Random Phased Subspace States

• Random phased subspace states are pure quantum state ensembles defined on a fixed support subspace with predetermined amplitudes and randomized phases.
• They provide a framework for analyzing entanglement concentration, canonical typicality, and volume-law behavior in many-body quantum systems.
• These states underpin practical implementations using diagonal circuits and serve as a basis for exploring pseudorandom state generation in quantum computation.

Searching arXiv for the cited papers to ground the article in current arXiv records. Random phased subspace states are most naturally understood as pure-state ensembles obtained by fixing an orthonormal basis and an amplitude profile on a chosen support subspace, and randomizing only the phases. In the formalism of phase-random states, one fixes $\{ |u_n\rangle \}_{n=1}^{2^N}$, amplitudes $r_n \ge 0$ with $\sum_n r_n^2=1$, and draws phases $\varphi_n$ independently and uniformly from $[0,2\pi]$, giving

$$|\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle.$$

If $r_n=0$ for $n\notin S$, then every state is supported on $\mathcal H_S:=\mathrm{span}\{|u_n\rangle:n\in S\}$, yielding the fixed-support specialization most closely aligned with the phrase “random phased subspace states” (Nakata et al., 2011). Closely related, but technically distinct, are full-support binary-phase ensembles, randomized subspace-covering ensembles for degenerate eigenspaces, and random subset-phase states generated by pseudochaotic dynamics (Brakerski et al., 2019, Scali et al., 10 Apr 2026, Lee et al., 2024).

1. Definition and ensemble structure

The foundational ensemble is the phase-random state ensemble

$$\Upsilon_{\rm phase}=\{\,|\phi\rangle\,\}_{\mathrm d\varphi}, \qquad |\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle,$$

with phase measure

$r_n \ge 0$0

The defining structural constraint is that amplitudes $r_n \ge 0$1 and basis vectors $r_n \ge 0$2 are fixed, whereas the phase vector is uniformly distributed on the torus. The ensemble-average density matrix is

$r_n \ge 0$3

so phase averaging removes off-diagonal terms in the chosen basis (Nakata et al., 2011).

This formalism already contains the subspace-supported case. If $r_n \ge 0$4 outside a prescribed support set $r_n \ge 0$5, then the states lie in

$r_n \ge 0$6

When amplitudes are equal on the support, $r_n \ge 0$7 for $r_n \ge 0$8, the ensemble becomes the natural equal-amplitude random-phase model on a fixed subspace. The construction is basis-dependent: the notion of “subspace support” is defined relative to the basis in which phases are randomized (Nakata et al., 2011).

The surrounding literature uses nearby constructions that should not be conflated.

Ensemble	State form	Structural distinction
Phase-random state	$r_n \ge 0$9	Fixed amplitudes, continuous random phases in a fixed basis
Binary-phase full-support state	$\sum_n r_n^2=1$0	Full support, real amplitudes, binary phases $\sum_n r_n^2=1$1
QRSI branch state	$\sum_n r_n^2=1$2	Random ambient-space rotations induce random directions in a degenerate eigenspace
Random subset-phase state	$\sum_n r_n^2=1$3	Support is a random subset of size $\sum_n r_n^2=1$4, not generally a linear subspace

A recurrent misconception is that all such ensembles are interchangeable. They are not. Fixed-support phase-random states randomize phases while preserving a predetermined amplitude profile; binary-phase constructions are full-support analogues; QRSI randomizes directions in a target eigenspace through branchwise unitaries rather than through phase-only perturbations; and random subset-phase states replace subspace support by a hidden computational-basis subset.

2. Subspace support, reduced states, and canonical typicality

An explicit subspace-restricted instance arises when the fixed basis is the Hamiltonian eigenbasis and the support is an energy shell

$\sum_n r_n^2=1$5

If the initial state lies in $\sum_n r_n^2=1$6, then the corresponding phase-random ensemble has support only on that shell. In this setting, long-time dynamics under a time-independent Hamiltonian

$\sum_n r_n^2=1$7

is related to phase randomization through the evolution

$\sum_n r_n^2=1$8

under the assumption of phase ergodicity (Nakata et al., 2011).

