Introduction and Problem Formulation
Let ( Z t ) t ≥ 0 (Z_t)_{t \ge0} ( Z t ) t ≥ 0 be a planar Brownian motion started at 0 0 0 with generator Δ / 2 \Delta/2 Δ/2 , and write
τ U = inf { t ≥ 0 : Z t ∉ U } . \tau_U = \inf\{t \ge0 : Z_t \notin U\}. τ U = inf { t ≥ 0 : Z t ∈ / U } .
In other words, τ U \tau_U τ U is the exit time of ( Z t ) t ≥ 0 (Z_t)_{t\geq0} ( Z t ) t ≥ 0 from U U U . Now, fix a real centered probability measure μ \mu μ with a finite second moment and look for a planar domain U U U such that
Law ( Re Z τ U ) = μ . \operatorname{Law}(\operatorname{Re} Z_{\tau_U})=\mu. Law ( Re Z τ U ) = μ .
This is the planar Skorokhod embedding problem PSEP, originally investigated by Gross in [5 ] . A solution is called μ \mu μ -domain [17 ] , or a planar embedding of μ \mu μ . In [5 ] , Gross constructed a canonical solution: a simply connected domain, which we denote always by U μ G U_{\mu}^{G} U μ G . Later on, Gross construction was shown to be applicable to any μ \mu μ having some finite p p p -th moment with a p > 1 p>1 p > 1 [20 ] (the endpoint p = 1 p=1 p = 1 was discussed in [21 ] ). The geometric transformation of a domain U ⊆ C U\subseteq\mathbb{C} U ⊆ C defined by
B : U ⟼ U μ U G (1) \mathcal{B}:U\longmapsto U_{\mu_U}^{G} \tag*{(1)} B : U ⟼ U μ U G (1)
where μ U = Law ( Re Z τ U ) \mu_U=\operatorname{Law}(\operatorname{Re} Z_{\tau_U}) μ U = Law ( Re Z τ U ) was introduced in [8 , 2 ] . The map 1.1 is referred to as Brownian symmetrization. This construction preserves a marginal probability distribution determined by harmonic measure at the origin. In [2 ] , the author showed that the Brownian approach provides a genuine extension of the classical Baernstein symmetrization studied by Pruss in [7 ] , i.e. encompassing a broader class of domains, namely unbounded ones. In the bounded simply connected setting, however, the two constructions coincide. The Baernstein symmetrization (attributed by Pruss to Baernstein) rearranges the real boundary values of a conformal map symmetrically and then reconstructs the imaginary part by harmonic conjugation, normalized to vanish at the origin. If f U : D → U f_U:\mathbb{D}\to U f U : D → U is conformal and onto with f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 , then μ U \mu_U μ U is the distribution of Re f U ( e i Θ ) \operatorname{Re} f_U(e^{\mathrm{i}\Theta}) Re f U ( e i Θ ) for a uniform angle Θ \Theta Θ . This identification connects the probabilistic construction with the classical theory of symmetrization. Brownian symmetrization does not increase the area. In other words
Area ( B ( U ) ) ≤ Area ( U ) . \operatorname{Area}(\mathcal{B}(U))\leq\operatorname{Area}(U). Area ( B ( U )) ≤ Area ( U ) .
In particular, Gross embedding has the least area. See [7 , 2 ] for the two different perspectives.
Although the area minimization problem may also be degenerate, its structure differs fundamentally from that of conformal-perimeter minimization. Indeed, Gross’s embedding always minimizes area, whereas it need not minimize conformal perimeter as it will be discussed. Moreover, finite area does not guarantee the existence of any finite-perimeter embedding: there exist centered, compactly supported probability measures for which the area minimization is nondegenerate while the perimeter minimization is degenerate.
Thus, perimeter optimization introduces both a stronger regularity requirement on admissible embeddings and a genuinely nontrivial minimization problem. Our results address these two aspects through existence and degeneracy criteria, explicit counterexamples to Gross minimality, and a convexity-based uniqueness theorem.
Throughout, μ \mu μ is a centered, nondegenerate probability measure on R \mathbb{R} R with a bounded support. In particular, μ \mu μ has finite moments of any order. Denote by C ( μ ) C(\mu) C ( μ ) the class of all proper simply connected domains U ⊆ C U\subseteq\mathbb{C} U ⊆ C such that 0 ∈ U 0\in U 0 ∈ U and
Law ( Re Z τ U ) = μ , \operatorname{Law}(\operatorname{Re} Z_{\tau_U})=\mu, Law ( Re Z τ U ) = μ ,
where, as usual, ( Z t ) t ≥ 0 (Z_t)_{t\geq0} ( Z t ) t ≥ 0 is a planar Brownian motion started at 0. An element of C ( μ ) \mathcal{C}(\mu) C ( μ ) is called admissible embedding of μ \mu μ . For each U ∈ C ( μ ) U\in\mathcal{C}(\mu) U ∈ C ( μ ) , write f U : D → U f_U:\mathbb{D}\to U f U : D → U the onto normalized conformal map
f U ( 0 ) = 0 , f U ′ ( 0 ) > 0. f_U(0)=0,\qquad f_U'(0)>0. f U ( 0 ) = 0 , f U ′ ( 0 ) > 0.
The conformal perimeter of U U U is defined as
Per c ( U ) = 2 π ∥ f U ′ ∥ H 1 , \operatorname{Per}_{c}(U)=2\pi\lVert f_U'\rVert_{H^1}, Per c ( U ) = 2 π ∥ f U ′ ∥ H 1 ,
where ∥ ⋅ ∥ H 1 \lVert\cdot\rVert_{H^1} ∥ ⋅ ∥ H 1 is the 1-Hardy norm, with the value + ∞ +\infty + ∞ when f U ′ ∉ H 1 f_U'\notin H^1 f U ′ ∈ / H 1 . In this manuscript, we consider the minimization of the conformal perimeter over C ( μ ) \mathcal{C}(\mu) C ( μ )
m c ( μ ) : = inf U ∈ C ( μ ) Per c ( U ) . (2) m_c(\mu):=\inf_{U\in\mathcal{C}(\mu)}\operatorname{Per}_{c}(U). \tag*{(2)} m c ( μ ) := U ∈ C ( μ ) inf Per c ( U ) . (2)
We shall call 1.2 the conformal perimeter optimization problem CPOP. If m c ( μ ) m_c(\mu) m c ( μ ) is finite we say that CPOP is nondegenerate, otherwise degenerate. The (increasing) quantile of μ \mu μ , which we denote by Q μ Q_\mu Q μ , is defined by
Q μ ( u ) = inf { x ∈ R ∣ μ ( ( − ∞ , x ] ) ≥ u } , u ∈ ( 0 , 1 ) . Q_\mu(u)=\inf\{x\in\mathbb{R}\mid\mu((-\infty,x])\geq u\},\qquad u\in(0,1). Q μ ( u ) = inf { x ∈ R ∣ μ (( − ∞ , x ]) ≥ u } , u ∈ ( 0 , 1 ) .
A famous statistical feature of Q μ Q_\mu Q μ is that, when fed with uniform values in ( 0 , 1 ) (0,1) ( 0 , 1 ) , it generates values sampling as μ \mu μ . In other words,
Q μ ( Uni ( 0 , 1 ) ) ∼ μ . Q_\mu(\operatorname{Uni}(0,1))\sim\mu. Q μ ( Uni ( 0 , 1 )) ∼ μ .
Set
q μ ( t ) = Q μ ( 1 − ∣ t ∣ π ) , − π < t < π , q_\mu(t)=Q_\mu\left(1-\frac{|t|}{\pi}\right),\qquad-\pi<t<\pi, q μ ( t ) = Q μ ( 1 − π ∣ t ∣ ) , − π < t < π ,
with a periodic extension and continuous endpoint values when available. Its mean is zero. In this convention, the Gross map is
g μ ( z ) = ∑ n ≥ 1 c n z n , c n = 2 π ∫ 0 π q μ ( t ) cos ( n t ) d t . (3) g_\mu(z)=\sum_{n\geq1}c_nz^n,\qquad c_n=\frac{2}{\pi}\int_0^\pi q_\mu(t)\cos(nt)\,\mathrm{d}t. \tag*{(3)} g μ ( z ) = n ≥ 1 ∑ c n z n , c n = π 2 ∫ 0 π q μ ( t ) cos ( n t ) d t . (3)
The original choice of Gross is the increasing datum
Q μ ( ∣ t ∣ π ) , − π < t < π Q_\mu\left(\frac{|t|}{\pi}\right),\qquad-\pi<t<\pi Q μ ( π ∣ t ∣ ) , − π < t < π
The two analytic completion maps differ by precomposition with z ↦ − z z\mapsto-z z ↦ − z and have the same image. Gross’s theorem gives the univalence of g g g , g ( 0 ) = 0 g(0)=0 g ( 0 ) = 0 , and g ( D ) = U μ G g(\mathbb{D})=U_\mu^G g ( D ) = U μ G is the Gross embedding.
To preserve the logical flow of the exposition, proofs containing intermediate results that are used subsequently throughout the paper are presented immediately after their statements. By contrast, purely technical proofs serving only as auxiliary tools are deferred to the appendix. When the symbol † is used next to a result, it means the proof is deferred. Also, when μ \mu μ is clear from the context, it is omitted as a subscript.
The results of this manuscript can be summarized as follows:
Figure 1. Overview of the main results and open problems in conformal-perimeter optimization.
Conformal and geometric perimeter
The notion of perimeter in the framework of simply connected domains, although intuitive, needs a bit of carefulness, since we need to define the notion of boundary length. We distinguish between the geometric perimeter and the conformal perimeter. Let 1 ≤ p < ∞ 1 \le p < \infty 1 ≤ p < ∞ . For an analytic function f : D → C f : \mathbb{D} \to\mathbb{C} f : D → C , the p p p -Hardy norm of f f f is defined by
∥ f ∥ H p = sup 0 < r < 1 ( 1 2 π ∫ 0 2 π ∣ f ( r e i t ) ∣ p d t ) 1 / p . (4) \lVert f\rVert_{H^p}=\sup_{0<r<1}\left(\frac{1}{2\pi}\int_0^{2\pi}\lvert f(re^{it})\rvert^p\,dt\right)^{1/p}. \tag*{(4)} ∥ f ∥ H p = 0 < r < 1 sup ( 2 π 1 ∫ 0 2 π ∣ f ( r e i t ) ∣ p d t ) 1/ p . (4)
That is, the set of f f f with finite p p p -Hardy norm is called p p p -Hardy space, denoted by H p H^p H p . A crucial property of Hardy spaces is that when f ∈ H p f\in H^p f ∈ H p then f f f admits a radial extension f ∗ f^* f ∗ . More precisely,
f ∗ ( θ ) : = lim r ↑ 1 f ( r e i θ ) f^*(\theta):=\lim_{r\uparrow1}f(re^{i\theta}) f ∗ ( θ ) := r ↑ 1 lim f ( r e i θ )
exists a.e. on the unit circle T \mathbb T T . Moreover, f ∗ ∈ L p ( T ) f^*\in L^p(\mathbb T) f ∗ ∈ L p ( T ) . We refer the reader to [4 , 9 ] for detailed treatments of the theory of Hardy spaces.
Conformal perimeter
For an analytic function h h h on D \mathbb D D , write
M 1 ( r , h ) : = 1 2 π ∫ 0 2 π ∣ h ( r e i t ) ∣ d t , 0 < r < 1. M_1(r,h):=\frac{1}{2\pi}\int_0^{2\pi}\lvert h(re^{it})\rvert\,dt,\qquad0<r<1. M 1 ( r , h ) := 2 π 1 ∫ 0 2 π ∣ h ( r e i t )∣ d t , 0 < r < 1.
The quantity r ⋅ M 1 ( r , h ) r\cdot M_1(r,h) r ⋅ M 1 ( r , h ) represents the arc length of the curve
γ r ( t ) = h ( r e i t ) , 0 ≤ t ≤ 2 π . \gamma_r(t)=h(re^{it}),\qquad0\le t\le2\pi. γ r ( t ) = h ( r e i t ) , 0 ≤ t ≤ 2 π .
We write
L ( γ r ) = 2 π r M 1 ( r , h ′ ) . L(\gamma_r)=2\pi rM_1(r,h'). L ( γ r ) = 2 π r M 1 ( r , h ′ ) .
Let f f f a normalized univalent function mapping D \mathbb D D onto U U U . We define the conformal perimeter of U U U by
Per c ( U ) : = 2 π ∥ f ′ ∥ H 1 , (5) \operatorname{Per}_{c}(U):=2\pi\lVert f'\rVert_{H^1}, \tag*{(5)} Per c ( U ) := 2 π ∥ f ′ ∥ H 1 , (5)
with the convention that Per c ( U ) = + ∞ \operatorname{Per}_{c}(U)=+\infty Per c ( U ) = + ∞ when f ′ ∉ H 1 f'\notin H^1 f ′ ∈ / H 1 . This definition is independent of the choice of normalized conformal map. Indeed, any two conformal maps from D \mathbb D D onto U U U differ by precomposition with an automorphism of the unit disk, and the boundary length represented by (2.2) is invariant under such a change of parametrization. Since the membership of f ′ f' f ′ in H 1 H^1 H 1 implies that of f f f , then by the radial extension in Hardy spaces, one has this direct geometric interpretation: Per c ( U ) \operatorname{Per}_{c}(U) Per c ( U ) is the length of the parametrized curve
ξ ⟼ ( f ∗ ( ξ ) (6) \xi\longmapsto(f^*(\xi) \tag*{(6)} ξ ⟼ ( f ∗ ( ξ ) (6)
where f ∗ f^* f ∗ is the radial limit of f f f . Thus the conformal perimeter is the length of the boundary as traversed by the conformal boundary parametrization, with any multiplicity of traversal retained. Note that if Per c ( U ) \operatorname{Per}_{c}(U) Per c ( U ) is finite then f f f extends continuously to D ‾ \overline{\mathbb D} D . However f f f may loose the injectivity in the closed disc.
Geometric perimeter
For a set E ⊂ C E \subset\mathbb{C} E ⊂ C , its one-dimensional Hausdorff measure is defined by
H 1 ( E ) = lim δ ↓ 0 H δ 1 ( E ) , \mathcal{H}^{1}(E) = \lim_{\delta\downarrow0} \mathcal{H}^{1}_{\delta}(E), H 1 ( E ) = δ ↓ 0 lim H δ 1 ( E ) ,
where
H δ 1 ( E ) : = inf { ∑ j = 1 ∞ diam ( E j ) : E ⊂ ⋃ j = 1 ∞ E j , diam ( E j ) ≤ δ } . \mathcal{H}^{1}_{\delta}(E) := \inf\left\{ \sum_{j=1}^{\infty} \operatorname{diam}(E_j) : E \subset\bigcup_{j=1}^{\infty} E_j,\ \operatorname{diam}(E_j) \le\delta\right\}. H δ 1 ( E ) := inf { j = 1 ∑ ∞ diam ( E j ) : E ⊂ j = 1 ⋃ ∞ E j , diam ( E j ) ≤ δ } .
Accordingly, for a simply connected domain U U U we define the geometric perimeter of U U U by
Per g ( U ) : = H 1 ( ∂ U ) (7) \operatorname{Per}_{g}(U) := \mathcal{H}^{1}(\partial U) \tag*{(7)} Per g ( U ) := H 1 ( ∂ U ) (7)
whenever this quantity is finite. Geometrically, H 1 ( ∂ U ) \mathcal{H}^{1}(\partial U) H 1 ( ∂ U ) is the total one-dimensional length of the boundary viewed simply as a subset of the plane. It counts each geometric boundary point only once. So it is the natural generalization of ordinary curve length to boundaries that may have corners, cusps, slits, or more complicated shapes. A crucial distinction is that
Per g ( U ) ≤ Per c ( U ) \operatorname{Per}_{g}(U) \le\operatorname{Per}_{c}(U) Per g ( U ) ≤ Per c ( U )
in general. Consider the slit domain
U = D ∖ [ 1 / 2 , 1 ) . U = \mathbb{D} \setminus[1/2,1). U = D ∖ [ 1/2 , 1 ) .
It is simply connected. We can show that
Per g ( U ) = 2 π + 1 2 < 2 π ∥ f ′ ∥ H 1 = 2 π + 2 ( 1 2 ) . \operatorname{Per}_{g}(U) = 2\pi+ \frac{1}{2} < 2\pi\lVert f' \rVert_{\mathcal{H}^{1}} = 2\pi+ 2\left(\frac{1}{2}\right). Per g ( U ) = 2 π + 2 1 < 2 π ∥ f ′ ∥ H 1 = 2 π + 2 ( 2 1 ) .
The strict inequality comes from the fact that Hausdorff measure counts the slit only once, whereas the conformal boundary parametrization sees its two sides separately.
Relation between the two notions
For a rectifiable simple closed curve, H 1 \mathcal{H}^{1} H 1 agrees with its ordinary arc length. Hence, if U U U is a rectifiable Jordan domain, Per g ( U ) \operatorname{Per}_{g}(U) Per g ( U ) is the usual Euclidean perimeter. The distinction between (2.2) and (2.4) is relevant when the boundary is not Jordan. The conformal boundary parametrization may traverse the same geometric boundary piece more than once. In that case the conformal perimeter records the multiplicity, whereas Hausdorff measure counts each geometric point only once. The following implications summarize the situation.
Proposition 2.1. [15 ] Let U ⊊ C U \subsetneq\mathbb{C} U ⊊ C be simply connected and let f : D → U f : \mathbb{D} \to U f : D → U be conformal with range U U U . We have
Per g ( U ) ≤ Per c ( U ) ≤ 2 Per g ( U ) . (8) \operatorname{Per}_{g}(U) \le\operatorname{Per}_{c}(U) \le2\operatorname{Per}_{g}(U). \tag*{(8)} Per g ( U ) ≤ Per c ( U ) ≤ 2 Per g ( U ) . (8)
In particular, both geometric and conformal perimeters share the same finiteness status. If, in addition, U U U is a Jordan domain, then
f ′ ∈ H 1 ⟺ ∂ U is rectifiable , f' \in H^1 \Longleftrightarrow\partial U \text{ is rectifiable}, f ′ ∈ H 1 ⟺ ∂ U is rectifiable ,
and in this case
Per g ( U ) = Per c ( U ) = 2 π ∥ f ′ ∥ H 1 . (9) \operatorname{Per}_{g}(U)=\operatorname{Per}_{c}(U)=2\pi\lVert f'\rVert_{H^1}. \tag*{(9)} Per g ( U ) = Per c ( U ) = 2 π ∥ f ′ ∥ H 1 . (9)
The estimate 2.5 is referred to as Bishop’s inequality. The equivalence between the rectifiability of ∂ U \partial U ∂ U and the membership of f ′ f' f ′ in H 1 H^1 H 1 is standard. The factor 2 in Bishop’s inequality reflects the fact that the conformal boundary parametrization traverses almost every boundary point at most twice. Higher multiplicities can occur only on an exceptional countable set and therefore do not contribute to boundary length. Note that a simply connected domain obtained by removing a slit may have a conformal map with derivative in H 1 H^1 H 1 , while the boundary is not Jordan. The two sides of the slit correspond to distinct boundary approaches and are traversed separately by the conformal boundary parametrization. Consequently the slit is counted twice by Per c \operatorname{Per}_{c} Per c but only once by Per g \operatorname{Per}_{g} Per g .
Pruss’s containment theorem
For a bounded open set E ⊂ C E \subset\mathbb{C} E ⊂ C , define the length of its vertical section at x ∈ R x \in\mathbb{R} x ∈ R by
ℓ E ( x ) = L 1 ( { y ∈ R : x + i y ∈ E } ) , \ell_E(x)=\mathcal{L}^{1}\left(\{y\in\mathbb{R}:x+iy\in E\}\right), ℓ E ( x ) = L 1 ( { y ∈ R : x + i y ∈ E } ) ,
where L 1 \mathcal{L}^{1} L 1 denotes one-dimensional Lebesgue measure. Its Steiner symmetrization about the real axis is
S ( E ) = { x + i y : ℓ E ( x ) > 0 , ∣ y ∣ < ℓ E ( x ) 2 } . \mathcal{S}(E)=\left\{x+iy:\ell_E(x)>0,\ |y|<\frac{\ell_E(x)}{2}\right\}. S ( E ) = { x + i y : ℓ E ( x ) > 0 , ∣ y ∣ < 2 ℓ E ( x ) } .
