Signum: Discontinuous Sign Operators

• Signum is a mathematical operation that determines the sign of a real number and underpins discontinuous dynamics across various applications.
• It is applied in control systems through RISE controllers and bang-bang methods to achieve robust tracking-error convergence under uncertainty.
• In optimization and field theory, signum drives momentum-based updates in gradient descent and defines models like the signum-Gordon equation with non-analytical potentials.

Searching arXiv for recent and foundational papers on “signum” across mathematics, control, optimization, optics, and field theory. Signum denotes, in its primary mathematical sense, the function that records the sign of a real argument, but contemporary research uses the term more broadly for sign-based operators, sign-splitting filters, sign-driven control laws, and sign-governed nonlinear field equations. Across these settings, the common structural feature is a discontinuous transition at zero or across a dividing manifold, together with a piecewise-constant or piecewise-specified action away from that singular set (Kamalapurkar et al., 2013, Nayak et al., 2023, Klimas et al., 2023).

1. Core mathematical meaning

In standard notation,

$$\operatorname{sgn}(z)= \begin{cases} 1, & z>0,\ 0, & z=0,\ -1, & z<0. \end{cases}$$

Its basic analytical importance lies in the heuristic identity

$\frac{d}{dy}|f(y)|=f'(y)\operatorname{sgn}(f(y)),$

which is valid away from zeros of $f$ and must be handled carefully at points where $f(y)=0$ (Kamalapurkar et al., 2013).

A precise form of that heuristic is central in RISE-based control analysis. If $f:\mathbb R_+\to\mathbb R$ is locally absolutely continuous, then

$$\int_0^x f'(y)\operatorname{sgn}(f(y))\,dy = |f(x)|-|f(0)|.$$

This identity converts a signum-weighted integral into an absolute-value boundary term, which is then used in Lyapunov arguments for asymptotic tracking-error convergence (Kamalapurkar et al., 2013).

The same paper isolates a second basic property of signum-related nonsmoothness. If $f\in C^1([0,\infty))$, then

$$\mu\bigl(\{x: f(x)=0 \text{ and } f'(x)\neq 0\}\bigr)=0.$$

Moreover, under weaker differentiability or slope assumptions, the corresponding set is countable. This matters because signum terms are nonsmooth exactly when their argument is zero, and almost-everywhere derivative statements are often sufficient for stability proofs based on differential inclusions (Kamalapurkar et al., 2013).

A common misconception is that signum manipulations in analysis are merely formal. The control-theoretic literature cited here shows that the relevant identities and negligibility statements can be proved rigorously, with the zero set treated measure-theoretically rather than heuristically (Kamalapurkar et al., 2013).

2. Generalized analytical frameworks

Ordinary pointwise differentiability is not the only meaningful notion available for signum. In spherical-coordinate distribution theory, the obstruction comes from the factor

$$\underline{\omega}=\frac{\underline{x}}{|\underline{x}|},$$

which is undefined at the origin. To handle multiplication by $\underline{\omega}$, radial differentiation, and angular differentiation, the paper on spherical-coordinate distributions introduces signumdistributions as continuous linear functionals on

$\frac{d}{dy}|f(y)|=f'(y)\operatorname{sgn}(f(y)),$0

a test-function space with a controlled singularity at the origin (Brackx, 2018).

In that framework, an ordinary distribution $\frac{d}{dy}|f(y)|=f'(y)\operatorname{sgn}(f(y)),$1 is associated with a signumdistribution $\frac{d}{dy}|f(y)|=f'(y)\operatorname{sgn}(f(y)),$2 through

$\frac{d}{dy}|f(y)|=f'(y)\operatorname{sgn}(f(y)),$3

This allows one to define operations such as multiplication by $\frac{d}{dy}|f(y)|=f'(y)\operatorname{sgn}(f(y)),$4, multiplication by $\frac{d}{dy}|f(y)|=f'(y)\operatorname{sgn}(f(y)),$5, and radial derivatives that are not legitimate on ordinary distributions alone. The paper also shows that the radial derivative of a general distribution is, in general, not unique but only determined up to an equivalence class (Brackx, 2018).

A distinct but related result appears in bang-bang control. Although scalar $\frac{d}{dy}|f(y)|=f'(y)\operatorname{sgn}(f(y)),$6 is not differentiable at $\frac{d}{dy}|f(y)|=f'(y)\operatorname{sgn}(f(y)),$7, the mapping

$\frac{d}{dy}|f(y)|=f'(y)\operatorname{sgn}(f(y)),$8

is Fréchet differentiable at functions $\frac{d}{dy}|f(y)|=f'(y)\operatorname{sgn}(f(y)),$9 whose zero level set is sufficiently regular and satisfies

$f$0

The derivative is the hypersurface functional

$f$1

so the first-order variation of signum is concentrated on the moving switching surface $f$2 (Wachsmuth et al., 29 Sep 2025).

