, Pontryagin, and Euler classes connect differential forms with obstruction theory. Chern–Weil theory shows appropriate curvature polynomials represent cohomology classes independent of the chosen connection. Gauge theory treats connections on principal bundles as fields and curvature as field strength. Electromagnetism is an abelian example, while Yang–Mills theory uses noncommutative structure groups. Local gauge potentials can differ while describing the same connection, making patching data essential. ```svg ``` **Computational differential geometry must approximate objects, not just coordinate formulas.** A mesh represents a surface, basis functions represent fields, and discrete operators should mimic identities such as boundary-of-boundary equals zero. Geometry approximation, field approximation, quadrature, and algebraic solver errors contribute separately. Triangulated surfaces approximate smooth geometry through vertices, edges, and faces. Face normals are discontinuous; vertex normals are weighted estimates rather than intrinsic data. Angle defects approximate integrated Gaussian curvature at vertices, and their global sum satisfies a discrete Gauss–Bonnet relation. Discrete mean curvature can be derived from area variation or Laplace operators. Cotangent formulas work well on suitable meshes but can produce negative weights or instability on poor triangles. Mesh quality, boundary treatment, orientation, and nonmanifold connectivity must be checked before interpreting curvature. **The discrete exterior calculus preserves topological incidence exactly.** Cochains assign values to mesh cells, coboundary matrices discretize exterior derivative, and consecutive coboundaries multiply to zero. A discrete Hodge star introduces metric information. This separation mirrors smooth topology versus metric and supports conservative field solvers. Finite element exterior calculus designs compatible spaces for differential forms so gradient, curl, and divergence relationships survive discretization. Stable mixed methods depend on exact sequences and inf-sup conditions. Using arbitrary nodal elements for every field can create spurious modes. Geodesic distance on meshes can be approximated by graph shortest paths, fast marching, heat methods, or variational solvers. Edge-path distance overestimates paths restricted to the graph and converges slowly under anisotropic meshes. The heat method converts short-time diffusion into a normalized gradient field and Poisson solve. Surface parameterization maps patches to planar coordinates for texture, meshing, or integration. Isometric parameterization preserves lengths when developable; conformal parameterization preserves angles; authalic maps target area. Most curved surfaces cannot preserve all properties simultaneously, so distortion metrics and seams are design choices. Curvature estimation from noisy point clouds is ill-conditioned because it uses second-order information. Neighborhood scale trades noise suppression against feature blurring. Polynomial fitting, normal cycles, integral invariants, and regularization make different assumptions. Report scale and uncertainty with curvature values. Manifold learning infers low-dimensional structure from high-dimensional samples. Local PCA estimates tangent spaces, graph Laplacians approximate differential operators, and diffusion maps use heat-like connectivity. Sampling density, noise, boundary, and metric choice bias the recovered geometry. A visually smooth embedding does not prove the data lie on one manifold. Riemannian optimization performs descent while respecting constraints such as spheres, rotation groups, Stiefel manifolds, and positive-definite matrices. A Riemannian gradient projects the differential through the metric; retractions approximate exponential-map steps; vector transports compare tangent directions. The chosen metric changes gradients and conditioning while leaving feasible points fixed. **Optimization on manifolds separates constraints from coordinates.** Instead of enforcing nonlinear constraints with penalties, iterates remain on the feasible manifold. Critical points have vanishing tangent gradient. Hessians include connection or curvature effects. Quotient manifolds remove nonunique parameterizations but require horizontal-space constructions. Shape optimization treats the domain or surface as the variable. Shape derivatives measure response under deformations, and adjoint equations make gradients affordable. Tangential deformations may be pure reparametrization while normal components change geometry. Mesh motion and topology change complicate numerical implementation. Computer graphics uses surface normals, curvature, geodesics, parameterization, and Laplace–Beltrami operators for shading, smoothing, remeshing, deformation, and texture. Naive smoothing shrinks geometry; curvature flow and constrained variants manage shape change deliberately. Sharp features require nonsmooth or piecewise-smooth models. Robotics uses configuration manifolds for rotations, rigid motions, joint constraints, and contact. Euler angles have coordinate singularities; rotation matrices are redundant but globally smooth under constraints; unit quaternions double-cover rotations. Interpolation should follow the selected geometry rather than componentwise Euclidean averages. Mechanics