The Split-Quaternion Quadratic Family and Its Julia Sets

Introduction

The quadratic map of the split-quaternion algebra is again $\tilde q \mapsto \tilde q^2+c$, and again the algebra determines the phenomena: the product is associative, so the iterates exist, but the norm is indefinite, of inertia $(2,2)$, and the elements of norm zero are zero divisors. The consequences are more violent than in the definite algebra. The escape lemma of the quaternion case fails, because the Euclidean norm is not bounded below by a positive constant times the product of the norms — the null cone collapses products — and with it the escape radius fails: for $c=0$ the whole null cone of the vector subspace consists of points that square to zero, so the filled Julia set contains a double cone of arbitrarily large Euclidean radius. At the other extreme the filled Julia set can be empty, which never happens in the definite case. What survives is the symmetry, now the Lorentz group of the vector subspace, and the reduction of the critical orbit to the two-dimensional subalgebra generated by $1$ and $c$, which is the complex plane for a timelike parameter, a split-complex plane for a spacelike one, and the dual numbers on the null cone.

This article defines the family, proves the failure of the escape radius, exhibits the bounded set of $c=0$ and the empty slice of $c=2\vec e_y$, determines the symmetry group and states precisely which group it is, and reduces the critical orbit to the complex or split-complex slice. It is the first article of the split-quaternion fractal thread.

The split-quaternion algebra, its conjugation, its norm, its idempotents and its subalgebras are from Split-Quaternion Algebra, Split-Quaternion Norm and Invertibility, Split-Quaternion Ideals and Peirce Decomposition and Split-Quaternion Subspaces and the Involutions; the adjoint action, the Lorentz group and the isomorphism $\operatorname{PSL}_2(\mathbb{R})\cong\operatorname{SO}^{+}(2,1)$ are from Split-Quaternion Rotations and the Lorentz Group; the hyperbolic model is from Split-Quaternions and Hyperbolic Geometry; the complex quadratic family is from The Julia Sets of a Complex Polynomial and the split-complex one from The Split-Complex Julia Sets; and the dimensions are Fractal Geometry's. The quaternion thread it neighbours is The Quaternion Quadratic Map and Its Julia Sets and its companions. No physics is invoked.

Throughout, $\tilde q=q_te_t+q_x\vec e_x+q_y\vec e_y+q_z\vec e_z$ with $\vec e_x^2=-e_t$ and $\vec e_y^2=\vec e_z^2=e_t$, $\vec e_x \vec e_y=\vec e_z=-\vec e_y \vec e_x$; the conjugation is $\tilde q^{\natural}=q_te_t-q_x\vec e_x-q_y\vec e_y-q_z\vec e_z$, the norm is $\langle\tilde q\rangle_{\natural}=\tilde q\tilde q^{\natural}=q_t^2+q_x^2-q_y^2-q_z^2$, and $|\tilde q|$ is the Euclidean norm $\sqrt{q_t^2+q_x^2+q_y^2+q_z^2}$, the positive-definite norm of the underlying space $\mathbb{R}^4$, kept distinct from $\langle\cdot\rangle_{\natural}$. The null cone is $\{\tilde q: \langle\tilde q\rangle_{\natural}=0\}$, the set of zero divisors. The scalar line is $S=\mathbb{R}e_t$ and the vector subspace is $V=\operatorname{span}\{\vec e_x,\vec e_y,\vec e_z\}$.

The Family and the Indefinite Geometry

Definition. The split-quaternion quadratic map of $c\in\mathbb{H}_{\mathrm{spl}}$ is $f_c(\tilde q)=\tilde q^2+c$, with orbit $\tilde q_n=f_c^{\,n}(\tilde q)$, and the filled Julia set and Julia set are

$$ K_c=\{\tilde q \in \mathbb{H}_{\mathrm{spl}} : \text{the orbit is bounded in } \mathbb{R}^4\} , \qquad J_c=\partial K_c , $$

the boundedness being with respect to the Euclidean norm $|\cdot|$.

