SYSTEMATIC MATHEMATICS

Rough Paths & Controlled Differential Equations

This 100-chapter Stage 5 doctoral specialization asks one direct question: when an input path is too irregular for ordinary integration, what extra information makes calculus stable? Twelve content-sized units build increments, p-variation, controls, tensor signatures, Chen composition, Young integration, and the sewing lemma before defining geometric rough paths, controlled paths, compensated rough integrals, and RDEs. The course then proves bounded Itô–Lyons continuity, flow and Jacobian equations, bracket and Wong–Zakai interfaces, Brownian and semimartingale lifts, finite-dimensional Gaussian and fractional-Brownian routes, signature numerics, and eight original reconstruction dossiers. Every chapter decodes symbols, proves one typed route, computes an exact finite model, tests a failure mutation, and provides a full bilingual solution without claiming the separate jump, branched, infinite-dimensional, rough-PDE, regularity-structure, singular-SPDE, or application theories.

Before this course: Advanced Probability & Martingale Theory; Stochastic Processes & Stochastic Calculus; Malliavin Calculus & Gaussian Stochastic Analysis; Real Analysis; Measure & Lebesgue Integration; Functional Analysis; Linear Algebra; Differential Equations & Dynamical Systems; Differential Geometry & Manifolds.

COURSE FACTSLevel, chapters, units, prerequisite, and outcome
Chapter 1

An increment remembers orientation, not only two endpoints

Objective: Without looking back, explain why “The additive identity reconstructs every coarse increment exactly from increments on any ordered refinement.” follows in this chapter, identify the decisive exponent or algebraic identity, and name the first failed line under “Reversing only one increment or sorting values instead of times destroys the telescoping identity.”

Unit 1, chapter 1 begins with a concrete pathwise question: “An increment remembers orientation, not only two endpoints.” Rough-path analysis is needed because an irregular driver can make ordinary Riemann–Stieltjes sums unstable: the missing information is not repaired by writing a differential symbol more confidently. We first name the time interval, state space, increment convention, regularity exponent, tensor level, partition, and norm before any calculation.

Read the central object one piece at a time: For a path x:[0,T]→E, write x_{s,t}=x_t−x_s for 0≤s≤t≤T; then x_{s,t}=x_{s,u}+x_{u,t}. A first-level increment records displacement. A second or higher level records ordered interaction accumulated inside the interval. These levels are constrained by algebraic identities such as Chen’s relation and by analytic estimates such as finite p-variation or Hölder control; an arbitrary list of arrays is not a rough path.

The bounded theorem target is: The additive identity reconstructs every coarse increment exactly from increments on any ordered refinement. The conclusion belongs to the stated topology and regularity range. Equality of endpoints, equality of first levels, convergence of piecewise-linear paths, convergence in probability, and convergence in rough-path distance are different statements and are never silently exchanged.

The proof route is to start from “For a path x:[0,T]→E, write x_{s,t}=x_t−x_s for 0≤s≤t≤T; then x_{s,t}=x_{s,u}+x_{u,t}.”; keep every increment and tensor level typed; apply the chapter estimate to the declared partition or control; and stop at “The additive identity reconstructs every coarse increment exactly from increments on any ordered refinement.”. Each arrow carries a quantitative exponent and a domain. The decisive step may be a control inequality, multiplicativity, the sewing lemma, a controlled-remainder estimate, a contraction on a short interval, or continuity of the Itô–Lyons map; naming the theorem without checking its exponent does not justify the arrow.

The exact finite checkpoint is: For times 0<1<3 and values x_0=2, x_1=−1, x_3=4, compute x_{0,1}, x_{1,3}, and x_{0,3}. Keep the ordered time stamps, signed increments, tensor products, partition cells, and rational values visible. The audited endpoint is The increments are −3, 5, and 2, with −3+5=2. A finite computation verifies one algebraic or analytic checkpoint; it does not by itself prove convergence over every refining partition or existence of a stochastic lift.

The first invalid step under changed assumptions is part of the mathematics: Reversing only one increment or sorting values instead of times destroys the telescoping identity. We identify the first estimate or identity that fails, preserve all earlier valid lines, and state the extra variation, geometricity, compatibility, smoothness, integrability, or enhancement hypothesis needed for repair.

Keep four layers separate. The algebra layer handles increments, tensor multiplication, and Chen identities. The analytic layer controls variation, remainders, and sewn limits. The dynamical layer solves an RDE and studies its flow. The probabilistic layer constructs a random lift and proves moment or convergence claims. Passing one layer never certifies the other three automatically.

