Random Induced Mixed States

• Random induced mixed states are density operators obtained by partially tracing Haar-random pure states, forming a standard ensemble with well-characterized spectral properties.
• They exhibit detailed eigenvalue distributions and exact formulas for coherence, subentropy, and large deviations that underpin quantum state geometry.
• Applications include studying separability probabilities, APPT thresholds, and symmetry-resolved entanglement, offering deep insights into quantum correlations.

Random induced mixed states are density operators obtained by choosing a Haar-random pure state on a bipartite Hilbert space $\mathbb C^m\otimes \mathbb C^n$ and tracing out one factor. This construction defines the standard induced ensemble on $D(\mathbb C^m)$, with the Hilbert–Schmidt ensemble as a balanced special case, and admits equivalent formulations in terms of fixed-trace Wishart matrices, explicit joint eigenvalue densities, and ancilla dimensions controlling typical mixedness (Zhang et al., 2015, Zhang, 2016). The ensemble has become a canonical model for typical reduced states, for the geometry of atypicality around the maximally mixed state, and for entanglement-theoretic questions ranging from separability and APPT thresholds to the negativity spectrum of mixed-state subsystems (Wakakuwa, 2023, Collins et al., 2011, Shapourian et al., 2020).

1. Ensemble definition and measure structure

Let $m\le n$, let $|\Psi\rangle\in \mathbb C^m\otimes\mathbb C^n$ be Haar-distributed, and define

$$\rho=\operatorname{Tr}_{\text{anc}}\!\left(|\Psi\rangle\langle\Psi|\right).$$

This pushes Haar measure on the pure-state sphere of the composite system to a probability measure on $D(\mathbb C^m)$. In spectral coordinates $\rho=U\Lambda U^\dagger$, with $$\Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_m)$$, the induced measure factorizes as

$$d\mu_{m(n)}(\rho)=d\nu_{m(n)}(\Lambda)\times d\mu_{\mathrm{Haar}}(U),$$

with

$$d\nu_{m(n)}(\Lambda)=C_{m(n)}K_1(\Lambda)\prod_{j=1}^m \lambda_j^{\,n-m}\,d\lambda_j,$$

$D(\mathbb C^m)$0

Equivalently, the joint eigenvalue density is

$D(\mathbb C^m)$1

This ensemble sits inside a broader two-parameter family,

$D(\mathbb C^m)$2

whose normalization is governed by the Selberg-type integral

$D(\mathbb C^m)$3

with

$D(\mathbb C^m)$4

The standard induced ensemble is the specialization

$D(\mathbb C^m)$5

and the Hilbert–Schmidt induced measure is the balanced case

$D(\mathbb C^m)$6

This embedding is structurally important because many exact averages are most naturally expressed in the $D(\mathbb C^m)$7 language rather than in terms of $D(\mathbb C^m)$8 alone (Zhang et al., 2015).

A complementary realization uses random matrices. If $D(\mathbb C^m)$9 is an $m\le n$0 complex Ginibre matrix with i.i.d. entries

$m\le n$1

then

$m\le n$2

is a complex Wishart matrix and the induced state is

$m\le n$3

The induced density may be written as

$m\le n$4

and in eigenvalues this becomes the fixed-trace Wishart law

$m\le n$5

The diagonal entries in a fixed basis are much simpler than the eigenvalues: for the reduced state,

$m\le n$6

with density

$m\le n$7

In the $m\le n$8 induced family relevant to generalized two-qubit states, the ancilla dimension is often parameterized as

$m\le n$9

so that

$|\Psi\rangle\in \mathbb C^m\otimes\mathbb C^n$0

The Hilbert–Schmidt measure is then the special case $|\Psi\rangle\in \mathbb C^m\otimes\mathbb C^n$1, rather than a separate construction (Zhang, 2016, Slater et al., 2014).

2. Exact averages, coherence, and subentropy

Several information-theoretic quantities admit closed forms on the induced ensemble. For the diagonal entropy in a fixed reference basis,

$|\Psi\rangle\in \mathbb C^m\otimes\mathbb C^n$2

while Page’s formula gives the average von Neumann entropy

$|\Psi\rangle\in \mathbb C^m\otimes\mathbb C^n$3

Their difference is the average relative entropy of coherence,

$|\Psi\rangle\in \mathbb C^m\otimes\mathbb C^n$4

For the Hilbert–Schmidt case $|\Psi\rangle\in \mathbb C^m\otimes\mathbb C^n$5,

$|\Psi\rangle\in \mathbb C^m\otimes\mathbb C^n$6

For fixed $|\Psi\rangle\in \mathbb C^m\otimes\mathbb C^n$7, increasing ancilla dimension suppresses coherence: $|\Psi\rangle\in \mathbb C^m\otimes\mathbb C^n$8

