Shift Vectors in Physics and Machine Learning

• Shift vectors are controlled displacements in representation spaces, defined by gauge-invariant quantities in quantum materials and measurable phase shifts in scattering processes.
• They quantify intrinsic dipoles and spatial displacements for many-body states and topological interfaces, linking Berry-curvature differences with observable optical effects.
• In machine learning and operator theory, shift vectors serve as additive control directions for steering activations and characterizing dense orbit behaviors in structured function spaces.

“Shift vector” is a field-dependent term rather than a single universal concept. In the literature considered here, it denotes at least five distinct classes of objects: a gauge-invariant geometric displacement associated with interband transitions in quantum materials; an anomalous spatial displacement of a reflected electron beam at an interface; a many-body intrinsic dipole of correlated electronic states; a linear direction added to hidden activations in machine learning models; and vectors whose dynamics are governed by shift operators or translations in functional analysis and frame theory. What unifies these usages is a common structural role: each shift vector encodes a controlled or measurable displacement in a representation space, whether that space is Bloch momentum, interface momentum, latent-feature space, or a sequence space (Wang et al., 2024).

1. Quantum-geometric shift vectors in band theory

In condensed-matter physics, the canonical shift vector is a gauge-invariant geometric quantity governing interband contributions to second-order nonlinear optical responses such as the bulk photovoltaic effect and the circular photogalvanic effect. Its standard form is

$$R_{mn}^{a;b}(\mathbf{k}) = A_{mm}^{a}(\mathbf{k}) - A_{nn}^{a}(\mathbf{k}) - \partial_{k_a}\, \arg r_{mn}^{b}(\mathbf{k}),$$

where $A_{nn}^{a}$ is the intraband Berry connection and $r_{mn}^{b}$ is the interband dipole matrix element. Under independent $U(1)$ gauge transformations of the Bloch states, the Berry-connection difference and the phase-gradient term transform oppositely, so the combination remains invariant (Wang et al., 2024).

The same work gives the shift vector a geometric interpretation: it is identified with a geodesic curvature on the parameter manifold defined by Bloch momentum. In a two-band model, the paper states

$$K_g(\mathbf{k}) = \frac{R_{21}(\mathbf{k})}{2\, g_{12}(\mathbf{k})},$$

linking the shift vector to the quantum metric $g_{12}$. It also derives the differential relation

$$\nabla_{\mathbf{k}}\times \mathbf{R}_{mn}(\mathbf{k}) = \Omega_m(\mathbf{k}) - \Omega_n(\mathbf{k}),$$

so the shift vector is tied directly to Berry-curvature differences rather than merely to transition amplitudes (Wang et al., 2024).

A further result is a Gauss–Bonnet-like quantization formula,

$$2\pi\, X_{mn} = \int_{M} \Omega_{mn}(\mathbf{k})\, d^{2}\mathbf{k} - \oint_{\partial M} \mathbf{R}_{mn}(\mathbf{k})\cdot d\mathbf{k},$$

where $X_{mn}\in\mathbb{Z}$. In this framework, the loop integral of the shift vector contributes the non-quantized component of the trace of the circular photogalvanic conductivity, whereas the Berry-flux term yields the quantized part when the selected surface encloses a topological singularity. The same formalism is extended to non-vertical transitions and bosonic phonon-drag shift vectors through matrix elements of the electron–phonon coupling (Wang et al., 2024).

2. Correlated and excitonic generalizations

Recent work generalizes the shift vector from a transition attribute to an intrinsic property of a single correlated many-body state. For a correlated state

$$|\Psi^S\rangle = \sum_\Lambda A^S_\Lambda |\Lambda\rangle,$$

the intrinsic dipole per charge is written as $A_{nn}^{a}$0, with

$A_{nn}^{a}$1

In this formulation, the phase-gradient term reflects many-body coherence, while the second term is inherited from single-particle Berry connections. The paper argues that the generalized shift vector is therefore the intrinsic electric dipole moment of a correlated state rather than merely a difference between two states (Hu et al., 22 May 2026).

