Symmetry-Projection Techniques

• Symmetry-projection techniques are methods that restore, enforce, or approximate the symmetry-adapted component of a system, enabling a clearer understanding of its intrinsic properties.
• They are applied in diverse fields like nuclear many-body theory, materials modeling, and quantum computing to recover exact quantum numbers and enhance simulation accuracy.
• Advanced formulations employ group averaging, orthogonal projections, and spectral methods to optimize variational methods and ensure invariance under symmetry operations.

Symmetry-projection techniques are methods for extracting, restoring, enforcing, or approximating the component of a mathematical or physical object that belongs to a specified symmetry class. Across the literature, the object being projected may be a many-body wave function, a material tensor, a graph Laplacian, a projection geometry, or even a probability distribution viewed through random one-dimensional marginals. The unifying idea is that one starts from an entity that either truly breaks symmetry or only appears to do so because of representation choice, finite-size effects, noise, or approximation, and then applies a projector, a symmetry-constrained minimization, or a projection-domain criterion to isolate the symmetry-adapted content (Yao, 2022, Caro, 2014, Yen et al., 2019, Preuhs et al., 2018, Xie, 18 May 2026, Cholaquidis et al., 24 Dec 2025).

1. Conceptual structure and scope

A recurring distinction in this literature is between exact symmetry, broken-symmetry representations, and symmetry restoration. In nuclear and electronic many-body theory, the exact Hamiltonian often commutes with the relevant symmetry generators, but practical mean-field or truncated correlated ansätze may break rotational symmetry, particle-number conservation, parity, spin, complex conjugation, or time reversal. In materials modeling, a tensor may appear to have lower symmetry either because symmetry is genuinely reduced by disorder or because the tensor is written in an inconvenient frame. In imaging and geometry, bilateral or plane symmetry may be present in the object while being only indirectly visible in the acquired projections (Yao, 2022, Caro, 2014, Preuhs et al., 2018).

This leads to two broad uses of projection. The first is restorative: a broken-symmetry intrinsic state is projected back onto a sector with good quantum numbers. The second is diagnostic or approximative: among all objects satisfying a target symmetry constraint, one selects the closest one in a specified metric or the one most consistent with observed projections. In the first use, projection is typically formulated by group averaging; in the second, it is often cast as orthogonal projection onto a symmetry-adapted subspace, a commutant, or a set defined by spectral or geometric constraints (Yao, 2022, Xie, 18 May 2026, Herrera, 2024).

A further conceptual distinction concerns whether symmetry breaking is physical or representational. For material tensors, rotations do not change the material property, only its representation, so the “closest” higher-symmetry tensor should be sought over all rotated representations, not just in the fixed coordinate frame (Caro, 2014). In finite nuclei, by contrast, spontaneous symmetry breaking is not fundamental in the same sense as in infinite systems; the breaking arises from approximation, so projection is needed to recover laboratory-frame states with the correct quantum numbers (Yao, 2022). In VQE and related quantum algorithms, ignoring symmetry can send optimization toward the wrong Fock-space sector altogether, making projection or exact symmetry constraints part of the definition of the target state rather than a post-processing refinement (Yen et al., 2019).

2. Mathematical formulations

The most classical formulation is the group-theoretic projector onto an irreducible representation. In the notation used for compact groups, one writes

$$\hat P^\lambda_{ab} = \frac{d_\lambda}{v_G}\int_G dg\, S^{\lambda *}_{ab}(g)\,\hat R(g),$$

and, for the notation specialized to nuclear symmetry restoration,

$$\hat{P}_{\mu\mu}^{\lambda} \equiv \frac{d_{\lambda}}{n_G}\int d\boldsymbol{\varphi}\, D_{\mu\mu}^{\lambda *}(\boldsymbol{\varphi})\,\hat R(\boldsymbol{\varphi}).$$

Applied to a broken-symmetry state, this extracts the desired component with good quantum numbers; projected energies are then ratios of Hamiltonian and norm kernels (Yao, 2022).

