Loop Hafnian Generating Function

• Generating Function for Loop Hafnians is a multivariate formal power series that packages loop hafnians from all blown-up replicas of a symmetric matrix with its loop vector.
• It employs a differential-operator approach that transforms combinatorial sums into coefficient extraction from Gaussian integrals, linking power series and determinants.
• The formulation extends beyond quantum-optical constraints, validating the identity for all symmetric matrices and emphasizing its relevance in Gaussian boson sampling.

Searching arXiv for related work on loop hafnians, generating functions, and the cited differential-operator representation. The generating function for loop hafnians is a multivariate formal power series that packages the loop hafnians of all “blown-up” replicas of a symmetric matrix and its associated loop vector into a single closed expression. In the formulation established in “Derivation of the Loop Hafnian Generating Function for Arbitrary Symmetric Matrices via Gaussian Integration” (Tarasov, 21 Jul 2025), if $S\in\mathbb C^{2m\times 2m}$ is symmetric, $\mathbf v\in\mathbb C^{2m}$, and $\mathbb Z$ is the block–counter-diagonal matrix with $\diag(z_1,\dots,z_m)$ in the off-diagonal blocks, then

$G(S,\mathbf v;\mathbf z) = \frac{ \exp\!\bigl(\tfrac12\,\mathbf v^T (I-\mathbb ZS)^{-1}\mathbb Z\,\mathbf v\bigr) }{ \sqrt{\det(I-\mathbb ZS)} } = \sum_{\mathbf n\ge 0} \lhaf(\tilde S_{\mathbf n},\tilde{\mathbf v}_{\mathbf n}) \frac{\mathbf z^{\mathbf n}}{\mathbf n!}.$

The principal significance of the 2025 derivation is that it removes the quantum-optical restrictions present in earlier formulations and proves validity for arbitrary symmetric matrices by means of Gaussian integration and analytic continuation alone (Tarasov, 21 Jul 2025).

1. Definition and combinatorial setting

For a complex symmetric matrix $S=(s_{ij})$ of size $\mu\times\mu$ and a vector $\mathbf v=(v_1,\dots,v_\mu)\in\mathbb C^\mu$, the loop hafnian $\lhaf(S,\mathbf v)$ is defined by summing over all partitions of $\{1,\dots,\mu\}$ into blocks that are either pairs $\mathbf v\in\mathbb C^{2m}$0 or singletons $\mathbf v\in\mathbb C^{2m}$1. Denoting this family of partitions by $\mathbf v\in\mathbb C^{2m}$2, one has

$\mathbf v\in\mathbb C^{2m}$3

This definition extends the ordinary hafnian by permitting singleton blocks, which are encoded through the loop vector $\mathbf v\in\mathbb C^{2m}$4 (Tarasov, 21 Jul 2025).

Two specializations are structurally important. If $\mathbf v\in\mathbb C^{2m}$5, the loop hafnian becomes a quantity determined entirely by $\mathbf v\in\mathbb C^{2m}$6. If $\mathbf v\in\mathbb C^{2m}$7, the singleton contributions vanish and the loop hafnian reduces to the ordinary hafnian $\mathbf v\in\mathbb C^{2m}$8, provided one is in the standard even-dimensional setting with diagonal entries treated accordingly (Tarasov, 21 Jul 2025). In this sense, the loop hafnian interpolates between pure pairing enumeration and pairing-with-singletons enumeration.

The combinatorial role of loop hafnians is closely related to the broader hafnian formalism developed for Gaussian boson sampling and Gaussian state amplitudes. The use of hafnians and loop hafnians as the relevant matrix functions for these amplitudes is central to the framework introduced by Quesada, Arrazola, and Killoran (Hatamizadeh et al., 2019). The 2025 generating-function result should therefore be understood as a structural identity for a matrix function already embedded in quantum-optical and combinatorial practice, but now detached from specifically physical assumptions (Tarasov, 21 Jul 2025).

2. Differential-operator representation and blown-up matrices

A central ingredient in the generating-function derivation is the differential-operator representation of loop hafnians. Let

$\mathbf v\in\mathbb C^{2m}$9

For any nonnegative integers $\mathbb Z$0, one forms a blown-up matrix $\mathbb Z$1 and blown-up vector $\mathbb Z$2 by replacing each index $\mathbb Z$3 with $\mathbb Z$4 copies, each entry $\mathbb Z$5 with the $\mathbb Z$6 all-ones block $\mathbb Z$7, and each $\mathbb Z$8 with $\mathbb Z$9 repeated entries. The loop hafnian of this enlarged data is then recovered through partial differentiation: $\diag(z_1,\dots,z_m)$0 The 2025 note treats this as a standard lemma and attributes it to earlier work, specifically Quesada et al. in 2019 (Tarasov, 21 Jul 2025).

