Restricted Coercivity: Theory & Applications

• Restricted coercivity is a framework that restricts coercivity to active subspaces after neutralizing degenerate directions, ensuring meaningful growth conditions.
• It underpins methods in regularized least-squares, Kähler geometry, and magnetic systems by isolating dynamical or variationally significant directions.
• This concept bridges theory and experiment by providing precise criteria for convergence and stability across optimization, operator theory, and physical learning.

Restricted coercivity denotes a family of non-identical but structurally related ideas in which coercivity is not imposed on an entire space without qualification, but only after accounting for directions that are neutralized by a kernel, a symmetry group, a solution manifold, a field projection, or a microstructural reversal pathway. In the supplied literature, this appears explicitly as coercivity of a convex function on a subspace in regularized least-squares, as coercivity of the K-energy modulo automorphisms in complex geometry, and as a local rank-controlled coercivity transverse to a solution manifold in physical learning; in magnetic systems, closely related analyses interpret the measured coercive field as restricted by domain-wall depinning, defect structure, dipolar geometry, or exchange-averaged disorder rather than by a full anisotropy barrier (Xue et al., 28 Jul 2025, Hisamoto, 2016, McGinnis et al., 13 Jun 2026, Zhu et al., 2023). This suggests a common theme: the coercive quantity is meaningful only on the directions that remain dynamically or variationally active after the ambient degeneracies have been factored out.

1. Terminological landscape and formal patterns

The terminology is not uniform across the literature. In convex optimization, restricted coercivity is defined directly: for $f\in\Gamma_0(\mathbb{R}^n)$ and a proper linear subspace $D\subset\mathbb{R}^n$, $f$ is coercive on $D$ if

$\lim_{\|P_Dx\|\to+\infty} f(x)=+\infty.$

In Kähler geometry, ordinary coercivity of the Mabuchi K-energy,

$M(\varphi)\ge \delta J(\varphi)-C,$

is replaced by a reduced form

$$M(\varphi)\ge \delta J_{C(G)}(\varphi)-C,\qquad J_T(\varphi):=\inf_{g\in T}J(\varphi_g).$$

In operator theory, a bounded operator $A$ is coercive if

$$|(A\phi,\phi)_{\mathcal H}| \ge \alpha \|\phi\|_{\mathcal H}^2,$$

while a weaker but decisive condition for Galerkin convergence is that $A$ be coercive plus compact, equivalently $D\subset\mathbb{R}^n$0. In physical learning for linear circuits, the relevant coercivity is local and is equivalent near the solution manifold to full row rank of $D\subset\mathbb{R}^n$1, or, in circuit form, to full column rank of $D\subset\mathbb{R}^n$2 (Xue et al., 28 Jul 2025, Hisamoto, 2016, Chandler-Wilde et al., 2021, McGinnis et al., 13 Jun 2026).

Setting	Restricted object	Representative criterion
Regularized least-squares	$D\subset\mathbb{R}^n$3 on a subspace $D\subset\mathbb{R}^n$4	$D\subset\mathbb{R}^n$5
K-energy with automorphisms	$D\subset\mathbb{R}^n$6 modulo torus action	$D\subset\mathbb{R}^n$7
Physical learning in circuits	Directions transverse to $D\subset\mathbb{R}^n$8	$D\subset\mathbb{R}^n$9
Ferrimagnetic switching	Perpendicular field component	$f$0
Boundary integral operators	Coercivity modulo compact terms	$f$1

These formulations are not interchangeable. A plausible implication is that “restricted coercivity” is best treated as a pattern of analysis rather than as a single cross-disciplinary definition.

2. Restricted switching fields in ferrimagnetic FeTb

In amorphous $f$2 films of thickness $f$3 nm, coercivity is interpreted not as an intrinsic anisotropy field in the Stoner–Wohlfarth sense, but as a restricted switching field set by thermally assisted domain-wall depinning and by the projection of the applied field onto the perpendicular easy axis. Thickness alone tunes the ferrimagnet from Tb-dominated to Fe-dominated regimes, with compensation thickness $f$4 nm at $f$5 K and about $f$6 nm at $f$7 K, and the coercivity strengthens near this crossover region. The low-temperature out-of-plane switching field exceeds $f$8 kOe below $f$9 K for the $D$0 nm film and below $D$1 K for the other three films; a concrete example is the $D$2 nm sample at $D$3 K, where $D$4 kOe (Zhu et al., 2023).

