Digression: The History of Tensor Norms

Tensor Norms for Quantum Entanglement

Recap: Norm Criterion for Separability

1. Setup

Fix a $k$-partite, $d_1 \times \cdots \times d_k$ Hilbert space: \[ \mathcal H = \bigotimes_{j=1}^{k}\mathbb C^{d_j} \]
A (normalised) density operator $\rho \in \bigotimes_{j=1}^{k}M_{d_j}(\mathbb C)$ satisfies: \[ \rho \ge 0 \quad \text{and} \quad \Tr\rho=1 \]

The Projective Norm

For a density operator $\rho \in \bigotimes_{j=1}^{k} M_{d_j}(\mathbb{C})$

The projective norm is defined over the L2 norms of these underlying vectors:

\[ \|\rho\|_{\pi} = \inf\left\{ \sum_i \prod_{j=1}^{k} \|\phi_i^{(j)}\|_2\,\|\psi_i^{(j)}\|_2 : \rho=\sum_i \bigotimes_{j=1}^{k} \ket{\phi_i^{(j)}}\!\bra{\psi_i^{(j)}} \right\} \]

The Projective Norm

For a density operator $\rho \in \bigotimes_{j=1}^{k} M_{d_j}(\mathbb{C})$

The projective norm is defined over the L2 norms of these underlying vectors:

\[ \|\rho\|_{\pi} = \inf\left\{ \sum_i \prod_{j=1}^{k} \|\phi_i^{(j)}\|_2\,\|\psi_i^{(j)}\|_2 : \rho=\sum_i \bigotimes_{j=1}^{k} \ket{\phi_i^{(j)}}\!\bra{\psi_i^{(j)}} \right\} \]

The Key Insight: The trace norm (or nuclear norm, $\|\cdot\|_1$) of a rank-one operator is precisely the product of the L2 norms of its constituent vectors.

\[ \bigl\| \ket{\phi}\!\bra{\psi} \bigr\|_{1} = \|\phi\|_2 \|\psi\|_2 \]

The Projective Norm

For a density operator $\rho \in \bigotimes_{j=1}^{k} M_{d_j}(\mathbb{C})$

The projective norm is defined over the L2 norms of these underlying vectors:

\[ \|\rho\|_{\pi} = \inf\left\{ \sum_i \prod_{j=1}^{k} \|\phi_i^{(j)}\|_2\,\|\psi_i^{(j)}\|_2 : \rho=\sum_i \bigotimes_{j=1}^{k} \ket{\phi_i^{(j)}}\!\bra{\psi_i^{(j)}} \right\} \]

The Key Insight: The trace norm (or nuclear norm, $\|\cdot\|_1$) of a rank-one operator is precisely the product of the L2 norms of its constituent vectors.

\[ \bigl\| \ket{\phi}\!\bra{\psi} \bigr\|_{1} = \|\phi\|_2 \|\psi\|_2 \]

By substituting this identity, the vector-based definition transforms into one based on the nuclear norms of the component matrices.

2. The Projective Norm (Rewritten)

For a simple tensor $X_1 \otimes \cdots \otimes X_k \in \bigotimes_{j=1}^{k} M_{d_j}(\mathbb{C})$, we define:

\[ \|X_1\otimes\cdots\otimes X_k\|_{\pi} := \prod_{j=1}^{k}\|X_j\|_{1}, \quad\text{where }\|\,\cdot\,\|_{1}\text{ is the trace norm.} \]

We extend this by infimum to all tensors $\rho$ in the space:

\[ \|\rho\|_{\pi} := \inf\left\{ \sum_{i} \|Y_i^{(1)}\|_{1}\cdots\|Y_i^{(k)}\|_{1} \; : \; \rho=\sum_{i}Y_i^{(1)}\otimes\cdots\otimes Y_i^{(k)} \right\} \]

3. The Separability Criterion

A density operator $\rho$ is separable if and only if its projective tensor norm is at most 1.

\[ \rho\text{ separable} \quad\Longleftrightarrow\quad \|\rho\|_{\pi}\le 1 \]

(Equivalently, $\|\rho\|_{\pi} > 1 \iff \rho$ is entangled).

