Recovery on Curves Condition (RC)

• RC is a geometric-analytic hypothesis that enables recovering a field’s value via finitely many derivatives along prescribed curves.
• It yields curve-supported Green’s functions and averaged pseudodifferential operators that provide optimal regularization for PDE solution operators.
• RC is characterized by the finite-dimensional cokernel condition (FC), ensuring full rank of principal symbols and enabling recovery identities in various operator regimes.

Searching arXiv for the explicitly named “Recovery on Curves condition (RC)” and closely related curve-based recovery constructions. Recovery on Curves condition (RC) is a structural hypothesis for linear differential operators introduced as the starting point for constructing explicit solution operators for underdetermined systems $Pu=f$ and, by duality, explicit representation formulas for the adjoint overdetermined systems $P^*v=g$. In its formal form, RC asserts that the value of a test field $\varphi(y)$ can be recovered from finitely many derivatives of $P^*\varphi$ evaluated along a prescribed curve $\gamma(y,y_1,\cdot)$, up to a correction term supported at the endpoint $y_1$; by duality, this yields Green’s functions supported on individual curves and, after averaging over families of curves, pseudodifferential solution operators with prescribed support properties (Isett et al., 4 Sep 2025).

1. Definition and geometric meaning

Let $U\subseteq \mathbb{R}^d$ be open, and let $P$ be an $r_0\times s_0$ matrix-valued differential operator on $U$, with $P^*v=g$0. The formal $P^*v=g$1-adjoint is denoted $P^*v=g$2. A family of curves is written

$P^*v=g$3

defined on an open set $P^*v=g$4, with endpoints

$P^*v=g$5

The formal condition introduced in the PDE setting is the endpoint-supported version of RC: $P^*v=g$6 where only finitely many multi-indices $P^*v=g$7 occur and

$P^*v=g$8

Thus, for every $P^*v=g$9, the value of $\varphi(y)$0 at the initial point $\varphi(y)$1 is reconstructed from the jet of $\varphi(y)$2 along the curve from $\varphi(y)$3 to $\varphi(y)$4, plus a correction concentrated at the endpoint (Isett et al., 4 Sep 2025).

The model example is the divergence operator. If

$\varphi(y)$5

then RC reduces to the fundamental theorem of calculus along the curve: $\varphi(y)$6 In this sense, RC generalizes one-dimensional transport-type recovery identities to broad classes of underdetermined and overdetermined PDE systems.

The paper also discusses a weaker version, $\varphi(y)$7, in which $\varphi(y)$8 is assumed to vanish near $\varphi(y)$9, so that the endpoint term disappears. The stronger endpoint-supported version is the central formal condition.

2. Curve-supported Green’s functions and averaged integral operators

The dual form of RC produces a distributional kernel supported on a single prescribed curve. For fixed $P^*\varphi$0, one defines $P^*\varphi$1 by

$P^*\varphi$2

This kernel satisfies

$P^*\varphi$3

with support property

$P^*\varphi$4

Accordingly, RC is equivalent to the existence of Green’s functions supported on prescribed curves, up to an endpoint error term (Isett et al., 4 Sep 2025).

These raw curve-supported kernels are generally too singular for direct Sobolev mapping theory. The paper therefore averages over endpoints $P^*\varphi$5 using a smooth weight $P^*\varphi$6 satisfying

$P^*\varphi$7

and defines

$P^*\varphi$8

Then

$P^*\varphi$9

and

$\gamma(y,y_1,\cdot)$0

The support of the resulting operator is therefore prescribed by the chosen family of curves.

A central analytic theorem shows that after smooth averaging the kernel becomes a pseudodifferential kernel of the correct order. If $\gamma(y,y_1,\cdot)$1 is the order of the scalar operator $\gamma(y,y_1,\cdot)$2, then the averaged operator gains $\gamma(y,y_1,\cdot)$3 derivatives: $\gamma(y,y_1,\cdot)$4 This is the sense in which the resulting solution operators are regularizing of optimal order.

