Lavrentiev's Gap: Approximation & Regularity

• Lavrentiev’s gap is the strict inequality between the infimum of a variational integral over a natural energy space and that over a denser but more regular subclass.
• It highlights cases where finite-energy maps cannot be approximated by smoother competitors without increasing energy, thereby obstructing classical approximation and mollification techniques.
• Examples from one-dimensional models and double-phase functionals illustrate its impact on regularity, discretization, and topological aspects in variational calculus.

Lavrentiev's gap is the strict inequality between the infimum of a variational integral over a natural “large” admissible class and the infimum over a denser but more regular subclass. In the standard abstract form, if $\mathcal G:X\to[0,\infty]$ and $Y\subset X$ is dense in the topology of $X$, the phenomenon occurs when

$$\inf_{u\in X}\mathcal G(u)\;<\;\inf_{u\in Y}\mathcal G(u),$$

equivalently when some finite-energy map in $X$ cannot be approximated by smoother competitors without increasing the energy. First pointed out by M.A. Lavrentiev in 1927, the gap has become a central obstruction in the calculus of variations, affecting approximation theory, regularity, numerical discretization, and the structure of admissible classes in problems with singular or nonuniform growth (Balci et al., 2023).

1. Definitions and admissible classes

The precise form of Lavrentiev's gap depends on the ambient variational setting. In one-dimensional scalar problems, a common formulation compares $W^{1,1}$ or $AC$ against $C^1$ or $W^{1,\infty}$. For the autonomous problem

$$\min F(y)=\int_0^1L(y(t),y'(t))\,dt,\qquad y(0)=0,\quad y\in W^{1,1}([0,1],\mathbb R),$$

one sets

$Y\subset X$0

and a strict Lavrentiev gap occurs when $Y\subset X$1 (Raphael et al., 2022). Equivalent formulations appear throughout the literature: $Y\subset X$2 versus $Y\subset X$3, $Y\subset X$4 versus $Y\subset X$5, Sobolev–Orlicz classes versus closures of smooth maps, or regular invertible deformation classes versus their weak closures (Feng et al., 2016, Balci et al., 2020, Barchiesi et al., 24 Mar 2026).

The functional-analytic content is always the same. The gap is not merely a failure of topological density, since the smaller class is often dense in the ambient topology. Rather, it is a failure of energy density: there is no sequence of smoother admissible maps whose energies converge to the infimum attained or approximated in the larger class. This is why the phenomenon directly obstructs mollification arguments, Euler–Lagrange approximation, and conforming discretizations.

In generalized-growth settings the definition is expressed relative to the natural energy space. For instance, for the double-phase functional

$Y\subset X$6

one asks whether

$Y\subset X$7

where $Y\subset X$8 is the corresponding Musielak–Orlicz–Sobolev class (Borowski et al., 2023). For manifold-valued maps, the same question becomes the modular or strong density of smooth maps in $Y\subset X$9 (Antonini et al., 20 Dec 2025). In nonlinear elasticity, the gap can arise between a regular invertible class and its weak $X$0-closure, showing that even the admissible class itself may fail to be weakly closed (Barchiesi et al., 24 Mar 2026).

2. Canonical one-dimensional examples

The classical prototype is Manià’s problem,

$X$1

The singular map $X$2 satisfies $X$3, hence it realizes the $X$4-infimum. However, $X$5 is unbounded near $X$6, and any Lipschitz competitor has strictly positive energy; a numerical study quoting Ferriero states that for any sequence of Lipschitz trajectories $X$7 a.e., one has $X$8 (Feng et al., 2016, Pereira et al., 2010). This example became the standard model for how a singular minimizer can be invisible to regular approximants.

