Completely isometric subspaces of noncommutative $\mathrm{L}^{p}$-spaces and contractive projections

The proof of the first main result relies on the classification of complete isometries of Junge, Ruan and Sherman [JRS05]. Let $P\colon\mathrm{L}^{p}(\mathcal{M})\to\mathrm{L}^{p}(\mathcal{M})$ be a positive contractive projection and suppose that $\Ran P$ is completely isometric to $\mathrm{L}^{p}(\mathcal{N})$. A complete isometry $T\colon\mathrm{L}^{p}(\mathcal{N})\to\Ran P$ admits a factorization of the form $T=w\widetilde{T}$, where $w$ is a partial isometry and $\widetilde{T}$ is a completely positive complete isometry associated with a normal $*$-homomorphism. The main point is to show that the partial-isometry factor $w$, which in general prevents $T$ from being positive, can be absorbed into the von Neumann algebraic part of the factorization. We first use the positivity of $P$, hence the selfadjointness of its range, together with left and right support projections, to show that $w$ is a unitary in the relevant corner. We then consider a positive element of $\Ran P$ having full support and use the bimodule structure of $\widetilde{T}$ to prove that $w$ belongs to the von Neumann algebra generated by the image of the underlying $*$-homomorphism. It follows that $\Ran T=\Ran\widetilde{T}$. Since the latter range is completely positively and completely contractively complemented by the Junge–Ruan–Sherman theorem, the uniqueness of a contractive projection onto a fixed subspace of the smooth Banach space $\mathrm{L}^{p}(\mathcal{M})$ implies that $P\colon\mathrm{L}^{p}(\mathcal{M})\to\mathrm{L}^{p}(\mathcal{M})$ itself is completely positive.

The proof of the second main result is based on a stabilization argument. After amplification by $S^{p}$, a rectangular noncommutative $\mathrm{L}^{p}$-space can be identified completely isometrically with an ordinary noncommutative $\mathrm{L}^{p}$-space. The classification theorem of Junge, Ruan and Sherman then yields a completely contractive projection onto the stabilized copy. The particular form of this projection shows that it is in fact contractively decomposable. By compressing to a matrix corner, we obtain a contractively decomposable projection onto the original subspace.

The paper is organized as follows. In Section 2, we recall some facts on $\mathrm{W}^{*}$-TROs, support projections, properly infinite projections and uniqueness of contractive projections. In Section 3, we discuss completely isometric copies of noncommutative $\mathrm{L}^{p}$-spaces and their complementability. In Section 4, we consider rectangular noncommutative $\mathrm{L}^{p}$-spaces and prove the characterization of the ranges of contractively decomposable projections described previously.

Let $p$ be an orthogonal projection in a von Neumann algebra $\mathcal{M}$. The smallest projection $c(p)$ in the center $\mathrm{Z}(\mathcal{M})$ containing $p$ as a subprojection is called the central support (or central carrier) [Bla06, III.1.1.5 p. 223] (see also [Li92, Definition 1.5.7 p. 34]) of $p$. By [Li92, Proposition 1.5.8 p. 34], if $p$ is a projection in $\mathcal{M}$, and $z$ is a central projection in $\mathcal{M}$. Then

Following [KaR97b, Definition 6.1.4 p. 402], two orthogonal projections $p$ and $q$ in a von Neumann algebra $\mathcal{M}$ are said to be (Murray-von Neumann) equivalent in $\mathcal{M}$, written $p\sim q$, if there exists a partial isometry $w\in\mathcal{M}$ such that $w^{*}w=p$ and $ww^{*}=q$. We say that $p$ is subordinate to $q$, denoted by $p\lesssim q$ if $p$ is equivalent to a subprojection of $q$, i.e., if there is a partial isometry $w\in\mathcal{M}$ with $w^{*}w=p$ and $ww^{*}\leqslant q$.

An orthogonal projection $p$ in a von Neumann algebra $\mathcal{M}$ is called properly infinite if there exist some orthogonal projections $p_{1}$ and $p_{2}$ such that $p\sim p_{1}$, $p\sim p_{2}$, $p_{1}\leqslant p$, $p_{2}\leqslant p$ and $p_{1}\perp p_{2}$. Roughly speaking, a properly infinite projection contains two orthogonal copies of itself.

