Theory of Functional Connections

• Theory of Functional Connections (TFC) is a framework that embeds linear constraints exactly into function approximations, ensuring admissible solutions.
• It separates constraint satisfaction from equation solving, transforming differential problems into unconstrained residual minimization tasks.
• TFC finds applications in solving ODEs, PDEs, optimal control, and astrodynamics using methods like polynomial expansions, deep neural networks, and analytic free-function choices.

The Theory of Functional Connections (TFC) is a functional interpolation framework centered on the constrained expression: a functional representation of the family of all functions that satisfy a prescribed set of linear constraints. In TFC, boundary conditions, initial conditions, derivative conditions, integral conditions, and related linear functional constraints are embedded analytically into the trial solution itself, so the remaining numerical task is no longer a constrained solve over admissible functions but an unconstrained approximation problem over a free function. Across its modern literature, TFC appears both as a mathematical theory of constraint embedding and as a computational methodology for ordinary differential equations, partial differential equations, optimal control, inverse problems, astrodynamics, homotopy continuation, and nonlinear dynamics (Leake, 2021, Schiassi et al., 2020).

1. Defining principle and conceptual structure

At its most basic level, TFC rewrites an unknown function $y$ as a sum of an arbitrary free function $g$ and a correction term chosen so that the required constraints hold identically. In univariate form, one standard representation is

$$y(x,g(x)) = g(x) + \sum_{j=1}^{k}\phi_j(x)\,\rho_j(x,g(x)),$$

where $\phi_j$ are switching functions and $\rho_j$ are projection functionals. The projection functionals quantify the mismatch between the free function and the prescribed constraints, while the switching functions are constructed so that each constraint is activated in the required way and annihilated at the others (Leake, 2021).

This mechanism is what makes TFC qualitatively different from penalty-based formulations. In a penalty method, admissibility is only encouraged numerically. In TFC, admissibility is built into the approximation space itself: every admissible choice of the free function already satisfies the constraints. The numerical solver therefore concentrates on the residual of the governing equation rather than balancing that residual against separate boundary-condition penalties (Schiassi et al., 2020, Leake, 2018).

A central implication is that TFC separates two tasks that are usually entangled: constraint satisfaction and equation satisfaction. The first is handled analytically through the constrained expression; the second is handled numerically through least-squares, nonlinear least-squares, residual minimization, or gradient-based optimization, depending on the approximation chosen for the free function and on the differential equation itself (Johnston et al., 2020, Leake, 2021).

2. Mathematical formulation and generalization

The univariate theory is commonly expressed in terms of linear constraint operators. If the constraints are written as

$\kappa_i = {}_i[y],$

then the projection functional is

$\rho_i(x,g(x)) = \kappa_i - {}_i[g(x)],$

and the switching functions are defined so that

${}_i[\phi_j(x)] = \delta_{ij}.$

With this construction, applying the $i$-th constraint operator to the constrained expression reproduces $\kappa_i$ exactly for any admissible $g$0 (Leake, 2021).

The multivariate theory extends the same idea to PDE-relevant settings. In that case, TFC uses tensorial bookkeeping for boundary data and their intersections, together with coordinatewise switching-function vectors. In the notation adopted in the multivariate literature, the constrained expression has the form

$g$1

where $g$2 stores constraint information and $g$3 are switching-function vectors associated with each coordinate direction. This construction supports value constraints, derivative constraints, and many mixed linear constraints on multidimensional domains (Leake, 2021, Schiassi et al., 2020).

Within that formalism, constrained expressions are described as surjective projection functionals onto the set of admissible functions. They are generally not injective: different free functions can map to the same constrained function. This nonuniqueness is not incidental; it is part of why basis functions overlapping the support-function span must be excluded in numerical implementations to avoid rank deficiency (Leake, 2021, Johnston et al., 2020).

A further theoretical generalization shows that TFC is not restricted to real-valued functions. The codomain of the constrained and free functions need only be a vector space over a field, while the support and switching functions remain scalar-valued over that field. In the univariate case, even the input domain need not be numerical; it may be an arbitrary set. This extends TFC from the usual real-valued setting to vector-valued, complex-valued, function-valued, and other vector-space-valued constructions without changing the core algebraic proof structure (Leake, 2023).