For finite groups, the standard character projector appears in both operator-theoretic and quantum-computing contexts: $$\hat P^\Gamma = \frac{d_\Gamma}{|G|}\sum_{k=1}^{|G|} \chi_\Gamma(\hat O_k)^*\, \hat O_k.$$ For a single Hermitian symmetry operator, projector construction can instead be expressed through operator-valued indicator functions, including the Fourier form

$$\hat P_j = \frac{1}{2\pi}\int_0^{2\pi} e^{i\phi(\hat O-o_j)}\, d\phi,$$

the Lagrange-interpolation form

$$\hat P_j^{(i)} = \prod_{n\neq j}\frac{\hat O_i-o_n^{(i)}}{o_j^{(i)}-o_n^{(i)}},$$

and the resolvent-contour form

$$\hat P_j^{(i)}= \frac{1}{2\pi i}\oint_{C(o_j^{(i)})} (\hat O_i-z)^{-1}\,dz.$$

The general methodology is to identify the algebraic structure of the symmetry operators—finite group, Lie group, or Lie algebra—then construct projectors only for a fully commuting label set, typically using Cartan generators and Casimir operators (Izmaylov, 2019).

A different but closely related formulation projects not onto an irrep of a group action on states, but onto a symmetry-constrained operator subspace. In "Prism" (Xie, 18 May 2026), the symmetry constraint is

$[L,P]=LP-PL=0,$

where $L$ is a graph Laplacian and $P$ is a symmetric involution. The nearest $L'$ satisfying that constraint is obtained by diagonalizing $$\hat{P}_{\mu\mu}^{\lambda} \equiv \frac{d_{\lambda}}{n_G}\int d\boldsymbol{\varphi}\, D_{\mu\mu}^{\lambda *}(\boldsymbol{\varphi})\,\hat R(\boldsymbol{\varphi}).$$0 and zeroing the off-block entries: $$\hat{P}_{\mu\mu}^{\lambda} \equiv \frac{d_{\lambda}}{n_G}\int d\boldsymbol{\varphi}\, D_{\mu\mu}^{\lambda *}(\boldsymbol{\varphi})\,\hat R(\boldsymbol{\varphi}).$$1 with $$\hat{P}_{\mu\mu}^{\lambda} \equiv \frac{d_{\lambda}}{n_G}\int d\boldsymbol{\varphi}\, D_{\mu\mu}^{\lambda *}(\boldsymbol{\varphi})\,\hat R(\boldsymbol{\varphi}).$$2 if $$\hat{P}_{\mu\mu}^{\lambda} \equiv \frac{d_{\lambda}}{n_G}\int d\boldsymbol{\varphi}\, D_{\mu\mu}^{\lambda *}(\boldsymbol{\varphi})\,\hat R(\boldsymbol{\varphi}).$$3, and $$\hat{P}_{\mu\mu}^{\lambda} \equiv \frac{d_{\lambda}}{n_G}\int d\boldsymbol{\varphi}\, D_{\mu\mu}^{\lambda *}(\boldsymbol{\varphi})\,\hat R(\boldsymbol{\varphi}).$$4 otherwise. The same paper defines the normalized commutator defect

$$\hat{P}_{\mu\mu}^{\lambda} \equiv \frac{d_{\lambda}}{n_G}\int d\boldsymbol{\varphi}\, D_{\mu\mu}^{\lambda *}(\boldsymbol{\varphi})\,\hat R(\boldsymbol{\varphi}).$$5

which measures symmetry violation in Frobenius norm (Xie, 18 May 2026).

A further generalization appears in the study of nearby commuting operators with prescribed symmetries. There, symmetry maps $$\hat{P}_{\mu\mu}^{\lambda} \equiv \frac{d_{\lambda}}{n_G}\int d\boldsymbol{\varphi}\, D_{\mu\mu}^{\lambda *}(\boldsymbol{\varphi})\,\hat R(\boldsymbol{\varphi}).$$6 act on a unital $$\hat{P}_{\mu\mu}^{\lambda} \equiv \frac{d_{\lambda}}{n_G}\int d\boldsymbol{\varphi}\, D_{\mu\mu}^{\lambda *}(\boldsymbol{\varphi})\,\hat R(\boldsymbol{\varphi}).$$7-algebra and one seeks nearby commuting approximants that preserve $$\hat{P}_{\mu\mu}^{\lambda} \equiv \frac{d_{\lambda}}{n_G}\int d\boldsymbol{\varphi}\, D_{\mu\mu}^{\lambda *}(\boldsymbol{\varphi})\,\hat R(\boldsymbol{\varphi}).$$8-symmetry, $$\hat{P}_{\mu\mu}^{\lambda} \equiv \frac{d_{\lambda}}{n_G}\int d\boldsymbol{\varphi}\, D_{\mu\mu}^{\lambda *}(\boldsymbol{\varphi})\,\hat R(\boldsymbol{\varphi}).$$9-antisymmetry, or $$\hat P^\Gamma = \frac{d_\Gamma}{|G|}\sum_{k=1}^{|G|} \chi_\Gamma(\hat O_k)^*\, \hat O_k.$$0-phase symmetry. The central device is again projection, now through spectral projections and block decompositions adapted to the symmetry maps (Herrera, 2024).

