Backward in Time Harnack Inequality

• Backward in time Harnack inequality is a set of parabolic comparison estimates where later time data controls earlier solution values, countering the typical diffusion direction.
• It arises in diverse contexts such as boundary vanishing caloric functions, compact-manifold analyses, and geometric flows, providing key insights into regularity and stability.
• This concept highlights how specific model features like nonlocal tails or mixed-type dynamics enable controlled reverse-time estimates despite classical irreversibility.

Backward in time Harnack inequality is a collective term for parabolic comparison estimates in which information at a later time controls a solution at an earlier time, or in which comparison is possible in both temporal directions. In parabolic analysis this stands in deliberate contrast to the classical parabolic Harnack inequality, whose standard form is time-lagged and reflects the irreversible character of diffusion. The modern literature shows that backward formulations arise in several distinct settings: boundary vanishing caloric functions in NTA domains, positive solutions on compact manifolds, nonlinear backward heat equations along geometric flows, mixed-type equations with sign-changing time density, and certain nonlocal or fractional models where the usual waiting-time structure is altered or fails outright (Munive, 2010, Lu et al., 27 Aug 2025).

1. Conceptual role and relation to the classical parabolic Harnack principle

For parabolic equations, the standard Harnack mechanism is asymmetric in time. In the nonlocal parabolic setting studied by Strömqvist, the main Harnack estimate has the usual time-lag form: a supremum over an earlier backward cylinder is controlled by an infimum over a later cylinder, together with a tail term involving the negative part of the solution. The paper emphasizes that these estimates are “of intrinsic parabolic type” and that there is no forward-in-time Harnack principle due to the parabolic nature of the equation (Strömqvist, 2018).

Backward in time Harnack inequalities are therefore exceptional rather than generic. The literature represented here uses the phrase in several technically different ways. In one class, the inequality is a direct pointwise estimate between an earlier and a later space-time point, often near a boundary or under compactness assumptions. In another, it is a differential Harnack inequality for a backward heat-type equation, which becomes a classical space-time comparison after integration along paths. In a third, it appears as a conditional stability estimate for an ill-posed backward problem, where the estimate is not pointwise but controls norms at an earlier time by norms at a later time under an a priori bound (Guo et al., 2014, Cannarsa et al., 2023).

This multiplicity of meanings is not accidental. It reflects the fact that “backward” may refer to the direction of comparison, to the sign of the time derivative in the PDE, or to the geometry of the cylinders used in the estimate. A careful reading of any given result therefore requires attention to whether the inequality is pointwise, differential, integral, boundary-based, or stability-theoretic.

2. Boundary-point backward Harnack in sub-Riemannian heat theory

A canonical local pointwise form appears for the sub-Riemannian heat operator

$$L=-\sum_{i=1}^m X_i^*X_i-\frac{\partial}{\partial t},$$

where $X=\{X_1,\ldots,X_m\}$ are $C^\infty$ vector fields satisfying Hörmander’s finite rank condition, and the domain $\Omega\subset \mathbb{R}^n$ is NTA with respect to the Carnot-Carathéodory metric. In the spacetime cylinder $D_T=\Omega\times(0,T)$, caloric measure $\omega^{(x,t)}$ is defined on the parabolic boundary $\partial_p D_T$ through the representation formula for solutions of $Lu=0$ (Munive, 2010).

The backward Harnack inequality proved in this setting concerns non-negative solutions that vanish on the lateral boundary. If $u\ge 0$ solves $Lu=0$ in $X=\{X_1,\ldots,X_m\}$0 and $X=\{X_1,\ldots,X_m\}$1 on $X=\{X_1,\ldots,X_m\}$2, then for $X=\{X_1,\ldots,X_m\}$3 with $X=\{X_1,\ldots,X_m\}$4 and $X=\{X_1,\ldots,X_m\}$5,

$X=\{X_1,\ldots,X_m\}$6

where $X=\{X_1,\ldots,X_m\}$7 and $X=\{X_1,\ldots,X_m\}$8 are interior corkscrew points at spatial distance $X=\{X_1,\ldots,X_m\}$9 from $C^\infty$0, located at times $C^\infty$1 and $C^\infty$2, respectively, and $C^\infty$3 depends only on the geometric and PDE data (Munive, 2010).