Proof ($\Rightarrow$): separable $\Longrightarrow \|\rho\|_{\pi}\le 1$

By definition, a separable state can be written as a convex combination: \[ \rho = \sum_{i}\lambda_i\, \rho_i^{(1)}\otimes\cdots\otimes\rho_i^{(k)} \] with $\lambda_i\ge 0$, $\sum_i\lambda_i=1$, and each $\rho_i^{(j)}$ being a single-system density operator (i.e., $\rho_i^{(j)}\ge 0$, $\Tr\rho_i^{(j)}=1$).

Proof ($\Rightarrow$): separable $\Longrightarrow \|\rho\|_{\pi}\le 1$

By definition, a separable state can be written as a convex combination: \[ \rho = \sum_{i}\lambda_i\, \rho_i^{(1)}\otimes\cdots\otimes\rho_i^{(k)} \] with $\lambda_i\ge 0$, $\sum_i\lambda_i=1$, and each $\rho_i^{(j)}$ being a single-system density operator (i.e., $\rho_i^{(j)}\ge 0$, $\Tr\rho_i^{(j)}=1$).
This is already a valid decomposition. Using the definition of $\|\cdot\|_\pi$, this specific decomposition gives an upper bound: \[ \|\rho\|_{\pi} \le \sum_{i} \prod_{j=1}^{k} \| \lambda_i^{1/k} \rho_i^{(j)} \|_{1} = \sum_{i} \prod_{j=1}^{k} \lambda_i^{1/k} \|\rho_i^{(j)}\|_{1} \]
Since $\|\rho_i^{(j)}\|_{1} = \Tr(\rho_i^{(j)}) = 1$ for all density operators, this simplifies to: \[ \|\rho\|_{\pi} \le \sum_{i} \lambda_i = 1 \]

Proof ($\Leftarrow$): $\|\rho\|_{\pi}\le 1 \Longrightarrow$ separable

By compactness, the infimum in the norm definition is attained. So there exists a decomposition such that: \[ \rho=\sum_{i}Y_i^{(1)}\otimes\cdots\otimes Y_i^{(k)}, \qquad \sum_{i} \prod_{j=1}^{k}\|Y_i^{(j)}\|_{1} = \|\rho\|_{\pi} \le 1 \]

Proof ($\Leftarrow$): $\|\rho\|_{\pi}\le 1 \Longrightarrow$ separable

By compactness, the infimum in the norm definition is attained. So there exists a decomposition such that: \[ \rho=\sum_{i}Y_i^{(1)}\otimes\cdots\otimes Y_i^{(k)}, \qquad \sum_{i} \prod_{j=1}^{k}\|Y_i^{(j)}\|_{1} = \|\rho\|_{\pi} \le 1 \]
Key Lemma: For any matrix $A$, its trace is less than or equal to its trace norm, $\Tr(A) \le \|A\|_{1}$. Equality holds if and only if $A$ is positive semi-definite ($A \ge 0$).

Proof ($\Leftarrow$): $\|\rho\|_{\pi}\le 1 \Longrightarrow$ separable

By compactness, the infimum in the norm definition is attained. So there exists a decomposition such that: \[ \rho=\sum_{i}Y_i^{(1)}\otimes\cdots\otimes Y_i^{(k)}, \qquad \sum_{i} \prod_{j=1}^{k}\|Y_i^{(j)}\|_{1} = \|\rho\|_{\pi} \le 1 \]
Key Lemma: For any matrix $A$, $\Tr(A) \le \|A\|_{1}$, with equality if and only if $A$ is positive semi-definite ($A \ge 0$).
Now, take the trace of $\rho$ and apply the lemma term-by-term: \[ 1 = \Tr\rho = \sum_{i} \Tr(Y_i^{(1)})\cdots \Tr(Y_i^{(k)}) \le \sum_{i} \prod_{j=1}^{k}\|Y_i^{(j)}\|_{1} = \|\rho\|_{\pi} \le 1 \]
This entire chain of inequalities must collapse to equality. Specifically, we must have $\sum \prod \Tr(Y_i^{(j)}) = \sum \prod \|Y_i^{(j)}\|_1$, which implies $\Tr(Y_i^{(j)}) = \|Y_i^{(j)}\|_1 \forall i,j$.
By the lemma, this means each factor $Y_i^{(j)}$ must be positive semi-definite.