The central angular law is

$D$5

with $D$6 the polar angle of the applied field relative to the film normal. Equivalently,

$D$7

The measured coercivity is therefore restricted to the useful perpendicular field component $D$8. This sharply contrasts with the macrospin coherent-rotation expression

$D$9

and the discrepancy is reinforced by the experimental fact that $\lim_{\|P_Dx\|\to+\infty} f(x)=+\infty.$0 at normal incidence. Square Hall loops, switching fields much smaller than $\lim_{\|P_Dx\|\to+\infty} f(x)=+\infty.$1, and the non-Stoner–Wohlfarth angular scaling together support reversal by nucleation plus wall propagation with a depinning barrier rather than coherent rotation over the full anisotropy barrier.

Temperature dependence supplies the second restriction. The coercivity increases quasi-exponentially upon cooling, but no activation energy or explicit fit is extracted. The established interpretation is that thermal fluctuations assist domain walls in overcoming pinning barriers at higher temperature, whereas cooling suppresses that assistance and raises the field required for depinning. In this sense, the measured coercive field is a thermally assisted depinning field, not an unrestricted anisotropy field.

3. Microstructural and intrinsic limits in magnetic materials

A distinct magnetic use of the idea appears when coercivity is treated as a value systematically limited below an anisotropy-controlled ideal by nonuniform reversal, demagnetizing fields, thermal activation, or defect structure. For ideal $\lim_{\|P_Dx\|\to+\infty} f(x)=+\infty.$2 nm cubes of $\lim_{\|P_Dx\|\to+\infty} f(x)=+\infty.$3, the upper benchmark is

$\lim_{\|P_Dx\|\to+\infty} f(x)=+\infty.$4

but the practical coercive field is written

$\lim_{\|P_Dx\|\to+\infty} f(x)=+\infty.$5

Even without structural defects or soft phases, the coercivity is reduced to $\lim_{\|P_Dx\|\to+\infty} f(x)=+\infty.$6 of the anisotropy field at room temperature and to $\lim_{\|P_Dx\|\to+\infty} f(x)=+\infty.$7 at $\lim_{\|P_Dx\|\to+\infty} f(x)=+\infty.$8 K because of misorientation, self-demagnetizing fields, and thermal fluctuations; for $\lim_{\|P_Dx\|\to+\infty} f(x)=+\infty.$9 at $M(\varphi)\ge \delta J(\varphi)-C,$0 K, the ideal-cube limit is $M(\varphi)\ge \delta J(\varphi)-C,$1 T (Fischbacher et al., 2017).

In ThMn$M(\varphi)\ge \delta J(\varphi)-C,$2-type $M(\varphi)\ge \delta J(\varphi)-C,$3 powders, the limiting mechanism is microstructural. Raising the reduction-diffusion temperature from $M(\varphi)\ge \delta J(\varphi)-C,$4 to $M(\varphi)\ge \delta J(\varphi)-C,$5 increases the coercivity from $M(\varphi)\ge \delta J(\varphi)-C,$6 T to $M(\varphi)\ge \delta J(\varphi)-C,$7 T. High-temperature processing annihilates grain boundaries and reduces twin-boundary density by about $M(\varphi)\ge \delta J(\varphi)-C,$8, and the paper identifies grain boundaries and twin boundaries as weak links for demagnetization in this nucleation-type magnet (Polin et al., 2024). The key point is that the realized coercivity is held below the intrinsic anisotropy potential by internal interfaces that facilitate reversal nucleation.

Au-assisted MBE-grown $M(\varphi)\ge \delta J(\varphi)-C,$9 nanostructures show a complementary situation. The room-temperature hysteresis contains a high-coercivity component $$M(\varphi)\ge \delta J_{C(G)}(\varphi)-C,\qquad J_T(\varphi):=\inf_{g\in T}J(\varphi_g).$$0 T associated primarily with the tetragonal $$M(\varphi)\ge \delta J_{C(G)}(\varphi)-C,\qquad J_T(\varphi):=\inf_{g\in T}J(\varphi_g).$$1 phase, but this remains far below the Stoner–Wohlfarth estimate

$$M(\varphi)\ge \delta J_{C(G)}(\varphi)-C,\qquad J_T(\varphi):=\inf_{g\in T}J(\varphi_g).$$2