This perspective recovers previously proposed shift vectors as special cases. For electron–hole states, the two-band reduction becomes

$A_{nn}^{a}$2

while optically induced correlations reproduce the conventional optical shift vector and standard shift-current expression. The same framework also yields electron–phonon-mediated and excitonic shift vectors, and interprets the result both as a displacement of the real-space joint probability density and as a linear Stark response

$A_{nn}^{a}$3

in a static electric field (Hu et al., 22 May 2026).

A distinct but closely related result concerns bound excitons. Using a many-body flux-threading definition,

$A_{nn}^{a}$4

the authors show that the excitonic shift vector is exponentially insensitive to the form of the light–matter interaction and hence independent of light polarization in the thermodynamic limit. They further prove that, for vertical excitonic transitions with $A_{nn}^{a}$5, the excitonic shift vector transforms as an ordinary vector under point-group operations. A sharp consequence is that, in noncentrosymmetric but non-polar materials, vertical excitonic transitions have vanishing shift vector and therefore zero excitonic shift photocurrent, in contrast to the finite and strongly polarization-dependent shift vectors of non-interacting delocalized particle-hole excitations (Yang et al., 9 Jul 2025).

These results challenge a common assumption that shift vectors are inherently polarization-tagged transition quantities. A plausible implication is that many-body localization in the electron–hole relative coordinate reorganizes the relevant geometry from a light-coupling-dependent phase structure to an intrinsic correlated-state geometry (Yang et al., 9 Jul 2025).

3. Anomalous scattering shift vectors in topological semimetals

In interface scattering, a shift vector can denote the anomalous spatial displacement of a reflected electron beam. For a planar interface at $A_{nn}^{a}$6, the reflected wavepacket acquires an in-plane displacement

$A_{nn}^{a}$7

with $A_{nn}^{a}$8, $A_{nn}^{a}$9, and $r_{mn}^{b}$0. When the incident medium is trivial or symmetry enforces $r_{mn}^{b}$1, this reduces to the Artmann-type form

$r_{mn}^{b}$2

The shift vector then becomes a direct phase-space map from the reflection phase to a real-space displacement (Li et al., 16 Dec 2025).

For mirror-symmetry-protected nodal-ring semimetals, the shift-vector field in the interface momentum plane encodes both geometry and topology. In the minimal continuum model, the nodal ring in the $r_{mn}^{b}$3 plane carries an integer mirror hemisphere charge $r_{mn}^{b}$4, and the circulation of the shift field over a symmetry-adapted semicircle is quantized:

$r_{mn}^{b}$5

For $r_{mn}^{b}$6, corresponding to a conventional mirror-protected nodal ring, $r_{mn}^{b}$7 and there is no phase winding. For vortex nodal rings with $r_{mn}^{b}$8, the paper states that $r_{mn}^{b}$9 winds by $U(1)$0 along the semicircle, giving $U(1)$1 (Li et al., 16 Dec 2025).

The shift magnitude is strongly enhanced near the projection of the nodal ring onto the interface momentum plane. In the conventional-ring limit near the Fermi level, the paper gives the anisotropy

$U(1)$2

with $U(1)$3, so the shift points nearly perpendicular to the mirror plane. Mapping the locus of large $U(1)$4 reconstructs the ring projection and radius $U(1)$5. The circulation $U(1)$6 also diagnoses topological phase transitions: it remains $U(1)$7 across the ring-to-Weyl-pair transition for loops enclosing the upper charged degeneracy, but can jump to zero when a compensating Weyl charge crosses the semicircle (Li et al., 16 Dec 2025).

The paper explicitly places this phenomenon in correspondence with electronic analogues of Goos–Hänchen and Imbert–Fedorov beam shifts. In that usage, the shift vector is not an abstract topological invariant by itself; it is a measurable scattering displacement whose circulation becomes topological once mirror symmetry and the appropriate interface geometry are imposed (Li et al., 16 Dec 2025).

