---
title: Bistochastic Normalization
url: https://www.emergentmind.com/topics/bistochastic-normalization
type: topic
---

# Bistochastic Normalization

Searching arXiv for recent and foundational papers on bistochastic normalization across matrix scaling, kernels, confusion matrices, and operator-theoretic usages.
Bistochastic normalization denotes a family of procedures or constraints that transform an object into a bistochastic form. In its most common matrix-analytic usage, the target is a square nonnegative matrix whose row sums and column sums are all equal to \(1\); in other literatures, the same phrase or closely related terminology refers to unital and trace-preserving positive maps, to unistochastic matrices obtained from unitary amplitudes, or to nonlinear stochastic operators that move probability vectors toward more mixed configurations under a prescribed order [0907.2393] [2509.04959] [1005.5532] [1601.04297]. Across these settings, the normalization is used to remove marginal scale effects, impose balanced conservation laws, or place the object in a class with controlled spectral, probabilistic, or dynamical structure.

## 1. Definition, scope, and terminological variation

In the standard matrix setting, a bistochastic, or doubly stochastic, matrix \(B=(b_{ij})\) is a square nonnegative matrix satisfying
\[
\sum_j b_{ij}=1,\qquad \sum_i b_{ij}=1.
\]
This is the common target of iterative proportional fitting, Sinkhorn scaling, and related matrix-balancing procedures in network analysis, confusion-matrix analysis, and kernel methods [0907.2393] [2509.04959] [2006.00402].

The same structural idea appears in several adjacent forms. A unistochastic matrix is obtained from a unitary \(U\) by taking entrywise squared moduli,
\[
\bar U_{ij}=|U_{ij}|^2,
\]
and is automatically bistochastic because the rows and columns of a unitary are orthonormal [1912.12991]. In operator theory, a linear map \(\Phi:M_2(\mathbb C)\to M_2(\mathbb C)\) is called bistochastic when it is positive, unital, and trace-preserving; in that setting, the normalization is imposed on the map rather than on a numerical matrix [1005.5532]. In nonlinear simplex dynamics, “bistochastic” may instead mean that an operator sends each probability vector to a more mixed one under classical majorization or under the \(b\)-order \( \leq^b \) [2001.10444] [1601.04297].

The phrase is therefore context-sensitive. In the complex Hadamard literature, “bistochastic form” can denote a specialized normal form, but the available content for “Deformed Fourier matrices in bistochastic form” does not suffice to recover that paper’s exact definition [1901.04941]. This ambiguity is not peripheral: it marks a real divergence between matrix balancing, operator-algebraic normalization, and nonlinear majorization-based usage.

## 2. Matrix balancing formulations and algorithms

The canonical balancing formulation starts from a nonnegative square matrix and seeks diagonal scalings that enforce the two marginal constraints simultaneously. In the migration-flow setting, the normalized matrix is written
\[
B=RFC,\qquad b_{ij}=r_i f_{ij} c_j,
\]
where \(F=(f_{ij})\) is the raw flow table and the diagonal multipliers are chosen so that all row and column sums of \(B\) equal \(1\) [0907.2393]. In the confusion-matrix setting, the same idea is expressed as a KL projection:
\[
bis(M)\in \arg\min_{P \in \mathbb{R}_{>0}^{C \times C} \,:\, P_{i+}=P_{+j}=1\ \forall i,j} D_{\mathrm{KL}}(P \,\|\, M),
\]
and the solution has diagonal-scaling form \(P=D_1MD_2\) [2509.04959].

The standard computational mechanism is iterative proportional fitting, also called IPFP, Sinkhorn-Knopp, matrix scaling, biproportional fitting, or the RAS method. One alternately rescales rows to the target row sums and columns to the target column sums until convergence [0907.2393] [2509.04959]. For positive matrices, the confusion-matrix study states that existence, uniqueness, and convergence are guaranteed, and that the converged bistochastic matrix is the unique positive KL-closest matrix with the prescribed marginals [2509.04959]. In sparse network tables, the situation is more delicate: if the table is “critically sparse,” some nonzero entries may be “unsupported,” may converge to zero, and the row and column multipliers may fail to converge [0907.2393].