The paper develops two principal regimes. In the conic or nontrapping setting, RC yields an exact right-inverse

$\gamma(y,y_1,\cdot)$5

together with the dual identity

$\gamma(y,y_1,\cdot)$6

In bounded star-shaped settings, one instead obtains a right-inverse up to finite rank: $\gamma(y,y_1,\cdot)$7 and the dual left-inverse formula

$\gamma(y,y_1,\cdot)$8

These identities yield Poincaré-, Friedrich-, and Korn-type inequalities modulo the formal cokernel.

3. Finite-dimensional cokernel condition and algebraic characterization

The paper pairs RC with a symbol-level criterion called the finite-dimensional cokernel condition (FC). Writing the principal symbol of $\gamma(y,y_1,\cdot)$9 as

$y_1$0

FC is

$y_1$1

Equivalently, the principal symbol $y_1$2 is surjective, or full rank, for every nonzero complex covector $y_1$3 (Isett et al., 4 Sep 2025).

This requirement is stronger than ellipticity in the usual underdetermined or overdetermined sense, because ellipticity tests only real covectors $y_1$4, whereas FC requires full rank for all nonzero complex $y_1$5. The paper’s interpretation is that finite-dimensionality of the formal cokernel is obstructed by complex characteristic modes

$y_1$6

not merely by real plane waves.

The structural results are as follows.

• If FC holds, then there exists a maximal graded augmented system for $y_1$7, and hence RC holds for any admissible family of curves.
• In the constant-coefficient principal-symbol case, FC, RC, and finite-dimensionality of the formal cokernel are equivalent.

• More explicitly, for constant coefficients the following are equivalent:

  1. $y_1$8 satisfies FC;
  2. there exists a maximal graded augmented system;
  3. $y_1$9 satisfies RC for straight segments on convex sets;
  4. $U\subseteq \mathbb{R}^d$0.

The proof that FC implies RC uses Hilbert’s Nullstellensatz. If the $U\subseteq \mathbb{R}^d$1 minors of $U\subseteq \mathbb{R}^d$2 vanish only at $U\subseteq \mathbb{R}^d$3, then for sufficiently large $U\subseteq \mathbb{R}^d$4,

$U\subseteq \mathbb{R}^d$5

for suitable polynomial matrix multipliers $U\subseteq \mathbb{R}^d$6. This permits all sufficiently high derivatives of $U\subseteq \mathbb{R}^d$7 to be rewritten in terms of derivatives of $U\subseteq \mathbb{R}^d$8 plus lower-order derivatives of $U\subseteq \mathbb{R}^d$9, which is then converted into RC along arbitrary admissible curves.

4. Operators known to satisfy RC

The paper gives a short algebraic verification that a broad collection of geometric operators satisfies FC and therefore RC. The list includes both underdetermined operators, for which one obtains right-inverses up to finite rank, and overdetermined adjoints, for which one obtains representation formulas and coercive inequalities (Isett et al., 4 Sep 2025).

Operator family	Principal-symbol feature in the paper	Consequence
Divergence / gradient	$P$0 injective for $P$1	FC, hence RC
Double divergence / Hessian	$P$2	FC, hence RC
Trace-free double divergence / trace-free Hessian	injective for $P$3	FC, hence RC
Symmetric divergence / Killing	injective by algebraic symmetry argument	FC, hence RC
Trace-free symmetric divergence / conformal Killing	injective for $P$4	FC, hence RC
Linearized Einstein vacuum constraints	block-diagonal principal symbol	FC, hence RC

The abstract of the same paper also identifies the principal application classes more explicitly: divergence, linearized scalar curvature, and linearized Einstein constraint operators on the underdetermined side, together with gradient, Hessian, trace-free Hessian, Killing, and conformal Killing operators on the overdetermined side.

From the dual formulas

$P$5

the paper derives the familiar functional-analytic consequences. Gradient-type operators yield Poincaré or Friedrich inequalities; Killing and conformal Killing operators yield Korn-type and trace-free Korn-type inequalities. A plausible implication is that RC functions as a unifying localization principle behind these integral formulas, rather than as a case-by-case identity tied to a single operator.