3. Computational realizations and free-function choices

The constrained expression is only half of TFC; the other half is the choice of free function. Early and classical implementations use orthogonal polynomial expansions, especially Chebyshev and Legendre polynomials, so that the constrained differential problem becomes a linear or nonlinear least-squares problem over basis coefficients (Johnston et al., 2020, Leake, 2021).

The main computational variants are summarized below.

Variant	Free function	Distinguishing feature
Classical TFC	Orthogonal polynomials	Exact constraints plus least-squares over basis coefficients
Deep TFC	Deep neural network	PDE residual minimization with analytically embedded BCs
X-TFC	Single-hidden-layer ELM	Random fixed hidden layer; output weights solved analytically
DFL-TFC	FLNN with orthogonal basis block	Domain mapping and exact BC embedding via TFC

Deep TFC replaces polynomial free functions with deep neural networks while retaining exact constraint embedding. In that setting, the neural network is not responsible for learning the constraints; it only spans the unconstrained degrees of freedom left after the constrained expression has been constructed. The resulting method trains on the PDE residual alone and produces a closed-form, differentiable approximation over the domain (Leake, 2018).

X-TFC goes further by choosing the free function to be a single-hidden-layer Extreme Learning Machine (ELM). The hidden weights and biases are randomly assigned and fixed, while the output weights are solved analytically through a Moore–Penrose pseudoinverse in the linearized ELM stage. In exact forward problems this can substantially reduce training cost relative to fully trainable deep architectures, while preserving the TFC property of exact boundary or initial condition satisfaction (Schiassi et al., 2020).

More recent work adapts the same architecture to Functional Link Neural Networks and explicit domain mapping. In DFL-TFC, the physical differential-equation domain is mapped to the orthogonality domain of the basis functions, the hidden layer is replaced by a functional expansion block, and the constrained expression still guarantees exact satisfaction of the beam boundary conditions (Sahu et al., 8 Apr 2026).

4. Differential-equation solving as unconstrained residual minimization

The standard TFC workflow for differential equations is structurally stable across ODEs, PDEs, and many inverse problems. First, one specifies the boundary or initial constraints and derives the constrained expression. Second, one chooses a parameterization for the free function. Third, one substitutes the constrained expression and its derivatives into the governing differential operator, obtaining a residual equation in the free-function parameters. Finally, one enforces that residual at collocation points and solves the resulting unconstrained least-squares or nonlinear least-squares problem (Schiassi et al., 2020, Johnston et al., 2020).

For linear problems, the residual often becomes algebraically linear in the free-function coefficients, so the solve reduces to a least-squares system. High-order boundary-value problems are a particularly clear example. In eighth-order ODEs with endpoint constraints on $g$4, $g$5, $g$6, and $g$7, TFC constructs switching functions that exactly enforce all eight conditions, after which the eighth-order differential equation becomes an unconstrained least-squares problem in the coefficients of the free-function expansion (Johnston et al., 2020).

For nonlinear problems, the residual is generally nonlinear in the free-function parameters. The literature then uses iterative nonlinear least-squares, Newton-style linearization, or, in neural variants, gradient-based optimization. The important point is that these optimizers are not tasked with recovering feasibility of the boundary conditions; feasibility is already guaranteed by the constrained expression (Leake, 2018, Schiassi et al., 2020).

TFC also supports piecewise or segmented constructions. In segmented transfer problems, each segment has its own constrained expression and free function, while continuity in state and first derivative is embedded analytically at segment interfaces. This allows long-horizon trajectories to be decomposed into smaller subintervals while keeping continuity exact by construction rather than numerically enforced through separate matching constraints (Junior, 16 Sep 2025).

5. Applications across scientific computing and control

TFC has been applied to a wide range of differential-equation classes. In benchmark ODEs and PDEs, the literature repeatedly uses it to transform boundary-constrained problems into unconstrained residual minimization over polynomial, kernel, ELM, or deep-network free functions. This includes first-order and second-order ODEs, systems of ODEs, Poisson and wave equations, biharmonic equations, radiative transfer, heat transfer, and high-order boundary-value problems (Leake et al., 2018, Schiassi et al., 2020, Johnston et al., 2020).

Inverse and data-driven problems use the same pattern. In epidemiological compartment models, X-TFC embeds the initial conditions analytically and then estimates both the neural output weights and the unknown physical parameters by minimizing a stacked residual containing data mismatch and ODE residual terms. The TFC role there remains exact initial-condition enforcement; the inverse problem is carried by residual minimization over the augmented parameter vector (Schiassi et al., 2020).

