Double Phase Functionals

• Double phase functionals are variational integrals characterized by a switching between p- and q-growth phases via a modulating coefficient, leading to non-uniform ellipticity and diverse analytical challenges.
• The variational framework employs Musielak–Orlicz spaces and sharp capacity estimates to analyze approximation properties, regularity issues, and Lavrentiev phenomena.
• Recent research extends these models to nonlocal, weighted, and obstacle settings, establishing robust regularity, precise gap criteria, and novel approximation scales.

Searching arXiv for recent and foundational papers on double phase functionals to ground the article in the current literature. Searching arXiv for recent and foundational papers on double phase functionals to ground the article in the current literature. Double phase functionals are variational integrals whose energy density switches between two polynomial regimes, typically a $p$-phase and a $q$-phase, through a nonnegative modulating coefficient $a(x)$. The canonical model is

$$u \mapsto \int_{\Omega}\big(|Du|^p+a(x)|Du|^q\big)\,dx, \qquad 1<p<q,$$

and its distinguishing feature is non-uniform ellipticity: the operative growth and ellipticity change according to the local size of $a(x)|Du|^{q-p}$. In current research, double phase functionals are treated within generalized Orlicz and Musielak–Orlicz frameworks, and their theory now spans sharp regularity, Lavrentiev-gap criteria, capacities, topological constraints, weighted estimates, and nonlocal extensions (Baroni et al., 2017, Borowski, 8 Sep 2025, Adamadze et al., 28 Jan 2026).

1. Canonical model and phase-switching mechanism

The prototype double phase energy is

$$\mathcal{H}(u,\Omega)=\int_{\Omega}\bigl(|Du|^p+a(x)|Du|^q\bigr)\,dx,$$

with $a:\Omega\to[0,\infty)$. In this form, the term $|Du|^p$ defines the $p$-phase, while $a(x)|Du|^q$ defines the $q$0-phase. Where $q$1, the functional behaves like $q$2-growth; where $q$3, mixed growth becomes active, and for large gradients the $q$4-term can dominate. This phase-switching viewpoint is standard across the local, nonlocal, vectorial, and weighted theories (Filippis et al., 14 Aug 2025, Fang et al., 2024).

A general structural version replaces $q$5 by a Carathéodory integrand $q$6 satisfying two-sided control by the double phase growth function

$q$7

In quasiconvex vectorial settings, one assumes

$q$8

together with Hessian bounds and weak uniform $q$9-quasiconvexity. In geometric variants, the integrand may depend on $a(x)$0, $a(x)$1, and $a(x)$2 through elliptic metric tensors, while still remaining equivalent to the model double phase density up to constants (Jeong et al., 20 Mar 2026, Filippis et al., 14 Aug 2025).

A central quantitative manifestation of the switching mechanism is the ellipticity ratio. In the classical divergence-form model,

$a(x)$3

the pointwise ellipticity ratio behaves like

$a(x)$4

This is precisely the source of the characteristic analytical difficulty: ellipticity is neither uniform in $a(x)$5 nor homogeneous in $a(x)$6, and it can degenerate or blow up across the transition set $a(x)$7 (Filippis et al., 14 Aug 2025).

2. Variational framework and Euler–Lagrange structures

The natural ambient setting is Musielak–Orlicz analysis. For the modular

$a(x)$8

one works with $a(x)$9, $$u \mapsto \int_{\Omega}\big(|Du|^p+a(x)|Du|^q\big)\,dx, \qquad 1<p<q,$$0, and their local or zero-trace variants. In Kirchhoff-type problems, for example, the space

$$u \mapsto \int_{\Omega}\big(|Du|^p+a(x)|Du|^q\big)\,dx, \qquad 1<p<q,$$1

is equipped with the Luxemburg norm, and the double phase operator

$$u \mapsto \int_{\Omega}\big(|Du|^p+a(x)|Du|^q\big)\,dx, \qquad 1<p<q,$$2

is of type $$u \mapsto \int_{\Omega}\big(|Du|^p+a(x)|Du|^q\big)\,dx, \qquad 1<p<q,$$3 (Fiscella et al., 2020).