For a bipartite split $\sum_n r_n^2=1$9, the reduced subsystem state associated with shell-supported phase-random states is

$\varphi_n$0

where $\varphi_n$1. The canonical comparison state is the reduced microcanonical state

$\varphi_n$2

The condition derived for $\varphi_n$3 to coincide with the canonical state is

$\varphi_n$4

with

$\varphi_n$5

This is presented as a trade-off relation between the amplitude profile $\varphi_n$6 on the shell and the structure of the eigenstates in the restricted subspace (Nakata et al., 2011).

Two limiting mechanisms satisfy the condition directly. One is equal amplitudes in the shell, $\varphi_n$7. The other is identical subsystem reductions of all eigenstates in the shell. This places random phased subspace states in direct contact with canonical typicality: for energy-shell support, phase randomization alone can make subsystem states typically thermal when the support weights or eigenstate structure satisfy the stated condition.

The same logic extends to any fixed support subspace specified in the phase-randomizing basis. A plausible implication is that “random phased subspace states” are best regarded not as a separate formal class, but as a specialization of phase-random states in which the support restriction carries the physically relevant geometry.

3. Entanglement structure, concentration, and circuit generation

For a bipartition $\varphi_n$8, the reduced state is

$\varphi_n$9

and entanglement is measured by the linear entropy

$[0,2\pi]$0

The von Neumann entanglement is

$[0,2\pi]$1

with lower bound

$[0,2\pi]$2

The exact phase-average formula is

$[0,2\pi]$3

where

$[0,2\pi]$4

This decomposition separates the contribution of the dephased mixed state $[0,2\pi]$5 from a correction weighted by the entanglement of the basis vectors themselves (Nakata et al., 2011).

Support size enters sharply. If the basis is separable and only $[0,2\pi]$6 amplitudes are nonzero, then

$[0,2\pi]$7

If the number of nonzero coefficients is $[0,2\pi]$8, then

$[0,2\pi]$9

This makes the support cardinality itself an upper-bound parameter for average linear entropy in the separable-basis case. A small support restricts the attainable phase-averaged entanglement unless the support basis states are already entangled (Nakata et al., 2011).

For equal amplitudes in a separable basis, the ensemble exhibits both high average entanglement and concentration. The paper proves

$$|\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle.$$0

where

$$|\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle.$$1

Combined with the lower bound

$$|\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle.$$2

this links phase-random subspace ensembles to volume-law entanglement and matrix-product-state hardness in the large-support regime (Nakata et al., 2011).

The same work also gives an explicit diagonal-circuit mechanism. Starting from

$$|\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle.$$3

one applies random two-qubit diagonal gates

$$|\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle.$$4

with $$|\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle.$$5 and $$|\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle.$$6, producing

$$|\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle.$$7

The asymptotic average linear entropy converges to that of the computational-basis phase-random ensemble: $$|\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle.$$8 This establishes that phase-random subspace behavior can be generated by diagonal random circuits, not only by abstract sampling over phase variables (Nakata et al., 2011).

4. Binary-phase full-support ensembles as a limiting contrast

A distinct but closely related ensemble replaces continuous phases on a fixed support by random binary signs on the entire computational basis: $$|\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle.$$9 Equivalently, the amplitudes are real and take values $r_n=0$0. This is a full-support, equal-magnitude, binary-phase ensemble. It is therefore a special full-space analogue of phased subspace states, not a subspace-state result (Brakerski et al., 2019).

The main theorem states that the distribution

$r_n=0$1

is a

$r_n=0$2

Writing

$r_n=0$3

and $r_n=0$4 for the $r_n=0$5-copy Haar moment state, the exact comparison given is

$r_n=0$6

Thus for any polynomial $r_n=0$7,

$r_n=0$8

The result shows that binary phases $r_n=0$9 already suffice for near-Haar $n\notin S$0-copy behavior in the full-support setting (Brakerski et al., 2019).

The contrast with random phased subspace states is structural rather than superficial. The support is always the entire computational basis, and the proof relies on that full-support structure. The entrywise moment formulas are expressed on tuples from $n\notin S$1, and the combinatorics distinguishes exact multiplicity matching for complex phases from parity-of-multiplicity matching for binary phases. A plausible implication is that analogous parity-based moment mechanisms may exist for a fixed subspace support, but that extension is not proved in the paper (Brakerski et al., 2019).