These caveats explain why smoothing is frequently introduced. In confusion matrices, zeros are handled by replacing \(M\) with \(M+\epsilon\), which restores positivity while preserving similar empirical behavior in practice [2509.04959]. In large sparse flow tables, the migration study discusses structural zeros versus sampling zeros and mentions remedies such as smoothing the zero pattern, taking powers of the bistochastic matrix, convex combinations with a uniform off-diagonal matrix, or even adding \(1\) to every cell [0907.2393].

Bistochastic normalization also admits distributed realizations. For a strongly connected directed communication graph, a local algorithm can iteratively adapt outgoing edge weights and a self-weight so that the weight matrix is column stochastic at every step and converges asymptotically to a bistochastic matrix, using only neighbor-to-neighbor information [1210.4383]. This version preserves the graph sparsity pattern and turns bistochastic normalization into a graph-constrained control procedure rather than a centralized matrix-balancing computation.

## 3. Removal of marginal effects in statistical and network data

In weighted directed networks, bistochastic normalization is used to remove node-size effects while retaining relative interaction structure. For intercounty U.S. migration, the normalized matrix forces each county to have the same total inflow and the same total outflow, so the remaining variation reflects relative interaction rather than raw scale [0907.2393]. The paper emphasizes that the normalization preserves cross-product ratios,
\[
\frac{f_{ij}f_{kl}}{f_{il}f_{kj}},
\]
described there as “relative odds,” and also interprets the bistochastic entries as maximum-entropy estimates subject to the row and column constraints [0907.2393]. The second stage of that workflow, strong component hierarchical clustering, treats the largest bistochastic entries as the dominant relative interactions and uses them to extract a multiscale directed backbone.

The migration example illustrates the empirical impact of this balancing. In the \(3{,}107\times 3{,}107\) county flow table with \(735{,}531\) nonzero off-diagonal entries, the first \(25{,}329\) largest bistochastic links were sufficient to place all counties in one strong component, whereas disparity filtering required between \(80{,}203\) and \(83{,}693\) links under the \(OR\) rule for the same strongly connected outcome [0907.2393]. The paper also reports weak correlations between raw flows and bistochasticized values, supporting the claim that the normalization is not merely reproducing the largest raw magnitudes.

In classification analysis, bistochastic normalization is used to separate class similarity from distribution bias. The confusion-matrix study argues that raw confusion counts mix two effects: intrinsic confusability of classes and marginal imbalance in true-label and predicted-label frequencies [2509.04959]. Row normalization removes only label imbalance, column normalization removes only prediction imbalance, and all-normalization removes only total sample size; bistochastic normalization removes both marginals simultaneously by enforcing unit row sums and unit column sums [2509.04959]. The formal invariance
\[
bis(\alpha M \beta)=bis(M)
\]
for positive diagonal \(\alpha,\beta\) captures exactly this insensitivity to arbitrary positive row and column reweightings.

That paper further reports that, under moderate or stronger heterogeneity, bistochastic normalization exceeded the competing normalizations by about \(10\) points in median overlap when approximating a balanced confusion matrix, and by up to \(19\) points on Rotated Fashion-MNIST at \(\alpha=0.1\) [2509.04959]. It also stresses an interpretational boundary: if actual deployment frequencies or actual model prediction bias are the quantities of interest, removing the marginals may suppress information that an evaluation protocol wishes to preserve.