Remark (the definition is by boundedness, not by a radius). In the definite algebra the two definitions agree, because a radius exists beyond which every point escapes. Here they do not agree, and the definition by a radius is not available; this is the content of the next section. The filled Julia set is therefore defined as the set of points whose orbit is bounded, and the escape is not detected by a norm test.

Proposition (the critical set). The differential of $f_c$ at $\tilde q$ is $\tilde q_2\mapsto\tilde q\tilde q_2+\tilde q_2\tilde q$, and it is singular if and only if $q_t=0$ or $\langle\tilde q\rangle_{\natural}=0$; equivalently the critical set is the union of the vector subspace $V$ and the null cone,

$$ \boxed{\; \mathcal{C}=\{\tilde q : q_t=0\}\cup\{\tilde q : \langle\tilde q\rangle_{\natural}=0\} \;}. $$

Proof. Under the matrix model $\mathbb{H}_{\mathrm{spl}}\cong M_2(\mathbb{R})$ the differential is $h\mapsto qh+hq$; if $q$ has eigenvalues $\lambda_1,\lambda_2$ then this map has eigenvalues $\lambda_i+\lambda_j$ on the basis $v_iv_j^{\!\top}$, so it is singular exactly when some pair of eigenvalues sums to zero. For real eigenvalues this means $\lambda_1+\lambda_2=2q_t=0$, or $\lambda_1\lambda_2=\langle\tilde q\rangle_{\natural}=0$; for a conjugate complex pair it means $2\operatorname{Re}\lambda=2q_t=0$. Hence the condition is $q_t=0$ or $\langle\tilde q\rangle_{\natural}=0$. The verification at the nilpotent $\vec e_x+\vec e_y$ (kernel containing $\vec e_x$) and at the invertible element $\vec e_y+\vec e_z$ (kernel containing $\vec e_x$) is recorded in the companion file.

Remark. The critical set of the split-quaternion quadratic map is the three-dimensional subspace $V$ together with the three-dimensional null cone, and not the single point $0$; there is no unique critical point, and the connectedness locus cannot be tested by iterating one point. The natural substitute is the critical orbit of the subalgebra $\mathbb{R}[c]$, which does have a distinguished starting point, and it is the one used in The Split-Quaternion Mandelbrot Set. The whole subspace $V$ is critical because a purely vector element has trace zero, and the whole null cone is critical because a zero divisor has vanishing determinant.

The Failure of the Escape Radius

Proposition (the Euclidean norm is not submultiplicative from below). There is no constant $\kappa>0$ with $|\tilde q_1\tilde q_2|\geq\kappa|\tilde q_1||\tilde q_2|$ for all $\tilde q_1,\tilde q_2\in\mathbb{H}_{\mathrm{spl}}$.

Proof. Take $\tilde q_1=e_t+\vec e_y$ and $\tilde q_2=e_t-\vec e_y$; the multiplication table gives $\tilde q_1\tilde q_2=0$, while $|\tilde q_1|=|\tilde q_2|=\sqrt2$. Hence $|\tilde q_1\tilde q_2|=0$ while $|\tilde q_1||\tilde q_2|=2$, and no positive constant $\kappa$ exists. (The opposite inequality with a constant does hold, since the product is a bounded bilinear map of a finite-dimensional space; it is the lower bound that fails, and the lower bound is the one the escape lemma needs.)

Theorem (no escape radius). For $c=0$ the filled Julia set is unbounded: every element of the null cone of the vector subspace, $\{\tilde q \in V: \langle\tilde q\rangle_{\natural}=0\}$, lies in $K_0$, and this set is a double cone of arbitrarily large Euclidean radius.

Proof. For $\tilde q\in V$ one has $\tilde q^2=2q_t\tilde q-\langle\tilde q\rangle_{\natural}e_t=-\langle\tilde q\rangle_{\natural}e_t$; if in addition $\langle\tilde q\rangle_{\natural}=0$ then $\tilde q^2=0$, so $f_0(\tilde q)=0$ and the orbit is $\tilde q\mapsto0\mapsto0\mapsto\cdots$, bounded. The set $\{\tilde q\in V:\langle\tilde q\rangle_{\natural}=0\}$ is the double cone $q_x^2=q_y^2+q_z^2$, which contains $t(\vec e_x+\vec e_y)$ for every $t$ and is unbounded. Hence $K_0$ is unbounded and contains points of arbitrarily large Euclidean radius.