Thus Steiner symmetrization centers each vertical section while preserving its length. Steiner symmetrization is known to preserve the area while not boosting the perimeter [19 ] . In [7 ] , Pruss established the following containment theorem, which will play a role to identify when Gross embedding is the minimizer of our perimeter optimization problem.
Theorem 2.2. Let U U U be a bounded simply connected domain. Then
B ( U ) ⊂ S ( U ) . \mathcal{B}(U)\subset\mathcal{S}(U). B ( U ) ⊂ S ( U ) .
Note that the original statement of Pruss’s containment theorem 2.2 is within a Hardy-space framework, which is automatic here because U U U is bounded. For a compact convex body K K K , we use the corresponding closed symmetral
S ( K ) = { x + i y : x ∈ proj R K , ∣ y ∣ ≤ ℓ K ( x ) 2 } . \mathcal{S}(K)=\left\{x+iy:x\in\operatorname{proj}_{\mathbb{R}}K,\ |y|\le\frac{\ell_K(x)}{2}\right\}. S ( K ) = { x + i y : x ∈ proj R K , ∣ y ∣ ≤ 2 ℓ K ( x ) } .
This is again a compact convex body. Indeed, the upper endpoint of a vertical section of K K K is a concave function of x x x , and the lower endpoint is convex, so their difference ℓ K \ell_K ℓ K is concave.
We emphasize that 2.2 does not, by itself, imply a perimeter inequality: inclusion does not order the perimeters of arbitrary nonconvex domains.
Existence of a conformal-perimeter minimizer
In this section, we establish our main existence theorem for the conformal-perimeter minimization problem. We then investigate necessary and sufficient conditions for the infimum to be finite, identifying both geometric and measure-theoretic criteria.
Existence theorem
Theorem 3.1 (Attainment of the conformal-perimeter infimum). Suppose that there exists U 0 ∈ C ( μ ) U_0 \in\mathcal{C}(\mu) U 0 ∈ C ( μ ) such that
Per c ( U 0 ) < ∞ . \operatorname{Per}_{\mathrm{c}}(U_0) < \infty. Per c ( U 0 ) < ∞. Then the conformal-perimeter infimum is finite and attained. More precisely, there exists U ∗ ∈ C ( μ ) U_\ast\in\mathcal{C}(\mu) U ∗ ∈ C ( μ ) satisfying
Per c ( U ∗ ) = min U ∈ C ( μ ) Per c ( U ) < ∞ . (10) \operatorname{Per}_{\mathrm{c}}(U_\ast) = \min_{U \in\mathcal{C}(\mu)} \operatorname{Per}_{\mathrm{c}}(U) < \infty. \tag*{(10)} Per c ( U ∗ ) = U ∈ C ( μ ) min Per c ( U ) < ∞. (10) The minimizing domain U ∗ U_\ast U ∗ is bounded and its normalized conformal map satisfies f ∗ ′ ∈ H 1 f_\ast' \in\mathrm{H}^1 f ∗ ′ ∈ H 1 . No Jordan-boundary assumption is imposed or concluded.
Proof. We divide the argument into four steps.
Step 1. Uniform coefficient bounds. By hypothesis,
m c ( μ ) ≤ Per c ( U 0 ) < ∞ . m_{\mathrm{c}}(\mu) \le\operatorname{Per}_{\mathrm{c}}(U_0) < \infty. m c ( μ ) ≤ Per c ( U 0 ) < ∞.
Choose a minimizing sequence ( U j ) (U_j) ( U j ) and let
f j ( z ) = ∑ n = 1 ∞ c j , n z n . f_j(z) = \sum_{n=1}^{\infty} c_{j,n} z^n. f j ( z ) = n = 1 ∑ ∞ c j , n z n .
be the corresponding normalized conformal maps. After discarding finitely many terms, there exists C < ∞ C < \infty C < ∞ such that
∥ f j ′ ∥ H 1 ≤ C ( j ≥ 1 ) . (11) \lVert f_j' \rVert_{\mathrm{H}^1} \le C \qquad(j \ge1). \tag*{(11)} ∥ f j ′ ∥ H 1 ≤ C ( j ≥ 1 ) . (11)
For every 0 < r < 1 0 < r < 1 0 < r < 1 , Cauchy’s coefficient formula applied to f j ′ f_j' f j ′ gives
n c j , n r n − 1 = 1 2 π ∫ 0 2 π f j ′ ( r e i t ) e − i ( n − 1 ) t d t . n c_{j,n} r^{n-1} = \frac{1}{2\pi}\int_0^{2\pi} f_j'(r e^{it}) e^{-i(n-1)t}\,dt. n c j , n r n − 1 = 2 π 1 ∫ 0 2 π f j ′ ( r e i t ) e − i ( n − 1 ) t d t .
Therefore
n ∣ c j , n ∣ r n − 1 ≤ 1 2 π ∫ 0 2 π ∣ f j ′ ( r e i t ) ∣ d t ≤ C . n |c_{j,n}| r^{n-1} \le\frac{1}{2\pi}\int_0^{2\pi} |f_j'(r e^{it})|\,dt \le C. n ∣ c j , n ∣ r n − 1 ≤ 2 π 1 ∫ 0 2 π ∣ f j ′ ( r e i t ) ∣ d t ≤ C .
Letting r ↑ 1 r \uparrow1 r ↑ 1 , we obtain
∣ c j , n ∣ ≤ C n , j , n ≥ 1. (12) |c_{j,n}| \le\frac{C}{n}, \qquad j,n \ge1. \tag*{(12)} ∣ c j , n ∣ ≤ n C , j , n ≥ 1. (12)
In particular,
∑ n > N ∣ c j , n ∣ 2 ≤ C 2 ∑ n > N 1 n 2 ≤ C 2 N . (13) \sum_{n>N} |c_{j,n}|^2 \le C^2 \sum_{n>N} \frac{1}{n^2} \le\frac{C^2}{N}. \tag*{(13)} n > N ∑ ∣ c j , n ∣ 2 ≤ C 2 n > N ∑ n 2 1 ≤ N C 2 . (13)
Thus the Taylor coefficients have uniformly small H 2 H^2 H 2 tails.
Step 2. Strong H 2 H^2 H 2 compactness. For every fixed n n n , the sequence ( c j , n ) j (c_{j,n})_j ( c j , n ) j is bounded. By a diagonal argument, we may extract a subsequence such that
c j , n ⟶ c n for every n ≥ 1. c_{j,n} \longrightarrow c_n \qquad\text{for every } n \ge1. c j , n ⟶ c n for every n ≥ 1.
Set
f ( z ) = ∑ n = 1 ∞ c n z n . f(z)=\sum_{n=1}^{\infty}c_nz^n. f ( z ) = n = 1 ∑ ∞ c n z n .
The estimate ∣ c n ∣ ≤ C / n |c_n|\le C/n ∣ c n ∣ ≤ C / n shows that this series defines an analytic function in D \mathbb{D} D . Moreover, convergence of individual coefficients, together with (3.3), gives
f j ⟶ f locally uniformly in D . f_j \longrightarrow f \qquad\text{locally uniformly in }\mathbb{D}. f j ⟶ f locally uniformly in D .
We claim that the convergence is in fact strong in H 2 H^2 H 2 . Indeed,
∣ c j , n − c n ∣ 2 ≤ 4 C 2 n 2 , |c_{j,n}-c_n|^2\le\frac{4C^2}{n^2}, ∣ c j , n − c n ∣ 2 ≤ n 2 4 C 2 ,
and the right-hand side is summable. Dominated convergence for series therefore gives
∥ f j − f ∥ H 2 2 = ∑ n = 1 ∞ ∣ c j , n − c n ∣ 2 ⟶ 0. (14) \lVert f_j-f\rVert_{H^2}^2=\sum_{n=1}^{\infty}|c_{j,n}-c_n|^2\longrightarrow0. \tag*{(14)} ∥ f j − f ∥ H 2 2 = n = 1 ∑ ∞ ∣ c j , n − c n ∣ 2 ⟶ 0. (14)
Step 3. Univalence and preservation of the exit law. By the classical Hardy-space boundary regularity theorem, the finite-perimeter competitor U 0 U_0 U 0 is bounded. Since it realizes μ \mu μ , the latter has bounded support. In particular,
v : = ∫ R x 2 μ ( d x ) v:=\int_{\mathbb{R}}x^2\,\mu(dx) v := ∫ R x 2 μ ( d x )
is finite and strictly positive. Let m m m denote normalized angular measure on T \mathbb{T} T , and put
u j = Re f j ∗ . u_j=\operatorname{Re} f_j^*. u j = Re f j ∗ .
Conformal invariance of harmonic measure gives
( u j ) ∗ m = μ for every j . (u_j)_*m=\mu\qquad\text{for every }j. ( u j ) ∗ m = μ for every j .
Since f j ( 0 ) = 0 f_j(0)=0 f j ( 0 ) = 0 , Parseval’s identity yields
∥ f j ∥ H 2 2 = 2 ∥ u j ∥ L 2 ( T , m ) 2 = 2 v . \lVert f_j\rVert_{H^2}^2=2\lVert u_j\rVert_{L^2(\mathbb{T},m)}^2=2v. ∥ f j ∥ H 2 2 = 2 ∥ u j ∥ L 2 ( T , m ) 2 = 2 v .
By (3.5),
∥ f ∥ H 2 2 = 2 v > 0. \lVert f\rVert_{H^2}^2=2v>0. ∥ f ∥ H 2 2 = 2 v > 0.
Thus f f f is nonconstant. The locally uniform limit of univalent analytic functions is either univalent or constant. Consequently, f f f is univalent. Furthermore,
f ( 0 ) = 0 , f ′ ( 0 ) = lim j → ∞ f j ′ ( 0 ) > 0 , f(0)=0,\qquad f'(0)=\lim_{j\to\infty}f_j'(0)>0, f ( 0 ) = 0 , f ′ ( 0 ) = j → ∞ lim f j ′ ( 0 ) > 0 ,
where strict positivity follows from univalence of the limit. Next, strong H 2 H^2 H 2 convergence implies
Re f j ∗ → Re f ∗ in L 2 ( T , m ) . \operatorname{Re} f_j^* \to\operatorname{Re} f^* \quad\text{in } L^2(\mathbb{T},m). Re f j ∗ → Re f ∗ in L 2 ( T , m ) .
In particular, these boundary functions converge in probability. Since all of them have the same law μ \mu μ , their limit has that law as well:
( Re f ∗ ) ∗ m = μ . (15) (\operatorname{Re} f^*)_*m=\mu. \tag*{(15)} ( Re f ∗ ) ∗ m = μ . (15)
Step 4. Lower semicontinuity and attainment. Local uniform convergence of the analytic maps implies
f j ′ → f ′ locally uniformly in D . f_j' \to f' \quad\text{locally uniformly in } \mathbb{D}. f j ′ → f ′ locally uniformly in D .
For every fixed 0 < r < 1 0<r<1 0 < r < 1 ,
1 2 π ∫ 0 2 π ∣ f ′ ( r e i t ) ∣ d t = lim j → ∞ 1 2 π ∫ 0 2 π ∣ f j ′ ( r e i t ) ∣ d t . \frac{1}{2\pi}\int_0^{2\pi}\lvert f'(re^{it})\rvert\,dt
=
\lim_{j\to\infty}\frac{1}{2\pi}\int_0^{2\pi}\lvert f_j'(re^{it})\rvert\,dt. 2 π 1 ∫ 0 2 π ∣ f ′ ( r e i t )∣ d t = j → ∞ lim 2 π 1 ∫ 0 2 π ∣ f j ′ ( r e i t )∣ d t .
Hence
1 2 π ∫ 0 2 π ∣ f ′ ( r e i t ) ∣ d t ≤ lim inf j → ∞ ∥ f j ′ ∥ H 1 . \frac{1}{2\pi}\int_0^{2\pi}\lvert f'(re^{it})\rvert\,dt
\le
\liminf_{j\to\infty}\lVert f_j'\rVert_{H^1}. 2 π 1 ∫ 0 2 π ∣ f ′ ( r e i t )∣ d t ≤ j → ∞ lim inf ∥ f j ′ ∥ H 1 .
Taking the supremum over r r r gives
∥ f ′ ∥ H 1 ≤ lim inf j → ∞ ∥ f j ′ ∥ H 1 . (16) \lVert f'\rVert_{H^1}
\le
\liminf_{j\to\infty}\lVert f_j'\rVert_{H^1}. \tag*{(16)} ∥ f ′ ∥ H 1 ≤ j → ∞ lim inf ∥ f j ′ ∥ H 1 . (16)
In particular, f ′ ∈ H 1 f'\in H^1 f ′ ∈ H 1 . Again, f f f extends continuously to D ‾ \overline{\mathbb{D}} D and its image
U ∗ : = f ( D ) U_*:=f(\mathbb{D}) U ∗ := f ( D )
is bounded. Conformal invariance of Brownian motion, together with (3.6), shows that
Law ( Re Z τ U ∗ ) = μ . \operatorname{Law}(\operatorname{Re} Z_{\tau_{U_*}})=\mu. Law ( Re Z τ U ∗ ) = μ .
Thus U ∗ ∈ C ( μ ) U_*\in\mathcal{C}(\mu) U ∗ ∈ C ( μ ) . Finally, (3.7) gives
Per c ( U ∗ ) = 2 π ∥ f ′ ∥ H 1 ≤ lim inf j → ∞ 2 π ∥ f j ′ ∥ H 1 = m c ( μ ) . \operatorname{Per}_{\mathrm{c}}(U_*)=2\pi\lVert f'\rVert_{H^1}
\le\liminf_{j\to\infty}2\pi\lVert f_j'\rVert_{H^1}
= m_{\mathrm{c}}(\mu). Per c ( U ∗ ) = 2 π ∥ f ′ ∥ H 1 ≤ j → ∞ lim inf 2 π ∥ f j ′ ∥ H 1 = m c ( μ ) .
The reverse inequality follows from the definition of m c ( μ ) m_{\mathrm{c}}(\mu) m c ( μ ) , since U ∗ U_* U ∗ is admissible. Therefore
Per c ( U ∗ ) = m c ( μ ) , \operatorname{Per}_{\mathrm{c}}(U_*)=m_{\mathrm{c}}(\mu), Per c ( U ∗ ) = m c ( μ ) ,
which proves (3.1).
Corollary 3.2. One has
∃ U ∈ C ( μ ) : Per c ( U ) < ∞ ⟺ m c ( μ ) < ∞ \exists U\in\mathcal{C}(\mu):\operatorname{Per}_{\mathrm{c}}(U)<\infty
\Longleftrightarrow
m_{\mathrm{c}}(\mu)<\infty ∃ U ∈ C ( μ ) : Per c ( U ) < ∞ ⟺ m c ( μ ) < ∞ and
min U ∈ C ( μ ) Per c ( U ) = inf U ∈ C ( μ ) Per c ( U ) . \min_{U\in\mathcal{C}(\mu)}\operatorname{Per}_{\mathrm{c}}(U)
=
\inf_{U\in\mathcal{C}(\mu)}\operatorname{Per}_{\mathrm{c}}(U). U ∈ C ( μ ) min Per c ( U ) = U ∈ C ( μ ) inf Per c ( U ) .
Theorem 3.1 does not require convexity, symmetry, or a prescribed number of boundary branches. In particular, it applies to the full class of planar embeddings of μ \mu μ . The existence theorem 3.1 requires the existence of at least a competitor. So far, only two embeddings are available: Gross and Maher-Greg embeddings. As Maher-Greg embedding is always unbounded [17 , 2 ] , then it can never have a finite conformal perimeter (it will be shown). Among these two canonical constructions, only the Gross embedding can provide a finite-perimeter competitor for the CPOP. This does not exclude finite-perimeter embeddings obtained by other constructions.
Corollary 3.3. If
Per c ( U μ G ) < ∞ \operatorname{Per}_{c}(U_{\mu}^{G}) < \infty Per c ( U μ G ) < ∞ then there exists an embedding U ∗ U_{*} U ∗ of μ \mu μ such that
Per c ( U ∗ ) = min U ∈ C ( μ ) Per c ( U ) . \operatorname{Per}_{c}(U_{*}) = \min_{U \in\mathcal{C}(\mu)} \operatorname{Per}_{c}(U). Per c ( U ∗ ) = U ∈ C ( μ ) min Per c ( U ) .
Remark 3.4 (Scope of the existence result). The minimizing domain need not be the Gross domain as it will be shown in the sequel. Moreover, the theorem concerns conformal perimeter, which counts boundary traversal with multiplicity. It does not establish that the minimizing domain has a Jordan boundary, nor does it automatically give attainment of the geometric-perimeter infimum within the restricted class of rectifiable Jordan domains.
Necessary conditions
In this section, we gather some conditions that must hold if the CPOP is nondegenerate. Hence, these conditions serve as a primary test to check before starting looking for the minimizer of the CPOP. Recall that in section 1, we assumed that μ \mu μ has bounded support. It is not a technical assumption, rather a necessity to avoid the degeneracy of the CPOP. A distribution with unbounded support generates always embeddings with infinite perimeters. This is based on the following lemma.
Lemma 3.5. [16 ] Let U U U be a simply connected domain. Then
2 diam ( U ) ≤ Per c ( U ) . (17) 2\operatorname{diam}(U) \le\operatorname{Per}_{c}(U). \tag*{(17)} 2 diam ( U ) ≤ Per c ( U ) . (17) In particular, if U U U has a finite conformal perimeter then it is bounded.
Lemma 3.5 implies that if our CPOP is nondegenerate then μ \mu μ has a bounded support. Moreover, from 3.8, we obtain
m c ( μ ) ≥ 2 diam ( supp μ ) . (18) m_{c}(\mu) \ge2\operatorname{diam}(\operatorname{supp}\mu). \tag*{(18)} m c ( μ ) ≥ 2 diam ( supp μ ) . (18)
The support bound in 3.9 is sharp as it is asymptotically attained by increasingly thin rectangular domains. We reiterate that μ \mu μ is always with a bounded support.
Remark 3.6. Note that the bounded support of μ \mu μ does not prevent from obtaining unbounded embeddings.
The classical isoperimetric inequality asserts that
4 π Area ( U ) ≤ Per g ( U ) 2 4\pi\operatorname{Area}(U) \le\operatorname{Per}_{g}(U)^2 4 π Area ( U ) ≤ Per g ( U ) 2
for every planar set of finite area and finite geometric perimeter. In our conformal setting, no boundedness assumption is needed. Indeed, finite conformal perimeter implies that the normalized conformal map has derivative in H 1 H^1 H 1 , and hence extends continuously to the closed unit disk. In particular, the domain is bounded and has finite area. Consequently,
4 max ( diam ( U ) 2 , π Area ( U ) ) ≤ Per c ( U ) 2 . (19) 4\max\left(\operatorname{diam}(U)^2,\pi\operatorname{Area}(U)\right) \le\operatorname{Per}_{c}(U)^2. \tag*{(19)} 4 max ( diam ( U ) 2 , π Area ( U ) ) ≤ Per c ( U ) 2 . (19)
Note that diam ( U ) 2 \operatorname{diam}(U)^2 diam ( U ) 2 and π Area ( U ) \pi\operatorname{Area}(U) π Area ( U ) are uncorrelated.
Proposition 3.7. If μ \mu μ has a genuine gap ( a , b ) (a,b) ( a , b ) , i.e. μ ( ( a , b ) ) = 0 \mu((a,b))=0 μ (( a , b )) = 0 for some a < b a<b a < b , with positive mass on both sides of this interval, then any planar embedding of μ \mu μ contains the vertical strip ( a , b ) + i R (a,b)+\mathrm{i}\mathbb{R} ( a , b ) + i R .
Proof. Let S = ( a , b ) + i R S=(a,b)+\mathrm{i}\mathbb{R} S = ( a , b ) + i R . Suppose, for a contradiction, that there exists ξ ∈ ∂ U \xi\in\partial U ξ ∈ ∂ U such that
a < Re ξ < b . a < \operatorname{Re}\xi< b. a < Re ξ < b .