A crucial refinement is that endpoint conditions matter. In Mariconda’s one-dimensional non-autonomous theory, the same Manià integrand yields a true gap for the two-endpoint problem $X$9, $$\inf_{u\in X}\mathcal G(u)\;<\;\inf_{u\in Y}\mathcal G(u),$$0, but no gap for the one-endpoint problem with only $$\inf_{u\in X}\mathcal G(u)\;<\;\inf_{u\in Y}\mathcal G(u),$$1. The truncations

$$\inf_{u\in X}\mathcal G(u)\;<\;\inf_{u\in Y}\mathcal G(u),$$2

satisfy $$\inf_{u\in X}\mathcal G(u)\;<\;\inf_{u\in Y}\mathcal G(u),$$3, converge to $$\inf_{u\in X}\mathcal G(u)\;<\;\inf_{u\in Y}\mathcal G(u),$$4 in $$\inf_{u\in X}\mathcal G(u)\;<\;\inf_{u\in Y}\mathcal G(u),$$5, and satisfy $$\inf_{u\in X}\mathcal G(u)\;<\;\inf_{u\in Y}\mathcal G(u),$$6 for all $$\inf_{u\in X}\mathcal G(u)\;<\;\inf_{u\in Y}\mathcal G(u),$$7 (Mariconda, 2022). This sharply separates the one-endpoint and two-endpoint geometries.

A second instructive one-dimensional example is due to Cerf–Mariconda. Define

$$\inf_{u\in X}\mathcal G(u)\;<\;\inf_{u\in Y}\mathcal G(u),$$8

and let $$\inf_{u\in X}\mathcal G(u)\;<\;\inf_{u\in Y}\mathcal G(u),$$9. Then $X$0, so $X$1 is a $X$2-minimizer with $X$3, while any Lipschitz $X$4 with $X$5 satisfies $X$6. Hence $X$7 (Raphael et al., 2022). The mechanism is explicit: along the singular minimizer the term $X$8 vanishes exactly, while regular competitors cannot reproduce that singular profile near $X$9.

These examples established two enduring facts. First, polynomial or even smooth dependence on the gradient does not preclude a gap. Second, the phenomenon is often driven by a singular minimizer whose derivative blows up on a small set while the integrand is engineered to vanish exactly along that singular trajectory.

3. One-dimensional non-occurrence theorems

One-dimensional scalar problems also admit some of the sharpest positive results on the absence of the Lavrentiev phenomenon. For autonomous functionals with one endpoint condition,

$W^{1,1}$0

Cerf–Mariconda introduced Condition (R): there exist locally Lipschitz functions $W^{1,1}$1 such that $W^{1,1}$2 for all $W^{1,1}$3, and for every bounded interval $W^{1,1}$4,

$W^{1,1}$5

If $W^{1,1}$6, $W^{1,1}$7, and $W^{1,1}$8 satisfies Condition (R) on $W^{1,1}$9, then there exists a sequence of Lipschitz functions $AC$0 with $AC$1, $AC$2 in $AC$3, and $AC$4. In particular, no gap occurs at $AC$5; if Condition (R) holds on every bounded interval, then

$AC$6

Condition (R) is strictly weaker than the classical rectangle-boundedness condition of Alberti–Serra Cassano, because it requires boundedness only along two graphs $AC$7, not on full state-velocity rectangles (Raphael et al., 2022).

Mariconda’s non-autonomous theory reaches a related conclusion under a different structural package. Writing

$AC$8

the hypotheses include radial convexity in $AC$9, a time-Lipschitz condition (S) on $C^1$0, boundedness of $C^1$1 along the image of $C^1$2, continuity of $C^1$3, and local boundedness of $C^1$4 near the graph of $C^1$5. Under these assumptions, one obtains a constructive sequence of Lipschitz reparametrizations $C^1$6 with the same one-endpoint data, $C^1$7 in $C^1$8, and $C^1$9; for two endpoints one needs additional hypotheses such as positivity of $W^{1,\infty}$0 and either real-valuedness of $W^{1,\infty}$1 or blow-up of $W^{1,\infty}$2 near $W^{1,\infty}$3 (Mariconda, 2022).