3. Symmetry breaking and restoration in many-body theory

In nuclear structure theory, symmetry projection is formulated as the restoration of good quantum numbers from intrinsic Hartree-Fock or Hartree-Fock-Bogoliubov states that deliberately break symmetry to capture deformation or pairing. The canonical examples are parity projection,

$$\hat P^\Gamma = \frac{d_\Gamma}{|G|}\sum_{k=1}^{|G|} \chi_\Gamma(\hat O_k)^*\, \hat O_k.$$1

particle-number projection,

$$\hat P^\Gamma = \frac{d_\Gamma}{|G|}\sum_{k=1}^{|G|} \chi_\Gamma(\hat O_k)^*\, \hat O_k.$$2

and angular-momentum projection,

$$\hat P^\Gamma = \frac{d_\Gamma}{|G|}\sum_{k=1}^{|G|} \chi_\Gamma(\hat O_k)^*\, \hat O_k.$$3

The projected states are non-orthogonal, so one solves generalized eigenvalue problems built from Hamiltonian and norm kernels rather than ordinary diagonalizations (Yao, 2022).

A central methodological divide is between projection after variation and variation after projection. In the quantum-computing implementation of symmetry restoration for the pairing Hamiltonian, the projected energy functional is

$$\hat P^\Gamma = \frac{d_\Gamma}{|G|}\sum_{k=1}^{|G|} \chi_\Gamma(\hat O_k)^*\, \hat O_k.$$4

while Q-VAP minimizes the projected functional itself, and the paper states the hierarchy

$$\hat P^\Gamma = \frac{d_\Gamma}{|G|}\sum_{k=1}^{|G|} \chi_\Gamma(\hat O_k)^*\, \hat O_k.$$5

In that setting, the projector is implemented through QPE as a symmetry filter for fixed particle number (Guzman et al., 2021).

The TAURUS_vap program realizes genuine variation after particle-number projection for general real Bogoliubov quasiparticle vacua in a spherical harmonic oscillator basis. The projected state is

$$\hat P^\Gamma = \frac{d_\Gamma}{|G|}\sum_{k=1}^{|G|} \chi_\Gamma(\hat O_k)^*\, \hat O_k.$$6

and the VAP functional is

$$\hat P^\Gamma = \frac{d_\Gamma}{|G|}\sum_{k=1}^{|G|} \chi_\Gamma(\hat O_k)^*\, \hat O_k.$$7

The paper emphasizes that VAP is variationally superior to PAV but much more expensive because projection must be performed at every iteration (Bally et al., 2020).

The Lipkin-model analysis of the merger between projected Hartree-Fock and coupled-cluster theory makes the complementarity especially explicit. There, parity projection of a broken-symmetry determinant yields

$$\hat P^\Gamma = \frac{d_\Gamma}{|G|}\sum_{k=1}^{|G|} \chi_\Gamma(\hat O_k)^*\, \hat O_k.$$8

and the broad conclusion is that symmetry projection and coupled cluster fail in different ways over different correlation limits, whereas their merger succeeds across weakly correlated, strongly correlated, and recoupling regimes (Wahlen-Strothman et al., 2016).

4. Projected coupled cluster and quantum-algorithmic variants

Projected coupled cluster replaces the ordinary ansatz by

$$\hat P^\Gamma = \frac{d_\Gamma}{|G|}\sum_{k=1}^{|G|} \chi_\Gamma(\hat O_k)^*\, \hat O_k.$$9

with amplitude equations

$$\hat P_j = \frac{1}{2\pi}\int_0^{2\pi} e^{i\phi(\hat O-o_j)}\, d\phi,$$0

The crucial point is that the cluster amplitudes are optimized in the presence of the projector, not optimized first and projected afterward. In the formulation of the disentangled-cluster approximation, one rewrites

$$\hat P_j = \frac{1}{2\pi}\int_0^{2\pi} e^{i\phi(\hat O-o_j)}\, d\phi,$$1

which reduces projected kernels to CC-like expressions with an angle-dependent disentangled operator $$\hat P_j = \frac{1}{2\pi}\int_0^{2\pi} e^{i\phi(\hat O-o_j)}\, d\phi,$$2 (Qiu et al., 2018).

