LightCROWN: Neural CBF Verification

• LightCROWN is a neural network verification method that tightens Jacobian bounds in neural control barrier functions by using exact interval analysis for smooth activations.
• It replaces generic derivative relaxations in CROWN’s backward propagation with analytically computed extrema, specifically benefiting activations like tanh, sigmoid, and Swish.
• Empirical evaluations on nonlinear control systems show that LightCROWN improves verification rates and reduces runtimes compared to standard CROWN and BBV methods.

Searching arXiv for LightCROWN and closely related CROWN/NCBF verification papers. LightCROWN is a verification method for neural control barrier functions (NCBFs) with smooth nonlinear activations, introduced as a Jacobian-bounding improvement within the CROWN family of neural network verification methods. Its central purpose is to tighten interval bounds on activation derivatives during backward Jacobian propagation, thereby improving certification of safety conditions that depend on the Lie derivative $\nabla h(x)^\top f(x,u)$. In the paper "Efficient Verification of Neural Control Barrier Functions with Smooth Nonlinear Activations" (Zhang et al., 8 May 2026), LightCROWN is presented as a plug-in refinement of existing CROWN-style pipelines rather than a new barrier condition or a wholly separate branch-and-bound framework. The name should also be distinguished from low-light image enhancement work such as "Towards Controllable Low-Light Image Enhancement: A Continuous Multi-illumination Dataset and Efficient State Space Framework" (Han et al., 26 Mar 2026), which introduces CLE-RWKV and Light100 and does not mention LightCROWN.

1. Problem setting and conceptual role

LightCROWN is formulated for continuous-time nonlinear control systems

$$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$

A neural control barrier function $h(x)$ defines the safe set

$$C = \{x \in D \mid h(x) \le 0\}, \qquad \partial C = \{x \in D \mid h(x)=0\}.$$

Safety is understood as forward invariance: if $x(0)\in C$, then $x(t)\in C$ for all $t\ge 0$ (Zhang et al., 8 May 2026).

The verification condition used in the paper is

$\mathcal{L}_f h(x)+\alpha(h(x)) \le 0,$

where $\alpha$ is an extended class-$\mathcal K$ function and

$$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$0

By Nagumo’s theorem, on the boundary $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$1, this reduces to

$$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$2

The exact boundary is conservatively covered by finitely many axis-aligned subregions $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$3, and certification proceeds regionwise (Zhang et al., 8 May 2026).

Within this formulation, the bottleneck is the Jacobian bound for $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$4, because the verification inequality depends directly on the Lie derivative. The paper’s core claim is that prior CROWN-based methods for smooth activations are unnecessarily conservative at the derivative-bounding stage, and that this conservatism propagates into weak or failed NCBF certificates (Zhang et al., 8 May 2026). This suggests that LightCROWN’s significance lies less in changing the verification objective than in improving one analytically critical inner step.

2. Neural barrier networks and the verification objective

The NCBF is modeled as a feedforward neural network

$$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$5

with parameters $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$6, $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$7, and $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$8 (Zhang et al., 8 May 2026). The pre-activations are defined recursively by

$$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$9

Using the chain rule, the Jacobian is written as

$h(x)$0

where

$h(x)$1

On each subregion $h(x)$2, the verifier computes interval bounds

$h(x)$3

The paper rewrites certification as

$h(x)$4

where $h(x)$5 is an upper bound on $h(x)$6 over the region. Using componentwise sign decomposition, the explicit computable upper bound becomes

$h(x)$7

The paper assumes $h(x)$8 is a convex polytope, so it suffices to test vertices of $h(x)$9 (Zhang et al., 8 May 2026).

The practical implication is straightforward: any looseness in the interval enclosure for $$C = \{x \in D \mid h(x) \le 0\}, \qquad \partial C = \{x \in D \mid h(x)=0\}.$$0 directly weakens the upper bound in (1), causing either verification failure or more recursive splitting. LightCROWN is designed precisely to reduce that source of looseness.