The underlying method is reparametrization rather than direct smoothing. High-derivative sets are slowed down, and compensating intervals are inserted so that the final map remains Lipschitz and preserves the endpoint constraints. In the autonomous one-endpoint case, Cerf–Mariconda refine this by replacing problematic slopes on bad intervals with the bounded velocities $W^{1,\infty}$4, then reconnecting the endpoint through short differential-equation arcs $W^{1,\infty}$5 (Raphael et al., 2022). These constructions show that, at least in one dimension, the absence of a gap can often be reduced to a geometric control of the integrand along selected velocity profiles.

4. Nonuniform growth, sharp thresholds, and density scales

In higher dimensions and nonstandard-growth settings, Lavrentiev’s gap is tightly linked to the density of smooth maps in Musielak–Orlicz-type energy spaces. For double-phase energies

$W^{1,\infty}$6

the classical criterion assumes $W^{1,\infty}$7, $W^{1,\infty}$8, and $W^{1,\infty}$9, which guarantees density of $$\min F(y)=\int_0^1L(y(t),y'(t))\,dt,\qquad y(0)=0,\quad y\in W^{1,1}([0,1],\mathbb R),$$0 in the natural space and hence absence of the gap. A sharp extension is the scale $$\min F(y)=\int_0^1L(y(t),y'(t))\,dt,\qquad y(0)=0,\quad y\in W^{1,1}([0,1],\mathbb R),$$1, defined by

$$\min F(y)=\int_0^1L(y(t),y'(t))\,dt,\qquad y(0)=0,\quad y\in W^{1,1}([0,1],\mathbb R),$$2

If $$\min F(y)=\int_0^1L(y(t),y'(t))\,dt,\qquad y(0)=0,\quad y\in W^{1,1}([0,1],\mathbb R),$$3 and $$\min F(y)=\int_0^1L(y(t),y'(t))\,dt,\qquad y(0)=0,\quad y\in W^{1,1}([0,1],\mathbb R),$$4, then no Lavrentiev gap arises; if competitors are $$\min F(y)=\int_0^1L(y(t),y'(t))\,dt,\qquad y(0)=0,\quad y\in W^{1,1}([0,1],\mathbb R),$$5, the weaker bound $$\min F(y)=\int_0^1L(y(t),y'(t))\,dt,\qquad y(0)=0,\quad y\in W^{1,1}([0,1],\mathbb R),$$6 suffices. Conversely, under $$\min F(y)=\int_0^1L(y(t),y'(t))\,dt,\qquad y(0)=0,\quad y\in W^{1,1}([0,1],\mathbb R),$$7, there exist $$\min F(y)=\int_0^1L(y(t),y'(t))\,dt,\qquad y(0)=0,\quad y\in W^{1,1}([0,1],\mathbb R),$$8 and boundary data $$\min F(y)=\int_0^1L(y(t),y'(t))\,dt,\qquad y(0)=0,\quad y\in W^{1,1}([0,1],\mathbb R),$$9 such that the gap occurs (Borowski et al., 2023).

The smoothness of the weight can be pushed further. For $Y\subset X$00, the sharp threshold becomes

$Y\subset X$01

If this holds, then

$Y\subset X$02

while if

$Y\subset X$03

there exist $Y\subset X$04, $Y\subset X$05, and $Y\subset X$06 with a genuine gap. In particular, if $Y\subset X$07, no additional restrictions on $Y\subset X$08 and $Y\subset X$09 are required (Borowski, 8 Sep 2025).

Generalized Orlicz and borderline double-phase models exhibit a complementary dichotomy. In the two-dimensional log-weighted model

$Y\subset X$10

with checkerboard coefficient $Y\subset X$11, one has $Y\subset X$12 and hence no gap if $Y\subset X$13, whereas $Y\subset X$14 and a genuine gap occurs if $Y\subset X$15 and $Y\subset X$16 (Balci et al., 2020). In anisotropic Musielak–Orlicz settings, Borowski–Chlebicka–Miasojedow prove absence of the gap under a local balance condition

$Y\subset X$17

recovering in particular the sharp double-phase bound $Y\subset X$18 in the isotropic $Y\subset X$19-case and its coordinatewise analogues in multi-phase anisotropic models (Borowski et al., 2022).