The numerical lesson of that work is that variation-after-projection at the CC level is materially different from projection-after-variation. For the half-filled Hubbard systems studied there, the paper reports that VAP-SUCCSD gives errors within about $$\hat P_j = \frac{1}{2\pi}\int_0^{2\pi} e^{i\phi(\hat O-o_j)}\, d\phi,$$3 per electron almost everywhere, and that VAP also accelerates convergence with respect to truncation of the disentangled operator $$\hat P_j = \frac{1}{2\pi}\int_0^{2\pi} e^{i\phi(\hat O-o_j)}\, d\phi,$$4 (Qiu et al., 2018). The earlier Lipkin analysis had already indicated the same qualitative pattern: projected formulations inherit strong-correlation physics from the broken-symmetry reference while recovering weak-correlation effects through cluster amplitudes (Wahlen-Strothman et al., 2016).

A different projected-CC route is the Monte Carlo "projection-after-coupled-cluster" scheme, where the symmetry projector is written as

$$\hat P_j = \frac{1}{2\pi}\int_0^{2\pi} e^{i\phi(\hat O-o_j)}\, d\phi,$$5

and the projected energy is evaluated as the Rayleigh–Ritz ratio

$$\hat P_j = \frac{1}{2\pi}\int_0^{2\pi} e^{i\phi(\hat O-o_j)}\, d\phi,$$6

The Monte Carlo estimator is built from

$$\hat P_j = \frac{1}{2\pi}\int_0^{2\pi} e^{i\phi(\hat O-o_j)}\, d\phi,$$7

with MCMC sampling in a spherical basis and a truncation over deformed-basis configurations. In the three-level Lipkin benchmark, the paper states that projected CCD recovers about $$\hat P_j = \frac{1}{2\pi}\int_0^{2\pi} e^{i\phi(\hat O-o_j)}\, d\phi,$$8 of the correlation energy (Mizusaki et al., 2021).

Symmetry projection has also been extended to discrete antiunitary and spin-parity symmetries in coupled cluster. The projected wave function may restore complex conjugation via

$$\hat P_j = \frac{1}{2\pi}\int_0^{2\pi} e^{i\phi(\hat O-o_j)}\, d\phi,$$9

or, after gauge fixing, $$\hat P_j^{(i)} = \prod_{n\neq j}\frac{\hat O_i-o_n^{(i)}}{o_j^{(i)}-o_n^{(i)}},$$0, and spin-flip via

$$\hat P_j^{(i)} = \prod_{n\neq j}\frac{\hat O_i-o_n^{(i)}}{o_j^{(i)}-o_n^{(i)}},$$1

with time reversal recovered through $$\hat P_j^{(i)} = \prod_{n\neq j}\frac{\hat O_i-o_n^{(i)}}{o_j^{(i)}-o_n^{(i)}},$$2. The resulting projected-CC equations are again of the form

$$\hat P_j^{(i)} = \prod_{n\neq j}\frac{\hat O_i-o_n^{(i)}}{o_j^{(i)}-o_n^{(i)}},$$3

with nonorthogonal-kernel evaluation handled by determinant transformations and disentangling (Song et al., 2024).

In VQE, the projected functional is

$$\hat P_j^{(i)} = \prod_{n\neq j}\frac{\hat O_i-o_n^{(i)}}{o_j^{(i)}-o_n^{(i)}},$$4

while exact and approximate projector constructions are compared against penalty-function approaches. The paper emphasizes a practical tradeoff: projectors usually introduce higher numbers of terms for measurement, but improve the accuracy of the variational ansatz without introducing additional unitary transformations, which is beneficial for reducing depths of quantum circuits (Yen et al., 2019).

5. Projection onto symmetry-constrained tensors and operators

In materials modeling, the relevant object is often not a wave function but a tensor whose components should reflect a crystal symmetry class. The projection problem studied in "Extended scheme for the projection of material tensors of arbitrary symmetry onto a higher symmetry tensor" (Caro, 2014) is: given a tensor $$\hat P_j^{(i)} = \prod_{n\neq j}\frac{\hat O_i-o_n^{(i)}}{o_j^{(i)}-o_n^{(i)}},$$5 of low apparent symmetry, find the tensor of a chosen higher symmetry class that is closest to it. The paper extends the Moakher–Norris framework by including explicit optimization over rotations, because a physically cubic or hexagonal tensor may look low-symmetry merely because it is expressed in a rotated frame. For elasticity, the rank-4 tensor is mapped to a 21-component vector with normalization factors chosen so that the Euclidean vector norm matches the tensor norm, and the closest higher-symmetry tensor is found only after orientation has been optimized (Caro, 2014).