3. The LightCROWN mechanism

LightCROWN is best understood as a drop-in tightening of the Jacobian propagation stage in CROWN-style verifiers (Zhang et al., 8 May 2026). The backward recursion is unchanged; only the way the diagonal derivative matrices

$$C = \{x \in D \mid h(x) \le 0\}, \qquad \partial C = \{x \in D \mid h(x)=0\}.$$1

are bounded is replaced.

The procedure is:

1. propagate input intervals forward to obtain neuronwise pre-activation bounds

$$C = \{x \in D \mid h(x) \le 0\}, \qquad \partial C = \{x \in D \mid h(x)=0\}.$$2

1. compute exact minimum and maximum values of $$C = \{x \in D \mid h(x) \le 0\}, \qquad \partial C = \{x \in D \mid h(x)=0\}.$$3 on each interval, using analytic properties of the activation rather than generic linear relaxation;
2. inject those exact derivative bounds into the standard backward Jacobian interval recursion (Zhang et al., 8 May 2026).

For $$C = \{x \in D \mid h(x) \le 0\}, \qquad \partial C = \{x \in D \mid h(x)=0\}.$$4,

$$C = \{x \in D \mid h(x) \le 0\}, \qquad \partial C = \{x \in D \mid h(x)=0\}.$$5

Since this derivative is symmetric, maximized at $$C = \{x \in D \mid h(x) \le 0\}, \qquad \partial C = \{x \in D \mid h(x)=0\}.$$6 with value $$C = \{x \in D \mid h(x) \le 0\}, \qquad \partial C = \{x \in D \mid h(x)=0\}.$$7, and decreases as $$C = \{x \in D \mid h(x) \le 0\}, \qquad \partial C = \{x \in D \mid h(x)=0\}.$$8 grows, the exact derivative interval over $$C = \{x \in D \mid h(x) \le 0\}, \qquad \partial C = \{x \in D \mid h(x)=0\}.$$9 is

$x(0)\in C$0

The paper emphasizes that these are exact interval extrema, not affine relaxations in the usual sense (Zhang et al., 8 May 2026).

A general smooth-activation version is also given. Let

$x(0)\in C$1

be the critical points of $x(0)\in C$2. Then over interval $x(0)\in C$3,

$x(0)\in C$4

where

$x(0)\in C$5

Thus the framework applies whenever derivative extrema can be characterized from endpoints and finitely many critical points (Zhang et al., 8 May 2026).

This suggests that LightCROWN’s conceptual novelty is narrow but precise: it removes a generic approximation where exact interval analysis is already available.

4. Backward propagation, dynamics bounds, and full certification workflow

The generic backward Jacobian propagation shared by CROWN and LightCROWN is initialized by

$x(0)\in C$6

For $x(0)\in C$7, it computes

$x(0)\in C$8

$x(0)\in C$9

followed by

$x(t)\in C$0

$x(t)\in C$1

At termination,

$x(t)\in C$2

The paper states explicitly that LightCROWN changes none of this recursion; it tightens only the interval inputs $x(t)\in C$3 (Zhang et al., 8 May 2026).

Dynamics are bounded on each subregion by affine Taylor models

$x(t)\in C$4

with

$x(t)\in C$5

and entrywise offsets

$x(t)\in C$6

$x(t)\in C$7

where $x(t)\in C$8 bounds the Hessian operator norm of the $x(t)\in C$9-th component over the region (Zhang et al., 8 May 2026).

The complete pipeline has four stages:

Stage	Operation	Role
1	Search subregions covering the boundary	Conservative cover of $t\ge 0$0
2	Bound dynamics on each subregion	Compute $t\ge 0$1
3	Bound the NCBF Jacobian using LightCROWN	Compute $t\ge 0$2
4	Verify the subregion condition	Check the upper bound in (1)

If a region is not certified, it may be split into $t\ge 0$3 smaller hyperrectangles by bisecting each dimension, up to a maximum split count $t\ge 0$4 (Zhang et al., 8 May 2026). A plausible implication is that LightCROWN is especially valuable in branch-and-bound settings because improved local Jacobian bounds can reduce the number of unresolved nodes.