Parabolic double-phase problems admit an analogous density theory. If $Y\subset X$20 and

$Y\subset X$21

then every finite-energy map in the natural parabolic class admits smooth approximants with convergence in $Y\subset X$22 and convergence of the stationary energy $Y\subset X$23. The threshold improves to

$Y\subset X$24

for bounded solutions, and to

$Y\subset X$25

under $Y\subset X$26-regularity (Kim et al., 15 Mar 2026).

A broader scalar theorem shows that even in non-autonomous, non-convex problems, the gap can be discarded under an anti-jump balance condition $Y\subset X$27, while dropping boundedness or convexity in the second variable and any $Y\subset X$28-type assumption in the last variable (Borowski et al., 2024). This suggests that the decisive issue is not regularity of the integrand in isolation, but a quantitative balance between its spatial oscillation and its growth with respect to the gradient.

5. Singular sets, fractal barriers, and topological obstructions

A recurring misconception is that the absence of a Lavrentiev gap forces the singular set of a minimizer to be small. Gratwick disproved this sharply: given any closed Lebesgue-null set $Y\subset X$29 and an arbitrary smooth strictly convex superlinearity $Y\subset X$30, there exist a smooth strictly convex Lagrangian $Y\subset X$31 and a unique minimizer $Y\subset X$32 whose singular set

$Y\subset X$33

coincides exactly with $Y\subset X$34, while still admitting smooth competitors $Y\subset X$35 with $Y\subset X$36 uniformly and $Y\subset X$37. Thus non-occurrence of the Lavrentiev phenomenon does not imply that the singular set is small (Gratwick, 2015).

Another common belief is that the gap is tied to an exponent crossing the ambient dimension. Fractal constructions show this is not generally correct. Balci–Diening–Surnachev replace a point singularity by a fractal contact set $Y\subset X$38 of prescribed Hausdorff dimension and construct gaps for variable-exponent, double-phase, and weighted $Y\subset X$39-energies, thereby showing that the dimensional threshold is not essential in the classical sense (Balci et al., 2019). In a related direction, variational problems with differential forms admit a unifying “separating pair” construction $Y\subset X$40 that violates Stokes’ theorem on a negligible set and produces non-density in double-phase, borderline double-phase, and variable-exponent models (Balci et al., 2023).

For manifold-valued maps, the issue is entangled with topology. Under growth conditions expressed by integral criteria on $Y\subset X$41, or under $Y\subset X$42-connectedness of the target manifold $Y\subset X$43 with upper-growth exponent $Y\subset X$44, smooth maps are strongly or weakly dense in $Y\subset X$45, and hence no Lavrentiev gap occurs. But when the local comparability condition fails, there are vectorial double-phase counterexamples with smooth boundary data and a genuine gap between $Y\subset X$46 and smooth $Y\subset X$47-valued competitors (Antonini et al., 20 Dec 2025). The scalar density problem and the topological obstruction problem are therefore distinct but structurally compatible facets of the same phenomenon.

6. Numerical analysis and discrete approximation

Lavrentiev’s gap is a decisive obstruction for conforming numerical schemes because such schemes typically minimize over spaces contained in $Y\subset X$48 or in smooth finite-dimensional trial classes. When a gap is present, the discrete minimum converges, at best, to the infimum over the regular subclass rather than to the true Sobolev infimum. This failure is already visible in Manià’s problem: piecewise-linear approximations and standard finite-difference discretizations drive the discrete integral away from zero as the mesh is refined, matching the theoretical obstruction (Pereira et al., 2010).