Corollary (the escape-radius test is unavailable). In the split-quaternion family there is no radius $R$ with the property that $|\tilde q_n|>R$ implies escape.

Proof. For $c=0$ the point $t(\vec e_x+\vec e_y)$ has bounded orbit and Euclidean norm $\sqrt2\,|t|$, unbounded in $t$; so no radius separates the bounded from the unbounded orbits of $f_0$.

Remark (what the escape is tested on). The menu's phrase "the escape is tested on the quadratic form and not on the norm" records the direction of the difficulty but not its resolution: the quadratic form $\langle\cdot\rangle_{\natural}$ is not sound either, because $|\langle\tilde q\rangle_{\natural}|\leq|\tilde q|^2$ and the equality is approached on the null cone, so $\langle\cdot\rangle_{\natural}$ can vanish on points of large Euclidean norm that are bounded, exactly as in the theorem. The honest definition is the boundedness of the orbit itself, and no single scalar quantity in the algebra detects the escape. The consequences are carried into The Split-Quaternion Mandelbrot Set and the rest of the thread.

The Empty Case and the Two Extremes

Proposition (an empty invariant slice). Let $c=2\vec e_y$. The critical orbit escapes and the intersection of $K_c$ with the split-complex plane $\mathbb{C}_{\mathrm{spl},2}=\operatorname{span}\{e_t,\vec e_y\}$ is empty.

Proof. The plane $\mathbb{C}_{\mathrm{spl},2}$ is the subalgebra generated by $1$ and $\vec e_y$, and $c\in\mathbb{C}_{\mathrm{spl},2}$; the critical orbit lies in it and is the orbit of the split-complex quadratic map with the parameter $2\vec e_y$. Under the idempotent decomposition of the next section the two real components of the parameter are $2$ and $-2$, and the real quadratic map $u\mapsto u^2+2$ has no bounded orbit, so the component of every orbit in the corresponding idempotent direction escapes; hence no point of $\mathbb{C}_{\mathrm{spl},2}$ has a bounded orbit.

Remark. A numerical search over the ball of radius $3$ finds no bounded orbit for $c=2\vec e_y$, so the filled Julia set is either empty or extremely thin; the example establishes that the invariant slice is empty, which cannot happen in the definite algebra and which shows that the family has a genuinely different connectedness theory. The comparison with the definite case is developed in the neighbouring articles.

The Symmetry

Proposition (equivariance, and the correct group). For $\tilde u$ an invertible element of the algebra and $\operatorname{Ad}_{\tilde u}(\tilde q)=\tilde u\tilde q\tilde u^{-1}$,

$$ \boxed{\; f_{\operatorname{Ad}_{\tilde u}(c)}\circ\operatorname{Ad}_{\tilde u}=\operatorname{Ad}_{\tilde u}\circ f_c \;}. $$

The inner automorphisms form the group of all automorphisms of the algebra over $\mathbb{R}$, isomorphic to $\operatorname{PSL}_2(\mathbb{R})\cong\operatorname{SO}^{+}(2,1)$, acting on the vector subspace by the adjoint representation of Split-Quaternion Rotations and the Lorentz Group. Consequently $K_{\operatorname{Ad}_{\tilde u}c}=\operatorname{Ad}_{\tilde u}K_c$.

Proof. An algebra automorphism preserves squares, so $\operatorname{Ad}_{\tilde u}(\tilde q^2)=(\operatorname{Ad}_{\tilde u}\tilde q)^2$; adding $\operatorname{Ad}_{\tilde u}(c)$ gives the identity. The group of automorphisms is the inner group because $\mathbb{H}_{\mathrm{spl}}\cong M_2(\mathbb{R})$ is central simple, and the rest is the theorem on the adjoint representation of Split-Quaternion Rotations and the Lorentz Group.

