Border Inputs in Algorithms and Systems

• Border inputs are a technical concept where decisive signals are derived from boundaries rather than global features, enhancing precision in detection and segmentation.
• The approach is applied across dense detection, semantic segmentation, and constrained control, with measurable gains in metrics like AP and mIoU.
• Extensions into algebraic geometry and physical systems show that border-based representations guide admissibility, safety, and efficient controller synthesis.

“Border inputs” (Editor’s term) denotes a recurring construction in which the decisive input to an algorithm, dynamical system, or representation is taken from a boundary, border, or admissibility frontier rather than from an interior or global description. The supplied literature suggests that this principle recurs across dense detection, semantic segmentation, figure–ground organization, constrained control, distributed-parameter systems, algebraic border-basis theory, computational geometry, stochastic growth, and literal fence-based monitoring. In these settings, the border is not merely a geometric outline; it is the locus at which localization, viability, thermodynamic exchange, rewriting, or intervention is decided (Qiu et al., 2020, Dhingra et al., 2021, Doná et al., 2013, Ramirez et al., 2021, Boffi et al., 2016).

1. Border-derived representations in visual recognition

In dense object detection, BorderDet is built around the claim that dense detectors should not rely only on the feature at a single grid point when predicting an object box. Instead, it explicitly uses features from the predicted box borders. Its core operator, Border-Align, takes border-sensitive feature maps together with a coarse box $(x_0,y_0,x_1,y_1)$, computes $w=x_1-x_0$ and $h=y_1-y_0$, subdivides each side into $N$ sample points with default $N=10$, and performs channel-wise max pooling along each side using bilinear interpolation. BorderDet embeds this in a Border Alignment Module that first applies a $1\times 1$ convolution with instance normalization to produce a $(4+1)C$-channel border-sensitive bank and then fuses the resulting border-enhanced features back into the detection head. The architecture extends FCOS with a two-stage prediction structure in which coarse scores and boxes are followed by border classification and border location refinements; the final classification score is the product of coarse and border scores. With ResNet-50-FPN, FCOS achieves $38.6$ AP and BorderDet reaches $41.4$ AP; with ResNeXt-101-DCN and multi-scale testing, BorderDet reaches $50.3$ AP (Qiu et al., 2020).

A closely related border-focused refinement strategy appears in semantic segmentation. Border-SegGCN first applies a base network such as U-Net or DeepLabV3+ to produce a pre-segmented mask, then derives a Boolean mask selecting pixels “on the border of different objects,” builds a graph whose nodes are pixels and whose edges connect nearby selected border nodes, and trains a 3-layer GCN with a masked loss so that non-border nodes do not contribute to the loss. Node features combine RGB, the output segmented image from the base algorithm, and intermediate features from the base network. On CamVid with DeepLabV3+, the base mIoU is $w=x_1-x_0$0, Border-SegGCN reaches $w=x_1-x_0$1, and border mIoU improves from $w=x_1-x_0$2 to $w=x_1-x_0$3; with U-Net, overall mIoU improves from $w=x_1-x_0$4 to $w=x_1-x_0$5 and border mIoU from $w=x_1-x_0$6 to $w=x_1-x_0$7 (Dhingra et al., 2021).

These two systems use different mechanisms—side sampling in BorderDet and graph refinement in Border-SegGCN—but the supplied results suggest a common pattern: border inputs are often generated from a coarse first-stage prediction and then used to improve the part of the task most sensitive to precise boundaries, especially high-IoU localization or ambiguous class transitions.