For unconstrained scalar minimizers, the Euler–Lagrange equation is the familiar non-uniformly elliptic PDE

$$u \mapsto \int_{\Omega}\big(|Du|^p+a(x)|Du|^q\big)\,dx, \qquad 1<p<q,$$4

In more general quasiconvex settings, the weak formulation takes the form

$$u \mapsto \int_{\Omega}\big(|Du|^p+a(x)|Du|^q\big)\,dx, \qquad 1<p<q,$$5

and the associated Legendre–Hadamard coercivity is quantified by $$u \mapsto \int_{\Omega}\big(|Du|^p+a(x)|Du|^q\big)\,dx, \qquad 1<p<q,$$6 (Jeong et al., 20 Mar 2026).

The theory becomes substantially richer under constraints. In vectorial obstacle problems, the admissible class consists of maps taking values in a target $$u \mapsto \int_{\Omega}\big(|Du|^p+a(x)|Du|^q\big)\,dx, \qquad 1<p<q,$$7, and local minimizers satisfy Euler–Lagrange identities with contact terms supported on $$u \mapsto \int_{\Omega}\big(|Du|^p+a(x)|Du|^q\big)\,dx, \qquad 1<p<q,$$8. In manifold-valued cases, these identities include the nearest-point retraction and the normal $$u \mapsto \int_{\Omega}\big(|Du|^p+a(x)|Du|^q\big)\,dx, \qquad 1<p<q,$$9 to $a(x)|Du|^{q-p}$0; when the target has no boundary, the boundary reaction term vanishes (Filippis et al., 14 Aug 2025).

3. Sharp regularity theory

A large part of the modern theory concerns identifying the exact interaction between the regularity of $a(x)|Du|^{q-p}$1, the gap $a(x)|Du|^{q-p}$2, and the regularity of minimizers. For general non-autonomous functionals controlled by

$a(x)|Du|^{q-p}$3

Baroni–Colombo–Mingione proved gradient Hölder continuity under the sharp alternatives

$a(x)|Du|^{q-p}$4

when $a(x)|Du|^{q-p}$5, together with intrinsic Morrey decay

$a(x)|Du|^{q-p}$6

for every $a(x)|Du|^{q-p}$7 (Baroni et al., 2017). These estimates are intrinsic in the sense that they are expressed directly in the modular $a(x)|Du|^{q-p}$8, rather than in separate $a(x)|Du|^{q-p}$9 and $$\mathcal{H}(u,\Omega)=\int_{\Omega}\bigl(|Du|^p+a(x)|Du|^q\bigr)\,dx,$$0 norms.

The vectorial theory is subtler because full regularity is generally unavailable. For quasiconvex double phase integrals with weak uniform $$\mathcal{H}(u,\Omega)=\int_{\Omega}\bigl(|Du|^p+a(x)|Du|^q\bigr)\,dx,$$1-quasiconvexity, Jeong–Ok established partial Hölder regularity of the gradient: if $$\mathcal{H}(u,\Omega)=\int_{\Omega}\bigl(|Du|^p+a(x)|Du|^q\bigr)\,dx,$$2 minimizes

$$\mathcal{H}(u,\Omega)=\int_{\Omega}\bigl(|Du|^p+a(x)|Du|^q\bigr)\,dx,$$3

under assumptions (C1)–(C7), then there exists an open set $$\mathcal{H}(u,\Omega)=\int_{\Omega}\bigl(|Du|^p+a(x)|Du|^q\bigr)\,dx,$$4 with $$\mathcal{H}(u,\Omega)=\int_{\Omega}\bigl(|Du|^p+a(x)|Du|^q\bigr)\,dx,$$5 such that

$$\mathcal{H}(u,\Omega)=\int_{\Omega}\bigl(|Du|^p+a(x)|Du|^q\bigr)\,dx,$$6

and the singular set is characterized through intrinsic excess quantities built from $$\mathcal{H}(u,\Omega)=\int_{\Omega}\bigl(|Du|^p+a(x)|Du|^q\bigr)\,dx,$$7 and frozen integrands $$\mathcal{H}(u,\Omega)=\int_{\Omega}\bigl(|Du|^p+a(x)|Du|^q\bigr)\,dx,$$8 (Jeong et al., 20 Mar 2026).