on manifolds formulates velocities in tangent bundles and momenta in cotangent bundles. Constraints restrict admissible directions, kinetic energy defines a metric, and geodesic motion models force-free systems. Lagrange–d'Alembert handles nonholonomic constraints. Coordinate choices can simplify equations but cannot remove curvature globally. Relativity uses Lorentzian differential geometry to make spacetime physics coordinate independent. Proper time, causal cones, geodesic deviation, curvature tensors, and volume forms are geometric objects. Numerical relativity discretizes constrained hyperbolic equations, where gauge choice and constraint preservation strongly affect stability. Continuum mechanics uses deformation maps between manifolds or Euclidean bodies. The deformation gradient pulls and pushes metric, area, volume, stress, and flux. Strain compares reference and current metrics. Objective constitutive laws remain invariant under rigid observer changes. Shell and membrane theory reduce three-dimensional elasticity to curved midsurfaces. The first fundamental form measures stretching and the second bending. Thin structures penalize stretch far more strongly than bend, producing buckling and geometric nonlinearity. Discrete shell models must represent both forms consistently. Optics interprets rays as geodesics of an optical metric in isotropic media, while anisotropic media require richer structures such as Finsler geometry. Fermat's principle is variational. Caustics occur where neighboring rays focus and the exponential map becomes singular. Information geometry equips statistical model families with metrics such as Fisher information. Geodesics and curvature describe local distinguishability and parameter coupling. Coordinate-invariant theory does not remove singular models, boundaries, or nonidentifiability where the metric degenerates. Optimal transport gives probability measures a geometric structure where distance is minimal transport cost. Wasserstein geodesics move mass rather than interpolate densities pointwise. Gradient flows in this space describe diffusion and related PDEs. The space is generally not a finite-dimensional smooth manifold. Geometric deep learning builds models respecting graph, group, manifold, or gauge structure. Equivariance constrains how features transform; invariant outputs ignore symmetry-related coordinates. Discretization and sampling can break exact symmetry, and learned latent geometry does not automatically correspond to physical curvature. Crystallography and materials science use manifolds and quotient spaces for orientation distributions, grain rotations, and order parameters. Defects can be classified by topology, while elastic distortion uses geometric incompatibility. Orientation averaging should respect rotation geometry rather than arithmetic component averages. Semiconductor processing has curved wafer, feature, and interface geometry, but that applied geometry is distinct from this mathematical discipline. Differential geometry contributes normals, curvature-driven evolution, surface PDEs, coordinate-free transport, and mesh operators. The pre-existing process-geometry article remains the appropriate route for manufacturing geometry rather than the canonical `differential geometry` query. The major geometric objects can be compared directly. | Object | Local data | Invariant role | Common coordinate trap | |---|---|---|---| | Tangent vector | curve velocity or derivation | direction of motion | treating components as invariant numbers | | Metric | positive-definite bilinear form | length, angle, volume | forgetting inverse metric when raising indices | | Connection | covariant derivative | compare vectors, define geodesics | treating Christoffel symbols as a tensor | | Curvature tensor | commutator of covariant derivatives | intrinsic non-Euclidean behavior | mixing sign and index conventions | | Differential form | alternating covariant tensor | oriented integration | omitting pullback or orientation | | Shape operator | derivative of surface normal | extrinsic principal curvature | confusing mean and Gaussian curvature | | Cohomology class | closed form modulo exact forms | global obstruction | inferring global exactness from local closure | ```flowchart st=>start: Identify the smooth space, dimension, topology, and regularity op1=>operation: Choose charts while naming coordinate-invariant objects cond1=>condition: Is the question intrinsic or embedding-dependent? op2=>operation: Use metric, connection, geodesics, and intrinsic curvature op3=>operation: Use normal bundle, shape operator, and second fundamental form cond2=>condition: Do transformations, orientations, and limiting cases agree? op4=>operation: Change chart, refine mesh, or test an invariant formulation e=>end: Report geometry with convention, domain, and error or regularity limits st->op1->cond1 cond1(intrinsic)->op2->cond2 cond1(extrinsic)->op3->cond2 cond2(yes)->e cond2(no)->op4->op1 ``` **A reliable differential-geometry workflow distinguishes invariant objects from representations.