Proof ($\Leftarrow$): $\|\rho\|_{\pi}\le 1 \Longrightarrow$ separable

By compactness, the infimum in the norm definition is attained. So there exists a decomposition such that: \[ \rho=\sum_{i}Y_i^{(1)}\otimes\cdots\otimes Y_i^{(k)}, \qquad \sum_{i} \prod_{j=1}^{k}\|Y_i^{(j)}\|_{1} = \|\rho\|_{\pi} \le 1 \]
Key Lemma: For any matrix $A$, $\Tr(A) \le \|A\|_{1}$, with equality if and only if $A$ is positive semi-definite ($A \ge 0$).
Now, take the trace of $\rho$ and apply the lemma term-by-term: \[ 1 = \Tr\rho = \sum_{i} \Tr(Y_i^{(1)})\cdots \Tr(Y_i^{(k)}) \le \sum_{i} \prod_{j=1}^{k}\|Y_i^{(j)}\|_{1} = \|\rho\|_{\pi} \le 1 \]
This entire chain of inequalities must collapse to equality. Specifically, we must have $\sum \prod \Tr(Y_i^{(j)}) = \sum \prod \|Y_i^{(j)}\|_1$, which implies $\Tr(Y_i^{(j)}) = \|Y_i^{(j)}\|_1 \forall i,j$.
By the lemma, this means each factor $Y_i^{(j)}$ must be positive semi-definite.
We can now construct the separable form. Define $\lambda_i := \prod_j \|Y_i^{(j)}\|_1$ and $\rho_i^{(j)} := Y_i^{(j)}/\|Y_i^{(j)}\|_1$. This gives the explicit separable decomposition $\rho = \sum_i \lambda_i \bigotimes_j \rho_i^{(j)}$, completing the proof.

History: Two Worlds, One Problem

Quantum Information (1990s-2000s)

Physicists sought a definitive, computable test to distinguish separable states from entangled ones.

Peres (1996) - PPT Criterion: For separable $\rho$, the partial transpose is positive: $\rho^{T_B} \ge 0$.
Rains (1999): Used $\|\rho^{T_B}\|_1$ (the "greatest cross norm") to bound distillable entanglement.
Rudolph (2000-2003): Formulated the CCN criterion, realizing it was equivalent to $\|\rho\|_{\pi} \le 1$ and was an exact criterion based on the Hahn-Banach theorem.

Functional Analysis (1950s)

Grothendieck sought a canonical way to define norms on tensor product spaces to linearize bilinear maps.

Goal: Given Banach spaces $X, Y$ and a bounded bilinear map $u: X \times Y \to \mathbb{C}$, build a space $X\otimes_\pi Y$ and a linear map $U$ such that $u(x,y) = U(x \otimes y)$ and $\|U\| = \|u\|$.
Solution: He defined the projective tensor norm $\|\cdot\|_\pi$ to achieve this "universal linearisation".
The separability problem was unknowingly solved decades earlier in a different context.

Definition: Tensor Norm

A tensor norm on $X \otimes Y$ is a norm $\|\cdot\|$ such that it agrees with the product of norms on simple tensors, at both the primal and dual levels:

$\|x \otimes y\| = \|x\|_X \cdot \|y\|_Y$
$\|\alpha \otimes \beta\|_* = \|\alpha\|_{X^*} \cdot \|\beta\|_{Y^*}$

Where $x \in X, y \in Y$ and $\alpha \in X^*, \beta \in Y^*$.

Historical Footnote

Grothendieck coined the term reasonable cross-norm for norms satisfying (i) and (ii). Axiom (i) makes the canonical bilinear map $j:X\times Y \to X\otimes_\alpha Y,\; (x,y)\mapsto x\otimes y$ an isometry on elementary tensors. Axiom (ii) ensures that the adjoint map embeds the space of rank-one bilinear forms isometrically into $(X\otimes_\alpha Y)^{*}$; hence every bounded bilinear form extends uniquely and continuously along $j$.