Remark (the menu's $SO(2,2)$). The isometry group of the norm form $\langle\cdot\rangle_{\natural}$ on $\mathbb{R}^{2,2}$ is $O(2,2)$, and the menu names $SO(2,2)$ as the invariance group acting by the adjoint representation. The group that actually intertwines the iteration is strictly smaller, namely the automorphism group $\operatorname{Inn}\cong \operatorname{PSL}_2(\mathbb{R})\cong\operatorname{SO}^{+}(2,1)$. An element of $O(2,2)$ need not preserve the product — the map $\tilde q\mapsto g\tilde q h^{-1}$ with $g,h\in\operatorname{SL}_2(\mathbb{R})$ preserves the norm, being the determinant, but satisfies $(g\tilde qh^{-1})^2=g\tilde qh^{-1}g\tilde qh^{-1}$, which is not $g\tilde q^2h^{-1}$ unless $g$ and $h$ commute with $\tilde q$; the orbit of the family under $O(2,2)$ is not one dynamical object. Only the automorphisms act on the parameter and the point simultaneously, so the equivariance group is the Lorentz group $SO^{+}(2,1)$ of the vector subspace, and the article states it.

Remark (the anti-automorphisms). The conjugation ${}^{\natural}$ and the reversal $\rho$ of Split-Quaternion Algebra are anti-automorphisms, and they too satisfy $\rho(\tilde q^2)=\rho(\tilde q)^2$, so they commute with $f_c$ whenever they fix $c$: the conjugation fixes a real parameter, and the reversal fixes the points of the plane $\operatorname{span}\{\vec e_x,\vec e_y\}$ by Split-Quaternion Subspaces and the Involutions. The full symmetry of a member is generated by the automorphisms and the anti-automorphisms fixing $c$, an enlargement of the Lorentz group by the reflections it does not contain.

The Slice Reductions

Proposition (the parameter's subalgebra). Let $\mathbb{R}[c]=\operatorname{span}\{e_t,c\}$. Then $\mathbb{R}[c]$ is a subalgebra, the critical orbit lies in it, and its isomorphism type is decided by the sign of $\langle \vec c\rangle_{\natural}$: it is a copy of $\mathbb{C}$ if $\langle \vec c\rangle_{\natural}>0$ (the vector part timelike), a copy of the split-complex algebra $\mathbb{C}_{\mathrm{spl}}$ if $\langle \vec c\rangle_{\natural}<0$ (the vector part spacelike), and the algebra of dual numbers if $\langle \vec c\rangle_{\natural}=0$ (the vector part null).

Proof. The span of $1$ and $c$ is closed under multiplication; the critical orbit starts at $0$ and stays in it as in the definite case. The minimal polynomial of $c$ is $x^2-2c_tx+\langle c\rangle_{\natural}$, with $\langle c\rangle_{\natural}=c_t^2+\langle \vec c\rangle_{\natural}$, so its discriminant is $4c_t^2-4\langle c\rangle_{\natural}=-4\langle\vec c\rangle_{\natural}$; the three cases correspond to a negative discriminant (complex roots, the algebra $\mathbb{C}$), a positive one (two real roots, the algebra $\mathbb{C}_{\mathrm{spl}}$), and a zero one (a double root, the dual numbers).

Proposition (the complex and split-complex slices). On $\mathbb{R}[c]$ the restriction of $f_c$ is the quadratic map of the subalgebra, conjugated to $z\mapsto z^2+\kappa$ in the complex case and to the split-complex map $A\mapsto A^2+C$ in the split-complex case, where the conjugate parameter is the image of $c$ under the isomorphism.

Proof. The restriction is a polynomial map with coefficients in the subalgebra, and the isomorphism carries it to the standard quadratic map of the target algebra, as in the definite case.

Proposition (the two real orbits of the idempotent decomposition). Let $c\in\mathbb{C}_{\mathrm{spl},2}=\operatorname{span}\{e_t,\vec e_y\}$ and let $\tilde\pi_{1,2}=\tfrac12(e_t\pm \vec e_y)$ be the idempotents, with $A=A_+\tilde\pi_1+A_-\tilde\pi_2$ for $A\in\mathbb{C}_{\mathrm{spl},2}$. Then the split-complex orbit of a point of $\mathbb{C}_{\mathrm{spl},2}$ is the pair of real orbits

$$ \boxed{\; A_{\pm}^{(n+1)}=\bigl(A_\pm^{(n)}\bigr)^2+C_\pm , \qquad A_\pm^{(0)}=\text{the components of } A , \quad C_\pm=\text{the components of } c \;}. $$

Thus a split-complex quadratic orbit is two real quadratic orbits, one in each idempotent direction; and in the complex case $\langle \vec c\rangle_{\natural}>0$ the corresponding statement is a single complex orbit.