This representation is important because it converts a combinatorial sum over single-pair matchings into coefficient extraction from an exponential with quadratic and linear terms. That mechanism places loop hafnians within the general orbit of Gaussian moment calculus. It also makes clear why blown-up matrices are natural: the multiplicities $\diag(z_1,\dots,z_m)$1 are encoded as derivative orders, and the replica structure appears because repeated differentiation with respect to the same variable corresponds combinatorially to repeated copies of the same vertex label.

In the hafnian literature, differential and generating identities of this type are closely tied to the matrix-function viewpoint used for Gaussian state probabilities and threshold detection formulas. In particular, the broader program of connecting Gaussian state observables to hafnians and loop hafnians was systematized in the work on Gaussian boson sampling with threshold detectors and displaced Gaussian states (Kumar et al., 2020, Hatamizadeh et al., 2019). The 2025 result refines that program at the level of formal generating structure rather than algorithmic output probabilities.

3. Gaussian-integral derivation

The derivation in (Tarasov, 21 Jul 2025) proceeds by taking $\diag(z_1,\dots,z_m)$2, grouping the indices into pairs $\diag(z_1,\dots,z_m)$3, and rewriting the replicated loop hafnian as

$\diag(z_1,\dots,z_m)$4

Assuming temporarily that $\diag(z_1,\dots,z_m)$5 is invertible, the exponential is represented by a Gaussian integral obtained by completing the square: $\diag(z_1,\dots,z_m)$6 In this integral representation, the $\diag(z_1,\dots,z_m)$7-derivatives act only on the term linear in $\diag(z_1,\dots,z_m)$8, which becomes a monomial in the components of $\diag(z_1,\dots,z_m)$9 after differentiation (Tarasov, 21 Jul 2025).

The next step introduces auxiliary variables $G(S,\mathbf v;\mathbf z) = \frac{ \exp\!\bigl(\tfrac12\,\mathbf v^T (I-\mathbb ZS)^{-1}\mathbb Z\,\mathbf v\bigr) }{ \sqrt{\det(I-\mathbb ZS)} } = \sum_{\mathbf n\ge 0} \lhaf(\tilde S_{\mathbf n},\tilde{\mathbf v}_{\mathbf n}) \frac{\mathbf z^{\mathbf n}}{\mathbf n!}.$0 through the block–counter-diagonal matrix

$G(S,\mathbf v;\mathbf z) = \frac{ \exp\!\bigl(\tfrac12\,\mathbf v^T (I-\mathbb ZS)^{-1}\mathbb Z\,\mathbf v\bigr) }{ \sqrt{\det(I-\mathbb ZS)} } = \sum_{\mathbf n\ge 0} \lhaf(\tilde S_{\mathbf n},\tilde{\mathbf v}_{\mathbf n}) \frac{\mathbf z^{\mathbf n}}{\mathbf n!}.$1

For $G(S,\mathbf v;\mathbf z) = \frac{ \exp\!\bigl(\tfrac12\,\mathbf v^T (I-\mathbb ZS)^{-1}\mathbb Z\,\mathbf v\bigr) }{ \sqrt{\det(I-\mathbb ZS)} } = \sum_{\mathbf n\ge 0} \lhaf(\tilde S_{\mathbf n},\tilde{\mathbf v}_{\mathbf n}) \frac{\mathbf z^{\mathbf n}}{\mathbf n!}.$2, each factor $G(S,\mathbf v;\mathbf z) = \frac{ \exp\!\bigl(\tfrac12\,\mathbf v^T (I-\mathbb ZS)^{-1}\mathbb Z\,\mathbf v\bigr) }{ \sqrt{\det(I-\mathbb ZS)} } = \sum_{\mathbf n\ge 0} \lhaf(\tilde S_{\mathbf n},\tilde{\mathbf v}_{\mathbf n}) \frac{\mathbf z^{\mathbf n}}{\mathbf n!}.$3 is recovered via

$G(S,\mathbf v;\mathbf z) = \frac{ \exp\!\bigl(\tfrac12\,\mathbf v^T (I-\mathbb ZS)^{-1}\mathbb Z\,\mathbf v\bigr) }{ \sqrt{\det(I-\mathbb ZS)} } = \sum_{\mathbf n\ge 0} \lhaf(\tilde S_{\mathbf n},\tilde{\mathbf v}_{\mathbf n}) \frac{\mathbf z^{\mathbf n}}{\mathbf n!}.$4