** State the manifold or surface, regularity, metric, orientation, and embedding if relevant. Use a chart to calculate but verify transformation behavior. Declare curvature and shape-operator sign conventions. Check singular points, boundaries, and global topology before extending a local result. Coordinate checks use overlap regions. Compute an object in two charts and transform components according to tensor type. Scalars should agree directly, vectors by Jacobian pushforward, covectors by pullback, and volume forms with orientation. Christoffel symbols acquire inhomogeneous terms because they represent a connection rather than a tensor. Dimensional analysis remains useful. Curvature has inverse-length units, Gaussian curvature inverse-length squared, and integrated Gaussian curvature is dimensionless. Metric components inherit coordinate units. Exponentiating or adding geometric quantities with inconsistent units signals a faulty model. Embedding checks compare intrinsic and extrinsic conclusions. A cylinder has zero intrinsic Gaussian curvature despite nonzero bending; a sphere has positive intrinsic curvature; a saddle has negative. If an alleged intrinsic quantity distinguishes a plane from an unrolled cylinder, the formula likely depends on embedding. Topological checks test whether a claimed global frame, normal, potential, or coordinate system can exist. Hairy-ball phenomena obstruct nowhere-vanishing tangent fields on even spheres, nonorientability obstructs global normals, and nontrivial cohomology obstructs global potentials. Local formulas cannot override these obstructions. Regularity checks matter because curvature uses second derivatives of a metric or embedding. A surface reconstructed only piecewise linearly has curvature as a discrete or weak object, not a classical pointwise tensor. Generalized notions should be named, and mesh-dependent estimates should be studied under refinement. Numerical verification should test known flat, spherical, cylindrical, and hyperbolic cases. Preserve exact incidence identities, monitor symmetry and positive definiteness, compare integrated curvature with topology, and refine geometry as well as solution fields. A converged algebraic solver on a fixed inaccurate mesh is not geometric convergence. Theorems have local and global scopes. Normal coordinates flatten first derivatives at one point but not a neighborhood. Geodesics locally minimize but may not globally. Closed forms are locally exact but may not globally. Curvature bounds yield global results only with completeness, compactness, or topology assumptions explicitly present. Convention mismatches are a frequent source of apparent disagreement. Authors may reverse the Riemann tensor, shape operator, mean curvature, Laplacian, or boundary orientation. State a defining equation and translate downstream formulas consistently rather than comparing one isolated sign. The subject's central local-to-global pattern appears repeatedly: derivatives define curvature locally, curvature integrates to topological information, infinitesimal symmetries generate conservation, and local charts assemble through transition maps. Global conclusions require compactness, completeness, orientation, or topology beyond coordinate calculation. MIT's differential-geometry curriculum begins with curves and surfaces, centers concrete geometry on curvature, and develops first and second fundamental forms, Christoffel symbols, intrinsic versus extrinsic geometry, Gauss's theorem, Gauss–Bonnet, geodesics, and hyperbolic space. Advanced MIT geometry adds smooth manifolds, differential forms, Lie groups, connections, Riemannian curvature, and Hodge theory. **Geometric intuition should be tested by invariant calculation.** A picture depends on projection, coordinates, and embedding. Length, angle, curvature, holonomy, topology, and spectral data provide checks that survive representation. When a result changes under harmless reparametrization, it is describing the coordinates rather than the geometry. **Local flatness does not mean zero curvature.** Normal coordinates can make the metric Euclidean and Christoffel symbols vanish at one point, just as a tangent plane matches a surface to first order. Curvature lives in second-order variation and cannot generally be transformed away over a neighborhood. **Every geometric computation needs a stated convention and domain.** Identify orientation, metric signature, tensor index order, connection, and curvature sign. Exclude coordinate and geometric singularities explicitly. Without those declarations, correct formulas from different sources can appear contradictory or be combined inconsistently. **Geodesic completeness and metric completeness must be checked globally.** A metric can look smooth in every displayed chart while an omitted boundary lies at finite distance. Conversely, coordinates can diverge while the manifold continues smoothly in another chart. Test whether Cauchy sequences converge in the space and whether geodesics extend for every affine time. **Curvature is an operator before it is a scalar.** Gaussian, sectional, Ricci, and scalar curvature are successive specializations or contractions suited to different questions. In dimensions above two, one scalar cannot reconstruct directional bending. Select the curvature object whose hypotheses and conclusion match distance, volume, topology, relativity, or embedding behavior. **Global geometry requires topology and analysis in addition to local calculus.