5. Relation to prior CROWN-style verification and supported activations

The paper contrasts LightCROWN with three categories of prior approaches (Zhang et al., 8 May 2026).

First, it distinguishes itself from ReLU-based symbolic derivative propagation methods such as Hu et al. These methods are described as efficient but limited to non-smooth activations. LightCROWN instead targets smooth nonlinear activations by exploiting the analytic structure of activation derivatives rather than piecewise-linearity.

Second, it is contrasted with BBV-style methods. The paper states that BBV supports smooth activations but relies on conservative generic approximations, whereas LightCROWN replaces derivative relaxation with exact interval extrema on each neuron’s pre-activation interval (Zhang et al., 8 May 2026).

Third, it is contrasted with more complex affine or McCormick relaxation approaches, which can reduce conservatism but require more propagation parameters and engineering choices. LightCROWN is described as simpler and parameter-free in the derivative stage (Zhang et al., 8 May 2026).

Supported activations are smooth nonlinear activations whose derivative extrema over an interval can be obtained analytically or from finitely many critical points. The paper explicitly mentions:

• $t\ge 0$5,
• sigmoid,
• Swish (Zhang et al., 8 May 2026).

The main practical emphasis is on $t\ge 0$6, because the derivative structure is especially transparent:

$t\ge 0$7

has a single global maximizer at $t\ge 0$8. The paper therefore presents $t\ge 0$9 as the cleanest case, while retaining a more general smooth-activation formulation (Zhang et al., 8 May 2026).

A common misconception would be to treat LightCROWN as a complete departure from the CROWN paradigm. The paper states the opposite: it is a plug-in improvement to CROWN/$\mathcal{L}_f h(x)+\alpha(h(x)) \le 0,$0-$\mathcal{L}_f h(x)+\alpha(h(x)) \le 0,$1-CROWN/auto_LiRPA-style Jacobian verification pipelines, with the same backward propagation machinery and only a modified derivative-bounding subroutine (Zhang et al., 8 May 2026).

6. Empirical results, scalability, and observed behavior

Experiments are reported on three nonlinear control systems:

• Inverted Pendulum (2D),
• Dubins Car (3D),
• Planar Quadrotor (6D) (Zhang et al., 8 May 2026).

The NCBF architecture is a two-layer $\mathcal{L}_f h(x)+\alpha(h(x)) \le 0,$2-activated MLP

$\mathcal{L}_f h(x)+\alpha(h(x)) \le 0,$3

Initial hidden widths are 6 for the pendulum, 64 for the car, and 8 for the quadrotor; a scalability study uses widths $\mathcal{L}_f h(x)+\alpha(h(x)) \le 0,$4 (Zhang et al., 8 May 2026). The verification margin is

$\mathcal{L}_f h(x)+\alpha(h(x)) \le 0,$5

Baselines are a general CROWN linear-relaxation baseline adapted from Hu et al. by removing ReLU-specific symbolic optimizations, and BBV (Zhang et al., 8 May 2026).

Under regular training, the inverted pendulum is described as easy, with all methods around $\mathcal{L}_f h(x)+\alpha(h(x)) \le 0,$6 verified rate. For Dubins car at $\mathcal{L}_f h(x)+\alpha(h(x)) \le 0,$7, the paper reports CROWN $\mathcal{L}_f h(x)+\alpha(h(x)) \le 0,$8, LightCROWN $\mathcal{L}_f h(x)+\alpha(h(x)) \le 0,$9, and BBV $\alpha$0. For planar quadrotor at $\alpha$1, it reports CROWN $\alpha$2, LightCROWN $\alpha$3, and BBV $\alpha$4, with runtimes of $\alpha$5 s, $\alpha$6 s, and $\alpha$7 s respectively (Zhang et al., 8 May 2026).