Implementation is unusually weak. If $n\notin S$2 is a keyed Boolean function, the state

$n\notin S$3

is generated by a Hadamard layer followed by reversible evaluation of $n\notin S$4 and phase kickback. If $n\notin S$5 is computable by a classical circuit of size $n\notin S$6 and depth $n\notin S$7, then the generator is computable by an HT circuit—one Hadamard layer followed by Toffoli gates—of size $n\notin S$8 and depth $n\notin S$9. The same construction yields pseudorandom states from post-quantum pseudorandom functions and $\mathcal H_S:=\mathrm{span}\{|u_n\rangle:n\in S\}$0-approximate state $\mathcal H_S:=\mathrm{span}\{|u_n\rangle:n\in S\}$1-designs from $\mathcal H_S:=\mathrm{span}\{|u_n\rangle:n\in S\}$2-wise independent Boolean functions (Brakerski et al., 2019).

5. Randomized subspace-covering ensembles and the phase-only boundary

The sharpest distinction between random phased subspace states and other randomized subspace ensembles is provided by quantum randomized subspace iteration. Here the target object is a degenerate eigenspace

$\mathcal H_S:=\mathrm{span}\{|u_n\rangle:n\in S\}$3

with ground-state specialization $\mathcal H_S:=\mathrm{span}\{|u_n\rangle:n\in S\}$4 and spectral gap

$\mathcal H_S:=\mathrm{span}\{|u_n\rangle:n\in S\}$5

The problem is to generate states $\mathcal H_S:=\mathrm{span}\{|u_n\rangle:n\in S\}$6 such that each has substantial overlap with $\mathcal H_S:=\mathrm{span}\{|u_n\rangle:n\in S\}$7 and the projected vectors span $\mathcal H_S:=\mathrm{span}\{|u_n\rangle:n\in S\}$8 (Scali et al., 10 Apr 2026).

QRSI does not randomize phases in a fixed basis. Instead it samples independent branchwise random unitaries $\mathcal H_S:=\mathrm{span}\{|u_n\rangle:n\in S\}$9, defines rotated Hamiltonians

$$\Upsilon_{\rm phase}=\{\,|\phi\rangle\,\}_{\mathrm d\varphi}, \qquad |\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle,$$0

runs an eigenstate-preparation primitive on each branch, and maps the outputs back: $$\Upsilon_{\rm phase}=\{\,|\phi\rangle\,\}_{\mathrm d\varphi}, \qquad |\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle,$$1 The resulting target manifold on branch $$\Upsilon_{\rm phase}=\{\,|\phi\rangle\,\}_{\mathrm d\varphi}, \qquad |\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle,$$2 is

$$\Upsilon_{\rm phase}=\{\,|\phi\rangle\,\}_{\mathrm d\varphi}, \qquad |\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle,$$3

The paper emphasizes that this is more general than random phases: each branch presents “a randomly oriented copy of $$\Upsilon_{\rm phase}=\{\,|\phi\rangle\,\}_{\mathrm d\varphi}, \qquad |\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle,$$4” to the preparation primitive (Scali et al., 10 Apr 2026).

The key subspace object is the foot-point matrix

$$\Upsilon_{\rm phase}=\{\,|\phi\rangle\,\}_{\mathrm d\varphi}, \qquad |\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle,$$5

so that

$$\Upsilon_{\rm phase}=\{\,|\phi\rangle\,\}_{\mathrm d\varphi}, \qquad |\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle,$$6

This gives each branch a random coefficient vector in $$\Upsilon_{\rm phase}=\{\,|\phi\rangle\,\}_{\mathrm d\varphi}, \qquad |\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle,$$7, but the coefficients do not arise from random phases over a fixed deterministic amplitude profile. They arise from the random orientation induced by $$\Upsilon_{\rm phase}=\{\,|\phi\rangle\,\}_{\mathrm d\varphi}, \qquad |\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle,$$8 together with the branchwise preparation dynamics (Scali et al., 10 Apr 2026).

Under independent Haar-random branch rotations and a preparation primitive with nonzero projection onto the rotated target manifold, Proposition 1a states that for $$\Upsilon_{\rm phase}=\{\,|\phi\rangle\,\}_{\mathrm d\varphi}, \qquad |\phi\rangle=\sum_{n=1}^{2^N} r_n e^{i\varphi_n}|u_n\rangle,$$9,

$r_n \ge 0$00

Equivalently, the projected ensemble spans $r_n \ge 0$01 almost surely. The proof uses a $r_n \ge 0$02-equivariance argument inside the degenerate eigenspace, not a random-phase ansatz. The normalized foot-point is $r_n \ge 0$03-invariant and hence uniform on the unit sphere $r_n \ge 0$04, so the construction yields independent continuous random directions in $r_n \ge 0$05 (Scali et al., 10 Apr 2026).