## 4. Kernel normalization, spectral structure, and unistochastic processes

Kernel methods provide a particularly sharp form of bistochastic normalization. For a Gaussian affinity matrix with zero diagonal,
\[
K_{i,j} =
\begin{dcases}
\exp(-\Vert \mathbf{x}_i - \mathbf{x}_j\Vert^2/\varepsilon), & i\neq j, \\
0, & i=j,
\end{dcases}
\]
the doubly-stochastic normalization is
\[
W^{(\mathrm d)}=\operatorname{diag}(d)\,K\,\operatorname{diag}(d),
\]
with \(d_i>0\) chosen so that \(W^{(\mathrm d)}\mathbf 1=\mathbf 1\) and \((W^{(\mathrm d)})^T\mathbf 1=\mathbf 1\) [2006.00402]. Under a fixed-\(n\), high-dimensional heteroskedastic-noise model satisfying \(\|\Sigma_i\|_2\le C_\eta m^{-1}\), the noisy doubly-stochastic affinity converges to the clean one at rate
\[
\|\widetilde W^{(\mathrm d),(m)}-W^{(\mathrm d),(m)}\|_F=\mathcal O_p(m^{-1/2}),
\]
whereas the row-stochastic and symmetric normalizations are reported to behave unfavorably under the same noise model [2006.00402]. The reason is structural: heteroskedastic noise induces an approximately left-and-right diagonal multiplicative bias in the kernel, and the bistochastic scaling absorbs that bias.

A related kernel construction defines bistochastic normalization by
\[
\mathbf{P} = \mathbf{D}^{-1}\mathbf{K}\mathbf{Q}^{-1}\mathbf{K}\mathbf{D}^{-1},
\]
with
\[
\mathbf{D}=\operatorname{diag}(\mathbf{K}\mathbf{1}_N), \qquad
\mathbf{Q}=\operatorname{diag}(\mathbf{K}\mathbf{D}^{-1}\mathbf{1}_N).
\]
The resulting matrix is symmetric and stochastic, its eigenvalues lie in \([0,1]\), its largest eigenvalue is \(1\), and the corresponding leading eigenvector is constant [2510.26574]. Because direct eigendecomposition is cubic in \(N\), the paper develops an \(O(Nr^2)\) approximation based on a pivoted partial Cholesky factorization \(\mathbf K\approx \mathbf F\mathbf F^\top\), followed by QR and small-matrix SVD steps that recover an approximate eigendecomposition of the bistochastic kernel without explicitly diagonalizing the full normalized matrix [2510.26574].

Quantum-information usage connects bistochastic normalization to measurement statistics. Given two orthonormal bases related by a unitary \(U\), the matrix
\[
\bar U_{ij}=|U_{ij}|^2
\]
is unistochastic and therefore bistochastic [1912.12991]. In that framework, \(\bar U\) and \(\bar U^T\) act as classical Markovian processes on measurement outcome distributions, and alternating measurement sequences correspond to powers or alternating products of these bistochastic matrices. The resulting sequence of entropic-uncertainty lower bounds is computable with resources polynomial in the Hilbert-space dimension [1912.12991]. Here bistochastic normalization is not imposed by scaling an arbitrary matrix; it is inherited canonically from the mod-squared unitary.

## 5. Operator-theoretic normalization on \(M_2(\mathbb C)\)

In operator algebra, bistochastic normalization refers to a normalization of maps rather than of numerical tables. A linear map \(\Phi:M_2(\mathbb C)\to M_2(\mathbb C)\) is bistochastic when it is positive, unital, and trace-preserving [1005.5532]. In the Pauli-basis representation,
\[
x=w_0 I+\mathbf w\cdot \sigma,
\]
such a map necessarily has the form
\[
\Phi(w_0 I+\mathbf w\cdot \sigma)=w_0 I+(T\mathbf w)\cdot \sigma,
\]
with \(T\in M_3(\mathbb R)\) for positive maps [1005.5532]. The normalization removes the affine translation term and fixes the scalar part, so the entire problem reduces to the action of a real \(3\times 3\) matrix on the Bloch vector.