Constrained vectorial and manifold-valued problems exhibit the same partial-regularity pattern. In the obstacle setting, under

$$\mathcal{H}(u,\Omega)=\int_{\Omega}\bigl(|Du|^p+a(x)|Du|^q\bigr)\,dx,$$9

and under the structural ellipticity assumptions on the frozen metric tensors, constrained local minimizers enjoy partial gradient regularity: $a:\Omega\to[0,\infty)$0 on an open full-measure subset $a:\Omega\to[0,\infty)$1. The same work proves higher integrability

$a:\Omega\to[0,\infty)$2

and Hausdorff-measure estimates for the singular set, including

$a:\Omega\to[0,\infty)$3

(Filippis et al., 14 Aug 2025).

Variable-exponent analogues preserve the same structural theme. For

$a:\Omega\to[0,\infty)$4

with $a:\Omega\to[0,\infty)$5, $a:\Omega\to[0,\infty)$6, and

$a:\Omega\to[0,\infty)$7

Ragusa–Tachikawa proved

$a:\Omega\to[0,\infty)$8

for local minimizers (Ragusa et al., 2020).

4. Lavrentiev phenomenon, density, and sharp approximation scales

The Lavrentiev phenomenon measures the failure of smooth competitors to approximate the variational infimum over the natural energy space. For the model functional

$a:\Omega\to[0,\infty)$9

one compares

$|Du|^p$0

Absence of the gap means $|Du|^p$1 (Borowski, 8 Sep 2025).

The classical Hölder-scale threshold is now understood as part of a broader picture. For $|Du|^p$2, the optimal range is

$|Du|^p$3

and this is sharp: if

$|Du|^p$4

there exist a bounded Lipschitz domain, a nonnegative $|Du|^p$5, and boundary data $|Du|^p$6 with finite energy such that the Lavrentiev gap occurs. In particular, if $|Du|^p$7, then no additional restrictions on $|Du|^p$8 and $|Du|^p$9 are required (Borowski, 8 Sep 2025).

A key mechanism behind these sharp results is the factorization

$p$0

with $p$1 controlling decay near the $p$2-phase and $p$3 providing weighted control of the $p$4-phase. This decomposes the modulating coefficient into a rate-of-decay component and a Muckenhoupt component, and it yields energy-preserving approximation by mollification (Borowski, 8 Sep 2025).

Earlier work of Bulíček–Gwiazda–Skrzeczkowski showed that, for Hölder weights, it suffices to regularize bounded functions and exploit the $p$5 bound on the function rather than an $p$6 estimate on the gradient. This gives absence of the Lavrentiev phenomenon in the range

$p$7

with sharpness for $p$8 (Bulíček et al., 2021).

A different sharp scale is furnished by the classes $p$9, defined by

$a(x)|Du|^q$0

If $a(x)|Du|^q$1 and

$a(x)|Du|^q$2

then the gap is absent, whereas for

$a(x)|Du|^q$3

one can construct $a(x)|Du|^q$4 and finite-energy boundary data for which the gap is present (Borowski et al., 2023). This corrects a common simplification: Hölder continuity of $a(x)|Du|^q$5 is not the only meaningful scale for approximation theory.

In the vectorial case, approximation requires extra structure. A recent energy-approximation theorem proves convergence of energies under mollification assuming the threshold

$a(x)|Du|^q$6

together with convexity in the gradient, one-sided comparability in $a(x)|Du|^q$7, and a $a(x)|Du|^q$8-independent local spatial minimizer condition (F4) (Carozza et al., 11 Jan 2025).