2. Border ownership, contrast orientation, and contextual signals in vision

In cortical vision models, “border input” appears as the local edge fragment together with the contextual signals needed to determine which side of that border belongs to the figure. A computational account of border ownership models paired BOS neurons for each local contour-selective complex cell: two neurons share receptive-field center, orientation selectivity, and local feature selectivity, but have opposite side-of-figure preference. The proposed architecture combines ventral simple and complex cells for local border evidence, early recurrence from dorsal MT for fast large-field context, and relaxation labeling for local consistency. The timing argument is central: border ownership signals in V1/V2 diverge almost from stimulus onset and reach half-peak at about $w=x_1-x_0$8–$w=x_1-x_0$9 ms, whereas MT is reported at about $h=y_1-y_0$0 ms at half-maximal response, making dorsal recurrence a plausible source of early contextual information. The initial BOS response is produced by multiplicative modulation of ventral complex-cell responses by asymmetric MT context sampled from the preferred side of the border, after which relaxation labeling sharpens the assignment (Mehrani et al., 2019).

Psychophysical work interpreted through FACADE theory and the 3D LAMINART model places the same issue in a surface-depth setting. The displays used fragmented inducers on a gray background with luminance $h=y_1-y_0$1 cd/m$h=y_1-y_0$2, black fragments at $h=y_1-y_0$3 cd/m$h=y_1-y_0$4, white fragments at $h=y_1-y_0$5 cd/m$h=y_1-y_0$6, and Weber contrasts $h=y_1-y_0$7 and $h=y_1-y_0$8. Inward-directed contrast edges yielded $h=y_1-y_0$9, $N$0, and $N$1; outward-directed edges yielded $N$2, $N$3, and $N$4. The ANOVA showed a highly significant effect of contrast edge direction on both “in front” and “behind,” but no significant effect of contrast sign. The paper explains this by a polarity-invariant boundary grouping stage—V1 complex cells pool over simple cells with opposite contrast polarities, and V2 bipole grouping cells complete collinear boundaries—followed by surface filling-in, surface contour feedback, and boundary pruning (Dresp-Langley et al., 2018).

Taken together, these studies indicate that border inputs in biological and biologically inspired vision are not reducible to local edge polarity alone. The supplied evidence suggests that contextual information outside the classical receptive field, and in particular information organized relative to the two sides of a border, is necessary for border ownership and relative depth.

3. Boundaries of admissibility in constrained control

For nonlinear control systems with state and input constraints, the border becomes the boundary of the set from which constraints can be satisfied for all future times. For the system $N$5 with $N$6 and state constraints $N$7, the admissible set $N$8 is closed and its boundary decomposes into a usable part on the state-constraint boundary and an interior barrier: $N$9 The usable part satisfies

$N=10$0

while the interior barrier is semipermeable and is generated by trajectories that hit the state-constraint boundary tangentially. Along the barrier, the paper derives a minimum-like principle with Hamiltonian $N=10$1: $N=10$2 This gives a constructive characterization of the barrier by backward integration from tangency points on $N=10$3 (Doná et al., 2013).

With bounded disturbances, the robust admissible set

$N=10$4

admits the analogous decomposition into a usable part and an interior barrier. At the point where a barrier trajectory reaches the active state constraint, the paper proves the ultimate tangentiality condition

$N=10$5

together with a Hamiltonian saddle-point condition

$N=10$6

In the adaptive cruise control example, the barrier control and worst disturbances are bang-bang selections determined by the signs of adjoint components (Rußwurm et al., 23 Sep 2025).

A more algorithmic line of work studies the input-space feasibility induced by multiple control barrier functions. For a control-affine system $N=10$7 with $N=10$8, the admissible controls at state $N=10$9 are

$1\times 1$0

The margin

$1\times 1$1

quantifies compatibility: $1\times 1$2 gives positive slack, $1\times 1$3 is the knife-edge case, and $1\times 1$4 means incompatibility. The paper proposes a Lipschitz-based grid refinement algorithm that verifies or falsifies compatibility offline over a state region (Tan et al., 2022).