Choose ε > 0 \varepsilon>0 ε > 0 sufficiently small that Ball ( ξ , ε ) ⊂ S \operatorname{Ball}(\xi,\varepsilon)\subset S Ball ( ξ , ε ) ⊂ S . By the classical full-support property of harmonic measure, for which Markowsky gave a simple probabilistic proof [23 ] , followed by the strong maximum principle, we have
P 0 ( Z τ U ∈ Ball ( ξ , ε ) ) > 0. \mathbb{P}_0\left(Z_{\tau_U}\in\operatorname{Ball}(\xi,\varepsilon)\right)>0. P 0 ( Z τ U ∈ Ball ( ξ , ε ) ) > 0.
Consequently,
μ ( ( a , b ) ) = P 0 ( a < Re Z τ U < b ) ≥ P 0 ( Z τ U ∈ Ball ( ξ , ε ) ) > 0 , \mu((a,b))=\mathbb{P}_0\left(a<\operatorname{Re}Z_{\tau_U}<b\right)
\ge\mathbb{P}_0\left(Z_{\tau_U}\in\operatorname{Ball}(\xi,\varepsilon)\right)>0, μ (( a , b )) = P 0 ( a < Re Z τ U < b ) ≥ P 0 ( Z τ U ∈ Ball ( ξ , ε ) ) > 0 ,
contradicting the assumption μ ( ( a , b ) ) = 0 \mu((a,b))=0 μ (( a , b )) = 0 . Hence
∂ U ∩ S = ∅ . \partial U\cap S=\varnothing. ∂ U ∩ S = ∅ .
Since S S S is connected and does not meet ∂ U \partial U ∂ U , we have either S ⊂ U S\subset U S ⊂ U or S ∩ U = ∅ S\cap U=\varnothing S ∩ U = ∅ . The latter is impossible: by connectedness, it would force U U U to lie entirely on one side of the strip, contradicting the assumption that μ \mu μ assigns positive mass to both ( − ∞ , a ] (-\infty,a] ( − ∞ , a ] and [ b , ∞ ) [b,\infty) [ b , ∞ ) . Therefore,
( a , b ) + i R ⊂ U . (a,b)+\mathrm{i}\mathbb{R}\subset U. ( a , b ) + i R ⊂ U .
Corollary 3.8. If μ \mu μ has a genuine gap ( a , b ) (a,b) ( a , b ) then the CPOP is degenerate.
Theorem 3.9 (Unbounded Gross domains). If the Gross embedding U μ G U_{\mu}^{G} U μ G is unbounded then
Per c ( U ) = + ∞ for every U ∈ C ( μ ) . (20) \operatorname{Per}_{c}(U)=+\infty\qquad\text{for every } U\in\mathcal{C}(\mu). \tag*{(20)} Per c ( U ) = + ∞ for every U ∈ C ( μ ) . (20) In particular, the CPOP is degenerate. Equivalently,
∃ U ∈ C ( μ ) : Per c ( U ) < ∞ ⟹ U μ G is bounded. \exists U\in\mathcal{C}(\mu):\operatorname{Per}_{c}(U)<\infty\quad\Longrightarrow\quad U_{\mu}^{G}\text{ is bounded.} ∃ U ∈ C ( μ ) : Per c ( U ) < ∞ ⟹ U μ G is bounded. Proof. Suppose that there exists an embedding U ∈ C ( μ ) U\in\mathcal{C}(\mu) U ∈ C ( μ ) such that Per c ( U ) < ∞ \operatorname{Per}_{c}(U)<\infty Per c ( U ) < ∞ . In particular, U U U is bounded by the isoperimetric inequality 3.10. By Pruss’s containment theorem 2.2,
B ( U ) = U μ G ⊂ S ( U ) , \mathcal{B}(U)=U_{\mu}^{G}\subset\mathcal{S}(U), B ( U ) = U μ G ⊂ S ( U ) , Since U U U is bounded, its Steiner symmetrization S ( U ) \mathcal{S}(U) S ( U ) is also bounded. Hence U μ G U_{\mu}^{G} U μ G is bounded as well, contrary to our assumption. Therefore every embedding of μ \mu μ is unbounded. Again by the isoperimetric inequality, we get
Per c ( U ) = + ∞ for every U ∈ C ( μ ) . \operatorname{Per}_{c}(U)=+\infty\qquad\text{for every } U\in\mathcal{C}(\mu). Per c ( U ) = + ∞ for every U ∈ C ( μ ) .
Comment. The preceding result carried by theorem 3.9 reveals a distinctive role of the Gross embedding in the perimeter optimization problem. Its unboundedness is not merely a geometric feature of a particular construction, but an intrinsic obstruction imposed by the prescribed exit law. This property is specific to the canonical role of the Gross embedding: the unboundedness of an arbitrary admissible domain carries no such implication. For example, Maher-Greg embedding is always unbounded, however it does not mean that the underlying CPOP is degenerate.
One might think that Gross embedding would be bounded if we pick a distribution μ \mu μ with no gaps, along with absolute continuity. Surprisingly, this does not prevent the unboundedness of Gross embedding. In their study of the boundedness of Gross domains, Boudabra, Kroumi, and Smii [12 ] established a characterization of the boundedness of Gross embedding when the underlying distribution is continuous and compactly supported. In particular, the following distribution μ \mu μ with density
ρ ( x ) = ( 2 ∫ 0 1 e − 1 / t d t ) − 1 exp ( − 1 ∣ x ∣ ) 1 ( − 1 , 1 ) ( x ) \rho(x)=\left(2\int_{0}^{1}e^{-1/t}\,dt\right)^{-1}\exp\left(-\frac{1}{\lvert x\rvert}\right)\mathbf{1}_{(-1,1)}(x) ρ ( x ) = ( 2 ∫ 0 1 e − 1/ t d t ) − 1 exp ( − ∣ x ∣ 1 ) 1 ( − 1 , 1 ) ( x )
generates an unbounded Gross domain. Hence, the associated CPOP is degenerate. Thus even for an absolutely continuous, compactly supported law with no gaps, the conformal-perimeter optimization problem may have no finite-valued solution. This example demonstrates that the finite-perimeter competitor hypothesis in Theorem 3.1 is a genuine restriction and cannot be replaced merely by regularity of the distribution or connectedness of its support.
We now turn to measure-theoretic obstructions to the existence of finite-perimeter embeddings. While boundedness of the Gross domain provides a necessary geometric condition, the regularity of the prescribed exit law imposes further restrictions. We first show that purely singular distributions cannot admit finite-perimeter embeddings. We then establish a stronger necessary condition: the absolutely continuous component of the exit law must have a strictly positive density almost everywhere on its support, although a singular component may still be present.
Two measures μ \mu μ and ν \nu ν are mutually singular, denoted by μ ⊥ ν \mu\perp\nu μ ⊥ ν , if there exists a measurable set E E E such that
μ ( E ) = 0 and ν ( E c ) = 0. \mu(E) = 0 \qquad\text{and}\qquad\nu(E^{\mathrm{c}}) = 0. μ ( E ) = 0 and ν ( E c ) = 0.
In other words, the two measures are concentrated on disjoint measurable sets. The notation ν ≪ μ \nu\ll\mu ν ≪ μ means that ν \nu ν is absolutely continuous with respect to μ \mu μ , namely,
μ ( E ) = 0 ⟹ ν ( E ) = 0 \mu(E) = 0 \quad\Longrightarrow\quad\nu(E) = 0 μ ( E ) = 0 ⟹ ν ( E ) = 0
for every measurable set E E E . In other words, every μ \mu μ -null set is also ν \nu ν -null. Recall that, by the Lebesgue decomposition theorem (see [24 ] ), every Borel probability measure μ \mu μ on R \mathbb{R} R admits a unique decomposition
μ ( d x ) = ρ ( x ) d x + μ s ( d x ) . \mu(\mathrm{d}x) = \rho(x)\,\mathrm{d}x + \mu_s(\mathrm{d}x). μ ( d x ) = ρ ( x ) d x + μ s ( d x ) .
where ρ ≥ 0 \rho\ge0 ρ ≥ 0 is integrable and μ s \mu_s μ s is singular with respect to Lebesgue measure.
Theorem 3.10. Let μ \mu μ be singular with respect to Lebesgue measure . Then U ∈ C ( μ ) U \in\mathcal{C}(\mu) U ∈ C ( μ ) satisfies
Per c ( U ) = + ∞ . \operatorname{Per}_{\mathrm{c}}(U) = +\infty. Per c ( U ) = + ∞. Consequently, the CPOP is degenerate
Proof. Suppose, by contradiction, that there exists U ∈ C ( μ ) U \in\mathcal{C}(\mu) U ∈ C ( μ ) with finite conformal perimeter. In particular, we have f U ′ ∈ H 1 f'_U \in H^1 f U ′ ∈ H 1 and hence f U f_U f U extends continuously to D ‾ \overline{\mathbb{D}} D , and its boundary trace is absolutely continuous. In particular,
u ( θ ) = Re f U ( e i θ ) u(\theta) = \operatorname{Re} f_U(e^{i\theta}) u ( θ ) = Re f U ( e i θ )
is a nonconstant absolutely continuous function on [ 0 , 2 π ] [0,2\pi] [ 0 , 2 π ] . Since μ \mu μ is singular, there exists a Borel set N ⊂ supp ( μ ) N \subset\operatorname{supp}(\mu) N ⊂ supp ( μ ) such that
L 1 ( N ) = 0 , μ ( N ) = 1. \mathcal{L}^1(N) = 0,\qquad\mu(N) = 1. L 1 ( N ) = 0 , μ ( N ) = 1.
Because μ \mu μ is the distribution of u ( Θ ) u(\Theta) u ( Θ ) , where Θ \Theta Θ is uniformly distributed on [ 0 , 2 π ] [0,2\pi] [ 0 , 2 π ] , the set
E = { θ : u ( θ ) ∉ N } E = \{\theta: u(\theta) \notin N\} E = { θ : u ( θ ) ∈ / N }
has Lebesgue measure zero. Absolute continuity implies Lusin’s N N N -property: the image under u u u of every Lebesgue-null set has Lebesgue measure zero (see [18 ] ). so
L 1 ( u ( E ) ) = 0. \mathcal{L}^1(u(E)) = 0. L 1 ( u ( E )) = 0.
Moreover,
u ( [ 0 , 2 π ] ) ⊂ N ∪ u ( E ) . u([0,2\pi]) \subset N \cup u(E). u ([ 0 , 2 π ]) ⊂ N ∪ u ( E ) .
It follows that u ( [ 0 , 2 π ] ) u([0,2\pi]) u ([ 0 , 2 π ]) has Lebesgue measure zero. This is impossible, since u u u is continuous and nonconstant, and therefore its image is a nondegenerate interval. Thus no embedding of μ \mu μ can have finite geometric perimeter. The equivalence of perimeter finiteness gives the corresponding conclusion for conformal perimeter. □
Example. Let ( ξ n ) n ≥ 1 (\xi_n)_{n\ge1} ( ξ n ) n ≥ 1 be independent Bernoulli random variables with
P ( ξ n = 1 ) = 1 4 , P ( ξ n = 0 ) = 3 4 , \mathbb{P}(\xi_n=1)=\frac{1}{4},\qquad\mathbb{P}(\xi_n=0)=\frac{3}{4}, P ( ξ n = 1 ) = 4 1 , P ( ξ n = 0 ) = 4 3 , and define
X = ∑ n = 1 ∞ ξ n 2 n , μ = Law ( X − 1 4 ) . X=\sum_{n=1}^{\infty}\frac{\xi_n}{2^n},\qquad\mu=\operatorname{Law}\left(X-\frac{1}{4}\right). X = n = 1 ∑ ∞ 2 n ξ n , μ = Law ( X − 4 1 ) . X X X is well defined and E ( X ) = 1 / 4 \mathbb{E}(X)=1/4 E ( X ) = 1/4 . The measure μ \mu μ is then centered. Moreover, μ \mu μ is nonatomic, singular continuous, and has support [ − 1 / 4 , 3 / 4 ] [-1/4,3/4] [ − 1/4 , 3/4 ] . Indeed, every dyadic interval has positive probability, so the support of X X X is [ 0 , 1 ] [0,1] [ 0 , 1 ] . On the other hand, by the strong law of large numbers, almost every realization of X X X has binary digit frequency 1 / 4 1/4 1/4 , whereas Lebesgue-almost every point of [ 0 , 1 ] [0,1] [ 0 , 1 ] has binary digit frequency 1 / 2 1/2 1/2 . Hence μ \mu μ is singular with respect to Lebesgue measure. Hence the underlying CPOP is degenerate by theorem 3.10. Note that we can show that the Gross embedding is nevertheless bounded.
Figure 2. Gross embedding
The next result reveals a necessary measure-theoretic condition for the existence of finite-perimeter embeddings: the prescribed exit law must dominate Lebesgue measure on its support, although it need not itself be absolutely continuous.
Theorem 3.11. If the CPOP is nondegenerate then
L supp ( μ ) 1 ≪ μ . \mathcal{L}^{1}_{\operatorname{supp}(\mu)} \ll\mu. L supp ( μ ) 1 ≪ μ .
Equivalently, if
μ ( d x ) = ρ ( x ) d x + μ s ( d x ) \mu(dx) = \rho(x)\,dx + \mu_s(dx) μ ( d x ) = ρ ( x ) d x + μ s ( d x )
is the Lebesgue decomposition of μ \mu μ , then
ρ ( x ) > 0 for Lebesgue-almost every x ∈ supp ( μ ) . \rho(x) > 0 \quad\text{for Lebesgue-almost every } x \in\operatorname{supp}(\mu). ρ ( x ) > 0 for Lebesgue-almost every x ∈ supp ( μ ) .
Proof. Denote by S S S the supp ( μ ) \operatorname{supp}(\mu) supp ( μ ) . Since m c ( μ ) < ∞ m_c(\mu) < \infty m c ( μ ) < ∞ , there exists U ∈ C ( μ ) U \in C(\mu) U ∈ C ( μ ) with finite conformal perimeter. Let f U : D → U f_U : \mathbb{D} \to U f U : D → U be its normalized conformal map. By definition, f U ′ ∈ H 1 f'_U \in H^1 f U ′ ∈ H 1 . It follows from the classical boundary regularity theorem for Hardy spaces that f f f extends continuously to D ‾ \overline{\mathbb{D}} D and that its boundary trace
γ ( t ) = f U ( e i t ) , 0 ≤ t ≤ 2 π , \gamma(t) = f_U(e^{it}), \qquad0 \le t \le2\pi, γ ( t ) = f U ( e i t ) , 0 ≤ t ≤ 2 π ,
is absolutely continuous. In particular,
u ( t ) = Re γ ( t ) u(t) = \operatorname{Re}\gamma(t) u ( t ) = Re γ ( t )
is an absolutely continuous real-valued function. By conformal invariance of harmonic measure,
μ = u ∗ ( d t 2 π ) . \mu= u_*\left(\frac{dt}{2\pi}\right). μ = u ∗ ( 2 π d t ) .
Since u u u is continuous, the support of this pushforward measure coincides with its image. Hence
S = u ( [ 0 , 2 π ] ) , S = u([0,2\pi]), S = u ([ 0 , 2 π ]) ,
which is a compact interval. Now let E ⊆ S E \subseteq S E ⊆ S be a Borel set such that μ ( E ) = 0 \mu(E) = 0 μ ( E ) = 0 . By the preceding identity,
L 1 ( u − 1 ( E ) ) = 2 π μ ( E ) = 0. \mathcal{L}^1(u^{-1}(E)) = 2\pi\mu(E) = 0. L 1 ( u − 1 ( E )) = 2 π μ ( E ) = 0.
Since absolutely continuous functions map Lebesgue-null sets to Lebesgue-null sets,
L 1 ( u ( u − 1 ( E ) ) ) = 0. \mathcal{L}^1(u(u^{-1}(E))) = 0. L 1 ( u ( u − 1 ( E ))) = 0.
But E ⊆ S = u ( [ 0 , 2 π ] ) E \subseteq S = u([0,2\pi]) E ⊆ S = u ([ 0 , 2 π ]) , so
u ( u − 1 ( E ) ) = E . u(u^{-1}(E)) = E. u ( u − 1 ( E )) = E .
Consequently,
L 1 ( E ) = 0 , \mathcal{L}^1(E) = 0, L 1 ( E ) = 0 ,
which proves
L 1 ∣ S ≪ μ . \mathcal{L}^1|_S \ll\mu. L 1 ∣ S ≪ μ .
Finally, write μ = ρ L 1 + μ s \mu= \rho\mathcal{L}^1 + \mu_s μ = ρ L 1 + μ s , where μ s \mu_s μ s is supported on a Lebesgue-null set N N N . If ρ \rho ρ vanished on a subset E ⊆ S E \subseteq S E ⊆ S of positive Lebesgue measure, then
L 1 ( E ∖ N ) > 0 , μ ( E ∖ N ) = 0 , \mathcal{L}^1(E \setminus N) > 0, \qquad\mu(E \setminus N) = 0, L 1 ( E ∖ N ) > 0 , μ ( E ∖ N ) = 0 ,
contradicting L 1 ∣ S ≪ μ \mathcal{L}^1|_S \ll\mu L 1 ∣ S ≪ μ . Therefore ρ > 0 \rho> 0 ρ > 0 almost everywhere on S S S .
Theorem 3.11 immediately implies a Hausdorff-dimension result. Recall that the Hausdorff dimension of μ \mu μ (a probability measure in general) is defined by
dim H μ = inf { dim H ( E ) : μ ( E ) = 1 } . \dim_{\mathcal{H}} \mu= \inf\{\dim_{\mathcal{H}}(E) : \mu(E) = 1\}. dim H μ = inf { dim H ( E ) : μ ( E ) = 1 } .
Corollary 3.13. If m c ( μ ) < ∞ m_c(\mu) < \infty m c ( μ ) < ∞ , then
dim H μ = 1. \dim_{\mathcal{H}} \mu= 1. dim H μ = 1. In particular, dim H μ < 1 ⟹ m c ( μ ) = + ∞ \dim_{\mathcal{H}} \mu< 1 \implies m_c(\mu) = +\infty dim H μ < 1 ⟹ m c ( μ ) = + ∞ .
Proof. By Theorem 3.11, S = supp μ S = \operatorname{supp} \mu S = supp μ is a nondegenerate compact interval and L 1 ∣ S ≪ μ \mathcal{L}^{1}|_S \ll\mu L 1 ∣ S ≪ μ . Thus every Borel set E E E with μ ( E ) = 1 \mu(E) = 1 μ ( E ) = 1 satisfies
H 1 ( E ∩ S ) = L 1 ( S ) > 0. \mathcal{H}^{1}(E \cap S) = \mathcal{L}^{1}(S) > 0. H 1 ( E ∩ S ) = L 1 ( S ) > 0.
Hence dim H ( E ) ≥ 1 \dim_{\mathcal{H}}(E) \ge1 dim H ( E ) ≥ 1 . Since μ \mu μ is supported on R \mathbb{R} R , the conclusion follows.
Sufficient conditions for existence of a perimeter minimizer
In [8 ] , the Skorokhod energy of U ∈ C ( μ ) U \in\mathcal{C}(\mu) U ∈ C ( μ ) is defined by
Λ ( U ) = 1 4 ∥ f U ∥ H 2 2 , \Lambda(U) = \frac{1}{4}\lVert f_U\rVert_{H^2}^{2}, Λ ( U ) = 4 1 ∥ f U ∥ H 2 2 ,
and it is shown that
Λ ( U μ G ) ≤ Λ ( U ) . \Lambda(U_\mu^G) \le\Lambda(U). Λ ( U μ G ) ≤ Λ ( U ) .
The quantity Λ ( U μ G ) \Lambda(U_\mu^G) Λ ( U μ G ) is referred to as the Skorokhod energy of μ \mu μ and denoted by Λ μ \Lambda_\mu Λ μ . When μ \mu μ has a density ρ \rho ρ and no gaps, then Λ μ \Lambda_\mu Λ μ admits the integral formula
Λ μ = 1 2 π 2 ∫ supp ( μ ) d x ρ ( x ) . \Lambda_\mu= \frac{1}{2\pi^2}\int_{\operatorname{supp}(\mu)} \frac{dx}{\rho(x)}. Λ μ = 2 π 2 1 ∫ supp ( μ ) ρ ( x ) d x .