Proof. The idempotents satisfy $\tilde\pi_1^2=\tilde\pi_1$, $\tilde\pi_2^2=\tilde\pi_2$ and $\tilde\pi_1\tilde\pi_2=0$ by Split-Quaternion Ideals and Peirce Decomposition, and the algebra $\mathbb{C}_{\mathrm{spl},2}$ is the direct sum $\mathbb{R}\tilde\pi_1\oplus\mathbb{R}\tilde\pi_2$; squaring is componentwise in this basis, $(A_+\tilde\pi_1+A_-\tilde\pi_2)^2=A_+^2\tilde\pi_1+A_-^2\tilde\pi_2$, and adding $c$ adds the components. The direct computation is recorded in the companion file. The complex case has no real idempotents, and its orbit is the single complex orbit.

Remark (no decomposition in the full algebra). The idempotent decomposition is a statement about the split-complex subalgebra and not about the algebra: for a general $\tilde q$ the components $\tilde q\tilde\pi_{1,2}$ do not multiply independently, since $\tilde q\tilde\pi_1\cdot\tilde q\tilde\pi_2=\tilde q\tilde\pi_1\tilde q\tilde\pi_2$ need not vanish, and the direct computation gives the $\vec e_x$ and $\vec e_z$ components of $\tilde q^2$ as $2q_tq_x$ and $2q_tq_z$, which vanish only on the coordinate planes. The decomposition of the menu is therefore the split-complex one and is not available for the whole family; it is the correct tool exactly when the parameter and the point lie in the same split-complex plane, and it is the tool used in the next article for the spacelike connectedness locus.

Comparison with the Neighbouring Families

Remark (the definite and the split-complex comparisons). The quaternion family has a positive-definite norm, a single critical point, an escape radius and a connectedness locus that is the solid of revolution of the complex Mandelbrot set; none of these holds here. The split-complex family has a two-dimensional algebra, the two real orbits of the idempotent decomposition, and a connectedness locus that is a parallelogram; it is the subalgebra theory of the spacelike parameters, and it is the subject of The Split-Complex Quadratic Family and The Split-Complex Julia Sets. The split-quaternion family is the four-dimensional indefinite one that contains both: its timelike parameters behave as the complex family, its spacelike parameters as the split-complex family, and its null parameters as the dual-number degeneration. The three are compared throughout the thread, and the definite quaternion case is the comparison of The Quaternion Quadratic Map and Its Julia Sets.

Worked Example

Example (three parameters). The computations are recorded in the companion file.

(a) The parameter $c=0$. The null cone of $V$ lies in $K_0$: the point $\vec e_x+\vec e_y$ has $N=0$ and squares to $0$, and $t(\vec e_x+\vec e_y)$ has Euclidean norm $\sqrt2|t|$ and bounded orbit for every $t$. So $K_0$ is unbounded and contains the double cone.

(b) The parameter $c=2\vec e_y$. The split-complex components are $C_+=2$ and $C_-=-2$; the real map $u\mapsto u^2+2$ has empty filled Julia set, so the invariant slice is empty; the critical orbit escapes, and the numerical search finds no bounded orbit.

(c) The parameter $c=\vec e_x$. Then $\langle \vec c\rangle_{\natural}=1>0$, the subalgebra $\mathbb{R}[\vec e_x]=\operatorname{span}\{e_t,\vec e_x\}\cong\mathbb{C}$, and the critical orbit is the complex orbit of $z\mapsto z^2+i$; it is bounded, and $\vec e_x$ lies in the connectedness locus.