Two later constructions move the barrier viewpoint directly into bounded-input controller synthesis. A data-driven input-output CBF method treats the extended input-output history

$1\times 1$5

as an extended state, computes the maximal input-output safe control invariant set $1\times 1$6 from Hankel data, and defines

$1\times 1$7

with an adaptive-decay safety filter that preserves recursive feasibility (Bajelani et al., 24 Feb 2025). A closed-form bounded-input backup CBF method instead interpolates between a nominal controller and a backup controller inside a convex polytope $1\times 1$8,

$1\times 1$9

and derives the optimal interpolation coefficient in closed form, thereby preserving input bounds by convexity while retaining bCBF-style safety guarantees (Wijk et al., 6 Oct 2025). A related hybrid safety synthesis constructs barrier pairs $(4+1)C$0 such that $(4+1)C$1 implies both $(4+1)C$2 and $(4+1)C$3, and then combines local quadratic certificates into a Min-Quadratic Barrier for input-limited nonlinear systems (Thomas et al., 2018).

4. Boundary actuation and thermodynamic ports in distributed systems

In wall-bounded shear flow, the physical actuator is mounted at the border of the domain, but the mathematical input need not appear as a modified boundary condition. An input-output framework for actuated boundary layers keeps the linearized Navier–Stokes state-space form in Fourier space and models dielectric-barrier discharge plasma actuators as localized near-wall body-force distributions. A single localized actuator element uses a narrow Gaussian wall-normal profile

$(4+1)C$4

with forcing applied at the grid point closest to the wall, corresponding to $(4+1)C$5. The method exploits linearity in time and space: pulse-width-modulated inputs are written with $(4+1)C$6, and arbitrary actuator geometry is assembled from shifted, weighted point-source responses. The steady response to a point source in direction $(4+1)C$7 is

$(4+1)C$8

The framework is validated on a constricted-discharge plasma actuator, a spanwise array of symmetric plasma actuators, and a serpentine actuator, and is presented as a low-computational-cost tool for extensive parametric studies (Gluzman et al., 2020).

A more structural treatment of boundary inputs appears in boundary controlled irreversible port-Hamiltonian systems on one-dimensional spatial domains. On $(4+1)C$9, the state consists of extensive variables $38.6$0 and entropy density $38.6$1, with total energy

$38.6$2

The paper defines boundary inputs and outputs by

$38.6$3

where the boundary port vector includes not only $38.6$4 but also irreversible transport terms. The central structural identities are

$38.6$5

for the first law and

$38.6$6

with nonnegative internal entropy production for the second law. In the examples, boundary variables become pressure/velocity pairs for fluid models, entropy flux/temperature for heat conduction, and matter-flux/chemical-potential pairings for diffusion-reaction systems (Ramirez et al., 2021).

These two formulations treat boundary-originated inputs differently—near-wall forcing in one case, thermodynamic port variables in the other—but both make the boundary the privileged interface between an external actuation mechanism and a distributed dynamical system.

5. Algebraic border data and border-basis geometry

In border-basis theory, the essential input is not a geometric contour but an order ideal of monomials. For a lattice ideal $38.6$7, a max-compatible order ideal $38.6$8 is one that is closed under division, contains no two distinct elements equivalent modulo $38.6$9, and represents every congruence class. Once such an $41.4$0 is fixed, the $41.4$1-border basis is canonical: $41.4$2 where $41.4$3 is the representative of $41.4$4 modulo $41.4$5. The paper proves that there are only finitely many max-compatible order ideals, hence only finitely many border bases, for every lattice ideal $41.4$6, even if $41.4$7 has positive dimension. It also proves that not all border bases of a lattice ideal come from Gröbner bases: in one example the Gröbner fan computation gives $41.4$8 reduced Gröbner bases, while the order-ideal construction finds $41.4$9 max-compatible order ideals (Boffi et al., 2016).