Following [Bla06, p. 226], a von Neumann algebra $\mathcal{M}$ is said to be locally countably decomposable if there is a family $(z_{i})_{i\in I}$ of mutually orthogonal central projections in $\mathcal{M}$ with $\sum_{i\in I}z_{i}=1_{\mathcal{M}}$ such that each von Neumann algebra $\mathcal{M}z_{i}$ is countably decomposable. The following is [Bla06, Corollary III.1.3.7 p. 229].

Rectangular operator spaces are naturally modeled by off-diagonal corners of von Neumann algebras. The intrinsic objects corresponding to such corners are ternary rings of operators, which we briefly recall. A ternary ring of operators (or simply $\mathrm{TRO}$) is a norm closed subspace $V$ of the space $\mathrm{B}(H,K)$, for some complex Hilbert spaces $H$ and $K$, which is closed under the triple product $(x,y,z)\mapsto xy^{*}z$, see, e.g., [BLM04, 4.4.1 p. 161]. A TRO $V$ is called a $\mathrm{W}^{*}$-$\mathrm{TRO}$ if it is weak* closed in the dual Banach space $\mathrm{B}(H,K)$. A sub-TRO of a TRO $V$ is a closed subspace $W$ of $V$ satisfying $WW^{*}W\subset W$. We refer to [BLM04], [BFT12], [BuT19], [BuT13a], [BuT13b], [DoR07], [EsM23], [Ham92], [Ham99], [KaR02], [NeR03], [PlR19], [SaS13], [SaS17] and [Zet83] for more information on TROs.

Since the results of this paper are formulated in terms of complete isometries, the canonical operator space structure of a TRO will play an essential role. Every TRO admits an operator space structure. Indeed, let us assume that $V$ is a TRO contained in $\mathrm{B}(H,K)$. Then for each integer $n\geqslant 1$, the matrix space $\mathrm{M}_{n}(V)$ can be identified with a TRO contained in $\mathrm{M}_{n}(\mathrm{B}(H,K))\cong\mathrm{B}(H^{n},K^{n})$. This provides a canonical operator space matrix norm on $V$ such that each matrix space $\mathrm{M}_{n}(V)$ is a TRO. By [KaR02, Proposition 2.1 p. 265] or [Ham99, Proposition 2.1 p. 83], the TRO-matrix norms are uniquely determined on each TRO and do not depend on the choice of the representing Hilbert spaces.

A $\mathrm{TRO}$-homomorphism (or triple morphism) is a linear map $T\colon V\to W$ between two $\mathrm{TRO}$s respecting the triple product, i.e.,

If, in addition, $T$ is an injection from $V$ onto $W$, we call $T$ a $\mathrm{TRO}$-isomorphism from $V$ onto $W$. By [EOR01, Proposition 2.4 p. 498], a TRO-isomorphism between $\mathrm{W}^{*}$-TROs is necessarily weak* continuous. By [EOR01, Proposition 2.1 p. 495], every TRO-homomorphism is completely contractive, and every injective $\mathrm{TRO}$-homomorphism is completely isometric. Finally, by [Ham99, Proposition 2.1 p. 83] (see also [BLM04, Corollary 4.4.6 p. 163]), if $T\colon V\to W$ is a linear map between TRO’s then $T$ is a surjective complete isometry if and only if $T$ is a surjective 2-isometry if and only if $T$ is a TRO-isomorphism. Thus, for TROs, the complete isometric structure already determines the triple product.

The linking algebra allows us to pass back and forth between the rectangular setting of TROs and the usual setting of von Neumann algebras. If $\mathcal{M}$ is a von Neumann algebra and $e$ is an orthogonal projection in $\mathcal{M}$, then $e\mathcal{M}e^{\perp}$ is a $\mathrm{W}^{*}$-TRO by Example 2.2. Conversely, if the subspace $V$ of the space $\mathrm{B}(H,K)$ is a $\mathrm{W}^{*}$-TRO, then we can consider the adjoint space $V^{*}\overset{\mathrm{def}}{=}\{x^{*}:x\in V\}$, which is a subspace of the space $\mathrm{B}(K,H)$ and the von Neumann subalgebras