Proposition 3.14 (Bounds for the optimal conformal perimeter). We have
2 π 2 Var ( μ ) ≤ m c ( μ ) ≤ 4 π Λ μ . (21) 2\pi\sqrt{2\operatorname{Var}(\mu)} \le m_c(\mu) \le4\pi\sqrt{\Lambda_\mu}. \tag*{(21)} 2 π 2 Var ( μ ) ≤ m c ( μ ) ≤ 4 π Λ μ . (21) Moreover, both bounds are sharp. In particular, if Λ μ \Lambda_\mu Λ μ is finite then 1.2 \mathord{1.2} 1.2 is nondegenerate.
Proof. Let U ∈ E ( μ ) U \in\mathcal{E}(\mu) U ∈ E ( μ ) of finite perimeter , and write
f U ( z ) = ∑ n = 1 ∞ c n z n . f_U(z)=\sum_{n=1}^{\infty}c_nz^n. f U ( z ) = n = 1 ∑ ∞ c n z n .
In particular, U U U is bounded by 3.5. Hence E ( τ U ) < ∞ \mathbb{E}(\tau_U)<\infty E ( τ U ) < ∞ , and so f U ∈ H 2 f_U\in H^2 f U ∈ H 2 by Burkholder equivalence theorem [3 ] . In this case
E ( τ U ) = 1 2 ∑ n = 1 ∞ ∣ c n ∣ 2 = Var ( μ ) . \mathbb{E}(\tau_U)=\frac{1}{2}\sum_{n=1}^{\infty}|c_n|^2=\operatorname{Var}(\mu). E ( τ U ) = 2 1 n = 1 ∑ ∞ ∣ c n ∣ 2 = Var ( μ ) .
See [2 ] . On the other hand, the area formula for univalent functions gives
Area ( U ) = π ∑ n = 1 ∞ n ∣ c n ∣ 2 ≥ 2 π Var ( μ ) . \operatorname{Area}(U)=\pi\sum_{n=1}^{\infty}n|c_n|^2\geq2\pi\operatorname{Var}(\mu). Area ( U ) = π n = 1 ∑ ∞ n ∣ c n ∣ 2 ≥ 2 π Var ( μ ) .
Combining the isoperimetric inequality with the domination of geometric perimeter by conformal perimeter yields
Per c ( U ) ≥ 2 π Area ( U ) \operatorname{Per}_{c}(U)\geq2\sqrt{\pi\operatorname{Area}(U)} Per c ( U ) ≥ 2 π Area ( U )
≥ 2 π 2 Var ( μ ) . \geq2\pi\sqrt{2\operatorname{Var}(\mu)}. ≥ 2 π 2 Var ( μ ) .
Taking the infimum over all admissible domains proves the lower bound. For the upper bound, Baunyakovsky–Cauchy–Schwarz inequality applied to Gross map g g g yields
m c ( μ ) ≤ Per c ( U μ G ) = 2 π ∥ g ′ ∥ H 1 \begin{aligned}
m_c(\mu)\leq\operatorname{Per}_{c}(U_\mu^G) \\
&=2\pi\|g'\|_{H^1}
\end{aligned} m c ( μ ) ≤ Per c ( U μ G ) = 2 π ∥ g ′ ∥ H 1
≤ 2 π ∥ g ′ ∥ H 2 = 4 π Λ μ . \begin{aligned}
\leq2\pi\|g'\|_{H^2} \\
&=4\pi\sqrt{\Lambda_\mu}.
\end{aligned} ≤ 2 π ∥ g ′ ∥ H 2 = 4 π Λ μ .
The sharpness of the bounds follow for example from the centered arcsine distribution on [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] , whose law is
d μ a s ( x ) = 1 ( − 1 , 1 ) ( x ) π 1 − x 2 d x . d\mu_{\mathrm{as}}(x)=\frac{\mathbf{1}_{(-1,1)}(x)}{\pi\sqrt{1-x^2}}\,dx. d μ as ( x ) = π 1 − x 2 1 ( − 1 , 1 ) ( x ) d x .
where in such a case we have
2 π 2 Var ( μ a s ) = m c ( μ a s ) = 2 π = 4 π Λ μ a s . 2\pi\sqrt{2\operatorname{Var}(\mu_{\mathrm{as}})}=m_c(\mu_{\mathrm{as}})=2\pi=4\pi\sqrt{\Lambda_{\mu_{\mathrm{as}}}}. 2 π 2 Var ( μ as ) = m c ( μ as ) = 2 π = 4 π Λ μ as .
If Λ μ \Lambda_\mu Λ μ is finite then Gross embedding would have a finite conformal perimeter, hence the existence theorem 3.1 applies to μ \mu μ . This completes the proof.
Corollary 3.15. If μ \mu μ has a density ρ \rho ρ bounded below by a positive threshold then 1.2 is nondegenerate.
We now establish a refined sufficient condition for the existence of finite-conformal-perimeter embeddings. Based on logarithmic integrability, this condition is strictly weaker than the one arising from finite Skorokhod energy, thereby extending the class of distributions for which the perimeter optimization problem admits a finite minimizer.
Proposition 3.16. Suppose that μ \mu μ has no gaps and a density ρ > 0 \rho> 0 ρ > 0 . If
∫ log ( 1 + 1 ρ ( x ) ) d x < ∞ , \int\log\left(1+\frac{1}{\rho(x)}\right)\,dx < \infty, ∫ log ( 1 + ρ ( x ) 1 ) d x < ∞ , then the Gross embedding has finite conformal perimeter. In particular, m c ( μ ) < ∞ m_c(\mu) < \infty m c ( μ ) < ∞ , and the conformal-perimeter infimum is attained.
Proof. Let Q Q Q be the quantile function of μ \mu μ . Since ρ > 0 \rho> 0 ρ > 0 almost everywhere, Q Q Q is absolutely continuous and satisfies
Q ′ ( t ) = 1 ρ ( Q ( t ) ) Q'(t)=\frac{1}{\rho(Q(t))} Q ′ ( t ) = ρ ( Q ( t )) 1
for almost every t ∈ ( 0 , 1 ) t \in(0,1) t ∈ ( 0 , 1 ) .
The symmetric decreasing boundary datum defining the Gross conformal map is
q ( θ ) = Q ( 1 − ∣ θ ∣ π ) , − π ≤ θ ≤ π . q(\theta)=Q\left(1-\frac{|\theta|}{\pi}\right), \qquad-\pi\leq\theta\leq\pi. q ( θ ) = Q ( 1 − π ∣ θ ∣ ) , − π ≤ θ ≤ π .
In particular, q q q is absolutely continuous and
∣ q ′ ( θ ) ∣ = 1 π Q ′ ( 1 − ∣ θ ∣ π ) |q'(\theta)|=\frac{1}{\pi}Q'\left(1-\frac{|\theta|}{\pi}\right) ∣ q ′ ( θ ) ∣ = π 1 Q ′ ( 1 − π ∣ θ ∣ )
for almost every θ \theta θ . A change of variables gives
∫ − π π ∣ q ′ ( θ ) ∣ log ( 1 + ∣ q ′ ( θ ) ∣ ) d θ = 2 ∫ 0 1 Q ′ ( t ) log ( 1 + Q ′ ( t ) π ) d t = 2 ∫ a b log ( 1 + 1 π ρ ( x ) ) d x < ∞ . \begin{aligned}
\int_{-\pi}^{\pi}|q'(\theta)|\log(1+|q'(\theta)|)\,d\theta
&=2\int_0^1 Q'(t)\log\left(1+\frac{Q'(t)}{\pi}\right)\,dt\\
&=2\int_a^b\log\left(1+\frac{1}{\pi\rho(x)}\right)\,dx\\
&<\infty.
\end{aligned} ∫ − π π ∣ q ′ ( θ ) ∣ log ( 1 + ∣ q ′ ( θ ) ∣ ) d θ = 2 ∫ 0 1 Q ′ ( t ) log ( 1 + π Q ′ ( t ) ) d t = 2 ∫ a b log ( 1 + π ρ ( x ) 1 ) d x < ∞.
Thus q ′ ∈ L log L ( T ) q' \in L\log L(\mathbb{T}) q ′ ∈ L log L ( T ) . By the classical Zygmund inequality for the periodic Hilbert transform H \mathcal{H} H [25 ] ,
q ′ ∈ L log L ( T ) ⟹ H q ′ ∈ L 1 ( T ) . q' \in L\log L(\mathbb{T}) \quad\Longrightarrow\quad\mathcal{H}q' \in L^1(\mathbb{T}). q ′ ∈ L log L ( T ) ⟹ H q ′ ∈ L 1 ( T ) .
The Gross conformal map has boundary values
g ∗ = q + i H q . g^*=q+i\mathcal{H}q. g ∗ = q + i H q .
Since the Hilbert transform commutes with distributional differentiation,
d d θ g ∗ = q ′ + i H q ′ ∈ L 1 ( T ) . \frac{d}{d\theta}g^*=q'+i\mathcal{H}q'\in L^1(\mathbb{T}). d θ d g ∗ = q ′ + i H q ′ ∈ L 1 ( T ) .
Writing P r P_r P r for the Poisson kernel, we obtain
i r e i θ g ′ ( r e i θ ) = P r ∗ ( q ′ + i H q ′ ) ( θ ) . ire^{i\theta}g'(re^{i\theta})=P_r*(q'+i\mathcal{H}q')(\theta). i r e i θ g ′ ( r e i θ ) = P r ∗ ( q ′ + i H q ′ ) ( θ ) .
Consequently,
r ∫ − π π ∣ g ′ ( r e i θ ) ∣ d θ ≤ ∫ − π π ∣ g ′ + i H g ′ ∣ d θ < ∞ . r\int_{-\pi}^{\pi} |g'(re^{i\theta})|\,d\theta\le\int_{-\pi}^{\pi} |g'+i\mathcal{H}g'|\,d\theta< \infty. r ∫ − π π ∣ g ′ ( r e i θ ) ∣ d θ ≤ ∫ − π π ∣ g ′ + i H g ′ ∣ d θ < ∞.
Letting r ↑ 1 r \uparrow1 r ↑ 1 and using the monotonicity of Hardy integral means yields
Therefore,
Per c ( U μ G ) = 2 π ∥ g ′ ∥ H 1 < ∞ . \operatorname{Per}_{c}(U_{\mu}^{G})=2\pi\|g'\|_{H^1}<\infty. Per c ( U μ G ) = 2 π ∥ g ′ ∥ H 1 < ∞.
Since U μ G ∈ C ( μ ) U_{\mu}^{G}\in\mathcal{C}(\mu) U μ G ∈ C ( μ ) , we conclude that m c ( μ ) < ∞ m_c(\mu)<\infty m c ( μ ) < ∞ . Attainment follows from the conformal-perimeter existence theorem 3.1.
The logarithmic condition accommodates a remarkably broad class of distributions, including densities with strong degeneracies. For instance, the centered probability densities
ρ α ( x ) = α + 1 2 ∣ x ∣ α 1 ( − 1 , 1 ) ( x ) , α > 0 , \rho_{\alpha}(x)=\frac{\alpha+1}{2}|x|^{\alpha}\mathbf{1}_{(-1,1)}(x),\qquad\alpha>0, ρ α ( x ) = 2 α + 1 ∣ x ∣ α 1 ( − 1 , 1 ) ( x ) , α > 0 ,
satisfy the condition for every α > 0 \alpha>0 α > 0 , although the corresponding Skorokhod-energy integrability condition fails whenever α ≥ 1 \alpha\ge1 α ≥ 1 . More strikingly, consider
ρ α ( x ) = c α exp ( − 1 ∣ x ∣ α ) 1 ( − 1 , 1 ) ( x ) , 0 < α < 1 , \rho_{\alpha}(x)=c_{\alpha}\exp\left(-\frac{1}{|x|^{\alpha}}\right)\mathbf{1}_{(-1,1)}(x),\qquad0<\alpha<1, ρ α ( x ) = c α exp ( − ∣ x ∣ α 1 ) 1 ( − 1 , 1 ) ( x ) , 0 < α < 1 ,
where c α c_{\alpha} c α is the normalizing constant and ρ α ( 0 ) = 0 \rho_{\alpha}(0)=0 ρ α ( 0 ) = 0 . Since
log ( 1 + 1 ρ α ( x ) ) ∼ 1 ∣ x ∣ α ( x → 0 ) , \log\left(1+\frac{1}{\rho_{\alpha}(x)}\right)\sim\frac{1}{|x|^{\alpha}}\qquad(x\to0), log ( 1 + ρ α ( x ) 1 ) ∼ ∣ x ∣ α 1 ( x → 0 ) ,
the logarithmic condition is satisfied. Nevertheless,
∫ − 1 1 ρ α ( x ) − ε d x = + ∞ for every ε > 0. \int_{-1}^{1}\rho_{\alpha}(x)^{-\varepsilon}\,dx=+\infty\qquad\text{for every }\varepsilon>0. ∫ − 1 1 ρ α ( x ) − ε d x = + ∞ for every ε > 0.
Thus, the logarithmic criterion guarantees the existence of a finite-conformal-perimeter minimizer even for distributions whose densities vanish faster than any polynomial and for which every positive inverse-power integrability condition fails.
A family of distributions where Gross embedding is not optimal
Let U U U be an embedding in C ( μ ) \mathcal{C}(\mu) C ( μ ) . For s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) the fractional boundary energy E s ( U ) \mathcal{E}_s(U) E s ( U ) of U U U is defined by
E s ( U ) = ∑ n = 1 + ∞ n 2 s ∣ c n ∣ 2 \mathcal{E}_s(U)=\sum_{n=1}^{+\infty}n^{2s}|c_n|^2 E s ( U ) = n = 1 ∑ + ∞ n 2 s ∣ c n ∣ 2
where the c n c_n c n ’s are the coefficients of f U f_U f U . In [2 , 7 ] , It is proved that Gross embedding represents the unique minimizer of E s \mathcal{E}_s E s . In particular, Gross embedding has the least area which corresponds to π E 1 1 / 2 \pi\mathcal{E}_1^{1/2} π E 1 1/2 . This further reinforced our earlier conjecture that the Gross embedding would also minimize perimeter. However, we were unable to show that. Recently, we thought to investigate the other scenario: what if Gross embedding does not the minimal perimeter? So we starting by selecting random univalent functions f f f and compute the perimeter of their ranges f ( D ) f(\mathbb{D}) f ( D ) , and then compare it with their Brownian symmetrizations. Surprisingly, some choices show that
Per c ( f ( D ) ) < Per c ( B ( f ( D ) ) ) ! \operatorname{Per}_{c}(f(\mathbb{D})) < \operatorname{Per}_{c}(\mathcal{B}(f(\mathbb{D})))! Per c ( f ( D )) < Per c ( B ( f ( D )))!
Repeating the computations yielded the same comparison. We succeeded to turn the observations into a rigorous work. The idea is to use a perturbation approach. Before constructing the "counterexamples", we begin by some technical tools. The classical Hardy point-evaluation inequality is
∣ h ( w ) ∣ ≤ ∥ h ∥ H 1 1 − ∣ w ∣ 2 , h ∈ H 1 , w ∈ D . (22) |h(w)| \le\frac{\|h\|_{\mathrm{H}^{1}}}{1-|w|^{2}}, \qquad h \in\mathrm{H}^{1}, \quad w \in\mathbb{D}. \tag*{(22)} ∣ h ( w ) ∣ ≤ 1 − ∣ w ∣ 2 ∥ h ∥ H 1 , h ∈ H 1 , w ∈ D . (22)
Recall its elementary mechanism. For H ( z ) = ∑ n ≥ 0 a n z n ∈ H 2 H(z)=\sum_{n\geq0}a_nz^n \in\mathrm{H}^{2} H ( z ) = ∑ n ≥ 0 a n z n ∈ H 2 , Parseval and Cauchy–Schwarz give
∣ H ( w ) ∣ 2 ≤ ( ∑ n ≥ 0 ∣ a n ∣ 2 ) ( ∑ n ≥ 0 ∣ w ∣ 2 n ) = ∥ H ∥ H 2 2 1 − ∣ w ∣ 2 . |H(w)|^{2} \le\left(\sum_{n\geq0}|a_n|^{2}\right)\left(\sum_{n\geq0}|w|^{2n}\right)=\frac{\|H\|_{\mathrm{H}^{2}}^{2}}{1-|w|^{2}}. ∣ H ( w ) ∣ 2 ≤ ( n ≥ 0 ∑ ∣ a n ∣ 2 ) ( n ≥ 0 ∑ ∣ w ∣ 2 n ) = 1 − ∣ w ∣ 2 ∥ H ∥ H 2 2 .
If h h h is zero-free, choose an analytic square root H 2 = h H^{2}=h H 2 = h ; then ∥ H ∥ H 2 2 = ∥ h ∥ H 1 \|H\|_{\mathrm{H}^{2}}^{2}=\|h\|_{\mathrm{H}^{1}} ∥ H ∥ H 2 2 = ∥ h ∥ H 1 , proving (4.1). For general h h h , the classical Riesz factorization h = A B h=AB h = A B with A , B ∈ H 2 A,B\in\mathrm{H}^{2} A , B ∈ H 2 and ∥ A ∥ H 2 ∥ B ∥ H 2 = ∥ h ∥ H 1 \|A\|_{\mathrm{H}^{2}}\|B\|_{\mathrm{H}^{2}}=\|h\|_{\mathrm{H}^{1}} ∥ A ∥ H 2 ∥ B ∥ H 2 = ∥ h ∥ H 1 gives the same estimate by applying the H 2 \mathrm{H}^{2} H 2 inequality twice; see [4 ] . It turns out that if h h h is even then we can refine the Hardy estimation 4.1 a tiny bit more.
Lemma 4.1. † If h ∈ H 1 h\in\mathrm{H}^{1} h ∈ H 1 is even, then, for 0 < r < 1 0<r<1 0 < r < 1 ,
∥ h ∥ H 1 ≥ ( 1 − r 4 ) ∣ h ( r ) ∣ . (23) \|h\|_{\mathrm{H}^{1}} \geq(1-r^{4})|h(r)|. \tag*{(23)} ∥ h ∥ H 1 ≥ ( 1 − r 4 ) ∣ h ( r ) ∣. (23)
Corollary 4.2. If h h h is odd univalent function on the unit disc then
Per c ( h ( D ) ) ≥ 2 π ( 1 − r 4 ) ∣ h ′ ( r ) ∣ . (24) \operatorname{Per}_{c}(h(\mathbb{D})) \geq2\pi(1-r^{4})|h'(r)|. \tag*{(24)} Per c ( h ( D )) ≥ 2 π ( 1 − r 4 ) ∣ h ′ ( r ) ∣. (24)
Lemma 4.3. † Suppose q q q is a real-valued, even, 2 π 2\pi 2 π -periodic C 3 C^{3} C 3 function of zero mean, strictly decreasing on [ 0 , π ] [0,\pi] [ 0 , π ] . Its analytic completion (1.3) extends to a C 1 C^{1} C 1 map on D \mathbb{D} D , is injective there, and maps D \mathbb{D} D onto a bounded rectifiable Jordan domain. In particular, 2.6 applies to g g g .