4. Representation-space shift vectors in machine learning

In contemporary machine learning, “shift vectors” often denote linear directions added to internal activations to steer model behavior at inference time. In multimodal LLMs, a vector $U(1)$8 derived from a text-only backbone can be added to residual-stream hidden states at layer $U(1)$9,

$$K_g(\mathbf{k}) = \frac{R_{21}(\mathbf{k})}{2\, g_{12}(\mathbf{k})},$$0

for selected image tokens, text tokens, or both. The paper “Textual Steering Vectors Can Improve Visual Understanding in Multimodal LLMs” states that mean shift boosts spatial relationship accuracy on CV-Bench by up to $$K_g(\mathbf{k}) = \frac{R_{21}(\mathbf{k})}{2\, g_{12}(\mathbf{k})},$$1 and counting accuracy by up to $$K_g(\mathbf{k}) = \frac{R_{21}(\mathbf{k})}{2\, g_{12}(\mathbf{k})},$$2, with strong out-of-distribution generalization and no multimodal data required to construct the vectors (Gan et al., 20 May 2025).

A related study analyzes fine-tuning-induced representation shifts in multimodal LLMs by defining concept shift vectors between an original model $$K_g(\mathbf{k}) = \frac{R_{21}(\mathbf{k})}{2\, g_{12}(\mathbf{k})},$$3 and a fine-tuned model $$K_g(\mathbf{k}) = \frac{R_{21}(\mathbf{k})}{2\, g_{12}(\mathbf{k})},$$4:

$$K_g(\mathbf{k}) = \frac{R_{21}(\mathbf{k})}{2\, g_{12}(\mathbf{k})},$$5

These directions can be added directly to residual states,

$$K_g(\mathbf{k}) = \frac{R_{21}(\mathbf{k})}{2\, g_{12}(\mathbf{k})},$$6

to recover fine-tuned concepts from the original model and to steer answer types, specific answers, and caption styles without any further training. The paper reports that deeper layers yield stronger and more targeted effects, and that $$K_g(\mathbf{k}) = \frac{R_{21}(\mathbf{k})}{2\, g_{12}(\mathbf{k})},$$7 is near optimal for concept recovery across several fine-tunings (Khayatan et al., 6 Jan 2025).

In language-model safety and behavior control, persona vectors are diff-in-means directions in residual space associated with traits such as evil, sycophancy, and hallucination. They are used both for monitoring and intervention through

$$K_g(\mathbf{k}) = \frac{R_{21}(\mathbf{k})}{2\, g_{12}(\mathbf{k})},$$8

The paper reports correlations of $$K_g(\mathbf{k}) = \frac{R_{21}(\mathbf{k})}{2\, g_{12}(\mathbf{k})},$$9–$g_{12}$0 between last-prompt projections and subsequent trait expression under system prompting, and correlations of $g_{12}$1–$g_{12}$2 between finetuning-induced activation shifts and post-finetuning trait scores. It also introduces preventative steering during finetuning, which reduces trait acquisition while preserving capability better than post-hoc inference steering (Chen et al., 29 Jul 2025).

In classifier editing, Class Vectors are defined as class-centroid differences between a pretrained and a fine-tuned encoder,

$g_{12}$3

They support latent steering,

$g_{12}$4

and a Jacobian-based mapping into weight space. The paper argues that Class Vectors capture each class’s semantic shift, support class arithmetic, and can be used for unlearning, environmental adaptation, adversarial defense, and adversarial trigger optimization (Kim et al., 13 Oct 2025).

The same activation-addition principle has also been extended to diffusion transformers. SHIFT learns concept-specific steering vectors for pooled text embeddings and intermediate transformer activations, then applies

$g_{12}$5

during denoising. The method is described as training-free at deployment time and is used for concept removal, style-domain shifts, object addition, and concept switching in Flux-family DiT models (Konovalova et al., 10 Apr 2026).

Across these machine-learning uses, the term denotes a linear representation-space displacement rather than a Berry-geometric quantity. This suggests a convergence of vocabulary around additive control directions, even when the underlying spaces and objectives differ sharply (Gan et al., 20 May 2025).