The paper further weakens Haar randomness to an anti-concentration condition on the output foot-point distribution. If for every hyperplane $r_n \ge 0$06,

$r_n \ge 0$07

then $r_n \ge 0$08 independent branches give

$r_n \ge 0$09

and $r_n \ge 0$10 success follows with

$r_n \ge 0$11

This establishes manifold coverage rather than exact Haar sampling on the subspace in every implementation (Scali et al., 10 Apr 2026).

A second misconception addressed by QRSI concerns spectral deformation. Randomization does not split the degeneracy or alter the branch difficulty: for any $r_n \ge 0$12,

$r_n \ge 0$13

The recovered manifold is then identified through the corrected-state coefficient matrix

$r_n \ge 0$14

or the Gram matrix

$r_n \ge 0$15

With branch leakage

$r_n \ge 0$16

Proposition 3 gives

$r_n \ge 0$17

The conceptual outcome is that QRSI is best understood as a randomized subspace-covering ensemble. It is relevant to random phased subspace states only in a broad sense: both produce random coefficient vectors in a target subspace, but the QRSI mechanism is random basis rotation in the ambient space, not phase-only randomization (Scali et al., 10 Apr 2026).

6. Random subset-phase states and the broader pseudorandom landscape

A further nearby notion is the random subset-phase state generated by random subsystem-embedded dynamics. The construction uses a subsystem size $r_n \ge 0$18, a random subset isometry

$r_n \ge 0$19

and a full unitary

$r_n \ge 0$20

For an initial computational basis state $r_n \ge 0$21, the evolved state is

$r_n \ge 0$22

Its support is the set $r_n \ge 0$23, of size $r_n \ge 0$24. This is a random computational-basis subset, not generally a linear subspace (Lee et al., 2024).

The especially important example is

$r_n \ge 0$25

where $r_n \ge 0$26 is a random sign operator and the elements of $r_n \ge 0$27 are “exactly $r_n \ge 0$28 up to signs.” In that case the evolved state is a uniform-magnitude random signed superposition over a subset of size $r_n \ge 0$29. The phases are discrete signs rather than continuous $r_n \ge 0$30, and the support is subset-structured rather than linear-subspace-structured. Accordingly, the closest description is “random subset-phase state,” not random phased linear-subspace state (Lee et al., 2024).

The paper’s formal guarantees are pseudorandom rather than exact-Haar. One theorem states that if $r_n \ge 0$31 is an ensemble of embedded unitaries satisfying, for $r_n \ge 0$32,

$r_n \ge 0$33

for some $r_n \ge 0$34, together with

$r_n \ge 0$35

for all $r_n \ge 0$36, then the corresponding RSED ensemble generates pseudorandom states. A second theorem states that “RSEDs with an ensemble of embedded unitary operators that generate a pseudorandom state ensemble in the subspace generate a pseudorandom state ensemble in the entire space” (Lee et al., 2024).

The same framework gives a pseudochaotic dynamical interpretation through the out-of-time-ordered correlator

$r_n \ge 0$37

For $r_n \ge 0$38, local-operator OTOCs are negligible with probability higher than $r_n \ge 0$39 if the maximum averaged magnitude of the embedded operator elements is $r_n \ge 0$40. Circuit complexity is also explicit: “a $r_n \ge 0$41 depth local circuit can implement a pseudochaotic RSED with an error negligible in $r_n \ge 0$42 up to time $r_n \ge 0$43,” and the overview states that a circuit implementation “requires $r_n \ge 0$44 depth with all-to-all connectivity” (Lee et al., 2024).

Within the broader taxonomy, random phased subspace states occupy the phase-only, fixed-support corner. Full-support binary-phase states show that very sparse phase alphabets can already yield near-Haar $r_n \ge 0$45-copy behavior, but only in the full-support regime. QRSI shows that randomized coverage of a target manifold can be achieved by branchwise ambient-space rotations rather than by phase-only perturbation. Random subset-phase states show that equal-magnitude random-sign superpositions over hidden support sets can be generated efficiently and can be computationally indistinguishable from Haar-random states under restricted access. The central invariant across these models is randomized coherence on a structured support; the central difference is whether the structure is a fixed subspace, the full space, a degenerate eigenspace reached by random rotation, or a random subset.