A closely analogous idea appears in graph symmetry diagnostics. There the target set is the commutant

$$\hat P_j^{(i)} = \prod_{n\neq j}\frac{\hat O_i-o_n^{(i)}}{o_j^{(i)}-o_n^{(i)}},$$6

and the projection of a Laplacian $$\hat P_j^{(i)} = \prod_{n\neq j}\frac{\hat O_i-o_n^{(i)}}{o_j^{(i)}-o_n^{(i)}},$$7 onto that symmetry-constrained set is

$$\hat P_j^{(i)} = \prod_{n\neq j}\frac{\hat O_i-o_n^{(i)}}{o_j^{(i)}-o_n^{(i)}},$$8

The solution is the block-diagonal projection in the eigenspace decomposition of the involution $$\hat P_j^{(i)} = \prod_{n\neq j}\frac{\hat O_i-o_n^{(i)}}{o_j^{(i)}-o_n^{(i)}},$$9, and the commutator norm $$\hat P_j^{(i)}= \frac{1}{2\pi i}\oint_{C(o_j^{(i)})} (\hat O_i-z)^{-1}\,dz.$$0 becomes a scalar defect measuring structural symmetry breaking (Xie, 18 May 2026).

Projection methods for nearby commuting operators introduce yet another operator-level formulation. Given a unitary or self-adjoint operator and an additional operator $$\hat P_j^{(i)}= \frac{1}{2\pi i}\oint_{C(o_j^{(i)})} (\hat O_i-z)^{-1}\,dz.$$1, the existence of nearby commuting approximants is characterized by families of spectral projections $$\hat P_j^{(i)}= \frac{1}{2\pi i}\oint_{C(o_j^{(i)})} (\hat O_i-z)^{-1}\,dz.$$2 satisfying localization conditions such as

$$\hat P_j^{(i)}= \frac{1}{2\pi i}\oint_{C(o_j^{(i)})} (\hat O_i-z)^{-1}\,dz.$$3

together with small commutators $$\hat P_j^{(i)}= \frac{1}{2\pi i}\oint_{C(o_j^{(i)})} (\hat O_i-z)^{-1}\,dz.$$4. Those projections can then be used to construct nearby commuting $$\hat P_j^{(i)}= \frac{1}{2\pi i}\oint_{C(o_j^{(i)})} (\hat O_i-z)^{-1}\,dz.$$5 or $$\hat P_j^{(i)}= \frac{1}{2\pi i}\oint_{C(o_j^{(i)})} (\hat O_i-z)^{-1}\,dz.$$6 that preserve the same symmetries, antisymmetries, or phase symmetries as the original pair (Herrera, 2024).

At a more foundational level, projector construction on eigensubspaces can itself be treated as an operator-synthesis problem. The work on "On construction of projection operators" (Izmaylov, 2019) organizes this systematically through finite-group character projectors, Lie-algebra label sets, and spectral-indicator functions such as Fourier, polynomial, and contour-integral formulas. A related but distinct line of work classifies symmetries of projections, effects, and self-adjoint operators by reducing nonlinear preservation problems to projective geometry, adjacency preservers, and Wigner–Uhlhorn-type theorems (Semrl, 30 Jul 2025). This suggests that symmetry-projection techniques are not confined to numerical approximation: they also encode deep structural relations between symmetry, spectral decomposition, and operator geometry.

6. Projection-domain and geometric methods

In projection-domain imaging, symmetry can be exploited without first reconstructing a volume. For plane-symmetric transmission objects, if $$\hat P_j^{(i)}= \frac{1}{2\pi i}\oint_{C(o_j^{(i)})} (\hat O_i-z)^{-1}\,dz.$$7 is a calibrated projection matrix and $$\hat P_j^{(i)}= \frac{1}{2\pi i}\oint_{C(o_j^{(i)})} (\hat O_i-z)^{-1}\,dz.$$8 is the reflection across the symmetry plane, the mirrored camera is

$$\hat P_j^{(i)}= \frac{1}{2\pi i}\oint_{C(o_j^{(i)})} (\hat O_i-z)^{-1}\,dz.$$9

The key claim is that a real view and its mirror-reflected counterpart yield the same projection image for an exactly plane-symmetric object. This permits estimation of the 3-D symmetry plane directly from projection data through Grangeat’s theorem and epipolar consistency, and then generation of a virtual mirrored trajectory. When the symmetry plane is oblique to the acquisition plane, the real and virtual trajectories form an X-shaped configuration that improves data completeness for motion analysis (Preuhs et al., 2018).