This trades the $G(S,\mathbf v;\mathbf z) = \frac{ \exp\!\bigl(\tfrac12\,\mathbf v^T (I-\mathbb ZS)^{-1}\mathbb Z\,\mathbf v\bigr) }{ \sqrt{\det(I-\mathbb ZS)} } = \sum_{\mathbf n\ge 0} \lhaf(\tilde S_{\mathbf n},\tilde{\mathbf v}_{\mathbf n}) \frac{\mathbf z^{\mathbf n}}{\mathbf n!}.$5-derivatives for $G(S,\mathbf v;\mathbf z) = \frac{ \exp\!\bigl(\tfrac12\,\mathbf v^T (I-\mathbb ZS)^{-1}\mathbb Z\,\mathbf v\bigr) }{ \sqrt{\det(I-\mathbb ZS)} } = \sum_{\mathbf n\ge 0} \lhaf(\tilde S_{\mathbf n},\tilde{\mathbf v}_{\mathbf n}) \frac{\mathbf z^{\mathbf n}}{\mathbf n!}.$6-derivatives outside the integral and converts the entire problem into evaluation of a shifted Gaussian integral with quadratic form $G(S,\mathbf v;\mathbf z) = \frac{ \exp\!\bigl(\tfrac12\,\mathbf v^T (I-\mathbb ZS)^{-1}\mathbb Z\,\mathbf v\bigr) }{ \sqrt{\det(I-\mathbb ZS)} } = \sum_{\mathbf n\ge 0} \lhaf(\tilde S_{\mathbf n},\tilde{\mathbf v}_{\mathbf n}) \frac{\mathbf z^{\mathbf n}}{\mathbf n!}.$7 (Tarasov, 21 Jul 2025).

The standard Gaussian integral identity

$G(S,\mathbf v;\mathbf z) = \frac{ \exp\!\bigl(\tfrac12\,\mathbf v^T (I-\mathbb ZS)^{-1}\mathbb Z\,\mathbf v\bigr) }{ \sqrt{\det(I-\mathbb ZS)} } = \sum_{\mathbf n\ge 0} \lhaf(\tilde S_{\mathbf n},\tilde{\mathbf v}_{\mathbf n}) \frac{\mathbf z^{\mathbf n}}{\mathbf n!}.$8

then yields the closed form after algebraic simplification. Specifically, setting $G(S,\mathbf v;\mathbf z) = \frac{ \exp\!\bigl(\tfrac12\,\mathbf v^T (I-\mathbb ZS)^{-1}\mathbb Z\,\mathbf v\bigr) }{ \sqrt{\det(I-\mathbb ZS)} } = \sum_{\mathbf n\ge 0} \lhaf(\tilde S_{\mathbf n},\tilde{\mathbf v}_{\mathbf n}) \frac{\mathbf z^{\mathbf n}}{\mathbf n!}.$9 and $S=(s_{ij})$0, one obtains

$S=(s_{ij})$1

and

$S=(s_{ij})$2

which together produce the stated generating function (Tarasov, 21 Jul 2025).

4. Closed-form identity and scope of validity

The final generating function is

$S=(s_{ij})$3

or equivalently,

$S=(s_{ij})$4

The paper proves this for any symmetric matrix $S=(s_{ij})$5 and any $S=(s_{ij})$6, with the identity understood in a neighborhood of $S=(s_{ij})$7 (Tarasov, 21 Jul 2025).

The proof is first carried out under the temporary assumptions that $S=(s_{ij})$8 is invertible and that $S=(s_{ij})$9 is positive definite for sufficiently small $\mu\times\mu$0. These restrictions are then removed by analyticity in the entries of $\mu\times\mu$1 and in $\mu\times\mu$2, so the formula extends by continuity to all symmetric matrices of even size (Tarasov, 21 Jul 2025). An odd-dimensional matrix is handled by embedding it into an even-dimensional one by adjoining a dummy row and column; the paper describes this as adding a trivial “1” in the upper-left corner.

The scope statement is the central conceptual advance of the note. Earlier derivations of the same formula worked by interpreting $\mu\times\mu$3, or a rearranged version of it, as data associated with a physical Gaussian quantum state, with $\mu\times\mu$4 serving as displacement. Those derivations therefore inherited block symmetries and positivity or uncertainty-principle constraints from quantum optics (Tarasov, 21 Jul 2025). The Gaussian-integral proof shows that such physical constraints are not mathematically necessary for the generating identity itself.

A common misconception is that loop-hafnian generating functions are intrinsically tied to covariance matrices of Gaussian states. The 2025 result directly contradicts that stronger claim: the identity is valid for every complex symmetric $\mu\times\mu$5, not merely those realizable as covariance data of physical Gaussian states (Tarasov, 21 Jul 2025). What remains quantum-optical is the interpretation of certain subclasses of matrices, not the algebraic identity.

5. Relation to quantum optics and Gaussian boson sampling

Loop hafnians arise naturally in the quantum optics of Gaussian states, especially in photon-number statistics for displaced Gaussian states. Earlier work established that Gaussian state amplitudes and related detection probabilities can be expressed through hafnians and loop hafnians (Hatamizadeh et al., 2019). Within that framework, the generating function organizes families of output amplitudes associated with replicated mode occupations.