of the algebras $\mathrm{B}(K)$ and $\mathrm{B}(H)$. Since $\mathrm{M}(V)\cdot V\subset V$ and $V\cdot\mathrm{N}(V)\subset V$, we can introduce the von Neumann algebra

which is called the linking von Neumann algebra of $V$, see [KaR02], [Rua04, p. 846] and [WCW24]. Then we have a TRO-isomorphism

where $e\overset{\mathrm{def}}{=}\begin{bmatrix}\mathrm{Id}_{K}&0\\ 0&0\\ \end{bmatrix}$ and $e^{\perp}\overset{\mathrm{def}}{=}\begin{bmatrix}0&0\\ 0&\mathrm{Id}_{H}\\ \end{bmatrix}$. So a $\mathrm{W}^{*}$-TRO $V$ can be identified with the off-diagonal corner at the (1,2) position of its linking von Neumann algebra. By [Rua04, Lemma 2.1 p. 847], the central covers $c(e)$ and $c(e^{\perp})$ of $e$ and $e^{\perp}$ in the linking von Neumann algebra $\mathrm{R}(V)$ are equal to $1$.

Conversely, it is worth noting that, according to [WCW24, Theorem 2.1], a von Neumann algebra is $*$-isomorphic to the linking von Neumann algebra of a $\mathrm{W}^{*}$-TRO if and only if it contains no abelian direct summand.

We next recall the support notation needed in Section 3. If $x$ is an unbounded operator on a Hilbert space, we have

If $x=u|x|$ is the polar decomposition of $x$, then $s_{r}(x)=u^{*}u$ and $s_{\ell}(x)=uu^{*}$. Thus the right and left supports are respectively the initial and final projections of the partial isometry $u$ occurring in the polar decomposition.

Following [Meg98, Definition 5.1.1 p. 426], we say that a normed linear space $X$ is strictly convex (or rotund) if for any $t\in(0,1)$ and any $x,y\in X$ with $\left\|x\right\|_{X}=\left\|y\right\|_{X}=1$ and $x\not=y$ we have $\left\|tx+(1-t)y\right\|_{X}<1$. A normed space $X$ is said to be smooth [Meg98, p. 480] if for any $x\in X$ there exists a unique $x^{*}\in X^{*}$ with $\left\|x^{*}\right\|_{X^{*}}=1$ such that $\langle x^{*},x\rangle=1$. According to [Meg98, Proposition 5.4.7 p. 481], a reflexive normed space is smooth if and only if its dual space is strictly convex.

We refer to [Pat18] and [Cio90] for more information on duality mappings and their relation with the geometry of Banach spaces. Let $X$ be a Banach space. For each $x\in X$, we define the subset

of the dual $X^{*}$. The multivalued map $\mathfrak{j}\colon X\to 2^{X^{*}}$ is called the duality mapping. By the Hahn--Banach theorem¹¹ 1 0. For any $x\in X$, there exists $y^{*}\in X^{*}$ with $\left\|y^{*}\right\|_{X^{*}}=1$ such that $\langle x,y^{*}\rangle_{X,X^{*}}=\left\|x\right\|_{X}$. Using $x^{*}=\left\|x\right\|_{X}y^{*}$, we conclude that $\mathfrak{j}(x)\not=\emptyset$ for each $x\in X$., $\mathfrak{j}(x)$ is nonempty for every $x\in X$, see [Cio90, Remark 4.2 p. 25]. The Banach space $X$ is smooth if and only if $\mathfrak{j}$ is single-valued, see [Cio90, Corollary 4.5 p. 27].

We will use the following observation [CoS70, Theorem 6] of Cohen and Sullivan. We include the short uniqueness argument. For any $x\in X$, let $\mathfrak{j}(x)$ denote its unique normalized norming functional, so that