Under adversarial training, the reported gains are larger. For Dubins car at $\alpha$8, the results are CROWN $\alpha$9, LightCROWN $\mathcal K$0, and BBV $\mathcal K$1; at $\mathcal K$2, they are CROWN $\mathcal K$3, LightCROWN $\mathcal K$4, and BBV $\mathcal K$5 (Zhang et al., 8 May 2026). For planar quadrotor at $\mathcal K$6, the paper reports CROWN $\mathcal K$7, LightCROWN $\mathcal K$8, and BBV $\mathcal K$9, with runtimes of $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$00 s, $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$01 s, and $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$02 s (Zhang et al., 8 May 2026). The stated headline is that verification success rates improve “up to 100\%” (Zhang et al., 8 May 2026).

The scalability study on Dubins car with hidden widths $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$03 reports that LightCROWN consistently attains the highest verified rate across widths and $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$04 values. At $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$05 and width $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$06, the reported verified rates are LightCROWN $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$07, CROWN $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$08, and BBV $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$09 (Zhang et al., 8 May 2026).

The paper also documents a notable phenomenon under regular training: at higher $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$10, LightCROWN’s verified rate can dip below CROWN’s in some cases. The authors attribute this to a distributional effect rather than weaker bounds: LightCROWN produces tighter upper bounds clustered just below zero, whereas CROWN produces looser but more dispersed bounds, some of which may happen to lie farther negative when the positive margin term $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$11 is added (Zhang et al., 8 May 2026). This is an important qualification, because it clarifies that tighter interval propagation need not monotonically improve every downstream verification statistic under every training regime.

7. Limitations, implementation details, and disambiguation

The paper acknowledges or implies several limitations (Zhang et al., 8 May 2026). The approach is strongest when activation derivative structure is analytically exploitable, as with $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$12. It does not eliminate conservatism arising from conservative boundary covers, affine dynamics bounds, or recursive subdivision. The existential control check assumes that $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$13 is a convex polytope and verifies over its vertices. The experiments use shallow two-layer $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$14 MLPs; width is varied, but depth is not broadly explored.

Implementation details reported in the paper include:

• hardware: Intel Xeon Platinum 8375C CPU and 512 GB RAM;
• boundary discretization: 600 grids/dim for pendulum, 50 grids/dim for Dubins car, and 6 grids/dim for planar quadrotor;
• branch-and-bound style splitting with maximum split count $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$15;
• regular training and adversarial training using PGD-style perturbations;
• code availability at github.com/Autonomous-Systems-and-Control-Lab/verify-neural-CBF (Zhang et al., 8 May 2026).

A practical conclusion drawn in the paper is that LightCROWN should help most when the activation is smooth, Jacobian bounds dominate conservatism, and tighter derivative intervals can reduce splitting. The paper also suggests hybrid pipelines in which LightCROWN certifies most regions quickly and exact SMT or MIP methods are reserved for hard residual regions (Zhang et al., 8 May 2026).

Finally, the term LightCROWN should be carefully disambiguated. The paper "Towards Controllable Low-Light Image Enhancement: A Continuous Multi-illumination Dataset and Efficient State Space Framework" (Han et al., 26 Mar 2026) introduces CLE-RWKV, Light100, L-RWKV blocks, an S2D design, and a noise-decoupled supervision strategy for controllable low-light image enhancement. It explicitly does not mention LightCROWN. The similarity is therefore nominal rather than substantive: LightCROWN belongs to neural verification for control, whereas CLE-RWKV and Light100 belong to controllable low-light image enhancement (Han et al., 26 Mar 2026).

In summary, LightCROWN is a CROWN-compatible Jacobian-bounding refinement for NCBF verification with smooth activations, especially $$\dot{x} = f(x,u), \quad x \in D \subset \mathbb{R}^n,\; u \in U \subset \mathbb{R}^k.$$16. Its defining contribution is to replace generic derivative relaxations with exact interval extrema of activation derivatives on pre-activation intervals, leaving the remainder of the backward CROWN machinery intact. The reported effect is tighter Lie-derivative bounds, higher verification success rates, and frequently improved runtime on difficult nonlinear control benchmarks (Zhang et al., 8 May 2026).