A direct remedy is the enhanced finite element method of Schnake–Feng. On a quasi-uniform mesh $Y\subset X$49, one defines the componentwise cut-off

$Y\subset X$50

and minimizes the modified energy

$Y\subset X$51

For Manià’s example, where the gradient growth exponent is $Y\subset X$52, any $Y\subset X$53 works; in numerical experiments, the enhanced method approaches $Y\subset X$54 well, while the standard FEM diverges (Feng et al., 2016).

The one-dimensional theory for Manià’s problem was later placed on a $Y\subset X$55-convergence foundation. Feng–Siktar consider the cut-off functional

$Y\subset X$56

on the piecewise affine space $Y\subset X$57, where

$Y\subset X$58

If

$Y\subset X$59

then $Y\subset X$60 $Y\subset X$61-converges to $Y\subset X$62 with respect to the strong $Y\subset X$63-topology, and the discrete minimizers converge strongly in $Y\subset X$64 to the unique singular minimizer $Y\subset X$65. The proof uses a “layered” interpolant $Y\subset X$66 with strong $Y\subset X$67-stability and approximation properties (Feng et al., 2024).

A different strategy is nonconformity. For convex non-autonomous integrands with non-standard growth, the Crouzeix–Raviart finite element scheme minimizes over a broken affine space $Y\subset X$68 with piecewise gradient $Y\subset X$69 and meshwise quadrature

$Y\subset X$70

Under either quadrature assumptions $Y\subset X$71–$Y\subset X$72 or the simpler condition $Y\subset X$73, one has

$Y\subset X$74

together with strong $Y\subset X$75-convergence of a subsequence and weak convergence of broken gradients. In the piecewise-constant exponent benchmark, conforming FEM converges to $Y\subset X$76, while the Crouzeix–Raviart scheme converges to $Y\subset X$77 (Balci et al., 2021). The numerical message is unambiguous: when the gap is genuine, conformity can be a liability rather than an advantage.

7. Nonlinear elasticity and current directions

The scope of Lavrentiev’s gap now extends beyond scalar and generalized-growth models into three-dimensional nonlinear elasticity. For the neo-Hookean energy

$Y\subset X$78

with $Y\subset X$79 convex, $Y\subset X$80, and $Y\subset X$81, a recent result constructs boundary data for which the infimum over the regular invertible class $Y\subset X$82 is strictly larger than the infimum over its sequential weak $Y\subset X$83-closure $Y\subset X$84. More precisely, for any $Y\subset X$85, there exists $Y\subset X$86 such that

$Y\subset X$87

The mechanism is a dipole-type singularity: a deformation creates self-contact across a sphere $Y\subset X$88, and the inverse deformation develops a jump whose singular gradient has total variation $Y\subset X$89. The gap therefore reflects inverse-regularity failure rather than classical cavitation (Barchiesi et al., 24 Mar 2026).

This development repositions the phenomenon in a broader landscape. In scalar theory, modern results increasingly identify structural “anti-jump” or “balance” conditions under which the gap disappears (Borowski et al., 2024). In nonuniform ellipticity, the sharpness of Hölder, $Y\subset X$90, $Y\subset X$91, and anisotropic balance thresholds has become a central theme (Borowski et al., 2023, Borowski, 8 Sep 2025, Borowski et al., 2022). In geometry and topology, density results for manifold-valued maps coexist with counterexamples based on fractal barriers and differential forms (Antonini et al., 20 Dec 2025, Balci et al., 2023).

A plausible implication is that Lavrentiev’s gap is best understood not as an isolated pathology, but as a precise indicator of mismatch between the natural energy space and the approximation class imposed by analysis, topology, or computation. In one dimension this mismatch can sometimes be neutralized by reparametrization; in double-phase and Musielak–Orlicz problems it is governed by sharp continuity scales; in numerical analysis it dictates whether conforming approximations are intrinsically blind; and in nonlinear elasticity it can encode physically meaningful singular mechanisms that are invisible to regular invertible competitors.