which gives an upper bound of about $$M(\varphi)\ge \delta J_{C(G)}(\varphi)-C,\qquad J_T(\varphi):=\inf_{g\in T}J(\varphi_g).$$3 T when literature values $$M(\varphi)\ge \delta J_{C(G)}(\varphi)-C,\qquad J_T(\varphi):=\inf_{g\in T}J(\varphi_g).$$4 and $$M(\varphi)\ge \delta J_{C(G)}(\varphi)-C,\qquad J_T(\varphi):=\inf_{g\in T}J(\varphi_g).$$5 are used for $$M(\varphi)\ge \delta J_{C(G)}(\varphi)-C,\qquad J_T(\varphi):=\inf_{g\in T}J(\varphi_g).$$6 at $$M(\varphi)\ge \delta J_{C(G)}(\varphi)-C,\qquad J_T(\varphi):=\inf_{g\in T}J(\varphi_g).$$7 K. The paper directly attributes the reduction to multiple-domain particles or strain, and it also reports phase coexistence, strain $$M(\varphi)\ge \delta J_{C(G)}(\varphi)-C,\qquad J_T(\varphi):=\inf_{g\in T}J(\varphi_g).$$8, coherent diffracting size $$M(\varphi)\ge \delta J_{C(G)}(\varphi)-C,\qquad J_T(\varphi):=\inf_{g\in T}J(\varphi_g).$$9 nm, and interaction domains spanning $A$0–$A$1 nanoparticles (Jamer et al., 2014).

Alnico provides a shape-anisotropy-dominated variant of the same pattern. Its coercivity is limited by curling rather than coherent rotation, by non-ellipsoidal rod geometry, by magnetostatic rod–rod interactions, and by exchange bridges between rods. The paper contrasts the Stoner–Wohlfarth-like estimate

$A$2

with a curling nucleation field

$A$3

and emphasizes that typical rod diameters in alnico are too large to reach coherent-rotation behavior. The practical design message is that coercivity is controlled by rod diameter, end shape, spacing, arrangement, and connectivity rather than by magnetocrystalline anisotropy alone (Ke et al., 2017).

4. Collective, dipolar, and disorder-controlled coercivity

Restricted coercivity also appears in systems where reversal is governed by collective fields or exchange-averaged disorder. In two-dimensional arrays of cobalt nanoparticles, the single-particle reference coercivity is $A$4, but dipolar coupling in a properly spaced array reduces it to about $A$5–$A$6, i.e. roughly a $A$7 reduction. For sufficiently large arrays, the optimal spacings are $A$8 nm for square, $A$9 nm for hexagonal, and $$|(A\phi,\phi)_{\mathcal H}| \ge \alpha \|\phi\|_{\mathcal H}^2,$$0 nm for triangular lattices. The square array gives the lowest reported saturated coercivity, while the triangular array is most robust against positional disorder, maintaining nearly the same optimum under random displacements up to $$|(A\phi,\phi)_{\mathcal H}| \ge \alpha \|\phi\|_{\mathcal H}^2,$$1 of the lattice spacing (Morales-Meza et al., 2014). Here the coercive field is restricted by a geometry-dependent cooperative dipolar environment rather than by isolated-particle anisotropy.

In the three-dimensional random-anisotropy Heisenberg model, coercivity is parametrically suppressed by exchange averaging. For weak random anisotropy $$|(A\phi,\phi)_{\mathcal H}| \ge \alpha \|\phi\|_{\mathcal H}^2,$$2, the ferromagnetic correlation length scales as

$$|(A\phi,\phi)_{\mathcal H}| \ge \alpha \|\phi\|_{\mathcal H}^2,$$3

and the paper argues that the coercive field is proportional to the effective anisotropy on that scale, giving

$$|(A\phi,\phi)_{\mathcal H}| \ge \alpha \|\phi\|_{\mathcal H}^2,$$4

When anisotropy axes are correlated within grains of size $$|(A\phi,\phi)_{\mathcal H}| \ge \alpha \|\phi\|_{\mathcal H}^2,$$5, the scaling becomes

$$|(A\phi,\phi)_{\mathcal H}| \ge \alpha \|\phi\|_{\mathcal H}^2,$$6

in the regime $$|(A\phi,\phi)_{\mathcal H}| \ge \alpha \|\phi\|_{\mathcal H}^2,$$7. In this regime, coercivity is not of order $$|(A\phi,\phi)_{\mathcal H}| \ge \alpha \|\phi\|_{\mathcal H}^2,$$8; it is strongly restricted by exchange smoothing of random local anisotropy, with metastability and hedgehogs controlling the irreversible reversal pathway (Proctor et al., 2014).