(d) The parameter $c=\vec e_y$. Then $\langle \vec c\rangle_{\natural}=-1<0$, the subalgebra is the split-complex plane $\mathbb{C}_{\mathrm{spl},2}$, and the idempotent components are $C_+=1$, $C_-=-1$; the real map $u\mapsto u^2+1$ has no bounded orbit, since $u^2+1>|u|$ for every real $u$, so the critical orbit escapes in the $+$ direction and $K_{\vec e_y}\cap\mathbb{C}_{\mathrm{spl},2}$ is empty. The parameter is outside the connectedness locus.

Summary

The split-quaternion quadratic map $f_c(\tilde q)=\tilde q^2+c$ is the pattern of the quaternion family read in an algebra of inertia $(2,2)$. The critical set is the union of the vector subspace and the null cone, and not the single point $0$. The Euclidean norm is not bounded below by a positive multiple of the product of the norms, because the null cone collapses products, and as a result no escape radius exists: for $c=0$ the double cone $q_x^2=q_y^2+q_z^2$ of the vector subspace squares to zero and lies in an unbounded filled Julia set. The filled Julia set can also be empty, as the invariant slice of $c=2\vec e_y$ shows. The filled Julia set is therefore defined by the boundedness of the orbit and not by a radius, and the escape is tested on the orbit and not on the norm or on the quadratic form. The symmetry is the adjoint action of the automorphism group $\operatorname{PSL}_2(\mathbb{R})\cong\operatorname{SO}^{+}(2,1)$ on the vector subspace; the norm isometry group $O(2,2)$, and the smaller group $SO(2,2)$ of the menu, do not preserve the product and do not intertwine the iteration, and the anti-automorphisms fixing $c$ add the reflections. The critical orbit lies in the subalgebra $\mathbb{R}[c]$, which is $\mathbb{C}$ for a timelike vector part, a split-complex plane for a spacelike one and the dual numbers on the null cone; the split-complex orbit reduces to two real orbits in the idempotent directions, and no such decomposition exists for the whole algebra. The quaternion, split-complex and definite comparisons are the neighbouring articles named above.

Summary of Notation

Symbol Meaning
$f_c(\tilde q)=\tilde q^2+c$ The split-quaternion quadratic map
$\langle\tilde q\rangle_{\natural}=q_t^2+q_x^2-q_y^2-q_z^2$ The indefinite norm, inertia $(2,2)$
$\lvert\cdot\rvert$ The Euclidean norm of $\mathbb{R}^4$, distinct from $\langle\cdot\rangle_{\natural}$
Null cone $\{\tilde q:\langle\tilde q\rangle_{\natural}=0\}$, the zero divisors, the critical set
$K_c$, $J_c=\partial K_c$ Filled Julia set (bounded orbits); Julia set
$\operatorname{Ad}_{\tilde u}$ Inner automorphism; group $\operatorname{PSL}_2(\mathbb{R})\cong\operatorname{SO}^{+}(2,1)$
$\mathbb{R}[c]=\operatorname{span}\{e_t,c\}$ Subalgebra of the parameter and the critical orbit
$\tilde\pi_{1,2}=\tfrac12(e_t\pm \vec e_y)$ Idempotents of the split-complex plane
$C_\pm$, $A_\pm$ Components in the idempotent directions

Further Reading

  • John C. Baez, "The octonions", Bulletin of the American Mathematical Society 39 (2002), 145–205. The split-quaternion algebra as $M_2(\mathbb{R})$, its norm and its zero divisors.
  • Garret Sobczyk, "The hyperbolic number plane", The College Mathematics Journal 26 (1995), 268–280. The split-complex quadratic map and the idempotent decomposition.
  • John W. Milnor, Dynamics in One Complex Variable, 3rd ed. (Princeton, 2006). The complex quadratic family that is the timelike slice.
  • Kenneth Falconer, Fractal Geometry: Mathematical Foundations and Applications, 3rd ed. (Wiley, 2014). The dimensions, cited to Fractal Geometry.
  • Robert Gilmore, Lie Groups, Lie Algebras, and Some of Their Applications (Wiley, 1974). The Lorentz group and $\operatorname{PSL}_2(\mathbb{R})\cong\operatorname{SO}^{+}(2,1)$, cited to Split-Quaternion Rotations and the Lorentz Group.