Suppose that $1<p<\infty$ with $p\not=2$. Let $\mathcal{M}$ be a $\sigma$-finite von Neumann algebra equipped with a normal faithful state. Let $\mathcal{N}$ be a $\sigma$-finite von Neumann algebra equipped with a normal faithful state $\varphi$ and let $T\colon\mathrm{L}^{p}(\mathcal{N})\to\mathrm{L}^{p}(\mathcal{M})$ be a positive map. According to [JRS05, Theorem 2 p. 287], [JRS05, Theorem 4.9 p. 308] and [JRS05, Remark 2 p. 310], the map $\mathrm{Id}\otimes T\colon\mathrm{L}^{p}(\mathrm{M}_{2}\overline{\otimes}\mathcal{N})\to\mathrm{L}^{p}(\mathrm{M}_{2}\overline{\otimes}\mathcal{M})$ is an isometry if and only if there exist an injective normal $*$-homomorphism $\pi\colon\mathcal{N}\to\mathcal{M}$ and a faithful normal conditional expectation $F\colon e\mathcal{M}e\to\pi(\mathcal{N})$, where $e\overset{\mathrm{def}}{=}\pi(1)$, such that, if we define $\mathbb{E}\colon\mathcal{M}\to\pi(\mathcal{N})$ by

then

Strictly speaking, the map $\mathbb{E}$ is not a conditional expectation from $\mathcal{M}$ onto $\pi(\mathcal{N})$, since in general $\mathbb{E}(1)=e=\pi(1)\not=1_{\mathcal{M}}$. We nevertheless keep the misleading notation and terminology of [JRS05], where $\mathbb{E}$ is referred to as a conditional expectation onto $\pi(\mathcal{N})$.

In this case, $T$ is automatically completely positive and completely isometric, and its range is the range of a completely positive and completely contractive projection on $\mathrm{L}^{p}(\mathcal{M})$. Moreover, $T$ is a right $\pi(\mathcal{N})$-module map:

Finally, by [JRS05, Theorem 2, (1.2) p. 288], for every $\omega\in\mathcal{N}_{*}^{+}$ we have

We first record a support property which will be used repeatedly below. It shows that a completely positive complete isometry transports the support projections exactly through the underlying $*$-homomorphism.

Conversely, consider a positive element $x\in\pi(s(a))\mathcal{M}\pi(s(a))$ and suppose that $\theta(x)=0$. Since $\mathbb{E}$ is a conditional expectation onto $\pi(\mathcal{N})$, it is $\pi(\mathcal{N})$-bimodular. Hence

Thus $\mathbb{E}(x)$ is a positive element of the von Neumann algebra $\pi(s(a))\pi(\mathcal{N})\pi(s(a))$. Since the functional $\omega$ is faithful on $s(a)\mathcal{N}s(a)$, the equality

implies $\pi^{-1}(\mathbb{E}(x))=0$, and hence $\mathbb{E}(x)=0$. Moreover, $\pi(s(a))\leqslant e$, so $x\in e\mathcal{M}e$. Since the restriction $\mathbb{E}|_{e\mathcal{M}e}=F$ is faithful by (3.1), we obtain $x=0$. Consequently, $\theta$ is faithful on $\pi(s(a))\mathcal{M}\pi(s(a))$. Since we already know that $s(\theta)\leqslant\pi(s(a))$, we conclude that

Since $T(a)=h_{\theta}^{\frac{1}{p}}$ and $s(h_{\theta}^{\frac{1}{p}})=s(h_{\theta})=s(\theta)$, we finally obtain

which proves (3.5).  

The second property that we need is the bimodule structure of the range. This will eventually allow us to absorb the partial-isometry factor once we know that it belongs to $\pi(\mathcal{N})$.

Let $Y$ be a subspace of $\mathrm{L}^{p}(\mathcal{M})$. We define its left and right support projections by

In particular, if $P\colon\mathrm{L}^{p}(\mathcal{M})\to\mathrm{L}^{p}(\mathcal{M})$ is a bounded projection, we write $s_{\ell}(P)\overset{\mathrm{def}}{=}s_{\ell}(\Ran P)$ and $s_{r}(P)\overset{\mathrm{def}}{=}s_{r}(\Ran P)$. Now, suppose that the projection $P\colon\mathrm{L}^{p}(\mathcal{M})\to\mathrm{L}^{p}(\mathcal{M})$ is positive. Following [ArR24, Section 6], we define the support projection of $P$ by