A thin-film counterpart appears in nanocrystalline $$|(A\phi,\phi)_{\mathcal H}| \ge \alpha \|\phi\|_{\mathcal H}^2,$$9 rectangular films. There the extended Random Anisotropy Model predicts a two-dimensional scaling law

$A$0

instead of the bulk $A$1 law, and with shape-induced anisotropy included,

$A$2

The coercive field follows

$A$3

For small grains, exchange averages the random anisotropy strongly, so coercivity approaches a finite baseline set by shape anisotropy; as grain size grows toward the renormalized exchange length of about $A$4–$A$5 nm, coercivity rises and reaches a maximum (Bachleitner-Hofmann et al., 2016).

A related bulk-alloy formulation appears in Fe–Ni. There the coercivity minimum occurs at permalloy $A$6, where the anisotropy constant is not zero: $A$7 The paper attributes this to the growth barrier of a spike-like reverse domain near a defect, controlled jointly by anisotropy, magnetostriction, elastic response, and magnetostatics rather than by anisotropy alone (Balakrishna et al., 2020). This suggests that “restriction” can also arise from the local metastable mode chosen as the physically relevant nucleus.

5. Restricted coercivity in regularized least-squares

The most explicit formal development of restricted coercivity in the supplied corpus occurs in the regularized least-squares problem

$A$8

with $A$9. The paper centers the orthogonal decomposition

$D\subset\mathbb{R}^n$00

so that the least-squares term already controls directions outside $D\subset\mathbb{R}^n$01. Restricted coercivity is then defined by coercivity of $D\subset\mathbb{R}^n$02 on a subspace $D\subset\mathbb{R}^n$03: $D\subset\mathbb{R}^n$04 and the relevant case for regularized least-squares is $D\subset\mathbb{R}^n$05 (Xue et al., 28 Jul 2025).

The main result is that nonempty compactness of the solution set is equivalent to several conditions, including ordinary coercivity of the full objective, coercivity of $D\subset\mathbb{R}^n$06 on $D\subset\mathbb{R}^n$07, positivity of the recession function on nonzero kernel directions,

$D\subset\mathbb{R}^n$08

and the recession-cone condition

$D\subset\mathbb{R}^n$09

Thus restricted coercivity on $D\subset\mathbb{R}^n$10 is neither merely heuristic nor merely sufficient; it is exactly the missing growth condition needed because $D\subset\mathbb{R}^n$11 is flat along $D\subset\mathbb{R}^n$12.

The paper also separates compactness from existence. Restricted coercivity does not characterize existence alone. For

$D\subset\mathbb{R}^n$13

the solution set is

$D\subset\mathbb{R}^n$14

which is nonempty but unbounded. For

$D\subset\mathbb{R}^n$15

with the same $D\subset\mathbb{R}^n$16, the solution set is

$D\subset\mathbb{R}^n$17

which is compact. For

$D\subset\mathbb{R}^n$18

the solution set is empty. The paper’s point is that failure of restricted coercivity may correspond either to unbounded minimizers or to nonexistence, and those cases should not be conflated.

A further contribution is the solution-set formula through the conjugate. If $D\subset\mathbb{R}^n$19 solves the problem and $D\subset\mathbb{R}^n$20, then

$D\subset\mathbb{R}^n$21

Compactness is characterized by

$D\subset\mathbb{R}^n$22

and the paper proves that this coincides with

$D\subset\mathbb{R}^n$23

Restricted coercivity therefore links solution geometry, recession analysis, and conjugate-subdifferential structure.

6. Reduced coercivity in Kähler geometry and coercive-plus-compact operator theory

In the cscK problem, Hisamoto reformulates coercivity in the presence of automorphisms by reducing the $D\subset\mathbb{R}^n$24-functional along torus orbits. For $D\subset\mathbb{R}^n$25,

$D\subset\mathbb{R}^n$26

and $D\subset\mathbb{R}^n$27-coercivity is expressed as

$D\subset\mathbb{R}^n$28

On the non-Archimedean side the reduced $D\subset\mathbb{R}^n$29-functional is

$D\subset\mathbb{R}^n$30

and the key slope formula is

$D\subset\mathbb{R}^n$31

In the toric case, reduced coercivity is equivalent to a reduced uniform stability condition: $D\subset\mathbb{R}^n$32 for every convex rational piecewise-linear $D\subset\mathbb{R}^n$33 on the moment polytope (Hisamoto, 2016). Here the restriction is explicitly quotienting by affine or torus directions.