A cubic family and the perimeter comparison
In this section we will provide examples where B ( U ) \mathcal{B}(U) B ( U ) has a larger perimeter than U U U . The approach is to perturb the arcsine distribution. Consider
F b ( z ) = z + b 3 z 3 , W a , b = F b ( D ) − F b ( i a ) , 0 < a < 1 , 0 < b ≤ 1. (25) F_b(z)=z+\frac{b}{3}z^{3}, \qquad W_{a,b}=F_b(\mathbb{D})-F_b(ia), \qquad0<a<1,\quad0<b\leq1. \tag*{(25)} F b ( z ) = z + 3 b z 3 , W a , b = F b ( D ) − F b ( ia ) , 0 < a < 1 , 0 < b ≤ 1. (25)
Write G a , b = B ( W a , b ) G_{a,b}=\mathcal{B}(W_{a,b}) G a , b = B ( W a , b ) and
k = 1 − a 2 1 + a 2 ∈ ( 0 , 1 ) . (26) k=\frac{1-a^2}{1+a^2}\in(0,1). \tag*{(26)} k = 1 + a 2 1 − a 2 ∈ ( 0 , 1 ) . (26)
Theorem 4.4. † The domains W a , b W_{a,b} W a , b and G a , b G_{a,b} G a , b are bounded rectifiable Jordan domains containing 0 0 0 and having the same symmetric real exit law. Their perimeters satisfy
Per c ( W a , b ) ≤ 2 π ( 1 − b ) + 8 b , (27) \operatorname{Per}_{c}(W_{a,b})\leq2\pi(1-b)+8b, \tag*{(27)} Per c ( W a , b ) ≤ 2 π ( 1 − b ) + 8 b , (27) Per c ( G a , b ) > 8 15 ( 1 + a 2 ) ( 5 + 3 b ) . (28) \operatorname{Per}_{c}(G_{a,b})>\frac{8}{15}(1+a^2)(5+3b). \tag*{(28)} Per c ( G a , b ) > 15 8 ( 1 + a 2 ) ( 5 + 3 b ) . (28) In particular, whenever
8 15 ( 1 + a 2 ) ( 5 + 3 b ) > 2 π ( 1 − b ) + 8 b , (29) \frac{8}{15}(1+a^2)(5+3b)>2\pi(1-b)+8b, \tag*{(29)} 15 8 ( 1 + a 2 ) ( 5 + 3 b ) > 2 π ( 1 − b ) + 8 b , (29) one has
Per c ( G a , b ) > Per c ( W a , b ) . (30) \operatorname{Per}_{c}(G_{a,b})>\operatorname{Per}_{c}(W_{a,b}). \tag*{(30)} Per c ( G a , b ) > Per c ( W a , b ) . (30) For b = 1 b=1 b = 1 , the perimeter in (27) is exactly 8. For 0 < b < 1 0<b<1 0 < b < 1 , both boundaries are real-analytic Jordan curves.
Remark 4.5. Note that Theorem 4.4 describes a family of counterexamples; the prescribed exit law varies with the parameters. Nevertheless, it does not compute m c ( μ a , b ) m_c(\mu_{a,b}) m c ( μ a , b ) where μ a , b \mu_{a,b} μ a , b is the law of Re ( Z τ W a , b ) \operatorname{Re}(Z_{\tau_{W_{a,b}}}) Re ( Z τ W a , b ) .
Figure 3. The region R = { ( a , b ) : 7 / 8 < a < 1 , f ( a ) < b ≤ 1 } \mathcal{R}=\{(a,b):\sqrt{7/8}<a<1,\ f(a)<b\leq1\} R = {( a , b ) : 7/8 < a < 1 , f ( a ) < b ≤ 1 } of parameters satisfying 8 15 ( 1 + a 2 ) ( 5 + 3 b ) > 2 π ( 1 − b ) + 8 b \frac{8}{15}(1+a^2)(5+3b)>2\pi(1-b)+8b 15 8 ( 1 + a 2 ) ( 5 + 3 b ) > 2 π ( 1 − b ) + 8 b , where f ( a ) = 30 π − 40 − 40 a 2 30 π − 96 + 24 a 2 f(a)=\frac{30\pi-40-40a^2}{30\pi-96+24a^2} f ( a ) = 30 π − 96 + 24 a 2 30 π − 40 − 40 a 2 . In the region R \mathcal{R} R , Per ( G a , b ) > Per ( W a , b ) \operatorname{Per}(G_{a,b})>\operatorname{Per}(W_{a,b}) Per ( G a , b ) > Per ( W a , b ) is guaranteed. The dashed curve b = f ( a ) b=f(a) b = f ( a ) and the line a = 1 a=1 a = 1 are excluded; the segment b = 1 b=1 b = 1 is included. The curve meets b = 1 b=1 b = 1 at a = 7 / 8 ≈ 0.935 a=\sqrt{7/8}\approx0.935 a = 7/8 ≈ 0.935 and a = 1 a=1 a = 1 at b 0 = 30 π − 80 30 π − 72 ≈ 0.640 b_0=\frac{30\pi-80}{30\pi-72}\approx0.640 b 0 = 30 π − 72 30 π − 80 ≈ 0.640 .
For illustration, we give two worked examples. For the first example take b = 1 b=1 b = 1 and a = 2 2 / 3 a=2\sqrt{2}/3 a = 2 2 /3 . Then a 2 = 8 / 9 a^2=8/9 a 2 = 8/9 and
F 1 ( i a ) = i a ( 1 − a 2 3 ) = 38 2 81 i . F_1(ia)=ia\left(1-\frac{a^2}{3}\right)=\frac{38\sqrt{2}}{81}i. F 1 ( ia ) = ia ( 1 − 3 a 2 ) = 81 38 2 i .
Thus W a , 1 = U W_{a,1}=U W a , 1 = U , and Proposition 4.4 gives
Per ( U ) = 8 , Per ( B ( U ) ) > 8 15 17 9 8 = 1088 135 . \operatorname{Per}(U)=8,\qquad\operatorname{Per}(\mathcal{B}(U))>\frac{8}{15}\frac{17}{9}8=\frac{1088}{135}. Per ( U ) = 8 , Per ( B ( U )) > 15 8 9 17 8 = 135 1088 .
In this case k = 1 / 17 k=1/17 k = 1/17 , and the quantile in (49) reduces to
q ( t ) = 4 cos 3 t 3 ( cos 2 t + 289 sin 2 t ) 3 / 2 . (31) q(t)=\frac{4\cos^{3}t}{3(\cos^{2}t+289\sin^{2}t)^{3/2}}. \tag*{(31)} q ( t ) = 3 ( cos 2 t + 289 sin 2 t ) 3/2 4 cos 3 t . (31)
For the second example take b = 3 / 4 b=3/4 b = 3/4 and a = 2 6 / 5 a=2\sqrt{6}/5 a = 2 6 /5 . Then a 2 = 24 / 25 a^{2}=24/25 a 2 = 24/25 , k = 1 / 49 k=1/49 k = 1/49 , and
F 3 / 4 ( i a ) = i a ( 1 − a 2 4 ) = 38 6 125 i . F_{3/4}(ia)=ia\left(1-\frac{a^{2}}{4}\right)=\frac{38\sqrt{6}}{125}i. F 3/4 ( ia ) = ia ( 1 − 4 a 2 ) = 125 38 6 i .
Hence W a , 3 / 4 = V W_{a,3/4}=V W a , 3/4 = V , and both boundaries are real-analytic. The bounds in Proposition 4.4 become
Per ( V ) ≤ 6 + π 2 < 53 7 , Per ( B ( V ) ) > 8 15 49 25 29 4 = 2842 375 . \operatorname{Per}(V)\leq6+\frac{\pi}{2}<\frac{53}{7},\qquad\operatorname{Per}(\mathcal{B}(V))>\frac{8}{15}\frac{49}{25}\frac{29}{4}=\frac{2842}{375}. Per ( V ) ≤ 6 + 2 π < 7 53 , Per ( B ( V )) > 15 8 25 49 4 29 = 375 2842 .
Here π < 22 / 7 \pi<22/7 π < 22/7 is sufficient; it follows, for instance, from the positive integral
22 7 − π = ∫ 0 1 x 4 ( 1 − x ) 4 1 + x 2 d x > 0. \frac{22}{7}-\pi=\int_{0}^{1}\frac{x^{4}(1-x)^{4}}{1+x^{2}}\,\mathrm{d}x>0. 7 22 − π = ∫ 0 1 1 + x 2 x 4 ( 1 − x ) 4 d x > 0.
Finally,
2842 375 − 53 7 = 19 2625 > 0. \frac{2842}{375}-\frac{53}{7}=\frac{19}{2625}>0. 375 2842 − 7 53 = 2625 19 > 0.
The supports and symmetry of the two exit laws follow from (49) .
Normalization and exact quadratic norms
The conformal perimeter is independent of the choice of conformal parametrization. Derivative H p H^{p} H p norms for p ≠ 1 p\ne1 p = 1 , however, need not be invariant under precomposition by arbitrary disc automorphisms. We therefore normalize at the Brownian starting point. For the family (25) , put
ϕ a ( z ) = z + i a 1 − i a z , f a , b ( z ) = F b ( ϕ a ( z ) ) − F b ( i a ) (32) \phi_{a}(z)=\frac{z+ia}{1-iaz},\qquad f_{a,b}(z)=F_{b}(\phi_{a}(z))-F_{b}(ia) \tag*{(32)} ϕ a ( z ) = 1 − ia z z + ia , f a , b ( z ) = F b ( ϕ a ( z )) − F b ( ia ) (32)
where again F b ( z ) = z + b z 3 / 3 F_{b}(z)=z+bz^{3}/3 F b ( z ) = z + b z 3 /3 (the same function in § 4.1 ). Then f a , b ( 0 ) = 0 f_{a,b}(0)=0 f a , b ( 0 ) = 0 and f a , b ′ ( 0 ) = ( 1 − b a 2 ) ( 1 − a 2 ) > 0 f'_{a,b}(0)=(1-ba^{2})(1-a^{2})>0 f a , b ′ ( 0 ) = ( 1 − b a 2 ) ( 1 − a 2 ) > 0 . The Gross map g a , b g_{a,b} g a , b also fixes 0 and has positive derivative there. All norms compared in this section use these maps.
Comment 1. Persistence near p = 1 p=1 p = 1 and an equality exponent. Let a , b a,b a , b for which (29) is satisfied. Then
∥ f a , b ′ ∥ H 1 < ∥ g a , b ′ ∥ H 1 . (33) \|f'_{a,b}\|_{H^{1}}<\|g'_{a,b}\|_{H^{1}}. \tag*{(33)} ∥ f a , b ′ ∥ H 1 < ∥ g a , b ′ ∥ H 1 . (33)
By the continuity of the H p H^p H p norms in p > 0 p>0 p > 0 , the inequality 4.12 persists near p = 1 + p=1^+ p = 1 + . There exists η > 0 \eta>0 η > 0 such that
∥ f a , b ′ ∥ H p < ∥ g a , b ′ ∥ H p ( 1 ≤ p < 1 + η ) . (34) \left\|f'_{a,b}\right\|_{H^p}<\left\|g'_{a,b}\right\|_{H^p}\qquad(1\leq p<1+\eta). \tag*{(34)} f a , b ′ H p < g a , b ′ H p ( 1 ≤ p < 1 + η ) . (34)
On the other hand, as the Skorokhod energy is exclusively minimal for Gross embedding, one has
∥ g a , b ′ ∥ H 2 < ∥ h a , b ′ ∥ H 2 . \left\|g'_{a,b}\right\|_{H^2}<\left\|h'_{a,b}\right\|_{H^2}. g a , b ′ H 2 < h a , b ′ H 2 .
For that same pair, there also exists at least one p ∗ ∈ ( 1 , 2 ) p_*\in(1,2) p ∗ ∈ ( 1 , 2 ) such that
∥ g ′ ∥ H p ∗ = ∥ f ′ ∥ H p ∗ . (35) \left\|g'\right\|_{H^{p_*}}=\left\|f'\right\|_{H^{p_*}}. \tag*{(35)} ∥ g ′ ∥ H p ∗ = ∥ f ′ ∥ H p ∗ . (35)
In particular, derivative H p H^p H p contraction under Gross symmetrization fails on an interval immediately to the right of 1, even for domains with real-analytic boundaries. This statement neither identifies p ∗ p_* p ∗ nor proves its uniqueness. It also does not assert a universal transition exponent independent of the exit law.
Comment 2. A failed rearrangement inequality
Let u ( t ) = Re f ( e i t ) u(t)=\operatorname{Re} f(e^{it}) u ( t ) = Re f ( e i t ) for either of the normalized maps above, and let u ∗ u^* u ∗ be its even decreasing rearrangement. Then u ∗ = q u^*=q u ∗ = q , the Gross boundary datum. Define the periodic Hilbert transform on Fourier modes by
H ( cos n t ) = sin n t , H ( sin n t ) = − cos n t , H 1 = 0. \mathcal{H}(\cos nt)=\sin nt,\qquad\mathcal{H}(\sin nt)=-\cos nt,\qquad\mathcal{H}1=0. H ( cos n t ) = sin n t , H ( sin n t ) = − cos n t , H 1 = 0.
Since f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 , analytic completion gives f ( e i t ) = u ( t ) + i H u ( t ) f(e^{it})=u(t)+i\mathcal{H}u(t) f ( e i t ) = u ( t ) + i H u ( t ) . Differentiation and the arc-length formula show that
J ( u ) : = ∫ 0 2 π u ′ ( t ) 2 + ( H u ′ ( t ) ) 2 d t = 2 π ∥ f ′ ∥ H 1 . (36) \mathcal{J}(u):=\int_{0}^{2\pi}\sqrt{u'(t)^2+(\mathcal{H}u'(t))^2}\,dt=2\pi\left\|f'\right\|_{H^1}. \tag*{(36)} J ( u ) := ∫ 0 2 π u ′ ( t ) 2 + ( H u ′ ( t ) ) 2 d t = 2 π ∥ f ′ ∥ H 1 . (36)
The same identity holds for u ∗ u^* u ∗ and g g g . Theorem 4.4 therefore gives the strict reverse comparison
J ( u ∗ ) > J ( u ) . (37) \mathcal{J}(u^*)>\mathcal{J}(u). \tag*{(37)} J ( u ∗ ) > J ( u ) . (37)
This disproves a general rearrangement-decreasing inequality for J \mathcal{J} J , even when both analytic completions are univalent and, using V V V , have regular analytic boundary traces.
In contrast, the scalar total variations agree. The function r b r_b r b in (8.3) traverses [ − 1 − b / 3 , 1 + b / 3 ] [-1-b/3,1+b/3] [ − 1 − b /3 , 1 + b /3 ] monotonically on each semicircle; the boundary automorphism in (4.11) does not change its total variation. The same is true of q q q . Hence
∫ 0 2 π ∣ u ′ ( t ) ∣ d t = ∫ 0 2 π ∣ ( u ∗ ) ′ ( t ) ∣ d t = 4 ( 1 + b 3 ) . \int_{0}^{2\pi}|u'(t)|\,dt=\int_{0}^{2\pi}|(u^*)'(t)|\,dt=4\left(1+\frac{b}{3}\right). ∫ 0 2 π ∣ u ′ ( t ) ∣ d t = ∫ 0 2 π ∣ ( u ∗ ) ′ ( t ) ∣ d t = 4 ( 1 + 3 b ) .
The failure in (4.16) concerns the full complex boundary, not an increase in variation of its real coordinate.
Comment 3. The isoperimetric ratio can increase. For a bounded rectifiable Jordan domain W W W , define
I ( W ) = Per ( W ) 2 4 π Area ( W ) . I(W)=\frac{\operatorname{Per}(W)^2}{4\pi\operatorname{Area}(W)}. I ( W ) = 4 π Area ( W ) Per ( W ) 2 .
The area-nonincreasing property
Area ( B ( W ) ) ≤ Area ( W ) (38) \operatorname{Area}(\mathcal{B}(W))\le\operatorname{Area}(W) \tag*{(38)} Area ( B ( W )) ≤ Area ( W ) (38)
is the case s = 1 / 2 s=1/2 s = 1/2 of [2 Theorem 3.4] , since
Area ( W ) = π ∑ n ≥ 1 n ∣ a n ∣ 2 \operatorname{Area}(W)=\pi\sum_{n\geq1}n|a_n|^2 Area ( W ) = π n ≥ 1 ∑ n ∣ a n ∣ 2
for the normalized univalent map f W ( z ) = ∑ n ≥ 1 a n z n f_W(z)=\sum_{n\geq1}a_nz^n f W ( z ) = ∑ n ≥ 1 a n z n . This cited area result is an additional input to the next consequence; it is not used in the counterexamples or Hardy-norm calculations.
Corollary 4.6 (Isoperimetric amplification in the cubic family). † For 0 < a < 1 0<a<1 0 < a < 1 and 0 < b ≤ 1 0<b\leq1 0 < b ≤ 1 , define
R ( a , b ) = 8 15 ( 1 + a 2 ) ( 5 + 3 b ) 2 π ( 1 − b ) + 8 b . R(a,b)=\frac{\frac{8}{15}(1+a^2)(5+3b)}{2\pi(1-b)+8b}. R ( a , b ) = 2 π ( 1 − b ) + 8 b 15 8 ( 1 + a 2 ) ( 5 + 3 b ) . Then the domains in (4.4) satisfy
I ( G a , b ) I ( W a , b ) > R ( a , b ) 2 . (39) \frac{I(G_{a,b})}{I(W_{a,b})}>R(a,b)^2. \tag*{(39)} I ( W a , b ) I ( G a , b ) > R ( a , b ) 2 . (39) In particular,
2 π ( 1 − b ) + 8 b ≤ 8 15 ( 1 + a 2 ) ( 5 + 3 b ) ⟹ I ( G a , b ) > I ( W a , b ) . (40) 2\pi(1-b)+8b\leq\frac{8}{15}(1+a^2)(5+3b)\quad\Longrightarrow\quad I(G_{a,b})>I(W_{a,b}). \tag*{(40)} 2 π ( 1 − b ) + 8 b ≤ 15 8 ( 1 + a 2 ) ( 5 + 3 b ) ⟹ I ( G a , b ) > I ( W a , b ) . (40) For the domain U = W 2 2 3 , 1 U=W_{\frac{2\sqrt{2}}{3},1} U = W 3 2 2 , 1 , this gives R ( 2 2 3 , 1 ) = 136 / 135 R\left(\frac{2\sqrt{2}}{3},1\right)=136/135 R ( 3 2 2 , 1 ) = 136/135 and hence
I ( B ( U ) ) I ( U ) > 1. (41) \frac{I(\mathcal{B}(U))}{I(U)}>1. \tag*{(41)} I ( U ) I ( B ( U )) > 1. (41)
Remark 4.7. Condition (4.19) is sufficient, not necessary, for an increase of the isoperimetric ratio. When it fails, estimate (4.18) alone does not determine whether I ( G a , b ) > I ( W a , b ) I(G_{a,b})>I(W_{a,b}) I ( G a , b ) > I ( W a , b ) . Thus Brownian symmetrization can increase perimeter and the isoperimetric ratio while not increasing area. This does not conflict with the classical isoperimetric inequality; it shows that this exit-law-preserving operation need not improve its ratio.
Figure 4. The region R = { ( a , b ) : 7 / 8 ≤ a < 1 , f ( a ) ≤ b ≤ 1 } \mathcal{R}=\{(a,b):\sqrt{7}/8\le a<1,\ f(a)\le b\le1\} R = {( a , b ) : 7 /8 ≤ a < 1 , f ( a ) ≤ b ≤ 1 } of parameters satisfying ( 8 15 ( 1 + a 2 ) ( 5 + 3 b ) 2 π ( 1 − b ) + 8 b ) 2 ≥ 1 \left(\frac{\frac{8}{15}(1+a^2)(5+3b)}{2\pi(1-b)+8b}\right)^2\ge1 ( 2 π ( 1 − b ) + 8 b 15 8 ( 1 + a 2 ) ( 5 + 3 b ) ) 2 ≥ 1 , where f ( a ) = 30 π − 40 − 40 a 2 30 π − 96 + 24 a 2 f(a)=\frac{30\pi-40-40a^2}{30\pi-96+24a^2} f ( a ) = 30 π − 96 + 24 a 2 30 π − 40 − 40 a 2 . The curve b = f ( a ) b=f(a) b = f ( a ) and the segment b = 1 b=1 b = 1 are included; the line a = 1 a=1 a = 1 is excluded. The curve meets b = 1 b=1 b = 1 at a = 7 / 8 ≈ 0.935 a=\sqrt{7}/8\approx0.935 a = 7 /8 ≈ 0.935 and a = 1 a=1 a = 1 at b 0 = 30 π − 80 30 π − 72 ≈ 0.640 b_0=\frac{30\pi-80}{30\pi-72}\approx0.640 b 0 = 30 π − 72 30 π − 80 ≈ 0.640 .
Case where Gross embedding is the minimal perimeter
The preceding counterexamples show that Gross symmetrization need not decrease perimeter. Nevertheless, we now give a sufficient geometric condition for perimeter contraction: convexity of the symmetrized domain. Unlike comparisons restricted to a prescribed family of boundary parametrizations, the result below applies to every bounded rectifiable Jordan competitor.