This clarifies another frequent misconception: the scalar non-differentiability of $f$3 at zero does not preclude differentiability as an operator between suitable infinite-dimensional spaces. What changes is the codomain and the nature of the derivative: it becomes a surface measure term rather than a pointwise multiplier (Wachsmuth et al., 29 Sep 2025).

3. Control, zero-finding, and robustness

In nonlinear control, signum is most prominently associated with RISE, short for Robust Integral of the Signum of the Error. RISE controllers are designed for nonlinear systems with exogenous disturbances and modeling uncertainties and are valued for asymptotic tracking-error convergence. Their Lyapunov analysis repeatedly uses the signum integral identity and almost-everywhere arguments around the nonsmooth point at zero (Kamalapurkar et al., 2013).

In optimal control, signum enters bang-bang structure directly. For control-constrained problems, optimal controls satisfy sign conditions of the form

$f$4

or, in semilinear settings,

$f$5

Using the Fréchet differentiability result above, the corresponding nonlinear operator equations admit Newton-type methods with local superlinear convergence under regularity and second-order conditions (Wachsmuth et al., 29 Sep 2025).

A separate line of work uses signum inside terminal zeroing neural networks and error-recurrence systems for time-variant zero-finding under uncertainty. There the error dynamics are split into a rectifying action and an uncertainty-compensation term,

$f$6

The exact compensation law is

$f$7

while the smoothing-signum regularization is

$f$8

The first is used for full rejection of bounded disturbance in the model, and the second is introduced to reduce chattering while preserving finite-duration stabilization into an adjustable residual set (Sun et al., 2024).

Across these control settings, signum is rarely the sole convergence mechanism. A typical pattern is that a nominal finite-time or fixed-time attracting law supplies the main error decay, while signum or smoothing-signum is added as a robustness layer against bounded uncertainty (Sun et al., 2024).

4. Sign-based optimization and learning dynamics

In optimization, Signum is the momentum sign-gradient method. In the continuous-time formulation used for homogeneous classification models, momentum is

$f$9

and the Signum update is

$f(y)=0$0

This is exactly normalized momentum steepest descent in $f(y)=0$1 geometry, so under smoothness, homogeneity, decaying learning rate, and directional convergence assumptions, its limiting direction is the direction of a KKT point of the $f(y)=0$2-max-margin problem (Gronich et al., 18 Feb 2026).

A complementary optimizer taxonomy treats Signum as the momentum-only, non-adaptive corner of the Adam family. In one formulation,

$f(y)=0$3

The same work decomposes Adam into an elementwise adaptation factor multiplied by $f(y)=0$4, identifying Signum as the sign-and-momentum backbone inside Adam (Eschenhagen et al., 10 Feb 2026).

Empirical findings are not uniform across settings. One language-model study reports that signed momentum “essentially recovers Adam’s performance and much of its hyperparameter robustness,” and also notes that Signum is equivalent to Lion with $f(y)=0$5 in its implementation (Zhao et al., 2024). Another controlled language-model study reports that AdamW outperforms Signum across all tested settings, framing Signum as the non-adaptive baseline in the analogy

$f(y)=0$6

(Eschenhagen et al., 10 Feb 2026). A further study argues that in a “narrowing valley” landscape with sufficiently large stochastic gradient noise, Signum can outperform Adam because it keeps fixed-magnitude directional steps, motivating FOCUS as a Signum-like optimizer with an attraction term toward moving-averaged parameters (Liu et al., 21 Jan 2025).

The geometry of Signum also changes how gradient noise should be measured. For sign-based steepest descent, the relevant non-Euclidean gradient noise scale is

$f(y)=0$7

not the Euclidean $f(y)=0$8 quantity used for SGD. An adaptive batch-size strategy based on this $f(y)=0$9 scale reduces training steps by 66.61\% for Signum on a 160M Llama 3 run while matching the better small-batch validation loss (Naganuma et al., 3 Feb 2026).

More recent hybrid optimizers treat Signum as a limiting case rather than an endpoint. In the LionMuon framework, Signum is the special case $f:\mathbb R_+\to\mathbb R$0, i.e. pure single-EMA sign descent. The same framework defines SignMuon by inserting periodic Muon steps into a Signum-style method and reports that LionMuon at $f:\mathbb R_+\to\mathbb R$1 Pareto-dominates Muon, Lion, Signum, and AdamW at 124M scale across all tested datasets and architectures (Bolatov et al., 19 May 2026).

5. Nonlinear wave and field-theoretic models

In relativistic field theory, the signum-Gordon model is defined by

$f:\mathbb R_+\to\mathbb R$2

Because the potential is $f:\mathbb R_+\to\mathbb R$3, the restoring force has constant magnitude away from $f:\mathbb R_+\to\mathbb R$4, the vacuum is non-analytic, and the model has no ordinary linear small-amplitude sector (Streibel et al., 19 Feb 2026, Klimas et al., 2018).