In this formulation, the earlier interior value is bounded by the later one. The estimate is localized near the boundary and is tied to the non-tangential geometry of NTA domains rather than to a global semigroup structure. The paper explicitly interprets the inequality as showing that, going backward in time, the solution cannot decrease too rapidly (Munive, 2010).

Its significance extends beyond the isolated inequality. The same work establishes the doubling property of $C^\infty$4-caloric measure, a Dahlberg-type estimate relating caloric measure to the Green function, and local and global comparison theorems for non-negative caloric functions vanishing on boundary portions. The backward Harnack inequality is described there as equivalent to the doubling property of caloric measure and as central to Fatou-type theorems and Hölder continuity up to the boundary (Munive, 2010).

3. Compact-manifold, geometric-flow, and curve-shortening formulations

A different pointwise regime occurs on compact manifolds without boundary. For a compact $C^\infty$5-dimensional Riemannian manifold $C^\infty$6 with $C^\infty$7, and a positive solution of

$C^\infty$8

an explicit backward Harnack inequality is established with constants depending only on $C^\infty$9 and $\Omega\subset \mathbb{R}^n$0. The ratio $\Omega\subset \mathbb{R}^n$1 is bounded by an explicit function of $\Omega\subset \mathbb{R}^n$2, the diameter, the distance $\Omega\subset \mathbb{R}^n$3, and the curvature bound. The paper stresses three features: comparison is possible in both time directions, the inequality remains meaningful when $\Omega\subset \mathbb{R}^n$4, and no assumptions such as boundedness or vanishing boundary value are imposed on the solution (Lu et al., 27 Aug 2025).

This compact-manifold result is sharply distinguished from the classical Li–Yau forward Harnack inequality. The latter requires a strict time gap $\Omega\subset \mathbb{R}^n$5 and becomes singular as $\Omega\subset \mathbb{R}^n$6, whereas the new inequality stays finite for $\Omega\subset \mathbb{R}^n$7 and yields a space-only Harnack inequality in that limit. The proof is based on Hamilton’s pointwise gradient estimates and Laplacian-log bounds for the heat equation, combined with heat kernel estimates and chaining arguments. The same paper also states that compactness is critical: the backward Harnack inequality need not hold on noncompact manifolds, and on compact domains with boundary such spatial-only or backward Harnack inequalities can fail for local solutions (Lu et al., 27 Aug 2025).

In geometric analysis, backward Harnack inequalities also appear in differential form. For a closed Riemannian manifold with evolving metric

$\Omega\subset \mathbb{R}^n$8

Guo and Ishida study positive solutions of the nonlinear backward heat-type equation

$\Omega\subset \mathbb{R}^n$9

Under assumptions expressed through the error term $D_T=\Omega\times(0,T)$0, they derive inequalities such as

$D_T=\Omega\times(0,T)$1

for $D_T=\Omega\times(0,T)$2. Integration along space-time paths yields classical backward Harnack inequalities for the nonlinear backward heat equation under general geometric flows, including Ricci flow and related coupled flows (Guo et al., 2014).

A nonlocal geometric analogue appears in curve shortening flow. For an initial proper curve in the plane with radial ends and no convexity assumption, Sobnack and Topping define the Harnack quantity

$D_T=\Omega\times(0,T)$3

where $D_T=\Omega\times(0,T)$4 is swept area and $D_T=\Omega\times(0,T)$5 is the turning function. Their alternative Harnack inequality asserts

$D_T=\Omega\times(0,T)$6

The paper describes this estimate as backward in time in flavor, because the bounds are determined by the initial swept area data. It yields an explicit time after which the flow becomes graphical and is presented as a new instance of delayed parabolic regularity (Sobnack et al., 20 Jan 2026).