The same calculation yields a compact formula for the average subentropy,

$|\Psi\rangle\in \mathbb C^m\otimes\mathbb C^n$9

Since

$$\rho=\operatorname{Tr}_{\text{anc}}\!\left(|\Psi\rangle\langle\Psi|\right).$$0

and

$$\rho=\operatorname{Tr}_{\text{anc}}\!\left(|\Psi\rangle\langle\Psi|\right).$$1

the average subentropy approaches the maximal possible subentropy value in high dimension. By contrast, the average coherence of random pure states in dimension $$\rho=\operatorname{Tr}_{\text{anc}}\!\left(|\Psi\rangle\langle\Psi|\right).$$2 is $$\rho=\operatorname{Tr}_{\text{anc}}\!\left(|\Psi\rangle\langle\Psi|\right).$$3, which diverges logarithmically, whereas induced mixed-state coherence remains uniformly bounded. This clean separation between pure and mixed typicality is one of the notable exact consequences of the fixed-trace Wishart description (Zhang et al., 2015, Zhang, 2016).

The coherence is also typical in the concentration-of-measure sense. If $$\rho=\operatorname{Tr}_{\text{anc}}\!\left(|\Psi\rangle\langle\Psi|\right).$$4 is the reduction of a Haar-random pure state on $$\rho=\operatorname{Tr}_{\text{anc}}\!\left(|\Psi\rangle\langle\Psi|\right).$$5, with $$\rho=\operatorname{Tr}_{\text{anc}}\!\left(|\Psi\rangle\langle\Psi|\right).$$6, then for every $$\rho=\operatorname{Tr}_{\text{anc}}\!\left(|\Psi\rangle\langle\Psi|\right).$$7,

$$\rho=\operatorname{Tr}_{\text{anc}}\!\left(|\Psi\rangle\langle\Psi|\right).$$8

The proof uses Lévy’s lemma together with Lipschitz bounds for both $$\rho=\operatorname{Tr}_{\text{anc}}\!\left(|\Psi\rangle\langle\Psi|\right).$$9 and $D(\mathbb C^m)$0. In the Hilbert–Schmidt case $D(\mathbb C^m)$1, this implies concentration near $D(\mathbb C^m)$2 for sufficiently large $D(\mathbb C^m)$3.

A structured entanglement application converts induced single-system mixed states into bipartite mixed states

$D(\mathbb C^m)$4

For this class,

$D(\mathbb C^m)$5

so both the average relative entropy of entanglement and the average distillable entanglement are exactly

$D(\mathbb C^m)$6

with the same exponential concentration bound as coherence. The operational significance is highly specific to this coherence-to-entanglement conversion map, not to generic bipartite mixed states (Zhang et al., 2015).

3. Large deviations and atypicality

The high-dimensional induced ensemble is not only concentrated near the maximally mixed state; it also obeys an exact large-deviation principle in the regime

$D(\mathbb C^m)$7

Writing

$D(\mathbb C^m)$8

the rate function is

$D(\mathbb C^m)$9

For any subset $\rho=U\Lambda U^\dagger$0,

$\rho=U\Lambda U^\dagger$1

and

$\rho=U\Lambda U^\dagger$2

If

$\rho=U\Lambda U^\dagger$3

then the exact exponential rate exists: $\rho=U\Lambda U^\dagger$4 Thus the traced-out dimension $\rho=U\Lambda U^\dagger$5 is the large-deviation speed, while the subsystem dimension $\rho=U\Lambda U^\dagger$6 multiplies the relative-entropy cost (Wakakuwa, 2023).

This perspective extends concentration-of-measure from small fluctuations to genuinely atypical regions of state space. If $\rho=U\Lambda U^\dagger$7, the probability of $\rho=U\Lambda U^\dagger$8 is exponentially small in $\rho=U\Lambda U^\dagger$9; if $$\Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_m)$$0, the exponent is $$\Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_m)$$1. The rate function is finite only for full-rank states, since

$$\Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_m)$$2

which formalizes the intuition that boundary states are extremely atypical at fixed $$\Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_m)$$3 and large ancilla dimension.