The general perimeter comparison
The convex hull of a domain U U U , denoted by conv ( U ) \operatorname{conv}(U) conv ( U ) is the smallest convex region containing U U U . A popular result from convex geometry is that the perimeter of conv ( U ‾ ) \operatorname{conv}(\overline{U}) conv ( U ) does not exceed that of U U U [22 ] . Recall that the convex hull of an open planar domain U U U is itself open. In particular,
conv ( U ) = int ( conv ( U ‾ ) ) . \operatorname{conv}(U)=\operatorname{int}(\operatorname{conv}(\overline{U})). conv ( U ) = int ( conv ( U )) .
Moreover, every bounded convex domain has a rectifiable Jordan boundary, and hence its conformal perimeter coincides with its geometric perimeter:
Per c ( V ) = Per g ( V ) \operatorname{Per}_{c}(V)=\operatorname{Per}_{g}(V) Per c ( V ) = Per g ( V )
for every bounded convex domain V V V . Consequently, the classical perimeter inequalities for convex bodies apply equally to their conformal perimeters which are the choice for this manuscript.
Theorem 5.1 (Convex-hull lower bound and uniqueness). Assume Gross embedding U μ G U_\mu^G U μ G is bounded, and set
K μ = conv ( U μ G ) . K_\mu= \operatorname{conv}(U_\mu^G). K μ = conv ( U μ G ) . Then
Per c ( K μ ) ≤ m c ( μ ) . \operatorname{Per}_{\mathrm{c}}(K_\mu) \le m_{\mathrm{c}}(\mu). Per c ( K μ ) ≤ m c ( μ ) . In particular, if U μ G U_\mu^G U μ G is convex, then
m c ( μ ) = Per c ( U μ G ) . m_{\mathrm{c}}(\mu) = \operatorname{Per}_{\mathrm{c}}(U_\mu^G). m c ( μ ) = Per c ( U μ G ) . Moreover, in this case, U μ G U_\mu^G U μ G is the unique minimizer: for every U ∈ C ( μ ) U \in\mathcal{C}(\mu) U ∈ C ( μ ) ,
Per c ( U ) = Per c ( U μ G ) ⟺ U = U μ G . \operatorname{Per}_{\mathrm{c}}(U) = \operatorname{Per}_{\mathrm{c}}(U_\mu^G)
\quad\Longleftrightarrow\quad
U = U_\mu^G. Per c ( U ) = Per c ( U μ G ) ⟺ U = U μ G . Proof. If m c ( μ ) = ∞ m_{\mathrm{c}}(\mu) = \infty m c ( μ ) = ∞ , the lower bound is immediate. Otherwise, consider U ∈ C ( μ ) U \in\mathcal{C}(\mu) U ∈ C ( μ ) such that Per c ( U ) < ∞ \operatorname{Per}_{\mathrm{c}}(U) < \infty Per c ( U ) < ∞ , and put
K U = conv ( U ) . K_U = \operatorname{conv}(U). K U = conv ( U ) . Since the normalized conformal map f U : D → U f_U : \mathbb{D} \to U f U : D → U satisfies f U ′ ∈ H 1 f_U' \in H^1 f U ′ ∈ H 1 , it admits a rectifiable boundary parametrization of length Per c ( U ) \operatorname{Per}_{\mathrm{c}}(U) Per c ( U ) . In particular, U U U is bounded. By the classical convex-hull perimeter inequality, we obtain and
Per c ( K U ) ≤ Per c ( U ) . \operatorname{Per}_{\mathrm{c}}(K_U) \le\operatorname{Per}_{\mathrm{c}}(U). Per c ( K U ) ≤ Per c ( U ) . By Pruss’s containment theorem Theorem 2.2, one has
U μ G ⊆ S ( U ) , U_\mu^G \subseteq\mathcal{S}(U), U μ G ⊆ S ( U ) , Since Steiner symmetrization preserves inclusion, we get
U μ G ⊆ S ( U ) ⊆ S ( K U ) . U_\mu^G \subseteq\mathcal{S}(U) \subseteq\mathcal{S}(K_U). U μ G ⊆ S ( U ) ⊆ S ( K U ) . The set S ( K U ) \mathcal{S}(K_U) S ( K U ) is convex, and therefore
K μ ⊆ S ( K U ) . K_\mu\subseteq\mathcal{S}(K_U). K μ ⊆ S ( K U ) . Using the monotonicity of perimeter under inclusion of planar convex bodies, together with the classical Steiner perimeter inequality, we obtain
Per c ( K μ ) ≤ Per c ( S ( K U ) ) ≤ Per c ( K U ) ≤ Per c ( U ) . \begin{aligned}
\operatorname{Per}_{\mathrm{c}}(K_\mu) &\le\operatorname{Per}_{\mathrm{c}}(\mathcal{S}(K_U)) \\
&\le\operatorname{Per}_{\mathrm{c}}(K_U) \\
&\le\operatorname{Per}_{\mathrm{c}}(U).
\end{aligned} Per c ( K μ ) ≤ Per c ( S ( K U )) ≤ Per c ( K U ) ≤ Per c ( U ) .
Taking the infimum over all admissible domains yields
Per c ( K μ ) ≤ m c ( μ ) . \operatorname{Per}_c(K_\mu) \le m_c(\mu). Per c ( K μ ) ≤ m c ( μ ) .
Suppose now that U μ G U_\mu^G U μ G is convex. Then K μ = U μ G K_\mu= U_\mu^G K μ = U μ G , and since U μ G U_\mu^G U μ G is a bounded convex domain, then
Per c ( K μ ) = Per c ( U μ G ) ≤ m c ( μ ) ≤ Per c ( U μ G ) , \operatorname{Per}_c(K_\mu) = \operatorname{Per}_c(U_\mu^G) \le m_c(\mu) \le\operatorname{Per}_c(U_\mu^G), Per c ( K μ ) = Per c ( U μ G ) ≤ m c ( μ ) ≤ Per c ( U μ G ) ,
which proves
m c ( μ ) = Per c ( U μ G ) . m_c(\mu) = \operatorname{Per}_c(U_\mu^G). m c ( μ ) = Per c ( U μ G ) .
It remains to prove uniqueness. Suppose that U ∈ C ( μ ) U \in C(\mu) U ∈ C ( μ ) is another minimizer. Then
Per c ( U ) = Per c ( U μ G ) . \operatorname{Per}_c(U) = \operatorname{Per}_c(U_\mu^G). Per c ( U ) = Per c ( U μ G ) .
The preceding chain of inequalities becomes
Per c ( U μ G ) = Per c ( S ( K U ) ) = Per c ( K U ) = Per c ( U ) . \operatorname{Per}_c(U_\mu^G) = \operatorname{Per}_c(\mathcal{S}(K_U)) = \operatorname{Per}_c(K_U) = \operatorname{Per}_c(U). Per c ( U μ G ) = Per c ( S ( K U )) = Per c ( K U ) = Per c ( U ) .
Since
U μ G ‾ ⊆ S ( K U ‾ ) \overline{U_\mu^G} \subseteq\mathcal{S}(\overline{K_U}) U μ G ⊆ S ( K U )
and both sets are compact convex bodies, strict monotonicity of perimeter under proper inclusion implies
U μ G ‾ = S ( K U ‾ ) . \overline{U_\mu^G} = \mathcal{S}(\overline{K_U}). U μ G = S ( K U ) .
Taking interiors, we conclude that
U μ G = S ( K U ) . U_\mu^G = \mathcal{S}(K_U). U μ G = S ( K U ) .
Since Steiner symmetrization preserves area, we obtain
Area ( U μ G ) = Area ( S ( K U ) ) = Area ( K U ) . \operatorname{Area}(U_\mu^G) = \operatorname{Area}(\mathcal{S}(K_U)) = \operatorname{Area}(K_U). Area ( U μ G ) = Area ( S ( K U )) = Area ( K U ) .
On the other hand, the area inequality for Brownian symmetrization gives
Area ( U μ G ) ≤ Area ( U ) ≤ Area ( K U ) . \operatorname{Area}(U_\mu^G) \le\operatorname{Area}(U) \le\operatorname{Area}(K_U). Area ( U μ G ) ≤ Area ( U ) ≤ Area ( K U ) .
Consequently,
Area ( U μ G ) = Area ( U ) . \operatorname{Area}(U_\mu^G) = \operatorname{Area}(U). Area ( U μ G ) = Area ( U ) .
Since both domains are images of univalent conformal maps, equality of their areas is equivalent to equality in the Dirichlet-norm inequality under Baernstein rearrangement. By Pruss’s rigidity theorem [7 ] , the corresponding conformal maps differ only by a rotation of the disk parameter. Their images therefore coincide:
U = U μ G . U = U_\mu^G. U = U μ G .
This completes the proof of uniqueness.
A straightforward corollary of Theorem 5.1 is that if the minimizer of 1.2 is not Gross embedding then this latter is necessarily nonconvex.
A sufficient condition ensuring the minimality of the perimeter of Gross embedding.
There is a classical coefficient criterion for convex analytic functions [13 ] . If f = ∑ n ≥ 1 a n z n f=\sum_{n\geq1}a_nz^n f = ∑ n ≥ 1 a n z n is univalent in the unit disc then a sufficient condition of the convexity of f ( D ) f(\mathbb{D}) f ( D ) is the following
∑ n ≥ 2 n 2 ∣ a n ∣ ≤ ∣ a 1 ∣ . (42) \sum_{n\geq2} n^2|a_n| \leq|a_1|. \tag*{(42)} n ≥ 2 ∑ n 2 ∣ a n ∣ ≤ ∣ a 1 ∣. (42)
Recall the Fourier expansion of the Gross map generated from the distribution μ \mu μ :
g μ ( z ) = ∑ n ≥ 1 a n z n , a n = 2 ∫ 0 1 q μ ( t ) cos ( n π t ) d t . (43) g_\mu(z)=\sum_{n\geq1}a_nz^n,\quad a_n=2\int_0^1q_\mu(t)\cos(n\pi t)\,dt. \tag*{(43)} g μ ( z ) = n ≥ 1 ∑ a n z n , a n = 2 ∫ 0 1 q μ ( t ) cos ( nπ t ) d t . (43)
Let
ξ ∼ Kumaraswamy ( 1 2 , 1 2 ) . \xi\sim\operatorname{Kumaraswamy}\left(\frac{1}{2},\frac{1}{2}\right). ξ ∼ Kumaraswamy ( 2 1 , 2 1 ) .
Its density on ( 0 , 1 ) (0,1) ( 0 , 1 ) is
ρ X ( x ) = 1 4 x 1 − x . \rho_X(x)=\frac{1}{4\sqrt{x}\sqrt{1-\sqrt{x}}}. ρ X ( x ) = 4 x 1 − x 1 .
In particular E ξ = 8 / 15 \mathbb{E}\xi=8/15 E ξ = 8/15 . See [14 ] . Let μ \mu μ be the law of ξ − 8 / 15 \xi-8/15 ξ − 8/15 . Computation of the underlying Fourier coefficients yields
a n = 48 ( − 1 ) n + 1 π 4 n 4 . a_n=\frac{48(-1)^{n+1}}{\pi^4n^4}. a n = π 4 n 4 48 ( − 1 ) n + 1 .
In particular
∑ n = 2 ∞ n 2 ∣ a n ∣ ∣ a 1 ∣ = π 2 6 − 1 ≈ 0.644934 < 1. \frac{\sum_{n=2}^{\infty}n^2|a_n|}{|a_1|}=\frac{\pi^2}{6}-1\approx0.644934<1. ∣ a 1 ∣ ∑ n = 2 ∞ n 2 ∣ a n ∣ = 6 π 2 − 1 ≈ 0.644934 < 1.
Hence the Gross embedding is convex and it is the minimizer for the perimeter optimization problem of μ \mu μ .
Illustrations
In this section, we provide worked examples associated with several distributions.
The uniform distribution on ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) .
For the uniform probability measure on ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) , the Gross embedding attains the bottom of perimeters in C ( μ ) \mathcal{C}(\mu) C ( μ ) . The increasing quantile is
Q μ ( t ) = 2 t − 1. Q_\mu(t)=2t-1. Q μ ( t ) = 2 t − 1.
In the increasing-quantile convention, the Gross map is given in [5 ] by
g ( z ) = − 8 π 2 ∑ n ≥ 1 n o d d z n n 2 . g(z)=-\frac{8}{\pi^2}\sum_{\substack{n\geq1\\ n\ \mathrm{odd}}}\frac{z^n}{n^2}. g ( z ) = − π 2 8 n ≥ 1 n odd ∑ n 2 z n .
The series converges uniformly on D ‾ \overline{D} D . Its two boundary branches can be written as
x ⟼ x ± i h ( x ) , − 1 ≤ x ≤ 1 , x \longmapsto x \pm i h(x), \qquad-1 \le x \le1, x ⟼ x ± ih ( x ) , − 1 ≤ x ≤ 1 ,
where
h ( x ) = 8 π 2 ∑ n ≥ 1 n o d d sin ( n θ ) n 2 , θ = π ( x + 1 ) 2 . h(x) = \frac{8}{\pi^2} \sum_{\substack{n\ge1\\ n\ \mathrm{odd}}} \frac{\sin(n\theta)}{n^2}, \qquad\theta= \frac{\pi(x+1)}{2}. h ( x ) = π 2 8 n ≥ 1 n odd ∑ n 2 sin ( n θ ) , θ = 2 π ( x + 1 ) .
The function h h h is continuous on [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] , with
h ( − 1 ) = h ( 1 ) = 0. h(-1) = h(1) = 0. h ( − 1 ) = h ( 1 ) = 0.
The identity
∑ n ≥ 1 n o d d z n n = 1 2 log 1 + z 1 − z , ∣ z ∣ < 1 , \sum_{\substack{n\ge1\\ n\ \mathrm{odd}}} \frac{z^n}{n} = \frac{1}{2}\log\frac{1+z}{1-z}, \qquad|z|<1, n ≥ 1 n odd ∑ n z n = 2 1 log 1 − z 1 + z , ∣ z ∣ < 1 ,
gives, by radial limits,
∑ n ≥ 1 n o d d cos ( n θ ) n = 1 2 log cot θ 2 , 0 < θ < π . \sum_{\substack{n\ge1\\ n\ \mathrm{odd}}} \frac{\cos(n\theta)}{n} = \frac{1}{2}\log\cot\frac{\theta}{2}, \qquad0<\theta<\pi. n ≥ 1 n odd ∑ n cos ( n θ ) = 2 1 log cot 2 θ , 0 < θ < π .
Consequently,
h ′ ( x ) = 2 π log cot π ( x + 1 ) 4 , h ′ ′ ( x ) = − 1 cos ( π x / 2 ) < 0 , − 1 < x < 1. h'(x) = \frac{2}{\pi}\log\cot\frac{\pi(x+1)}{4}, \qquad h''(x) = -\frac{1}{\cos(\pi x/2)}<0, \qquad-1<x<1. h ′ ( x ) = π 2 log cot 4 π ( x + 1 ) , h ′′ ( x ) = − cos ( π x /2 ) 1 < 0 , − 1 < x < 1.
Thus h h h is strictly concave and positive on ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) . It follows that
U μ G = { x + i y : − 1 < x < 1 , ∣ y ∣ < h ( x ) } U_{\mu}^{G} = \{x+iy:-1<x<1,\ |y|<h(x)\} U μ G = { x + i y : − 1 < x < 1 , ∣ y ∣ < h ( x )}
is bounded and convex. Theorem 5.1 applies. In particular, the minimum perimeter in the stated class is
Per ( U μ G ) = 2 ∫ − 1 1 1 + 4 π 2 ( log cot π ( x + 1 ) 4 ) 2 d x . \operatorname{Per}(U_{\mu}^{G}) = 2\int_{-1}^{1}\sqrt{1+\frac{4}{\pi^2}\left(\log\cot\frac{\pi(x+1)}{4}\right)^2}\,dx. Per ( U μ G ) = 2 ∫ − 1 1 1 + π 2 4 ( log cot 4 π ( x + 1 ) ) 2 d x .
The logarithmic endpoint singularities are integrable, so this quantity is finite. Using 3.12, one has
2 π 2 3 ≈ 5.130 ≤ Per ( U μ G ) = m c ( μ ) ≤ 4 2 ≈ 5.657 2\pi\sqrt{\frac{2}{3}} \approx5.130 \le\operatorname{Per}(U_{\mu}^{G}) = m_c(\mu) \le4\sqrt{2} \approx5.657 2 π 3 2 ≈ 5.130 ≤ Per ( U μ G ) = m c ( μ ) ≤ 4 2 ≈ 5.657
Numerically, we have
Per ( U μ G ) ≈ 5.3261068066989535. \operatorname{Per}(U_{\mu}^{G}) \approx5.3261068066989535. Per ( U μ G ) ≈ 5.3261068066989535.
Note that Gross map satisfies
∑ n ≥ 1 n 2 ∣ a n ∣ = ∞ . \sum_{n\ge1} n^2|a_n| = \infty. n ≥ 1 ∑ n 2 ∣ a n ∣ = ∞.
Nevertheless, Gross embedding is convex as shown. Hence the condition 5.1 is not necessary.
The arcsine distribution on ( − 1 , 1 ) (-1,1) ( − 1 , 1 )
Recall the arcsine law
d μ a s ( x ) = 1 ( − 1 , 1 ) ( x ) π 1 − x 2 d x . \mathrm{d}\mu_{\mathrm{as}}(x)=\frac{\mathbf{1}_{(-1,1)}(x)}{\pi\sqrt{1-x^2}}\,\mathrm{d}x. d μ as ( x ) = π 1 − x 2 1 ( − 1 , 1 ) ( x ) d x .
Gross embedding is the unit disk. Thus it is the unique perimeter-minimizing domain among all bounded rectifiable Jordan domains containing 0 0 0 with real Brownian exit law μ a s \mu_{\mathrm{as}} μ as . The minimum perimeter is
m c ( μ a s ) = 2 π . m_c(\mu_{\mathrm{as}})=2\pi. m c ( μ as ) = 2 π .
The increasing quantile is
Q μ a s ( t ) = − cos ( π t ) , Q_{\mu_{\mathrm{as}}}(t)=-\cos(\pi t), Q μ as ( t ) = − cos ( π t ) ,
so its Gross map is g ( z ) = − z g(z)=-z g ( z ) = − z and its Gross domain is D \mathbb{D} D . Since D \mathbb{D} D is convex, theorem 5.1 proves perimeter minimality. However, we give a separate coefficient argument to establish uniqueness. Let
f U ( z ) = ∑ n ≥ 1 c n z n f_U(z)=\sum_{n\geq1}c_nz^n f U ( z ) = n ≥ 1 ∑ c n z n
be a conformal map onto an admissible competitor U U U , with f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 . We have
1 2 ∑ n ≥ 1 ∣ c n ∣ 2 = Var ( μ a s ) = 1 2 , \frac{1}{2}\sum_{n\geq1}|c_n|^2=\operatorname{Var}(\mu_{\mathrm{as}})=\frac{1}{2}, 2 1 n ≥ 1 ∑ ∣ c n ∣ 2 = Var ( μ as ) = 2 1 ,
and therefore
∑ n ≥ 1 ∣ c n ∣ 2 = 1. \sum_{n\geq1}|c_n|^2=1. n ≥ 1 ∑ ∣ c n ∣ 2 = 1.
The conformal area formula implies
Area ( f ( D ) ) = π ∑ n ≥ 1 n ∣ c n ∣ 2 ≥ π . \operatorname{Area}(f(\mathbb{D}))=\pi\sum_{n\geq1}n|c_n|^2\geq\pi. Area ( f ( D )) = π n ≥ 1 ∑ n ∣ c n ∣ 2 ≥ π .
By the isoperimetric inequality,
Per c ( f ( D ) ) 2 ≥ 4 π Area ( f ( D ) ) ≥ 4 π 2 . \operatorname{Per}_{c}(f(\mathbb{D}))^2\geq4\pi\operatorname{Area}(f(\mathbb{D}))\geq4\pi^2. Per c ( f ( D ) ) 2 ≥ 4 π Area ( f ( D )) ≥ 4 π 2 .
If Per c ( f ( D ) ) = 2 π \operatorname{Per}_{c}(f(\mathbb{D}))=2\pi Per c ( f ( D )) = 2 π , equality holds throughout, and hence
∑ n ≥ 1 ( n − 1 ) ∣ c n ∣ 2 = 0. \sum_{n\geq1}(n-1)|c_n|^2=0. n ≥ 1 ∑ ( n − 1 ) ∣ c n ∣ 2 = 0.