A complementary algorithmic advance removes the need for any term ordering. “Computing Border Bases without using a Term Ordering” replaces leading terms by marked terms. A marking chooses, for each polynomial $50.3$0, a support term $50.3$1 with

$50.3$2

Marked Interreduction then performs Gaussian-elimination-style cleanup so that marked terms become pairwise distinct. The Border Basis Algorithm with Term Marking Strategy takes a set of marked generators of a zero-dimensional ideal, builds a candidate order ideal

$50.3$3

tests whether $50.3$4 is an order ideal and whether $50.3$5, and either stops or returns an $50.3$6-border basis. The paper’s motivating example computes a border basis for $50.3$7 that cannot be induced by any term ordering (Kaspar, 2011).

The geometric counterpart is the border basis scheme. For an order ideal $50.3$8 with border $50.3$9, the generic $w=x_1-x_0$00-border prebasis has the form

$w=x_1-x_0$01

and the affine scheme is defined by commutator equations for the formal multiplication matrices: $w=x_1-x_0$02 The corresponding universal family is flat, border basis schemes cover the punctual Hilbert scheme $w=x_1-x_0$03, and the principal component admits an explicit determinantal presentation through ratios $w=x_1-x_0$04. Near each radical point, the paper constructs explicit $w=x_1-x_0$05-parameter flat families, proving that the principal component is rational and smooth at radical points (Kreuzer et al., 2010).

This algebraic literature uses “border” in a technically different sense from detection or control, but the common structure is unmistakable: a border is the minimal frontier across which one rewrites, reduces, or deforms a more complicated object.

6. Literal borders, boundary layouts, and inward aggregation

Some of the supplied work treats borders literally. In constrained boundary labeling, a set of sites inside a rectangle must be labeled from its boundary using non-crossing po-leaders, while grouping constraints require all labels in a group to be consecutive and ordering constraints enforce a partial order along a side. The paper proves that constrained 2-sided boundary labeling is NP-complete even for uniform-height labels and fixed ports, that 1-sided sliding-port labeling with arbitrary label heights is weakly NP-hard, and that two important 1-sided cases are polynomial-time solvable: fixed ports with arbitrary heights in

$w=x_1-x_0$06

time, and sliding ports with uniform-height labels in

$w=x_1-x_0$07

time (Depian et al., 2024).

The border aggregation model makes the boundary the state variable of a stochastic process. Starting from a finite connected graph, an origin $w=x_1-x_0$08, and an initial border $w=x_1-x_0$09, one sets $w=x_1-x_0$10. Each released particle performs a random walk from $w=x_1-x_0$11 until it first reaches graph distance $w=x_1-x_0$12 from the current sticky set,

$w=x_1-x_0$13

sticks at $w=x_1-x_0$14, and enlarges the sticky set by one vertex: $w=x_1-x_0$15 The process stops at

$w=x_1-x_0$16

with deterministic bounds $w=x_1-x_0$17. The model covers OK Corral and the erosion model. On a $w=x_1-x_0$18-ray star, the residual nonsticky mass is of order $w=x_1-x_0$19; on the comb lattice, $w=x_1-x_0$20 is of order $w=x_1-x_0$21 with high probability; on $w=x_1-x_0$22 for $w=x_1-x_0$23, $w=x_1-x_0$24 is at least order $w=x_1-x_0$25 with high probability (Thacker et al., 2017).

A final, explicitly physical use of border inputs appears in an IoT border intrusion warning system. The proposed device is installed “over the fences” and “across the pillars of the border,” uses an infrared obstacle sensor with range $w=x_1-x_0$26 cm to $w=x_1-x_0$27 cm, sends notifications through a NodeMCU ESP32/ESP8266-class Wi-Fi MCU and the Blynk app, and applies the simple threshold rule

$w=x_1-x_0$28

When the threshold is crossed, the system sends a “Motion detected” notification and allows a soldier to activate a buzzer and emergency lights from the smartphone app (Tasnim et al., 2020).

These cases show the literal end of the spectrum. Here the border is neither an inferred box side nor an admissible-set frontier, but a rectangle edge, a growing sticky set, or a geopolitical fence. This suggests that “border inputs” is less a single theory than a recurrent technical move: elevate the boundary from a passive outline to the primary source of representation, action, or decision.