4. Nonlocal and fractional equations: tails, no-waiting-time phenomena, and failure mechanisms

In parabolic nonlocal equations of the form

$D_T=\Omega\times(0,T)$7

with symmetric kernel satisfying the stated ellipticity condition, the Harnack theory must account for the influence of data outside the ball of interest. Strömqvist introduces averaged and supremum parabolic tails and proves a Harnack inequality of the form

$D_T=\Omega\times(0,T)$8

together with a weak Harnack inequality and local boundedness. The result requires only local nonnegativity of the solution, not global positivity, and the tail of the negative part $D_T=\Omega\times(0,T)$9 is essential (Strömqvist, 2018).

These estimates are not genuine backward-in-time Harnack inequalities in the strict sense; rather, they retain the intrinsic time-lag structure of parabolic Harnack theory. The significance for backward theory lies in the contrast: the paper explicitly notes that the main results are “with time lag” and that no forward-in-time Harnack principle exists for such equations because of parabolic irreversibility (Strömqvist, 2018).

A markedly different phenomenon occurs for global nonlocal solutions. Liao and Weidner establish time-insensitive nonlocal parabolic Harnack estimates for general nonlocal operators with bounded measurable coefficients. For global weak solutions $\omega^{(x,t)}$0, they prove supremum and infimum estimates tied to a common weighted spatial integral, leading to a Harnack inequality on overlapping time intervals. The constants do not blow up as the time intervals approach one another, so no waiting time is required. The paper explicitly contrasts this with the local second-order parabolic case and interprets the result as showing that nonlocality can neutralize the waiting-time phenomenon for global solutions (Liao et al., 2024).

The fractional-time case exhibits the opposite behavior. For

$\omega^{(x,t)}$1

Felsinger, Kassmann, and Voigt prove that the classical local parabolic Harnack inequality fails for $\omega^{(x,t)}$2. Their counterexamples produce smooth positive solutions $\omega^{(x,t)}$3 such that

$\omega^{(x,t)}$4

for any $\omega^{(x,t)}$5 and $\omega^{(x,t)}$6. The paper attributes this failure to the memory effect of the fractional time derivative and the singularity of the fundamental solution at the origin for all $\omega^{(x,t)}$7 when $\omega^{(x,t)}$8. In the same work, a non-local, potential-type Harnack inequality is recovered under suitable assumptions on the initial data, and the classical local Harnack principle is shown to hold if $\omega^{(x,t)}$9 (Dier et al., 2018).

Taken together, these results show that “backward” behavior in nonlocal theory is highly model-dependent. Space-nonlocality may remove waiting time for global solutions, while time-nonlocality may destroy the local Harnack principle altogether.

5. Mixed-type equations, sign changes, and conditional stability for backward problems

Backward Harnack phenomena also arise when the sign of the time coefficient changes. For the mixed-type equation

$\partial_p D_T$0

where $\partial_p D_T$1 may be positive, zero, or negative, Paronetto, Vespri, and coauthors introduce a weighted parabolic De Giorgi class and prove a unified Harnack inequality covering forward parabolic, elliptic, and backward parabolic regions. In the backward parabolic regime $\partial_p D_T$2, the estimate takes the explicit form

$\partial_p D_T$3

so positivity propagates toward earlier times. The same framework yields local boundedness, Hölder continuity, and a maximum principle, including at the interface where $\partial_p D_T$4 changes sign (Paronetto, 2015).