In $f:\mathbb R_+\to\mathbb R$5 dimensions, the model is scale invariant under

$f:\mathbb R_+\to\mathbb R$6

and this leads to self-similar spherically symmetric solutions of the form

$f:\mathbb R_+\to\mathbb R$7

The reduced equation in $f:\mathbb R_+\to\mathbb R$8 dimensions is

$f:\mathbb R_+\to\mathbb R$9

and one class of self-similar solutions describes wiping out the initial field, while another describes accumulation of field energy in a finite and growing region of space (Arodz et al., 2012).

The $$\int_0^x f'(y)\operatorname{sgn}(f(y))\,dy = |f(x)|-|f(0)|.$$0-dimensional model supports exact compact oscillons and their Lorentz-boosted traveling versions. Their scattering produces outgoing quasi-oscillons, jet-like cascades, and occasional shock-wave-like intermediate states; the paper conjectures that radiation in the signum-Gordon model has a fractal-like nature (Hahne et al., 2019).

Higher-dimensional shock waves behave differently. In $$\int_0^x f'(y)\operatorname{sgn}(f(y))\,dy = |f(x)|-|f(0)|.$$1 and $$\int_0^x f'(y)\operatorname{sgn}(f(y))\,dy = |f(x)|-|f(0)|.$$2 dimensions, discontinuous shock-wave profiles solve not the homogeneous signum-Gordon equation alone but a forced equation with a delta source supported on the light cone. For these solutions, the energy trapped inside the cone satisfies

$$\int_0^x f'(y)\operatorname{sgn}(f(y))\,dy = |f(x)|-|f(0)|.$$3

so the interior energy grows proportionally to the $$\int_0^x f'(y)\operatorname{sgn}(f(y))\,dy = |f(x)|-|f(0)|.$$4-dimensional volume and the discontinuity at the front (Klimas et al., 2023).

A perturbed signum-Gordon model,

$$\int_0^x f'(y)\operatorname{sgn}(f(y))\,dy = |f(x)|-|f(0)|.$$5

arises from a dimensional reduction of the first BPS submodel of the Skyrme model. It preserves long-lived, spatially compact oscillatory states well approximated by signum-Gordon breathers at small amplitude, but it breaks exact scale invariance, introduces a second vacuum, and permits compact kink–antikink creation above threshold (Klimas et al., 2018).

The notion of mass is also altered. Because $$\int_0^x f'(y)\operatorname{sgn}(f(y))\,dy = |f(x)|-|f(0)|.$$6 is non-analytic at the vacuum, the perturbative mass is undefined. One recent study therefore defines a spectral mass by fitting dominant nonlinear spectral peaks to the Klein–Gordon dispersion law

$$\int_0^x f'(y)\operatorname{sgn}(f(y))\,dy = |f(x)|-|f(0)|.$$7

It finds that the control parameter is $$\int_0^x f'(y)\operatorname{sgn}(f(y))\,dy = |f(x)|-|f(0)|.$$8, with an approximately massless regime at large values and an ultra-massive nonlinear regime near unity, and identifies the amplitude

$$\int_0^x f'(y)\operatorname{sgn}(f(y))\,dy = |f(x)|-|f(0)|.$$9

as yielding a spectral mass of unity (Streibel et al., 19 Feb 2026).

6. Optical, number-theoretic, and other specialized usages

In Fourier optics, “Signum” names a specific phase operation rather than the scalar function itself. Signum phase mask differential microscopy places a glass cover slip in the Fourier plane of a standard $f\in C^1([0,\infty))$0 imaging system so that spatial frequencies with $f\in C^1([0,\infty))$1 acquire phase $f\in C^1([0,\infty))$2 and those with $f\in C^1([0,\infty))$3 acquire phase $f\in C^1([0,\infty))$4. This implements a one-dimensional signum filter in $f\in C^1([0,\infty))$5, yielding a full-field Hilbert-transform-based spatial differentiator that can produce simultaneous differential amplitude, phase, and polarization-gradient imaging (Nayak et al., 2023).

In analytic number theory, the term appears in Moser’s “law of asymptotic equality of signum-areas” for Hardy’s function $f\in C^1([0,\infty))$6. The relevant quantities are not $f\in C^1([0,\infty))$7, but the positive and negative graph areas

$f\in C^1([0,\infty))$8

On suitable disconnected sets built from Gram-type points, these satisfy

$f\in C^1([0,\infty))$9

so the weighted area above the axis is asymptotically equal to the weighted area below it (Moser, 2013).

The term also appears as nomenclature rather than as a sign operator. Ecce Signum (sigex) is an R package for multivariate structural time series modeling, model-based signal extraction, imputation/casting, and forecasting; its use of “Signum” belongs to package naming rather than to a new mathematical definition of the signum function (McElroy et al., 2022).

Taken together, these usages show that “Signum” functions as a unifying label for discontinuous sign selection, Fourier-domain sign splitting, hypersurface-supported variation, and piecewise-constant nonlinear forcing. The recurring technical theme is not a single application domain, but a shared structure: zero-crossings, switching surfaces, and the mathematics required to make sign-based dynamics rigorous.