A stronger statement is conditional concentration. Defining

$$\Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_m)$$4

and the trace-distance neighborhood

$$\Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_m)$$5

one has, for every $$\Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_m)$$6, some $$\Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_m)$$7 such that

$$\Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_m)$$8

Conditioned on the rare event $$\Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_m)$$9, the conditional law therefore concentrates exponentially near the minimizing set $$d\mu_{m(n)}(\rho)=d\nu_{m(n)}(\Lambda)\times d\mu_{\mathrm{Haar}}(U),$$0. If the minimizer is unique, the conditioned ensemble localizes near a single state $$d\mu_{m(n)}(\rho)=d\nu_{m(n)}(\Lambda)\times d\mu_{\mathrm{Haar}}(U),$$1.

Several constraint sets admit explicit exponents. For

$$d\mu_{m(n)}(\rho)=d\nu_{m(n)}(\Lambda)\times d\mu_{\mathrm{Haar}}(U),$$2

the exact rate is

$$d\mu_{m(n)}(\rho)=d\nu_{m(n)}(\Lambda)\times d\mu_{\mathrm{Haar}}(U),$$3

For

$$d\mu_{m(n)}(\rho)=d\nu_{m(n)}(\Lambda)\times d\mu_{\mathrm{Haar}}(U),$$4

the rate is $$d\mu_{m(n)}(\rho)=d\nu_{m(n)}(\Lambda)\times d\mu_{\mathrm{Haar}}(U),$$5, and for $$d\mu_{m(n)}(\rho)=d\nu_{m(n)}(\Lambda)\times d\mu_{\mathrm{Haar}}(U),$$6, $$d\mu_{m(n)}(\rho)=d\nu_{m(n)}(\Lambda)\times d\mu_{\mathrm{Haar}}(U),$$7,

$$d\mu_{m(n)}(\rho)=d\nu_{m(n)}(\Lambda)\times d\mu_{\mathrm{Haar}}(U),$$8

Entropy-deficiency events

$$d\mu_{m(n)}(\rho)=d\nu_{m(n)}(\Lambda)\times d\mu_{\mathrm{Haar}}(U),$$9

likewise reduce to minimizing $$d\nu_{m(n)}(\Lambda)=C_{m(n)}K_1(\Lambda)\prod_{j=1}^m \lambda_j^{\,n-m}\,d\lambda_j,$$0 over the constrained set. In this regime, atypicality is therefore governed exactly by the relative-entropy geometry centered at $$d\nu_{m(n)}(\Lambda)=C_{m(n)}K_1(\Lambda)\prod_{j=1}^m \lambda_j^{\,n-m}\,d\lambda_j,$$1 (Wakakuwa, 2023).

4. Separability, partial transpose, and APPT thresholds

For $$d\nu_{m(n)}(\Lambda)=C_{m(n)}K_1(\Lambda)\prod_{j=1}^m \lambda_j^{\,n-m}\,d\lambda_j,$$2 systems, the Peres–Horodecki criterion makes the determinant of the partial transpose a complete separability diagnostic: $$d\nu_{m(n)}(\Lambda)=C_{m(n)}K_1(\Lambda)\prod_{j=1}^m \lambda_j^{\,n-m}\,d\lambda_j,$$3 This fact underlies the induced-measure analysis of generalized two-qubit states, where the central moments are

$$d\nu_{m(n)}(\Lambda)=C_{m(n)}K_1(\Lambda)\prod_{j=1}^m \lambda_j^{\,n-m}\,d\lambda_j,$$4

In the $$d\nu_{m(n)}(\Lambda)=C_{m(n)}K_1(\Lambda)\prod_{j=1}^m \lambda_j^{\,n-m}\,d\lambda_j,$$5 setting, $$d\nu_{m(n)}(\Lambda)=C_{m(n)}K_1(\Lambda)\prod_{j=1}^m \lambda_j^{\,n-m}\,d\lambda_j,$$6 parameterizes the ancilla dimension and the same $$d\nu_{m(n)}(\Lambda)=C_{m(n)}K_1(\Lambda)\prod_{j=1}^m \lambda_j^{\,n-m}\,d\lambda_j,$$7-based determinantal-moment formula used for Hilbert–Schmidt measure ($$d\nu_{m(n)}(\Lambda)=C_{m(n)}K_1(\Lambda)\prod_{j=1}^m \lambda_j^{\,n-m}\,d\lambda_j,$$8) extends to the full induced family. The physically prominent Dyson-index-like values are

$$d\nu_{m(n)}(\Lambda)=C_{m(n)}K_1(\Lambda)\prod_{j=1}^m \lambda_j^{\,n-m}\,d\lambda_j,$$9