Thus c n = 0 c_n=0 c n = 0 for every n ≥ 2 n\geq2 n ≥ 2 and ∣ c 1 ∣ = 1 |c_1|=1 ∣ c 1 ∣ = 1 . The image domain is therefore exactly D \mathbb{D} D .
Infinite Skorokhod energy does not imply degeneracy of 1.2.
In virtue of Proposition 3.14, we saw that if the Skorokhod energy of μ \mu μ is finite then 1.2 is nondegenerate. However, such a condition is not necessary. Let
Ξ ( z ) = 1 − 1 − z , z ∈ D . \Xi(z)=1-\sqrt{1-z},\qquad z\in\mathbb{D}. Ξ ( z ) = 1 − 1 − z , z ∈ D .
We can check that
Ξ ′ ∈ H 1 ∖ H 2 . \Xi' \in H^1 \setminus H^2. Ξ ′ ∈ H 1 ∖ H 2 .
Let μ Ξ \mu_{\Xi} μ Ξ be the law of Re Ξ ( e i Θ ) \operatorname{Re}\Xi(e^{\mathrm{i}\Theta}) Re Ξ ( e i Θ ) where Θ \Theta Θ is uniform in ( − π , π ) (-\pi,\pi) ( − π , π ) . On the other hand, Ξ ( D ) \Xi(D) Ξ ( D ) is exactly the Gross embedding of μ Ξ \mu_{\Xi} μ Ξ in virtue of the uniqueness criterion in [20 ] . In particular
Λ μ Ξ = 1 4 ∥ Ξ ′ ∥ H 2 2 = + ∞ . \Lambda_{\mu_{\Xi}} = \frac{1}{4}\|\Xi'\|_{H^2}^2 = +\infty. Λ μ Ξ = 4 1 ∥ Ξ ′ ∥ H 2 2 = + ∞.
We can also show that Ξ ( D ) \Xi(D) Ξ ( D ) is convex and
m c ( μ Ξ ) = Per c ( Ξ ( D ) ) = 2 2 Beta ( 1 4 , 1 2 ) ≈ 3.70815 m_c(\mu_{\Xi}) = \operatorname{Per}_c(\Xi(D)) = \frac{\sqrt{2}}{2}\operatorname{Beta}\left(\frac{1}{4},\frac{1}{2}\right) \approx3.70815 m c ( μ Ξ ) = Per c ( Ξ ( D )) = 2 2 Beta ( 4 1 , 2 1 ) ≈ 3.70815
where Beta is Euler’s beta function.
Figure 5. Gross embedding of μ Ξ \mu_{\Xi} μ Ξ , which has the least perimeter.
Boundedness of the Gross domain does not imply that the perimeter optimization problem has finite value.
The classical Cantor-Lebesgue function C : [ 0 , 1 ] → [ 0 , 1 ] C : [0,1] \to[0,1] C : [ 0 , 1 ] → [ 0 , 1 ] is the continuous, nondecreasing function that is constant on each complementary interval of the middle-third Cantor set, satisfies C ( 0 ) = 0 C(0) = 0 C ( 0 ) = 0 and C ( 1 ) = 1 C(1) = 1 C ( 1 ) = 1 , and has derivative zero almost everywhere. Define
Q ( t ) = t + C ( t ) − 1 , 0 ≤ t ≤ 1. Q(t) = t + C(t) - 1,\qquad0 \leq t \leq1. Q ( t ) = t + C ( t ) − 1 , 0 ≤ t ≤ 1.
Let T T T be uniformly distributed on [ 0 , 1 ] [0,1] [ 0 , 1 ] and set
μ = Law ( Q ( T ) ) . \mu= \operatorname{Law}(Q(T)). μ = Law ( Q ( T )) .
We claim that μ \mu μ is centered, absolutely continuous with bounded density, and has support [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] , while
U μ G is bounded , m c ( μ ) = + ∞ . U_{\mu}^{G}\text{ is bounded},\qquad m_c(\mu) = +\infty. U μ G is bounded , m c ( μ ) = + ∞.
Properties of the distribution. Since C C C is continuous and nondecreasing, Q Q Q is continuous and strictly increasing, with
Q ( 0 ) = − 1 , Q ( 1 ) = 1. Q(0) = -1,\qquad Q(1) = 1. Q ( 0 ) = − 1 , Q ( 1 ) = 1.
Moreover, using the symmetry of the Cantor function,
∫ 0 1 Q ( t ) d t = 1 2 + 1 2 − 1 = 0. \int_{0}^{1} Q(t)\,\mathrm{d}t = \frac{1}{2} + \frac{1}{2} - 1 = 0. ∫ 0 1 Q ( t ) d t = 2 1 + 2 1 − 1 = 0.
Thus μ \mu μ is centered and supp μ = [ − 1 , 1 ] \operatorname{supp}\mu=[-1,1] supp μ = [ − 1 , 1 ] .
For 0 ≤ s < t ≤ 1 0 \le s < t \le1 0 ≤ s < t ≤ 1 , we have
Q ( t ) − Q ( s ) = ( t − s ) + ( C ( t ) − C ( s ) ) ≥ t − s . Q(t)-Q(s)=(t-s)+(C(t)-C(s))\ge t-s. Q ( t ) − Q ( s ) = ( t − s ) + ( C ( t ) − C ( s )) ≥ t − s .
Consequently, Q − 1 Q^{-1} Q − 1 is 1-Lipschitz. Since the distribution function of μ \mu μ on [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] is precisely Q − 1 Q^{-1} Q − 1 , the measure μ \mu μ is absolutely continuous:
μ ( d x ) = ρ ( x ) d x , 0 ≤ ρ ( x ) ≤ 1 a.e. \mu(\mathrm{d}x)=\rho(x)\,\mathrm{d}x,\qquad0\le\rho(x)\le1\quad\text{a.e.} μ ( d x ) = ρ ( x ) d x , 0 ≤ ρ ( x ) ≤ 1 a.e.
Boundedness of the Gross embedding. The Cantor function is Hölder continuous with exponent α = log 2 / log 3 \alpha=\log2/\log3 α = log 2/ log 3 . Hence Q Q Q has the same regularity. The symmetric decreasing boundary datum associated with the Gross construction is
q ( θ ) = Q ( 1 − ∣ θ ∣ π ) , − π ≤ θ ≤ π . q(\theta)=Q\left(1-\frac{|\theta|}{\pi}\right),\qquad-\pi\le\theta\le\pi. q ( θ ) = Q ( 1 − π ∣ θ ∣ ) , − π ≤ θ ≤ π .
Thus q ∈ C α ( T ) q\in C^\alpha(\mathbb{T}) q ∈ C α ( T ) . Since the periodic Hilbert transform preserves Hölder spaces C α C^\alpha C α , 0 < α < 1 0<\alpha<1 0 < α < 1 , its harmonic conjugate is bounded. It follows that the Gross conformal map is bounded as well, and therefore G μ G_\mu G μ is bounded.
Degeneracy of the perimeter problem. Let K ⊂ [ 0 , 1 ] K\subset[0,1] K ⊂ [ 0 , 1 ] denote the middle-third Cantor set and put E = Q ( K ) E=Q(K) E = Q ( K ) . Since L 1 ( K ) = 0 \mathcal{L}^{1}(K)=0 L 1 ( K ) = 0 and Q Q Q is injective,
μ ( E ) = L 1 ( K ) = 0. \mu(E)=\mathcal{L}^{1}(K)=0. μ ( E ) = L 1 ( K ) = 0.
On the other hand, L 1 ( E ) = 1 \mathcal{L}^{1}(E)=1 L 1 ( E ) = 1 . Indeed, the complementary intervals of K K K have total length 1, and C C C is constant on each of these intervals. Hence Q Q Q preserves their lengths. Since Q ( [ 0 , 1 ] ) = [ − 1 , 1 ] Q([0,1])=[-1,1] Q ([ 0 , 1 ]) = [ − 1 , 1 ] , whose length is 2, we obtain
L 1 ( Q ( K ) ) = 2 − 1 = 1. \mathcal{L}^{1}(Q(K))=2-1=1. L 1 ( Q ( K )) = 2 − 1 = 1.
Thus E ⊂ supp μ E\subset\operatorname{supp}\mu E ⊂ supp μ satisfies
μ ( E ) = 0 , L 1 ( E ) > 0. \mu(E)=0,\qquad\mathcal{L}^{1}(E)>0. μ ( E ) = 0 , L 1 ( E ) > 0.
Therefore,
L 1 ∣ supp μ ≪̸ μ . \mathcal{L}^{1}|_{\operatorname{supp}\mu}\not\ll\mu. L 1 ∣ supp μ ≪ μ .
By Theorem 3.11 , no admissible embedding can have finite conformal perimeter. Consequently,
m c ( μ ) = + ∞ . m_c(\mu)=+\infty. m c ( μ ) = + ∞.
Figure 6. Gross embedding.
Concluding remarks and open problems
The planar Skorokhod embedding problem provides a natural setting in which to investigate how a prescribed probability law constrains the geometry of a domain. From this perspective, conformal perimeter is not merely a boundary functional: it measures the geometric cost of realizing the prescribed real-part exit distribution. The results of this paper show that this cost is governed by an interplay between conformal mapping, rearrangement inequalities, and the measure-theoretic structure of the exit law.
Our findings clarify the role of the Gross embedding. Although it is distinguished by its area-minimizing property, our explicit counterexamples show that it need not minimize conformal perimeter. Nevertheless, it remains central to the optimization problem: its convex hull provides a universal geometric lower bound, its unboundedness rules out every finite-cost competitor, and, when it is bounded and convex, it is the unique perimeter minimizer. Thus, the Gross construction retains significant extremal information even when it does not itself solve the perimeter problem.
The existence theorem supplies a core result: once a single finite-conformal-perimeter embedding exists, a minimizer exists as well. No Jordan-boundary assumption is needed, and every finite-cost minimizer is bounded and admits an absolutely continuous conformal boundary parametrization. The principal issue is therefore not attainment in the finite-cost regime, but identifying which probability laws belong to that regime. Our necessary condition
L 1 ∣ supp μ ≪ μ L^{1}\big|_{\operatorname{supp}\mu} \ll\mu L 1 supp μ ≪ μ
and our logarithmic sufficient condition provide two different descriptions of this question. The examples demonstrate that neither connected support nor absolute continuity, even together with boundedness of the Gross domain, suffices. Conversely, logarithmic integrability allows finite-perimeter embeddings for densities with substantial degeneracies, well beyond the class covered by finite Skorokhod energy. These results leave several natural questions open.
Preservation of finite conformal perimeter. The central unresolved question is whether
m c ( μ ) < ∞ ⟹ Per c ( U μ G ) < ∞ . m_{\mathrm{c}}(\mu) < \infty\quad\Longrightarrow\quad\operatorname{Per}_{\mathrm{c}}(U_{\mu}^{G}) < \infty. m c ( μ ) < ∞ ⟹ Per c ( U μ G ) < ∞.
The reverse implication is immediate. A positive answer would therefore characterize finite-cost solvability through the canonical Gross embedding, without asserting that Gross minimizes perimeter. A negative answer would exhibit a different phenomenon: Gross symmetrization could transform an embedding of finite conformal perimeter into one of infinite conformal perimeter while preserving the exit law.
Analytically, the difficulty lies at the Hardy-space endpoint. If u = Re f μ ⋆ u = \operatorname{Re} f_{\mu}^{\star} u = Re f μ ⋆ and u # u^{\#} u # is its symmetric decreasing rearrangement, finite conformal perimeter gives
u ′ ∈ L 1 ( T ) , H u ′ ∈ L 1 ( T ) . u' \in L^{1}(\mathbb{T}), \qquad\mathcal{H}u' \in L^{1}(\mathbb{T}). u ′ ∈ L 1 ( T ) , H u ′ ∈ L 1 ( T ) .
Rearrangement controls ( u # ) ′ (u^{\#})' ( u # ) ′ in L 1 L^{1} L 1 , but the additional conclusion
H ( ( u # ) ′ ) ∈ L 1 ( T ) \mathcal{H}\bigl((u^{\#})'\bigr) \in L^{1}(\mathbb{T}) H ( ( u # ) ′ ) ∈ L 1 ( T )
does not follow from the usual L 1 L^{1} L 1 estimates. Understanding whether the conformal and equimeasurability constraints supply the missing control would substantially advance the characterization of finite-cost exit laws.
Optimality beyond the convex case. Bounded convex Gross domains are uniquely perimeter minimizing. Is convexity also necessary for finite-cost perimeter minimality of the Gross embedding? More generally, can one characterize those probability laws for which
m c ( μ ) = Per c ( G μ ) ? m_{\mathrm{c}}(\mu) = \operatorname{Per}_{\mathrm{c}}(G_{\mu})? m c ( μ ) = Per c ( G μ )?
Our counterexamples show that this equality does not hold universally, but they do not determine the full class of laws for which it does. A characterization would distinguish the distributions for which the canonical rearrangement is geometrically optimal from those requiring a different boundary arrangement.
Geometry and uniqueness of minimizers. Outside the convex-Gross setting, the geometry of minimizers remains largely undetermined. Must at least one minimizer be vertically convex? Must every minimizer have a Jordan boundary, or can a minimizing conformal boundary parametrization necessarily traverse some boundary portions more than once? What additional conditions on the exit law imply stronger boundary regularity?
Uniqueness is closely linked to these questions. Reflection about the real axis preserves both the exit law and conformal perimeter, so a unique minimizer must possess this symmetry. If it is also vertically convex, it is Steiner symmetric and hence coincides with the Gross embedding. Consequently, for a law whose Gross embedding is not minimizing, the existence of a vertically convex minimizer would imply nonuniqueness. This connects the geometry of optimal domains with the structure of the entire minimizing family.
Quantitative estimates, stability, and approximation. The geometric lower bounds, the Skorokhod-energy upper bound, and the logarithmic sufficient condition provide initial quantitative control of m c ( μ ) m_c(\mu) m c ( μ ) . It remains to determine their sharpness and to understand how the optimal value depends on the prescribed law. Under what assumptions does convergence of probability measures imply convergence of their optimal perimeters? When minimizers are unique, can one obtain corresponding convergence of their normalized conformal maps? Does the p p p -Brownian convergence concept, developed in [1 ] , hold?
These stability questions are also relevant to the construction of rigorous numerical methods. A useful approximation scheme must control the exit-law constraint and conformal perimeter while preserving univalence. Establishing convergence to a genuine minimizer, rather than merely producing low-perimeter approximate embeddings, would turn the existence theory into a constructive approach to optimal design.
Taken together, these questions suggest that perimeter optimization is not a direct extension of area minimization. It introduces a distinct sensitivity to boundary variation, for which the canonical rearrangement may preserve some geometric features while worsening others. Understanding this distinction offers a route toward a broader theory of optimal planar Skorokhod embeddings, connecting the distribution of probability mass with the regularity, geometry, and multiplicity of its optimal realizations.
Proofs of the technical tools
Proof. of lemma 4.1 Write h ( z ) = H ( z 2 ) h(z)=H(z^2) h ( z ) = H ( z 2 ) . A change of angle in the defining circle means shows H ∈ H 1 H\in H^1 H ∈ H 1 and ∥ H ∥ H 1 = ∥ h ∥ H 1 \lVert H\rVert_{H^1}=\lVert h\rVert_{H^1} ∥ H ∥ H 1 = ∥ h ∥ H 1 . Apply (4.1) to H H H at r 2 r^2 r 2 .
For the zero-free functions used below, there is a direct square-root proof. If K 2 = h K^2=h K 2 = h , then K ( − z ) / K ( z ) K(-z)/K(z) K ( − z ) / K ( z ) is constant with value 1 at 0, so K K K is even and K ( z ) = ∑ n ≥ 0 d n z 2 n K(z)=\sum_{n\geq0}d_nz^{2n} K ( z ) = ∑ n ≥ 0 d n z 2 n . Consequently
∣ h ( r ) ∣ = ∣ K ( r ) ∣ 2 ≤ ( ∑ n ≥ 0 ∣ d n ∣ 2 ) ( ∑ n ≥ 0 r 4 n ) = ∥ h ∥ H 1 1 − r 4 . |h(r)|=|K(r)|^2\leq\left(\sum_{n\geq0}|d_n|^2\right)\left(\sum_{n\geq0}r^{4n}\right)=\frac{\lVert h\rVert_{H^1}}{1-r^4}. ∣ h ( r ) ∣ = ∣ K ( r ) ∣ 2 ≤ ( n ≥ 0 ∑ ∣ d n ∣ 2 ) ( n ≥ 0 ∑ r 4 n ) = 1 − r 4 ∥ h ∥ H 1 .
The zero-free assumption is needed for this particular square-root argument, not for the lemma itself. The factor 1 − r 4 1-r^4 1 − r 4 instead of 1 − r 2 1-r^2 1 − r 2 is an elementary symmetry improvement of a classical estimate, not a new general Hardy inequality.
Proof. of lemma 4.3 The function q q q is the decreasing angular quantile of the law of q ( Θ ) q(\Theta) q ( Θ ) for Θ \Theta Θ uniform on ( 0 , π ) (0,\pi) ( 0 , π ) . This law is centered and nondegenerate, so Gross’s theorem applies. Three integrations by parts in the periodic Fourier formula give c n = O ( n − 3 ) c_n = O(n^{-3}) c n = O ( n − 3 ) . Hence the series for g g g and g ′ g' g ′ converge uniformly on D ‾ \overline{D} D , and R e g ( e i t ) = q ( t ) \mathop{\mathrm{Re}} g(e^{it}) = q(t) Re g ( e i t ) = q ( t ) . The coefficients are real, and
g ′ ( 0 ) = c 1 = − 2 π ∫ 0 π q ′ ( t ) sin t d t > 0. g'(0) = c_1 = -\frac{2}{\pi}\int_0^\pi q'(t)\sin t\,dt > 0. g ′ ( 0 ) = c 1 = − π 2 ∫ 0 π q ′ ( t ) sin t d t > 0.
Since g ′ g' g ′ never vanishes in D D D , it is positive on ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) . The endpoint values are g ( 1 ) = q ( 0 ) g(1) = q(0) g ( 1 ) = q ( 0 ) and g ( − 1 ) = q ( π ) g(-1) = q(\pi) g ( − 1 ) = q ( π ) ; thus
g ( ( − 1 , 1 ) ) = ( q ( π ) , q ( 0 ) ) . (44) g((-1,1)) = (q(\pi),q(0)). \tag*{(44)} g (( − 1 , 1 )) = ( q ( π ) , q ( 0 )) . (44)
For z z z in the upper half-disc, g ( z ) g(z) g ( z ) cannot be real, since otherwise g ( z ) = g ( z ‾ ) g(z) = g(\overline{z}) g ( z ) = g ( z ) would contradict univalence. Its imaginary part has a constant sign there, and g ′ ( 0 ) > 0 g'(0) > 0 g ′ ( 0 ) > 0 makes that sign positive. If ζ ∈ T \zeta\in\mathbb{T} ζ ∈ T , continuity implies g ( ζ ) ∈ g ( D ‾ ) g(\zeta) \in g(\overline{D}) g ( ζ ) ∈ g ( D ) . It cannot lie in g ( D ) g(D) g ( D ) : continuity of the conformal inverse at that point, applied to g ( r ζ ) g(r\zeta) g ( r ζ ) , would contradict r ζ → ζ ∉ D r\zeta\to\zeta\notin D r ζ → ζ ∈ / D . Consequently g ( ζ ) ∈ ∂ g ( D ) g(\zeta) \in\partial g(D) g ( ζ ) ∈ ∂ g ( D ) . For 0 < t < π 0 < t < \pi 0 < t < π , the imaginary part of g ( e i t ) g(e^{it}) g ( e i t ) cannot be zero, because its real part q ( t ) q(t) q ( t ) lies in the interior interval (8.1).