5. Shift vectors in linear dynamics and operator theory

In operator theory, “shift vectors” typically mean vectors whose orbits under backward shifts or shift-like operators have dense, frequent, or algebraically structured behavior. For the unweighted backward shift $g_{12}$6 on $g_{12}$7, Tsirivas studies common hypercyclic vectors for the family $g_{12}$8. The central threshold is the reciprocal series $g_{12}$9: if it converges, there are no common hypercyclic vectors for any real interval $$\nabla_{\mathbf{k}}\times \mathbf{R}_{mn}(\mathbf{k}) = \Omega_m(\mathbf{k}) - \Omega_n(\mathbf{k}),$$0; if it diverges, the intersection over all real $$\nabla_{\mathbf{k}}\times \mathbf{R}_{mn}(\mathbf{k}) = \Omega_m(\mathbf{k}) - \Omega_n(\mathbf{k}),$$1 is residual in $$\nabla_{\mathbf{k}}\times \mathbf{R}_{mn}(\mathbf{k}) = \Omega_m(\mathbf{k}) - \Omega_n(\mathbf{k}),$$2, and for a $$\nabla_{\mathbf{k}}\times \mathbf{R}_{mn}(\mathbf{k}) = \Omega_m(\mathbf{k}) - \Omega_n(\mathbf{k}),$$3 dense full-measure subset of complex $$\nabla_{\mathbf{k}}\times \mathbf{R}_{mn}(\mathbf{k}) = \Omega_m(\mathbf{k}) - \Omega_n(\mathbf{k}),$$4 with $$\nabla_{\mathbf{k}}\times \mathbf{R}_{mn}(\mathbf{k}) = \Omega_m(\mathbf{k}) - \Omega_n(\mathbf{k}),$$5 the same remains true (Tsirivas, 2015).

A different construction treats operators that “act like a shift” on a dense bi-infinite family $$\nabla_{\mathbf{k}}\times \mathbf{R}_{mn}(\mathbf{k}) = \Omega_m(\mathbf{k}) - \Omega_n(\mathbf{k}),$$6 satisfying

$$\nabla_{\mathbf{k}}\times \mathbf{R}_{mn}(\mathbf{k}) = \Omega_m(\mathbf{k}) - \Omega_n(\mathbf{k}),$$7

Under suitable summability hypotheses, the random series

$$\nabla_{\mathbf{k}}\times \mathbf{R}_{mn}(\mathbf{k}) = \Omega_m(\mathbf{k}) - \Omega_n(\mathbf{k}),$$8

converges almost surely, is frequently hypercyclic for $$\nabla_{\mathbf{k}}\times \mathbf{R}_{mn}(\mathbf{k}) = \Omega_m(\mathbf{k}) - \Omega_n(\mathbf{k}),$$9, and induces a $$2\pi\, X_{mn} = \int_{M} \Omega_{mn}(\mathbf{k})\, d^{2}\mathbf{k} - \oint_{\partial M} \mathbf{R}_{mn}(\mathbf{k})\cdot d\mathbf{k},$$0-invariant strongly mixing probability measure with full support. For weighted shifts on sequence spaces, the canonical choice is $$2\pi\, X_{mn} = \int_{M} \Omega_{mn}(\mathbf{k})\, d^{2}\mathbf{k} - \oint_{\partial M} \mathbf{R}_{mn}(\mathbf{k})\cdot d\mathbf{k},$$1, which reduces the shift-like hypothesis to the definition of the cumulative products $$2\pi\, X_{mn} = \int_{M} \Omega_{mn}(\mathbf{k})\, d^{2}\mathbf{k} - \oint_{\partial M} \mathbf{R}_{mn}(\mathbf{k})\cdot d\mathbf{k},$$2 (Agneessens, 2022).