Universal Property

\[ \begin{array}{ccc} X\times Y & \xrightarrow{\;j\;} & X\otimes_\alpha Y\\[8pt] {\scriptstyle b}\!\!\!\! & \searrow & \!\!\!\!\swarrow {\scriptstyle\tilde b}\\[4pt] & \mathbb C & \end{array} \qquad \|j\|=1; b\mapsto \tilde b\text{ is an isometry on simple tensors}. \]

Projective Norm

The projective norm $\|\cdot\|_\pi$ is defined by decomposition:

\[ \|z\|_\pi := \inf \left\{ \sum_{i=1}^k \|x_i\| \|y_i\| : z = \sum_{i=1}^k x_i \otimes y_i \right\} \]

Historical Footnote

Grothendieck selected $\|\cdot\|_\pi$ because it is the norm for which $(X\otimes_\pi Y)^{*}\simeq B(X,Y)$ holds isometrically, turning every bounded bilinear form into a bounded linear functional.

Isometry

\[ \begin{array}{ccc} B(X,Y) & \xrightarrow[\text{iso}]{\;\;\;\;\cong\;\;\;\;} & (X\otimes_\pi Y)^{*} \\[6pt] b & \longmapsto & \tilde b \end{array} \quad \|b\|=\|\tilde b\|. \]

Theorem: Maximal Tensor Norm

If a norm $\|\cdot\|$ on $X \otimes Y$ is a tensor norm then it is upper bounded by the projective norm.

For all $z \in X \otimes Y$:

\[ \|z\| \le \|z\|_\pi \]

Theorem: Duality & Injective Norm

The injective and projective norms are dual to each other. For Banach spaces $X$ and $Y$:

$(X \otimes_\varepsilon Y)^* = X^* \otimes_\pi Y^*$
$(X \otimes_\pi Y)^* = X^* \otimes_\varepsilon Y^*$

Historical Footnote

The injective norm $\|\cdot\|_\varepsilon$ was introduced as the dual companion of $\|\cdot\|_\pi$.

Duality and bilinear maps

\[ (X\otimes_\pi Y)^{*}\;\cong\;B(X,Y) \] \[ (X\otimes_\varepsilon Y)^{*}\;\cong\;L(X,Y^{*}) \]

Theorem: Minimal Tensor Norm

If a norm $\|\cdot\|$ on $X \otimes Y$ is a tensor norm then it is lower bounded by the injective norm.

For all $z \in X \otimes Y$:

\[ \|z\|_\varepsilon \le \|z\| \]

Theorem: Characterization

A norm $\|\cdot\|$ on $X \otimes Y$ is a tensor norm iff is sandwiched between these two norms.

\[ \|z\|_\varepsilon \le \|z\| \le \|z\|_\pi \]

Example: Matrix Norms

The operator (spectral) norm and trace (nuclear) norm on matrices arise naturally as tensor norms.

For the space of $n \times n$ matrices $\mathcal{M}_n(\mathbb{R}) \cong \ell_2^n \otimes \ell_2^n$:

Operator Norm ($s_\infty$):
$(\mathcal{M}_n, \|\cdot\|_{s_\infty}) = \ell_2^n \otimes_\varepsilon \ell_2^n$
Trace Norm ($s_1$):
$(\mathcal{M}_n, \|\cdot\|_{s_1}) = \ell_2^n \otimes_\pi \ell_2^n$

Historical Footnote

This identification goes back to Schatten (1950) and traces its conceptual clarity to Grothendieck’s framework: the nuclear norm of a matrix is simply the projective tensor norm coming from the self-dual Hilbert space $\ell_2^n$, while the spectral norm is the corresponding injective norm.

Other examples

Schatten $p$-norms: For $1 \leq p \leq \infty$, the Schatten $p$-norm on $\mathcal{M}_n$ is defined as \[ \|A\|_{s_p} = \left(\sum_{i=1}^n \sigma_i(A)^p\right)^{1/p} \] where $\sigma_i(A)$ are the singular values of $A$.
The Schatten $p$-norms interpolate between the trace (nuclear) norm ($p=1$) and the operator (spectral) norm ($p=\infty$): \[ \|A\|_{s_\infty} \leq \|A\|_{s_p} \leq \|A\|_{s_1} \] for all $A \in \mathcal{M}_n$ and $1 \leq p \leq \infty$.

Thank You!

Any Questions?