The 2025 note explicitly distinguishes its result from the quantum-optical route. In earlier work, the same formula had been obtained by interpreting $\mu\times\mu$6 and its rearrangements as covariance matrices of physical Gaussian states and $\mu\times\mu$7 as displacement, which imposed special structural constraints on the blocks of $\mu\times\mu$8 as well as positivity and uncertainty-principle conditions (Tarasov, 21 Jul 2025). By replacing that route with classical Gaussian integration, the note severs the mathematical statement from physical realizability.

This distinction matters for Gaussian boson sampling. The hafnian and loop hafnian are the matrix functions governing amplitudes in standard and displaced Gaussian boson sampling, and algorithmic work in the area typically inherits its matrices from physically admissible states. The unrestricted generating function suggests a broader analytic domain in which asymptotic expansions, symbolic manipulations, or formal identities can be studied without enforcing physical admissibility. This suggests that some tools developed for Gaussian-state calculations may extend to more general symmetric-matrix ensembles, although such an extension is an implication rather than an explicit theorem in (Tarasov, 21 Jul 2025).

Related quantum-optical advances support this broader context. Threshold Gaussian boson sampling has been formulated in terms of the torontonian, an analogue of the hafnian for threshold detectors (Kumar et al., 2020), while displaced-state probability formulas continue to rely on loop-hafnian structures (Hatamizadeh et al., 2019). The generating function in (Tarasov, 21 Jul 2025) fits into this ecosystem as a closed-form coefficient extractor for replicated loop-hafnian data.

6. Probabilistic, combinatorial, and analytic consequences

The paper identifies a probabilistic interpretation: for a non-centered Gaussian vector $\mu\times\mu$9, the mixed moments are given by loop hafnians of blown-up covariance data,

$\mathbf v=(v_1,\dots,v_\mu)\in\mathbb C^\mu$0

In this reading, the generating function acts as a moment-generating device for non-centered Gaussian vectors (Tarasov, 21 Jul 2025). This places the result in direct continuity with Wick-type expansions, with the loop vector accounting for singleton contributions induced by nonzero means.

From a combinatorial perspective, ordinary hafnian generating functions encode pairings. The loop-hafnian version augments this by admitting fixed points, that is, singleton blocks. The paper notes that such structures occur in matchings of graphs with loops and in settings involving unpaired particles (Tarasov, 21 Jul 2025). The significance is not merely terminological: permitting singletons changes the admissible partition class from perfect matchings to single-pair matchings, which is precisely the combinatorial content of the loop hafnian.

Analytically, the formula provides a compact bridge between formal power series in $\mathbf v=(v_1,\dots,v_\mu)\in\mathbb C^\mu$1, determinant expansions, and Gaussian integral identities. The determinant factor $\mathbf v=(v_1,\dots,v_\mu)\in\mathbb C^\mu$2 is the pairing component, while the exponential

$\mathbf v=(v_1,\dots,v_\mu)\in\mathbb C^\mu$3

encodes the singleton structure induced by $\mathbf v=(v_1,\dots,v_\mu)\in\mathbb C^\mu$4. This separation mirrors the familiar decomposition between centered and non-centered Gaussian moment formulas. A plausible implication is that analytic methods developed for determinant-based generating functions may be adapted to loop-hafnian settings with nonzero linear terms, although (Tarasov, 21 Jul 2025) does not develop this direction explicitly.

7. Position within the literature

The 2025 note is concise and technical, but it occupies a specific place in the development of hafnian methods. Earlier literature established loop hafnians as the natural matrix function for displaced Gaussian state amplitudes and related boson-sampling observables (Hatamizadeh et al., 2019). The differential-operator representation used as the starting lemma in (Tarasov, 21 Jul 2025) was already standard in that line of work. What the note contributes is not a new matrix function, but a proof strategy showing that the multivariate generating function is fundamentally a Gaussian-integration identity rather than a consequence of quantum-state structure.

This repositioning alters how the formula should be interpreted. Instead of being regarded primarily as a byproduct of covariance-matrix calculus for Gaussian states, it becomes a general identity for arbitrary symmetric matrices. That shift clarifies which parts of the earlier formalism are genuinely physical and which are purely algebraic.

Within the broader research landscape, hafnians, loop hafnians, and related functions such as the torontonian form a family of matrix functions encoding Gaussian-state statistics under different measurement models (Hatamizadeh et al., 2019, Kumar et al., 2020). The generating function proved in (Tarasov, 21 Jul 2025) strengthens the algebraic foundations of this family by showing that at least one of its core identities survives complete removal of quantum-optical admissibility conditions.