The upper boundary arc therefore lies strictly above the real axis and has a strictly decreasing real coordinate. It is injective. The lower arc is its reflection, and the two arcs meet only at their distinct endpoints. This proves injectivity on D ‾ \overline{D} D . Their union is the whole boundary: any boundary point is the limit of interior images, and compactness of D ‾ \overline{D} D supplies a preimage on T \mathbb{T} T . Finally, the C 1 C^1 C 1 boundary parametrization has length
∫ 0 2 π ∣ d d t g ( e i t ) ∣ d t = ∫ 0 2 π ∣ g ′ ( e i t ) ∣ d t = 2 π ∥ g ′ ∥ H 1 . \int_0^{2\pi}\left|\frac{d}{dt}g(e^{it})\right|\,dt = \int_0^{2\pi}|g'(e^{it})|\,dt = 2\pi\|g'\|_{H^1}. ∫ 0 2 π d t d g ( e i t ) d t = ∫ 0 2 π ∣ g ′ ( e i t ) ∣ d t = 2 π ∥ g ′ ∥ H 1 .
We will also use the following derivative formula. Integration by parts in (3) gives
n c n = − 2 π ∫ 0 π q ′ ( t ) sin ( n t ) d t . nc_n = -\frac{2}{\pi}\int_0^\pi q'(t)\sin(nt)\,dt. n c n = − π 2 ∫ 0 π q ′ ( t ) sin ( n t ) d t .
For 0 < r < 1 0 < r < 1 0 < r < 1 , summing the uniformly convergent geometric series
∑ n ≥ 1 r n − 1 sin ( n t ) = sin t 1 − 2 r cos t + r 2 \sum_{n\geq1}r^{n-1}\sin(nt) = \frac{\sin t}{1-2r\cos t+r^2} n ≥ 1 ∑ r n − 1 sin ( n t ) = 1 − 2 r cos t + r 2 sin t
therefore yields
g ′ ( r ) = − 2 π ∫ 0 π q ′ ( t ) sin t 1 − 2 r cos t + r 2 d t . (45) g'(r) = -\frac{2}{\pi}\int_0^\pi\frac{q'(t)\sin t}{1-2r\cos t+r^2}\,dt. \tag*{(45)} g ′ ( r ) = − π 2 ∫ 0 π 1 − 2 r cos t + r 2 q ′ ( t ) sin t d t . (45)
Proof. of corollary 4.6. The area inequality (38) gives
Area ( W a , b ) Area ( G a , b ) ≥ 1. \frac{\operatorname{Area}(W_{a,b})}{\operatorname{Area}(G_{a,b})} \geq1. Area ( G a , b ) Area ( W a , b ) ≥ 1.
Combining this with (27) and (28) , we obtain
I ( G a , b ) I ( W a , b ) = ( Per ( G a , b ) Per ( W a , b ) ) 2 Area ( W a , b ) Area ( G a , b ) ≥ ( Per ( G a , b ) Per ( W a , b ) ) 2 > ( 8 15 ( 1 + a 2 ) ( 5 + 3 b ) 2 π ( 1 − b ) + 8 b ) 2 = R ( a , b ) 2 . \begin{aligned}
\frac{I(G_{a,b})}{I(W_{a,b})}
&= \left(\frac{\operatorname{Per}(G_{a,b})}{\operatorname{Per}(W_{a,b})}\right)^2 \frac{\operatorname{Area}(W_{a,b})}{\operatorname{Area}(G_{a,b})}\\
&\ge\left(\frac{\operatorname{Per}(G_{a,b})}{\operatorname{Per}(W_{a,b})}\right)^2\\
&> \left(\frac{\frac{8}{15}(1+a^2)(5+3b)}{2\pi(1-b)+8b}\right)^2 = R(a,b)^2.
\end{aligned} I ( W a , b ) I ( G a , b ) = ( Per ( W a , b ) Per ( G a , b ) ) 2 Area ( G a , b ) Area ( W a , b ) ≥ ( Per ( W a , b ) Per ( G a , b ) ) 2 > ( 2 π ( 1 − b ) + 8 b 15 8 ( 1 + a 2 ) ( 5 + 3 b ) ) 2 = R ( a , b ) 2 .
This proves (39) . If R ( a , b ) ≥ 1 R(a,b) \ge1 R ( a , b ) ≥ 1 , the strict inequality implies I ( G a , b ) > I ( W a , b ) I(G_{a,b}) > I(W_{a,b}) I ( G a , b ) > I ( W a , b ) . □
Proof. of Theorem 4.4. The original domain. Since i a ∈ D ia \in\mathbb{D} ia ∈ D , the translation in (25) puts 0 0 0 inside W a , b W_{a,b} W a , b . If b < 1 b < 1 b < 1 , then
R e F b ′ ( z ) = R e ( 1 + b z 2 ) ≥ 1 − b > 0 ( z ∈ D ‾ ) . \mathop{\mathrm{Re}} F_b'(z)=\mathop{\mathrm{Re}}(1+bz^2)\ge1-b>0 \qquad(z\in\overline{\mathbb{D}}). Re F b ′ ( z ) = Re ( 1 + b z 2 ) ≥ 1 − b > 0 ( z ∈ D ) .
For distinct z , w ∈ D ‾ z,w\in\overline{\mathbb{D}} z , w ∈ D , integration along their joining segment shows that
R e F b ( z ) − F b ( w ) z − w = ∫ 0 1 R e F b ′ ( w + s ( z − w ) ) d s > 0. \mathop{\mathrm{Re}}\frac{F_b(z)-F_b(w)}{z-w}=\int_0^1\mathop{\mathrm{Re}}F_b'(w+s(z-w))\,\mathrm{d}s>0. Re z − w F b ( z ) − F b ( w ) = ∫ 0 1 Re F b ′ ( w + s ( z − w )) d s > 0.
Thus F b F_b F b is injective on D ‾ \overline{\mathbb{D}} D and its derivative is nonvanishing there. When b = 1 b=1 b = 1 , equality F 1 ( z ) = F 1 ( w ) F_1(z)=F_1(w) F 1 ( z ) = F 1 ( w ) with z ≠ w z\ne w z = w would imply z 2 + z w + w 2 = − 3 z^2+zw+w^2=-3 z 2 + z w + w 2 = − 3 . Equality in the triangle inequality then forces z 2 = z w = w 2 = − 1 z^2=zw=w^2=-1 z 2 = z w = w 2 = − 1 , and hence z = w z=w z = w , a contradiction. Also F 1 ′ F_1' F 1 ′ does not vanish in D \mathbb{D} D . This proves the Jordan and conformal assertions for the original domains. Translation does not affect length, and
∣ 1 + b e 2 i t ∣ = ∣ ( 1 − b ) + b ( 1 + e 2 i t ) ∣ ≤ ( 1 − b ) + 2 b ∣ cos t ∣ . \left|1+be^{2it}\right|=\left|(1-b)+b(1+e^{2it})\right|\le(1-b)+2b|\cos t|. 1 + b e 2 i t = ( 1 − b ) + b ( 1 + e 2 i t ) ≤ ( 1 − b ) + 2 b ∣ cos t ∣.
Integration proves (27) , with equality 8 8 8 when b = 1 b=1 b = 1 .
The exit law and its quantile. The map F b − F b ( i a ) F_b-F_b(ia) F b − F b ( ia ) sends i a ia ia to 0 0 0 . Conformal invariance of planar Brownian motion therefore identifies the exit law from W a , b W_{a,b} W a , b with the pushforward of harmonic measure at i a ia ia in the disc; see [6 , 10 , 11 ] . The corresponding angular density is the Poisson kernel
p a ( t ) = 1 − a 2 2 π ( 1 + a 2 − 2 a sin t ) . p_a(t)=\frac{1-a^2}{2\pi(1+a^2-2a\sin t)}. p a ( t ) = 2 π ( 1 + a 2 − 2 a sin t ) 1 − a 2 .
The real boundary coordinate is
r b ( t ) = R e F b ( e i t ) = cos t + b 3 cos ( 3 t ) = ( 1 − b ) cos t + 4 b 3 cos 3 t . (46) r_b(t)=\mathop{\mathrm{Re}}F_b(e^{it})=\cos t+\frac{b}{3}\cos(3t)=(1-b)\cos t+\frac{4b}{3}\cos^3t. \tag*{(46)} r b ( t ) = Re F b ( e i t ) = cos t + 3 b cos ( 3 t ) = ( 1 − b ) cos t + 3 4 b cos 3 t . (46)
It is even and strictly decreasing on [ 0 , π ] [0,\pi] [ 0 , π ] . Indeed,
r b ′ ( t ) = − sin t ( 1 − b + 4 b cos 2 t ) , r_b'(t)=-\sin t(1-b+4b\cos^2t), r b ′ ( t ) = − sin t ( 1 − b + 4 b cos 2 t ) ,
which is negative on ( 0 , π ) (0,\pi) ( 0 , π ) for b < 1 b<1 b < 1 ; for b = 1 b=1 b = 1 the function 4 3 cos 3 t \frac{4}{3}\cos^{3}t 3 4 cos 3 t is still strictly decreasing. Folding the angles t t t and − t -t − t gives, for 0 < t < π 0<t<\pi 0 < t < π ,
p a ( t ) + p a ( − t ) = k π ( cos 2 t + k 2 sin 2 t ) . (47) p_a(t)+p_a(-t)=\frac{k}{\pi(\cos^2 t+k^2\sin^2 t)}. \tag*{(47)} p a ( t ) + p a ( − t ) = π ( cos 2 t + k 2 sin 2 t ) k . (47)
Define the increasing continuous angle θ = θ ( t ) \theta=\theta(t) θ = θ ( t ) on [ 0 , π ] [0,\pi] [ 0 , π ] by
tan θ = k tan t , θ ( 0 ) = 0 , θ ( π / 2 ) = π / 2. (48) \tan\theta=k\tan t,\qquad\theta(0)=0,\qquad\theta(\pi/2)=\pi/2. \tag*{(48)} tan θ = k tan t , θ ( 0 ) = 0 , θ ( π /2 ) = π /2. (48)
The continuous branch is used across π / 2 \pi/2 π /2 . Since
θ ′ ( t ) = k cos 2 t + k 2 sin 2 t , \theta'(t)=\frac{k}{\cos^2 t+k^2\sin^2 t}, θ ′ ( t ) = cos 2 t + k 2 sin 2 t k ,
(47) becomes d θ / π d\theta/\pi d θ / π . Therefore the decreasing angular quantile is
q a , b ( θ ) = ( 1 − b ) C k ( θ ) + 4 b 3 C k ( θ ) 3 , C k ( θ ) = cos θ cos 2 θ + k − 2 sin 2 θ . (49) q_{a,b}(\theta)=(1-b)C_k(\theta)+\frac{4b}{3}C_k(\theta)^3,\qquad C_k(\theta)=\frac{\cos\theta}{\sqrt{\cos^2\theta+k^{-2}\sin^2\theta}}. \tag*{(49)} q a , b ( θ ) = ( 1 − b ) C k ( θ ) + 3 4 b C k ( θ ) 3 , C k ( θ ) = cos 2 θ + k − 2 sin 2 θ cos θ . (49)
Its even periodic extension is real-analytic (in particular C 3 C^3 C 3 ), strictly decreasing on [ 0 , π ] [0,\pi] [ 0 , π ] , and satisfies
q a , b ( π − θ ) = − q a , b ( θ ) . (50) q_{a,b}(\pi-\theta)=-q_{a,b}(\theta). \tag*{(50)} q a , b ( π − θ ) = − q a , b ( θ ) . (50)
The underlying law is consequently centered and symmetric, with support [ − 1 − b / 3 , 1 + b / 3 ] [-1-b/3,1+b/3] [ − 1 − b /3 , 1 + b /3 ] .
Let g = g a , b g=g_{a,b} g = g a , b be the analytic completion of q = q a , b q=q_{a,b} q = q a , b . Lemma 4.3 shows that G a , b = g ( D ) G_{a,b}=g(\mathbb{D}) G a , b = g ( D ) is a bounded rectifiable Jordan domain. By construction, its real exit law agrees with that of W a , b W_{a,b} W a , b . Equation (50) makes every even-indexed coefficient of g g g vanish. Thus g g g is odd and g ′ g' g ′ is even and zero-free.
A lower bound for one derivative value. Pairing t t t and π − t \pi-t π − t in (45) , and using q ′ ( π − t ) = q ′ ( t ) q'(\pi-t)=q'(t) q ′ ( π − t ) = q ′ ( t ) , gives
g ′ ( a ) = − 4 ( 1 + a 2 ) π ∫ 0 π / 2 q ′ ( t ) sin t ( 1 − 2 a cos t + a 2 ) ( 1 + 2 a cos t + a 2 ) d t . (51) g'(a)=-\frac{4(1+a^2)}{\pi}\int_0^{\pi/2}\frac{q'(t)\sin t}{(1-2a\cos t+a^2)(1+2a\cos t+a^2)}\,dt. \tag*{(51)} g ′ ( a ) = − π 4 ( 1 + a 2 ) ∫ 0 π /2 ( 1 − 2 a cos t + a 2 ) ( 1 + 2 a cos t + a 2 ) q ′ ( t ) sin t d t . (51)
Use the substitution t = arctan ( k x ) t=\arctan(kx) t = arctan ( k x ) . Then
q ( arctan ( k x ) ) = 1 − b 1 + x 2 + 4 b 3 ( 1 + x 2 ) 3 / 2 , q(\arctan(kx))=\frac{1-b}{\sqrt{1+x^2}}+\frac{4b}{3(1+x^2)^{3/2}}, q ( arctan ( k x )) = 1 + x 2 1 − b + 3 ( 1 + x 2 ) 3/2 4 b ,
q ′ ( t ) d t = − x [ 1 − b ( 1 + x 2 ) 3 / 2 + 4 b ( 1 + x 2 ) 5 / 2 ] d x , q'(t)\,dt=-x\left[\frac{1-b}{(1+x^2)^{3/2}}+\frac{4b}{(1+x^2)^{5/2}}\right]dx, q ′ ( t ) d t = − x [ ( 1 + x 2 ) 3/2 1 − b + ( 1 + x 2 ) 5/2 4 b ] d x ,
sin t = k x 1 + k 2 x 2 , \sin t=\frac{kx}{\sqrt{1+k^2x^2}}, sin t = 1 + k 2 x 2 k x ,
( 1 − 2 a cos t + a 2 ) ( 1 + 2 a cos t + a 2 ) = ( 1 − a 2 ) 2 cos 2 t + ( 1 + a 2 ) 2 sin 2 t = ( 1 − a 2 ) 2 1 + x 2 1 + k 2 x 2 . \begin{aligned}
(1-2a\cos t+a^2)(1+2a\cos t+a^2)=(1-a^2)^2\cos^2t+(1+a^2)^2\sin^2t \\
&=(1-a^2)^2\frac{1+x^2}{1+k^2x^2}.
\end{aligned} ( 1 − 2 a cos t + a 2 ) ( 1 + 2 a cos t + a 2 ) = ( 1 − a 2 ) 2 cos 2 t + ( 1 + a 2 ) 2 sin 2 t = ( 1 − a 2 ) 2 1 + k 2 x 2 1 + x 2 .
Substitution into (51) , together with ( 1 + a 2 ) k = 1 − a 2 (1+a^2)k=1-a^2 ( 1 + a 2 ) k = 1 − a 2 , yields the exact identity
g ′ ( a ) = 4 π ( 1 − a 2 ) ∫ 0 ∞ x 2 1 + k 2 x 2 [ 1 − b ( 1 + x 2 ) 5 / 2 + 4 b ( 1 + x 2 ) 7 / 2 ] d x . (52) g'(a)=\frac{4}{\pi(1-a^2)}\int_0^\infty x^2\sqrt{1+k^2x^2}\left[\frac{1-b}{(1+x^2)^{5/2}}+\frac{4b}{(1+x^2)^{7/2}}\right]\,dx. \tag*{(52)} g ′ ( a ) = π ( 1 − a 2 ) 4 ∫ 0 ∞ x 2 1 + k 2 x 2 [ ( 1 + x 2 ) 5/2 1 − b + ( 1 + x 2 ) 7/2 4 b ] d x . (52)
All terms are nonnegative, and 1 + k 2 x 2 > 1 \sqrt{1+k^2x^2}>1 1 + k 2 x 2 > 1 for x > 0 x>0 x > 0 . The substitution u = x / 1 + x 2 u=x/\sqrt{1+x^2} u = x / 1 + x 2 gives
∫ 0 ∞ x 2 ( 1 + x 2 ) 5 / 2 d x = ∫ 0 1 u 2 d u = 1 3 , \int_0^\infty\frac{x^2}{(1+x^2)^{5/2}}\,dx=\int_0^1u^2\,du=\frac{1}{3}, ∫ 0 ∞ ( 1 + x 2 ) 5/2 x 2 d x = ∫ 0 1 u 2 d u = 3 1 ,
∫ 0 ∞ x 2 ( 1 + x 2 ) 7 / 2 d x = ∫ 0 1 u 2 ( 1 − u 2 ) d u = 2 15 . \int_0^\infty\frac{x^2}{(1+x^2)^{7/2}}\,dx=\int_0^1u^2(1-u^2)\,du=\frac{2}{15}. ∫ 0 ∞ ( 1 + x 2 ) 7/2 x 2 d x = ∫ 0 1 u 2 ( 1 − u 2 ) d u = 15 2 .
It follows that
g ′ ( a ) > 4 π ( 1 − a 2 ) ( 1 − b 3 + 8 b 15 ) = 4 ( 5 + 3 b ) 15 π ( 1 − a 2 ) . (53) g'(a)>\frac{4}{\pi(1-a^2)}\left(\frac{1-b}{3}+\frac{8b}{15}\right)=\frac{4(5+3b)}{15\pi(1-a^2)}. \tag*{(53)} g ′ ( a ) > π ( 1 − a 2 ) 4 ( 3 1 − b + 15 8 b ) = 15 π ( 1 − a 2 ) 4 ( 5 + 3 b ) . (53)
Apply (24) with r = a r=a r = a to conclude
Per ( G a , b ) ≥ 2 π ( 1 − a 4 ) g ′ ( a ) > 8 15 ( 1 + a 2 ) ( 5 + 3 b ) , \operatorname{Per}(G_{a,b})\ge2\pi(1-a^4)g'(a)>\frac{8}{15}(1+a^2)(5+3b), Per ( G a , b ) ≥ 2 π ( 1 − a 4 ) g ′ ( a ) > 15 8 ( 1 + a 2 ) ( 5 + 3 b ) ,
which proves (28) .
Analytic boundaries when b < 1 b<1 b < 1 . The assertion for W a , b W_{a,b} W a , b was established above. The real-analytic periodic function q q q has exponentially decreasing Fourier coefficients, so g g g extends holomorphically across T \mathbb{T} T . For 0 < t < π 0<t<\pi 0 < t < π , strict negativity of q ′ ( t ) q'(t) q ′ ( t ) implies
ℜ ( i e i t g ′ ( e i t ) ) = g ′ ( t ) < 0 ; \Re\left(ie^{it}g'(e^{it})\right)=g'(t)<0; ℜ ( i e i t g ′ ( e i t ) ) = g ′ ( t ) < 0 ;
hence the boundary derivative does not vanish on the upper arc, or on the lower arc by reflection. At 1 1 1 , let r ↑ 1 r\uparrow1 r ↑ 1 in (45) . Since q ′ ( t ) = O ( sin t ) q'(t)=O(\sin t) q ′ ( t ) = O ( sin t ) at the endpoints, dominated convergence is applicable and gives
g ′ ( 1 ) = − 1 π ∫ 0 π q ′ ( t ) cot ( t / 2 ) d t > 0. g'(1)=-\frac{1}{\pi}\int_0^\pi q'(t)\cot(t/2)\,dt>0. g ′ ( 1 ) = − π 1 ∫ 0 π q ′ ( t ) cot ( t /2 ) d t > 0.
For example, for r ≥ 1 / 2 r\ge1/2 r ≥ 1/2 the denominator in (45) is at least 1 − cos t 1-\cos t 1 − cos t , which supplies a bounded majorant near 0 0 0 . The evenness of g ′ g' g ′ gives g ′ ( − 1 ) = g ′ ( 1 ) > 0 g'(-1)=g'(1)>0 g ′ ( − 1 ) = g ′ ( 1 ) > 0 . Thus g ′ g' g ′ is nonvanishing on the entire circle. Together with boundary injectivity, this proves that ∂ G a , b \partial G_{a,b} ∂ G a , b is a regular real-analytic Jordan curve.