Optimal control is another major application area. In powered descent guidance, TFC parameterizes state and costate trajectories so that endpoint and inter-segment continuity constraints are satisfied analytically, converting the indirect optimal-control TPBVP into an unconstrained least-squares problem over basis coefficients and switching-time-related variables (Johnston et al., 2020). In infinite-horizon optimal control, X-TFC is used to represent the value function in the stationary Hamilton–Jacobi–Bellman equation while exactly embedding the equilibrium condition $g$8; the optimal policy is then recovered analytically from the learned value gradient (Srinivasa et al., 31 Oct 2025).

Astrodynamics has become one of the most active TFC domains. Orbit-transfer work shows several distinct TFC strategies. One line uses change of variables so that mission constraints that are nonlinear and coupled in Cartesian coordinates become linear and decoupled in polar coordinates, making them compatible with standard TFC embeddings (Jr et al., 2023). Another line uses TFC directly for two-point boundary-value orbit transfers, where endpoint conditions are embedded analytically and the resulting nonlinear residual is solved by least-squares (Junior et al., 2021). More recent work introduces tangential-velocity constraints as TFC-embeddable terminal conditions for cislunar transfers and lunar swing-bys, and segmented vector formulations embed continuity and overdetermined endpoint geometry across multiple trajectory arcs (Jr. et al., 2024, Junior, 16 Sep 2025).

Beyond differential-equation solving in the narrow sense, TFC has also been used to parameterize homotopy paths between auxiliary and target nonlinear systems, producing families of admissible homotopies that satisfy the endpoint conditions exactly while leaving path-shaping freedom in a free function (Wang et al., 2019). It has likewise been used to generate continuous constrained trajectory representations for Poincaré-map construction, periodic-orbit computation, and first-recurrence maps without relying on conventional integration-plus-interpolation event detection (Jr et al., 31 Oct 2025).

6. Limitations, misconceptions, and open directions

A recurring misconception is that TFC solves a differential equation merely by embedding the constraints. It does not. TFC removes the need to enforce the constraints numerically, but the governing equation still has to be solved approximately through collocation, least-squares, residual minimization, or training. The difficult part of approximation has been changed, not eliminated (Schiassi et al., 2020, Leake, 2018).

A second misconception is that TFC is already a general theory of nonlinear constraint embedding. The literature is much more specific. Standard TFC is developed for linear constraints. When physically meaningful constraints are nonlinear or coupled, successful applications typically rely on reformulation: changing coordinates so the constraints become linear, using vector parameterizations that preserve linearity at the functional level, or embedding only a linearized or structurally equivalent constraint set (Jr et al., 2023, Jr. et al., 2024, Junior, 16 Sep 2025).

The present theory also has structural limits. Some multivariate integral-constraint configurations remain unresolved, especially when the order of recursive embedding becomes contradictory. General irregular domains are harder than rectangular or simply mapped ones. High-dimensional PDEs remain challenging for polynomial TFC, and even though neural variants mitigate some scalability issues, they do not by themselves provide convergence guarantees or remove sensitivity to architecture, initialization, or collocation design (Leake, 2021, Leake, 2018, Schiassi et al., 2020).

The literature also presents a nuanced computational picture. Classical polynomial TFC can attain extremely high accuracy on many benchmark ODEs and PDEs, sometimes near machine precision, but more complex multidimensional residuals may favor Deep TFC over polynomial TFC or X-TFC. In that sense, TFC is best understood as a constraint-embedding layer or admissible-function architecture rather than as a single fixed numerical solver (Leake, 2021).

Current open directions include broader nonlinear-constraint handling without special coordinate changes, non-rectangular multidimensional domains, more scalable distributed or segmented formulations, harder PDE systems such as Navier–Stokes-type problems, and extensions of TFC-based control formulations to settings where asymptotic or boundary-at-infinity conditions must be represented more directly (Srinivasa et al., 31 Oct 2025, Schiassi et al., 2020, Leake, 2021). A plausible implication is that the long-term significance of TFC lies less in any one basis or optimizer than in its reparameterization principle: feasibility is shifted from numerical enforcement into analytical structure.