** Compactness enables extrema and spectral discreteness, completeness controls geodesic extension, fundamental groups classify loops and coverings, and cohomology detects closed forms without global potentials. Ignoring these structures turns local coordinate identities into false global claims. The injectivity radius measures how far exponential coordinates remain uniquely minimizing and nonsingular. It is limited by conjugate points and multiple geodesics. Small injectivity radius can arise from high curvature or thin topology. Numerical algorithms using logarithm maps or geodesic neighborhoods should remain below a justified radius. The cut locus of a point marks endpoints where minimizing geodesics cease to be unique or minimizing. Distance from the point is smooth away from the point and its cut locus but nonsmooth on it. Gradient-based methods using squared geodesic distance must account for this domain. Comparison of manifold-valued data requires a mean definition. The Fréchet mean minimizes expected squared geodesic distance and can be nonunique on positively curved or broad distributions. The logarithm-map average is local and chart-center dependent. Euclidean component averaging can leave the manifold or violate symmetry. Parallel transport provides one way to compare tangent vectors from different data points. The result depends on path when curvature is present. Choosing shortest geodesics can still be ambiguous across cut loci. Algorithms that aggregate gradients on manifolds must specify transport and handle nonuniqueness. Shape spaces treat curves or surfaces modulo translation, rotation, scaling, or reparametrization. Quotient geometry removes irrelevant transformations, but singular shapes can have larger symmetry groups and create stratified spaces. Distance and geodesic computation then require alignment as well as deformation. Finsler geometry generalizes Riemannian length by allowing a direction-dependent norm not necessarily derived from an inner product. Travel time in anisotropic media and direction-dependent cost are natural examples. Geodesics remain variational but connections, curvature, and reversibility become richer. Sub-Riemannian geometry permits motion only along a bracket-generating distribution and measures lengths of admissible curves. Lie brackets can recover inaccessible directions over finite maneuvers, as in nonholonomic vehicles. Distances have anisotropic scaling and geodesics may include abnormal extremals. Alexandrov and metric geometry extend curvature bounds to nonsmooth spaces through triangle comparison. Ricci curvature also has synthetic formulations using optimal transport and measure. These theories distinguish geometric conclusions that truly require differentiability from those encoded by distance and volume alone. Ricci flow evolves the metric by its Ricci curvature and redistributes geometry rather than moving a surface through an ambient space. It can develop singularities requiring rescaling and surgery. Mean-curvature flow instead evolves an embedding by mean curvature. Similar names conceal different unknowns and invariances. Geometric evolution equations often contain gauge freedom or weak parabolicity. Choosing coordinates or a DeTurck-type correction exposes a well-posed analytic system without changing underlying geometry. Numerical methods must control both physical geometry and gauge artifacts. Spectral geometry studies how eigenvalues of Laplace-type operators reflect volume, curvature, boundary, and topology. Isospectral nonisometric spaces show that the spectrum does not determine every detail. Discretized spectra are sensitive to mesh, boundary conditions, and mass-matrix choice. Morse theory uses critical points of smooth functions to infer topology. A nondegenerate critical point has an index counting negative Hessian directions, and sublevel topology changes by attaching a cell. Degenerate critical points require perturbation or more general singularity theory. Index theorems connect analytical indices of differential operators with topological characteristic classes. They are far-reaching descendants of Gauss–Bonnet. Boundary conditions and ellipticity are central, and the index is stable under suitable perturbations even when individual kernel dimensions change. Singular geometry appears at cones, corners, self-intersections, defects, and topology changes. Classical manifold definitions exclude these points, but stratified spaces, currents, varifolds, and weak curvature extend selected operations. One should not assign smooth curvature formulas directly at a singularity without a limiting or generalized definition. Currents generalize oriented submanifolds as linear functionals on differential forms and retain a boundary operator compatible with Stokes. Varifolds retain unoriented geometric measure and suit area variation. These tools describe weak limits of surfaces, minimal interfaces, and concentrated defects. Reproducible geometric computation should record mesh generation, coordinate conventions, orientation repair, unit scaling, boundary classification, discrete operator, solver tolerances, and refinement evidence. Small implementation choices can reverse normals, change curvature signs, or alter topology through disconnected or nonmanifold elements. Read differential geometry through a tangent-metric-connection-curvature-and-topology lens rather than a coordinate-formula-and-surface-picture lens.