The algebraic structure of such vectors is subtle. For multiples $$2\pi\, X_{mn} = \int_{M} \Omega_{mn}(\mathbf{k})\, d^{2}\mathbf{k} - \oint_{\partial M} \mathbf{R}_{mn}(\mathbf{k})\cdot d\mathbf{k},$$3 on $$2\pi\, X_{mn} = \int_{M} \Omega_{mn}(\mathbf{k})\, d^{2}\mathbf{k} - \oint_{\partial M} \mathbf{R}_{mn}(\mathbf{k})\cdot d\mathbf{k},$$4 or $$2\pi\, X_{mn} = \int_{M} \Omega_{mn}(\mathbf{k})\, d^{2}\mathbf{k} - \oint_{\partial M} \mathbf{R}_{mn}(\mathbf{k})\cdot d\mathbf{k},$$5 under coordinatewise multiplication, it is shown that if $$2\pi\, X_{mn} = \int_{M} \Omega_{mn}(\mathbf{k})\, d^{2}\mathbf{k} - \oint_{\partial M} \mathbf{R}_{mn}(\mathbf{k})\cdot d\mathbf{k},$$6, then there exists $$2\pi\, X_{mn} = \int_{M} \Omega_{mn}(\mathbf{k})\, d^{2}\mathbf{k} - \oint_{\partial M} \mathbf{R}_{mn}(\mathbf{k})\cdot d\mathbf{k},$$7 such that $$2\pi\, X_{mn} = \int_{M} \Omega_{mn}(\mathbf{k})\, d^{2}\mathbf{k} - \oint_{\partial M} \mathbf{R}_{mn}(\mathbf{k})\cdot d\mathbf{k},$$8 for every $$2\pi\, X_{mn} = \int_{M} \Omega_{mn}(\mathbf{k})\, d^{2}\mathbf{k} - \oint_{\partial M} \mathbf{R}_{mn}(\mathbf{k})\cdot d\mathbf{k},$$9. Consequently, no frequently hypercyclic algebra exists under the natural coordinatewise product. By contrast, weaker recurrence notions based on other Furstenberg families do allow algebras, and tailor-made products can restore algebrability or strong algebrability for general frequently hypercyclic operators (Falcó et al., 2019).

For hypercyclicity without the frequent-return requirement, weighted backward shifts on Fréchet sequence algebras can have algebrable sets of hypercyclic vectors under coordinatewise or Cauchy products. In particular, the paper on algebrability of hypercyclic vectors for backward shift operators shows that the classical Rolewicz operator $X_{mn}\in\mathbb{Z}$0 and MacLane’s differentiation operator $X_{mn}\in\mathbb{Z}$1 fit into a constructive scheme yielding infinitely generated subalgebras of hypercyclic vectors (Falcó et al., 2018).

A common misconception is to conflate these operator-theoretic shift vectors with geometric displacement vectors. Here the term refers not to a gauge-invariant positional shift, but to vectors in a topological vector space whose iterates under a shift operator realize dense-orbit phenomena (Tsirivas, 2015).

6. Shift-generated vectors, shift-invariant spaces, and adjacent mathematical uses

In frame theory on finitely generated shift-invariant subspaces $X_{mn}\in\mathbb{Z}$2, the relevant vectors are generators whose translates form a structured family. Given $X_{mn}\in\mathbb{Z}$3, the associated shift-generated system is

$X_{mn}\in\mathbb{Z}$4

with $X_{mn}\in\mathbb{Z}$5. The paper develops measurable Schur–Horn criteria, generalized eigensteps, and waterfilling constructions to characterize when such systems have prescribed fine spectral structure or minimal convex spectral spread. In this setting, the “shift vectors” are the finite family of generators $X_{mn}\in\mathbb{Z}$6 rather than a single geometric direction (Benac et al., 2015).

The broader mathematical literature in the supplied corpus also contains neighboring but distinct uses of “shift.” In monotone circuit complexity, the shift operator on Boolean vectors has complexity $X_{mn}\in\mathbb{Z}$7, with special case $X_{mn}\in\mathbb{Z}$8, but there the central object is an operator rather than a vector (Sergeev, 2019). In field theory on $X_{mn}\in\mathbb{Z}$9, shift-symmetric spin-1 theories admit transformations

$$|\Psi^S\rangle = \sum_\Lambda A^S_\Lambda |\Lambda\rangle,$$0

where $$|\Psi^S\rangle = \sum_\Lambda A^S_\Lambda |\Lambda\rangle,$$1 is an $$|\Psi^S\rangle = \sum_\Lambda A^S_\Lambda |\Lambda\rangle,$$2 Killing vector and the special mass is $$|\Psi^S\rangle = \sum_\Lambda A^S_\Lambda |\Lambda\rangle,$$3; the operative notion is symmetry under a vector-valued shift, not a shift vector as an observable quantity (Bonifacio et al., 2019).

These adjacent usages delimit the semantic range of the term. In some areas, a shift vector is a measurable displacement; in others, it is a latent steering direction, a generator of translated systems, or simply a vector acted on by a shift. The literature therefore supports no field-independent definition. Instead, the precise meaning must be read from the ambient structure: Berry geometry, scattering kinematics, residual-stream representations, or shift-operator dynamics (Benac et al., 2015).