Then by [ArR24, Section 6] we have

We now turn to the main argument. Let $P\colon\mathrm{L}^{p}(\mathcal{M})\to\mathrm{L}^{p}(\mathcal{M})$ be a positive contractive projection with $\Ran P\not=\{0\}$. Suppose that there exists a complete isometry $T\colon\mathrm{L}^{p}(\mathcal{N})\to\mathrm{L}^{p}(\mathcal{M})$ with range $\Ran P$. Since $T$ is a complete isometry, it is in particular a $2$-isometry. By the classification theorem of Junge, Ruan and Sherman [JRS05, Theorem 2 p. 287] and [JRS05, Remark 2 p. 310], there exist a normal injective $*$-homomorphism $\pi\colon\mathcal{N}\to\mathcal{M}$, a partial isometry $w\in\mathcal{M}$ and a normal conditional expectation $\mathbb{E}\colon\mathcal{M}\to\mathcal{M}$ onto $\pi(\mathcal{N})$ such that

and such that the map

is a completely positive complete isometry. Moreover, in the construction of [JRS05, Theorem 4.9], the restriction $F=\mathbb{E}|_{e\mathcal{M}e}\colon e\mathcal{M}e\to\pi(\mathcal{N})$ is a faithful normal conditional expectation. By [JRS05, Theorem 2 p. 287], we have

Since $\Ran T=\Ran P$, it follows that

The first step is to compare the initial and final projections of $w$. Since $P$ is positive, its range is selfadjoint and therefore has identical left and right support projections. We now combine this observation with the support-preserving property of $\tilde{T}$. Since the von Neumann algebra $\mathcal{N}$ is $\sigma$-finite, choose with [KaR97b, Exercise 7.6.46 p. 500], a faithful normal state $\rho$ on $\mathcal{N}$ and consider the positive element $a_{0}\overset{\mathrm{def}}{=}h_{\rho}^{\frac{1}{p}}$ belonging to the space $\mathrm{L}^{p}(\mathcal{N})$. Then $s(a_{0})=1$, and we obtain

By Proposition 3.2, for every $z\in\Ran\tilde{T}$ we have $ez=ze=z$. Hence $s_{\ell}(z)\leqslant e$ and $s_{r}(z)\leqslant e$ for any $z\in\Ran\tilde{T}$. On the other hand, $s(\tilde{T}(a_{0}))\overset{\eqref{inter-5001}}{=}e$. Consequently, we have

For any $z\in\Ran\tilde{T}$, using $ez=ze=z$, we have

and

For any $z\in\Ran\tilde{T}$, we deduce with (2.5) that

Moreover, since $s_{\ell}(z)\leqslant e=w^{*}w$, we have

Taking suprema and using $s_{\ell}(\Ran\tilde{T})=s_{r}(\Ran\tilde{T})=e$, we obtain using (3.17)

On the other hand, the contractive projection $P\colon\mathrm{L}^{p}(\mathcal{M})\to\mathrm{L}^{p}(\mathcal{M})$ is positive, hence it preserves adjoints. Thus $\Ran P$ is invariant under the involution. It follows that its left and right support projections coincide. Consequently, (3.19) yields

Thus $w$ is a unitary element of the reduced von Neumann algebra $e\mathcal{M}e$. Now, we use the $\sigma$-finiteness assumption in an essential way. Since $P\colon\mathrm{L}^{p}(\mathcal{M})\to\mathrm{L}^{p}(\mathcal{M})$ is a positive contractive projection and since the von Neumann algebra $\mathcal{M}$ is $\sigma$-finite, the support result [ArR24, Proposition 6.1] provides a positive element $h\in\Ran P$ such that

For a positive projection $P$, its support $s(P)$ coincides with the left and right support projections of its range. Hence, by (3.19),

Consider the element $\xi\overset{\mathrm{def}}{=}T^{-1}(h)$ in the space $\mathrm{L}^{p}(\mathcal{N})$ and write its polar decomposition in the form

where $v\in\mathcal{N}$ is a partial isometry.

At this point we only know that $w$ is a unitary in the corner $e\mathcal{M}e$. This is not yet sufficient to identify $\Ran T$ with $\Ran\tilde{T}$. Indeed, for this purpose we need to prove that $w$ belongs to $\pi(\mathcal{N})$, since $\Ran\tilde{T}$ is a $\pi(\mathcal{N})$-bimodule. The next lemma establishes precisely this stronger conclusion. Its proof uses a positive element of $\Ran P$ with full support and the uniqueness of its polar decomposition.