5. Relation to earlier curve-based recovery in coding theory

Although the named condition RC appears explicitly in the PDE literature only in (Isett et al., 4 Sep 2025), several earlier papers on locally recoverable codes over algebraic curves develop closely related fiberwise recovery mechanisms without using that name. In these works, recovery occurs on fibers of morphisms of curves, and the operative condition is that the restriction of a codeword to one fiber lies in a low-dimensional interpolation space (Barg et al., 2015, Barg et al., 2016, Barg et al., 2017, Haymaker et al., 2016, Haymaker et al., 2023).

In the basic algebraic-geometric LRC construction, one starts with a separable morphism

$P$6

and evaluation points arranged in split fibers

$P$7

The code space is generated by functions of the form

$P$8

with $P$9. Because the $r_0\times s_0$0 are constant on each fiber, the restriction to one fiber is a univariate polynomial in the auxiliary function $r_0\times s_0$1 of degree at most $r_0\times s_0$2. If the $r_0\times s_0$3-values on the fiber are distinct, one erased coordinate is recovered from the other $r_0\times s_0$4 by interpolation (Barg et al., 2015, Barg et al., 2016).

The 2017 paper on algebraic curves and surfaces states the corresponding recovery criterion in a sharper linear-algebraic form. For a helper set

$r_0\times s_0$5

one forms the $r_0\times s_0$6 matrix

$r_0\times s_0$7

If every $r_0\times s_0$8 submatrix of $r_0\times s_0$9 is invertible, then the value at any point of the helper set can be calculated from the values at the other points of the same helper set (Barg et al., 2017). Fiber-product constructions extend the same idea to several disjoint recovery sets by varying one factor while keeping the others fixed (Haymaker et al., 2016).

These coding-theoretic results do not define RC as a named condition. This suggests a distinction between two uses of “recovery on curves.” In algebraic coding theory, the phrase refers to interpolation on fibers or line/curve sections. In the PDE paper, RC is an operator-theoretic recovery identity along prescribed curves, with dual Green’s functions and symbol-level criteria.

6. Terminological ambiguity and later usage

The acronym “RC” is not stable across the arXiv literature. In the nonconvex-optimization paper “Analytical Convergence Regions of Accelerated Gradient Descent in Nonconvex Optimization under Regularity Condition,” RC means Regularity Condition,

$U$0

and has no connection to curves in the PDE sense (Xiong et al., 2018). In algebraic geometry, “RC varieties” refers to rationally connected varieties, again unrelated (Karzhemanov, 2017). The paper on Bayesian recovery curves in medicine does not use the name RC, though it formalizes a class of monotone post-event trajectories (Wang et al., 2015). Signal-processing work on time-scale-chirp-rate recovery studies recovery of components along lifted parameter-space curves, but not the named PDE condition (Chui et al., 2020).

Against that background, the 2025 PDE paper is the first source in the provided literature that explicitly introduces the term “Recovery on Curves condition (RC)” and ties it to a comprehensive operator-theoretic program (Isett et al., 4 Sep 2025). Its defining features are the curvewise recovery formula for $U$1, the dual construction of Green’s functions supported on chosen curves, the smooth averaging procedure that produces operators of optimal regularizing order, and the equivalence, in the constant-coefficient principal-symbol setting, between RC, FC, and finite-dimensionality of the formal cokernel.

In that sense, RC is best understood as a geometric-analytic condition interpolating between one-dimensional curve integrals and multidimensional PDE inversion. Earlier coding-theoretic work supplies an instructive parallel: in both settings, the decisive step is restriction to a curve or fiber on which the object to be recovered lies in a controlled finite-dimensional class. The PDE formulation, however, is distinguished by its endpoint-supported identity, its duality with curve-supported Green’s functions, and its symbol-theoretic characterization through FC (Isett et al., 4 Sep 2025).