Operator theory supplies a different but related variant. For the Laplace double-layer operator $D\subset\mathbb{R}^n$34 on $D\subset\mathbb{R}^n$35, the decisive condition for universal Galerkin convergence is that

$D\subset\mathbb{R}^n$36

be coercive plus compact, equivalently

$D\subset\mathbb{R}^n$37

The paper settles negatively long-standing questions by constructing Lipschitz domains and Lipschitz polyhedra for which these operators are not coercive plus compact. It gives examples in $D\subset\mathbb{R}^n$38D and $D\subset\mathbb{R}^n$39D with Lipschitz constant two for which $D\subset\mathbb{R}^n$40 fails this property, and for every $D\subset\mathbb{R}^n$41 it constructs Lipschitz polyhedra for which $D\subset\mathbb{R}^n$42 and $D\subset\mathbb{R}^n$43 is not a compact perturbation of a coercive operator for any $D\subset\mathbb{R}^n$44 (Chandler-Wilde et al., 2021). The paper does not present a positive restricted-coercivity theorem, but it shows that local or compressed pieces of the operator already obstruct any global coercive-plus-compact structure.

7. Local restricted coercivity in physical learning

In physical learning for linear circuits, the state $D\subset\mathbb{R}^n$45 is the input-constrained minimizer of the circuit energy, the residual is

$D\subset\mathbb{R}^n$46

and the solution set

$D\subset\mathbb{R}^n$47

is generically an $D\subset\mathbb{R}^n$48-dimensional manifold because the number of trainable parameters $D\subset\mathbb{R}^n$49 usually exceeds the number of outputs $D\subset\mathbb{R}^n$50. The paper therefore does not seek global strong convexity in parameter space. Instead, it identifies a local coercivity condition near $D\subset\mathbb{R}^n$51, equivalent to full row rank of $D\subset\mathbb{R}^n$52, or, in circuit form, full column rank of

$D\subset\mathbb{R}^n$53

where $D\subset\mathbb{R}^n$54 is the active-incidence matrix built from edges with nonzero free-state voltage drop (McGinnis et al., 13 Jun 2026).

The key identity is

$D\subset\mathbb{R}^n$55

with $D\subset\mathbb{R}^n$56 the diagonal matrix of squared free-state voltage drops on active edges. If $D\subset\mathbb{R}^n$57 has full column rank, then

$D\subset\mathbb{R}^n$58

For Equilibrium Propagation this yields

$D\subset\mathbb{R}^n$59

so the loss decays exponentially. Coupled Learning obeys an analogous inequality with an additional cubic correction,

$D\subset\mathbb{R}^n$60

and near the solution manifold this again implies exponential decay. The paper then proves local convergence: if $D\subset\mathbb{R}^n$61 is sufficiently close to a regular point $D\subset\mathbb{R}^n$62, then $D\subset\mathbb{R}^n$63 converges to the solution manifold and $D\subset\mathbb{R}^n$64 decays exponentially.

The coercivity is restricted in a precise geometric sense. Because $D\subset\mathbb{R}^n$65 is generically a manifold rather than an isolated minimizer, flat tangent directions remain. The paper’s coercivity controls only directions detected by the residual map, i.e. directions transverse to $D\subset\mathbb{R}^n$66. This interpretation is reinforced by the local error bound

$D\subset\mathbb{R}^n$67

A symmetry-based failure example is given by the kite circuit. There the coercivity constant degenerates on part of the solution manifold because free-state and clamped-state voltage drops have disjoint support. For scalar output, the EP dissipation takes the form

$D\subset\mathbb{R}^n$68

so coercivity fails exactly when $D\subset\mathbb{R}^n$69. The paper then proves, via Sard’s theorem, that such degeneracies are non-generic: for almost every desired output $D\subset\mathbb{R}^n$70, the solution manifold is smooth of dimension $D\subset\mathbb{R}^n$71 and the coercivity condition holds at every point of it (McGinnis et al., 13 Jun 2026).