The chapter closes with reconstruction rather than recognition. Starting from raw path rows, you must rebuild the definition, justify the theorem’s decisive exponent, reproduce the exact endpoint, activate the failure mutation, and state the repaired conclusion. A familiar diagram, one matching decimal, or the phrase “rough path” is not completion.

Bounded theorem for this chapter: The additive identity reconstructs every coarse increment exactly from increments on any ordered refinement. The claim has exactly the algebraic compatibility, regularity, and convergence meaning stated here; no stronger stochastic or frontier conclusion is silently included.

Type every object in “For a path x:[0,T]→E, write x_{s,t}=x_t−x_s for 0≤s≤t≤T; then x_{s,t}=x_{s,u}+x_{u,t}..” Record whether it is a point, a two-index increment, a tensor level, a control, a path with finite variation, a controlled pair, a vector field, or a solution flow. State its interval and norm so that addition, multiplication, integration, and limits are meaningful.

Verify the finite algebra before estimating. Compute each increment with the same orientation, expand every tensor product in order, and check the relevant additive, multiplicative, shuffle, symmetric-part, or inverse identity on one three-time configuration s<u<t.

Build the analytic budget. Choose a control or a Hölder bound, raise increments to the required exponent, and show that the sum or supremum is finite on the declared interval. When two exponents interact, write the inequality—such as 1/p+1/q>1 or 3/p>1—before invoking convergence.

Apply the decisive theorem arrow: start from “For a path x:[0,T]→E, write x_{s,t}=x_t−x_s for 0≤s≤t≤T; then x_{s,t}=x_{s,u}+x_{u,t}.”; keep every increment and tensor level typed; apply the chapter estimate to the declared partition or control; and stop at “The additive identity reconstructs every coarse increment exactly from increments on any ordered refinement.”. If sewing is used, compute the three-index defect; if a fixed point is used, identify the short-interval contraction factor; if a stochastic lift is used, state the mode of convergence and the moment estimate that supports it.

Reconstruct the finite checkpoint independently: For times 0<1<3 and values x_0=2, x_1=−1, x_3=4, compute x_{0,1}, x_{1,3}, and x_{0,3}. Preserve exact values and verify The increments are −3, 5, and 2, with −3+5=2. Then repeat the endpoint by a second route—direct expansion, Chen composition, a telescoping sum, or an ordinary differential equation when the driver is smooth.

Activate the mutation: Reversing only one increment or sorting values instead of times destroys the telescoping identity. Locate the first proof line that loses meaning or its quantitative margin. Keep weaker statements that remain valid, and name one concrete additional assumption that restores the lost line instead of forcing the original conclusion.

These steps prove exactly “The additive identity reconstructs every coarse increment exactly from increments on any ordered refinement..” They do not prove a canonical lift for every random signal, a full support or large-deviation theorem, a singular-SPDE solution, a regularity structure, a production numerical guarantee, or a data-science claim unless those additional objects and estimates are separately constructed.

Worked path model to reconstruct from exact inputs: For times 0<1<3 and values x_0=2, x_1=−1, x_3=4, compute x_{0,1}, x_{1,3}, and x_{0,3}. Preserve ordered increments, tensor components, partition cells, and exponents until the final comparison.

  1. List the complete finite input for “For times 0<1<3 and values x_0=2, x_1=−1, x_3=4, compute x_{0,1}, x_{1,3}, and x_{0,3}.”: ordered times, path values or increments, tensor convention, partition, exponent, vector fields, and every parameter. Reject any row whose time order or dimension is ambiguous.
  2. Compute first-level increments and their variation or Hölder contributions without rounding. If a second level is present, expand ordered products component by component and check its symmetric and antisymmetric parts separately.
  3. Apply the chapter construction: form the control, signature product, sewing defect, compensated Riemann sum, controlled remainder, Euler step, Jacobian update, or stochastic approximation required by the model.
  4. Check the decisive exponent and recompute the endpoint by an independent route. Compare object types and tensor levels as well as numbers; two equal scalars do not validate an equality between differently typed objects.
  5. Both valid routes meet at The increments are −3, 5, and 2, with −3+5=2. State the hypothesis that made them agree, then activate the boundary mutation and record the first discrepancy or missing estimate.

Result: Audited answer: The increments are −3, 5, and 2, with −3+5=2. The calculation certifies this finite checkpoint under the stated convention and assumptions; it does not certify every refinement, random lift, RDE, or application.