To extract separability probabilities, the distribution of $D(\mathbb C^m)$00 is reconstructed numerically from large moment sets using a high-precision Legendre-polynomial procedure due to Provost. The support is effectively

$D(\mathbb C^m)$01

and the separable region is

$D(\mathbb C^m)$02

The computations use as many as $D(\mathbb C^m)$03 determinantal moments. For $D(\mathbb C^m)$04, the inferred separability probabilities are rational for $D(\mathbb C^m)$05, and the following all-$D(\mathbb C^m)$06 formulas reproduce the observed values: $D(\mathbb C^m)$07

$D(\mathbb C^m)$08

$D(\mathbb C^m)$09

At $D(\mathbb C^m)$10, these reduce to the Hilbert–Schmidt values

$D(\mathbb C^m)$11

The formulas are conjectural rather than rigorously proved, but they are numerically supported by extensive moment reconstruction. A finer Hilbert–Schmidt symmetry,

$D(\mathbb C^m)$12

splitting the separable region into $D(\mathbb C^m)$13 and $D(\mathbb C^m)$14, does not persist for $D(\mathbb C^m)$15; for example, when $D(\mathbb C^m)$16 and $D(\mathbb C^m)$17,

$D(\mathbb C^m)$18

with complementary separable contribution $D(\mathbb C^m)$19 (Slater et al., 2014).

In higher-dimensional bipartite systems

$D(\mathbb C^m)$20

a different threshold problem concerns the absolute positive partial transpose property. A random induced state

$D(\mathbb C^m)$21

is APPT if it remains PPT under every global unitary. The threshold occurs at environment dimension of order

$D(\mathbb C^m)$22

More precisely, APPT holds with high probability for $D(\mathbb C^m)$23 above a constant multiple of $D(\mathbb C^m)$24 and fails with high probability below a constant multiple of $D(\mathbb C^m)$25. In the regime

$D(\mathbb C^m)$26

the sharp leading constant is $D(\mathbb C^m)$27: $D(\mathbb C^m)$28

$D(\mathbb C^m)$29

For fixed $D(\mathbb C^m)$30 and $D(\mathbb C^m)$31, the sharp transition constant is

$D(\mathbb C^m)$32

The ordinary PPT threshold is much smaller, of order $D(\mathbb C^m)$33; APPT is therefore a substantially stronger property than PPT for random induced states (Collins et al., 2011).

5. Internal entanglement, negativity spectrum, and symmetry-resolved induced states

A natural refinement starts from a Haar-random pure state on

$D(\mathbb C^m)$34

and studies the induced mixed state on $D(\mathbb C^m)$35,

$D(\mathbb C^m)$36

The entanglement between $D(\mathbb C^m)$37 and $D(\mathbb C^m)$38 is probed by the partial transpose $D(\mathbb C^m)$39. An exact moment formula is

$D(\mathbb C^m)$40

where $D(\mathbb C^m)$41, $D(\mathbb C^m)$42 is the cycle number, and $D(\mathbb C^m)$43. The third PT moment already displays the tripartite symmetry: $D(\mathbb C^m)$44

The large-dimension phase diagram has three regimes. When the bath is larger than the system, the logarithmic negativity vanishes at leading order: $D(\mathbb C^m)$45 When $D(\mathbb C^m)$46 and both $D(\mathbb C^m)$47 and $D(\mathbb C^m)$48 are below half the total system size, a saturated regime appears: $D(\mathbb C^m)$49 independent, at leading order, of how $D(\mathbb C^m)$50 is partitioned. This plateau has no analogue for random pure states. If one subsystem dominates, the negativity becomes maximal in the smaller subsystem: $D(\mathbb C^m)$51 The partial-transpose spectrum is semicircular in the saturated and PPT regimes: $D(\mathbb C^m)$52 and the support crosses the origin precisely when

$D(\mathbb C^m)$53

Hence the PPT/NPT threshold is

$D(\mathbb C^m)$54

or, in qubit language, $D(\mathbb C^m)$55 at leading order. Despite the semicircle law, the ensemble is not GUE: the one-point Green function has an $D(\mathbb C^m)$56 correction

$D(\mathbb C^m)$57

and the connected two-point correlator also differs from the GUE form (Shapourian et al., 2020).