In category-specific 3-D reconstruction under orthography, bilateral reflection symmetry is encoded by

$[L,P]=LP-PL=0,$0

and the key decoupling is

$[L,P]=LP-PL=0,$1

The antisymmetric part $[L,P]=LP-PL=0,$2 depends only on the reflected coordinate, while the symmetric part $[L,P]=LP-PL=0,$3 depends only on the invariant coordinates. In the single-image case, Manhattan structure identifies the camera and symmetry then makes the shape uniquely recoverable; in the multiple-image case, the same decomposition yields a surrogate factorization into independent rank-1 and rank-2 blocks (Gao et al., 2016).

Random projections provide a probabilistic symmetry-identification mechanism for multivariate distributions. For a candidate axis $[L,P]=LP-PL=0,$4 and a projection direction $[L,P]=LP-PL=0,$5, the Kolmogorov discrepancy is

$[L,P]=LP-PL=0,$6

and for a finite set $[L,P]=LP-PL=0,$7,

$[L,P]=LP-PL=0,$8

In $[L,P]=LP-PL=0,$9, for a non-spherically symmetric distribution satisfying the Carleman condition, the paper proves that agreement on two independent random projections is enough, almost surely, to recover exactly the true set of axes of symmetry. It then proposes the level-set estimator

$L$0

and proves Hausdorff consistency in the plane (Cholaquidis et al., 24 Dec 2025).

These geometric examples show that “projection” may refer either to a linear projector in a function space or to the act of passing to lower-dimensional observations. The shared mechanism is that symmetry information is preserved in carefully chosen projected observables, and can therefore be restored or identified without full inversion of the original high-dimensional problem.

7. Approximations, tradeoffs, and recurrent limitations

Several tensions recur across the literature. The first is the distinction between exact and approximate projectors. In VQE, exact projectors constructed from full spectra or continuous group averages are idempotent, but approximate projectors obtained by removing only selected unwanted sectors are in general not idempotent: $L$1 The motivation is practical: exact number or spin projection can be measurement-expensive, whereas approximate projectors may already suppress the wrong sectors sufficiently for the variational search to find the desired state (Yen et al., 2019).

The second is the tension between projection after variation and variation after projection. VAP is repeatedly described as more accurate or variationally superior, but also much more expensive because the projection machinery enters every optimization step rather than a single post-processing evaluation (Guzman et al., 2021, Bally et al., 2020, Qiu et al., 2018). In projected coupled cluster, this is compounded by the need to handle gauge modes associated with the Goldstone manifold of rotated broken-symmetry references (Qiu et al., 2018).

A third recurrent limitation is that projection does not eliminate all model deficiencies. Symmetry-projected Hartree-Fock is not size extensive, and projected mean-field alone is not ideal for dynamic correlation (Wahlen-Strothman et al., 2016). Spin-flip projection in complex-conjugation projected CC is only half-spin projection, not full restoration of $L$2 (Song et al., 2024). Monte Carlo projection-after-coupled-cluster introduces statistical error and deformed-basis truncation, even though it avoids deterministic manipulation of the full projected state (Mizusaki et al., 2021).

A fourth issue is that many applications rely on approximate symmetry rather than exact symmetry. In transmission imaging, the plane symmetry of a head phantom is only approximate, so mirrored views are not exactly equivalent, even though the method still estimates a useful symmetry plane (Preuhs et al., 2018). In object reconstruction, bilateral symmetry and subtype rigidity are modeling assumptions rather than exact geometric facts (Gao et al., 2016). In random-projection symmetry estimation, the sharp planar identification theorem depends on the Carleman condition, and the higher-dimensional analogue remains conjectural (Cholaquidis et al., 24 Dec 2025).

Taken together, these works suggest that symmetry projection is best viewed not as a single algorithmic recipe but as a family of constructions that trade representational flexibility against exact symmetry, then recover the desired symmetry by explicit projection, constrained minimization, or projection-domain inference. Its success depends on how faithfully the chosen symmetry describes the object, how expensive exact restoration is, and whether the projected formulation is embedded into optimization itself or applied only after an initial broken-symmetry approximation has already been fixed.