5. Capacities, topology, and weighted analytic frameworks

Double phase analysis has acquired its own capacity theory. For singular solutions of

$a(x)|Du|^q$9

in $q$00, the intrinsic capacity

$q$01

is better adapted than the classical $q$02-capacity. Under

$q$03

convexity and symmetry of the domain, and $q$04, positive singular solutions are symmetric with respect to the symmetry hyperplane and monotone in the distinguished direction (Biagi et al., 2022).

Boundary regularity and global higher integrability can also be expressed through capacity density conditions. For

$q$05

and under local uniform fatness of $q$06, one obtains global higher integrability for quasiminimizers, integral Hardy inequalities, and equivalence between local $q$07-fatness, infimal $q$08-fatness, boundary Poincaré inequalities, and pointwise Hardy inequalities (Bäuerlein et al., 27 Mar 2025). A particularly important clarification is that the expected Maz’ya-type inequality fails with the raw $q$09-capacity; the correct boundary notion is an infimal capacity equivalent, at the fatness level, to $q$10-capacity (Bäuerlein et al., 27 Mar 2025).

A complementary functional-analytic direction is provided by the generalized Muckenhoupt theory for the double phase density

$q$11

The condition

$q$12

is equivalent to boundedness of the Hardy–Littlewood maximal operator on $q$13 and on $q$14, to generalized Jensen inequalities, and to Sobolev–Poincaré estimates. Under $q$15, weak solutions of the double phase Euler–Lagrange equation are locally Hölder continuous (Adamadze et al., 28 Jan 2026). This shows that regularity can be recovered under assumptions on averaging stability rather than on classical Hölder smoothness of $q$16.

Topological constraints enter when the target is not linear. In vectorial obstacle problems, the admissible target $q$17 is assumed to be $q$18-connected, meaning $q$19 for $q$20. This ensures the existence of finite-energy extensions and Lipschitz retractions needed for comparison arguments, linking the analytic exponent $q$21 directly to the topology of the target (Filippis et al., 14 Aug 2025).

6. Variants, applications, and current directions

The double phase paradigm has diversified well beyond the isotropic local model. In orthotropic problems, one considers

$q$22

where degeneracy is componentwise rather than isotropic. Under

$q$23

local minimizers enjoy higher integrability, and under natural Sobolev regularity of $q$24 they are locally Lipschitz (Almi et al., 24 Jul 2025).

Nonlocal and mixed local/nonlocal versions replace one or both phases by fractional interactions. For the nonlocal double phase functional

$q$25

weak Harnack inequalities, expansion of positivity, local boundedness, and quantitative sup-estimates have been established under the structural relations

$q$26

with tail terms encoding the far-field contribution (Fang et al., 2024). In the mixed setting

$q$27

one obtains local boundedness, Hölder continuity, and Harnack inequalities under near-sharp conditions such as

$q$28

when $q$29 and $q$30 (Byun et al., 2023).

The framework also reaches applied variational models. In image restoration, a BV-type double phase energy

$q$31

combines edge-preserving linear growth with adaptive quadratic smoothing. The associated regularized functionals $q$32-converge to the BV double phase functional, and the latter arises as the relaxation of a Sobolev-only model (Harjulehto et al., 2019).

Other directions include Kirchhoff-type problems on $q$33, where the nonlocal coefficient depends on the total double phase energy and existence and multiplicity follow from mountain-pass and Fountain arguments (Fiscella et al., 2020), and anisotropic Baouendi–Grushin models from transonic flow, where the energy

$q$34

exhibits variable exponent double phase behavior tied to mixed-type geometry (Bahrouni et al., 2019).

Taken together, these developments show that “double phase functionals” no longer denote only the model $q$35, but a larger analytical program. The common thread is the interaction of heterogeneous growth, switching ellipticity, and structure in $q$36: smoothness, decay, topology, capacity, or weighted averaging can each become the decisive mechanism governing regularity, approximation, and qualitative behavior (Baroni et al., 2017, Borowski, 8 Sep 2025, Adamadze et al., 28 Jan 2026).