Symmetry constraints produce a further deformation of the induced ensemble. If the global random pure state is restricted to a fixed Abelian charge sector $D(\mathbb C^m)$58, then tracing out $D(\mathbb C^m)$59 gives a block-diagonal reduced state

$D(\mathbb C^m)$60

Each block $D(\mathbb C^m)$61 is itself an induced mixed state, but now inside a symmetry sector. The relevant entanglement measure is the symmetry-resolved logarithmic negativity

$D(\mathbb C^m)$62

For $D(\mathbb C^m)$63, sector dimensions are equal,

$D(\mathbb C^m)$64

and the projected partial-transpose spectrum is the sum of two semicircle-type pieces, one centered at $D(\mathbb C^m)$65. Consequently, unlike the symmetry-free induced ensemble, there is no strictly PPT phase for the projected block $D(\mathbb C^m)$66: every block has negative eigenvalues. In the replica-symmetry-breaking plateau regime,

$D(\mathbb C^m)$67

while for

$D(\mathbb C^m)$68

the negativity is exponentially small but nonzero,

$D(\mathbb C^m)$69

For $D(\mathbb C^m)$70, the sector dimensions depend combinatorially on charge,

$D(\mathbb C^m)$71

and the maximal-entanglement transition broadens from a sharp line into a critical region. In the plateau regime,

$D(\mathbb C^m)$72

where

$D(\mathbb C^m)$73

while deep in the maximal-entanglement region,

$D(\mathbb C^m)$74

The symmetry effect is therefore not merely quantitative: it protects residual NPT entanglement, removes the sharp projected-block PPT transition, and reshapes phase boundaries into charge-dependent critical regions (Hejazi et al., 2021).

6. Generalizations, contrasts, and nonstandard related constructions

Not every natural ensemble of random mixed states is induced in the standard partial-trace sense. A useful contrast is provided by generalized Bures states, whose eigenvalue density is

$D(\mathbb C^m)$75

with the physical Bures case at $D(\mathbb C^m)$76. This family is linked to the $D(\mathbb C^m)$77 matrix model rather than to a Wishart partial-trace construction. Its purity

$D(\mathbb C^m)$78

has a three-regime large-$D(\mathbb C^m)$79 distribution with two phase transitions, including a detached-eigenvalue regime. For ordinary Bures states,

$D(\mathbb C^m)$80

whereas for Hilbert–Schmidt induced states

$D(\mathbb C^m)$81

Typical Bures states are therefore less mixed than Hilbert–Schmidt induced states. The comparison is valuable because it isolates which mixed-state phenomena are specific to induced/Wishart statistics and which persist across other unitarily invariant ensembles (Borot et al., 2011).

A different extension imposes a linear constraint on random mixed states,

$D(\mathbb C^m)$82

Starting from a unitarily invariant base measure $D(\mathbb C^m)$83, one defines

$D(\mathbb C^m)$84

For Hilbert–Schmidt measure, which the construction explicitly notes may be interpreted as an induced measure, this yields a constrained modification of the usual induced ensemble. The averaged state

$D(\mathbb C^m)$85

always commutes with $D(\mathbb C^m)$86, but is generally not Gibbsian in dimensions larger than two. The ensemble supports a statistical-mechanical formalism with

$D(\mathbb C^m)$87

and response relation

$D(\mathbb C^m)$88

For HS, the integrated density of states $D(\mathbb C^m)$89 is a piecewise polynomial of degree $D(\mathbb C^m)$90. This suggests that induced-state techniques can be carried into constrained thermodynamic settings, though the resulting ensembles are no longer standard induced ensembles in the narrow partial-trace sense (Miller, 1 Aug 2025).

A more remote but conceptually related mechanism appears in open quantum random walks. There, the internal density matrix of a monitored open system evolves by normalized Kraus maps,

$D(\mathbb C^m)$91

with

$D(\mathbb C^m)$92

For a two-level internal system, writing $D(\mathbb C^m)$93, one has

$D(\mathbb C^m)$94

unless the dynamics is essentially classical, so $D(\mathbb C^m)$95 converges almost surely to pure states along trajectories. The limiting pure state is random because it depends on the measurement record. This is not an induced ensemble from partial tracing Haar-random bipartite states, but it provides a distinct route from mixed states to random pure states through conditioning rather than through spectral sampling (Bauer et al., 2013).

Taken together, these developments place random induced mixed states within a broader landscape of random density-matrix ensembles. The standard induced ensemble remains distinguished by its operational origin as a reduced state of a Haar-random pure state, by its fixed-trace Wishart realization, and by the unusually large number of exact formulas it supports. At the same time, separability thresholds, APPT transitions, internal negativity plateaux, symmetry-protected residual entanglement, and fixed-energy deformations show that even within this ostensibly simple model, the geometry of mixed-state typicality is highly structured and sensitive to the choice of constraints, symmetries, and observables.