This reduction supports a detailed classification of Kadison–Schwarz behavior. A general bistochastic map is a Kadison–Schwarz operator iff \(T\) satisfies a contraction condition together with a cross-product inequality involving \(T[\mathbf w,\overline{\mathbf w}]\) and \([T\mathbf w,\overline{T\mathbf w}]\) for all \(\mathbf w\in\mathbb C^3\) [1005.5532]. Up to unitary conjugation, every bistochastic map reduces to a diagonal form with parameters \((\lambda_1,\lambda_2,\lambda_3)\in[-1,1]^3\), and in that representation Kadison–Schwarz is characterized by explicit necessary-and-sufficient inequalities in \(\lambda_1,\lambda_2,\lambda_3\) [1601.00719]. Complete positivity is stricter, and the papers exhibit the strict containments
\[
\text{CP} \subsetneq \text{KS} \subsetneq \text{positive}.
\]

The examples are structurally important. The transposition map, corresponding to \(\Phi_{(1,-1,1)}\), is positive but not Kadison–Schwarz [1005.5532]. Conversely, within the symmetric family \(\Phi_{(\lambda,\lambda,\lambda)}\), the interval
\[
1-\sqrt2 \le \lambda < -\frac13
\]
produces bistochastic Kadison–Schwarz maps that are not completely positive [1005.5532]. In this domain, bistochastic normalization is not an estimation step but a rigid algebraic constraint that exposes the geometry of positive, Kadison–Schwarz, and completely positive regions.

## 6. Nonlinear generalizations and dynamical systems

A broader line of work extends bistochasticity from linear matrices to nonlinear maps on the simplex. For quadratic stochastic operators
\[
(Vx)_k=\sum_{i,j=1}^m P_{ij,k}x_i x_j,
\]
one notion of bistochasticity requires
\[
V(x)\prec x
\]
for all simplex points \(x\), with \(\prec\) the majorization order [2001.10444]. The resulting set of quadratic bistochastic operators forms a convex compact polytope whose extreme points include all coordinate permutation operators, and every operator in the relative interior of that polytope is strictly regular: it has a unique fixed point, the barycenter, and no periodic points other than that fixed point [2001.10444]. Strictly regular quadratic bistochastic operators are also dense in the full quadratic bistochastic class.

A related refinement replaces classical majorization by the \(b\)-order \( \leq^b \), defined by cumulative partial sums
\[
\mathcal U_k(x)=\sum_{i=1}^k x_i,\qquad
x\leq^b y \iff \mathcal U_k(x)\le \mathcal U_k(y)\ \text{for }k=1,\dots,n-1.
\]
A quadratic stochastic operator is then \(b\)-bistochastic if \(V(x)\leq^b x\) for all \(x\) [1601.04297]. Every such operator fixes \(\mathbb P=(0,\dots,0,1)\), every trajectory converges to a fixed point, and a sufficient condition for uniqueness of the fixed point is
\[
P_{kk,k}<1,\qquad P_{kj,k}<\frac12
\]
for all \(k\in\{1,\dots,n-1\}\) and \(j>k\) [1601.04297]. The same paper associates a nonhomogeneous Markov measure to the operator and proves that, when the fixed point is unique, the associated measure is mixing; for a specific \(S^1\) family, the associated measures are mutually absolutely continuous [1601.04297].

These nonlinear theories are not normalization procedures in the Sinkhorn sense. They nevertheless preserve the core intuition of bistochastic normalization: the map is constrained so that the output is no less mixed than the input under an appropriate order, and the constraint has strong consequences for fixed points, asymptotics, and induced stochastic processes. In that sense, bistochastic normalization is best understood not as a single algorithm but as a recurring two-sided balancing principle that reappears in matrix analysis, spectral kernels, operator algebras, and nonlinear dynamics.

Source: https://www.emergentmind.com/topics/bistochastic-normalization