An inhomogeneous extension is obtained for equations of the form

$\partial_p D_T$5

where $\partial_p D_T$6 may change sign. Paronetto defines a homogeneous De Giorgi class of order $\partial_p D_T$7 and proves an inhomogeneous Harnack inequality whose cylinder heights depend on local averages of $\partial_p D_T$8. For forward-backward parabolic equations, the result is stated for solutions rather than the full De Giorgi class. At an interface point $\partial_p D_T$9, the estimate simultaneously couples a forward-time cylinder in the region $Lu=0$0 with a backward-time cylinder in the region $Lu=0$1. The paper emphasizes that when $Lu=0$2, a Harnack-type inequality across cylinders backward in time holds, with the roles of sup and inf adjusted to reflect the sign of the time coefficient. Hölder continuity at interfaces where $Lu=0$3 changes sign is a stated byproduct (Paronetto, 2023).

A different use of the phrase appears in inverse-type backward problems for degenerate parabolic equations. For

$Lu=0$4

in $Lu=0$5, with possibly degenerate diffusion matrix $Lu=0$6, Kian, Yamamoto, and collaborators study reconstruction of $Lu=0$7 from $Lu=0$8. Their main tool is a Carleman-type weighted $Lu=0$9 estimate with weight $u\ge 0$0. Under an a priori bound on $u\ge 0$1, they prove Hölder-type stability for $u\ge 0$2,

$u\ge 0$3

and logarithmic-type stability for $u\ge 0$4. The paper explicitly states that these stability estimates serve as backward Harnack inequalities in the sense that they quantify how the solution can grow, in a controlled manner, when moving backward in time under a priori upper bounds (Cannarsa et al., 2023).

6. Consequences, obstructions, and related semigroup formulations

Where it holds, a backward in time Harnack inequality is a structural statement about regularity, boundary behavior, or long-time comparison. In sub-Riemannian heat theory it underpins doubling of caloric measure, Dahlberg estimates, Fatou-type theorems, and local and global comparison principles (Munive, 2010). In the compact-manifold setting it supplies a space-time comparison valid even at equal times and yields a space-only Harnack inequality as a limiting case (Lu et al., 27 Aug 2025). In curve shortening flow it gives explicit geometric regularization times and describes how a polar graphical flow settles down to an expanding solution (Sobnack et al., 20 Jan 2026). In global nonlocal theory it leads to a Liouville-type theorem stating that a global solution bounded below or above everywhere must be constant (Liao et al., 2024).

The obstructions are equally instructive. The compact-manifold backward Harnack inequality need not extend to noncompact manifolds, and local solutions on domains with boundary may violate backward or space-only Harnack comparison (Lu et al., 27 Aug 2025). For nonlocal equations, no-waiting-time Harnack estimates hold for global solutions, but waiting time is still required for local solutions (Liao et al., 2024). In time-fractional diffusion, memory can destroy the local Harnack principle in the critical and supercritical dimension range $u\ge 0$5 (Dier et al., 2018). In mixed-type equations, backward Harnack comparison is contingent on the weighted De Giorgi framework, the solution class, and the geometry of the interface (Paronetto, 2015, Paronetto, 2023). In degenerate backward problems, the estimates are conditional rather than unconditional, reflecting the intrinsic ill-posedness of parabolic time reversal (Cannarsa et al., 2023).

A semigroup-dual perspective appears in recent work on Langevin dynamics with non-globally dissipative potentials. There, a dimension-free, uniform-in-time reverse transportation inequality controls the Kullback-Leibler divergence between endpoint laws started from different initial points and is presented as the dual version of the Harnack inequality. The paper further states that this reverse transportation inequality is equivalent to a log-Harnack inequality and extends the duality with Harnack inequalities to nonconvex, non-log-concave targets (Lu et al., 21 Dec 2025). This does not replace the PDE notion of backward Harnack inequality, but it shows that “backward” comparison also has an entropy-cost formulation in Markov semigroup theory.

In the modern literature, backward in time Harnack inequality is therefore best understood not as a single theorem but as a family of finely structured comparison principles. Their common content is control against the natural temporal direction of diffusion; their differences lie in the mechanism that makes such control possible—boundary vanishing, compactness, geometric monotonicity, weighted mixed-type structure, global nonlocal interaction, or conditional Carleman stability.