Iterative Proportional Mean-Matching (IPMM)

• Iterative Proportional Mean-Matching (IPMM) is a method that iteratively enforces linear moment or mean constraints via Kullback–Leibler projections, ensuring convergence in entropy minimization problems.
• It is applied both in classical iterative proportional fitting procedures and as a reparameterized training objective in Schrödinger Bridge models, aligning with diffusion-model mean prediction.
• The method’s exponential convergence is governed by the geometry of constraint spaces, with factors like the Friedrichs angle affecting the contraction rate.

Searching arXiv for papers on Iterative Proportional Mean-Matching and related IPFP/Schrödinger Bridge work. Iterative Proportional Mean-Matching (IPMM) denotes a family of iterative proportional procedures in which successive updates enforce mean or moment conditions. In the most general variational sense, IPMM is iterative proportional fitting specialized to linear expectation constraints, so that each step is a Kullback–Leibler projection onto a set of measures matching a prescribed collection of moments (Eckstein et al., 27 Feb 2025). In recent generative-modeling work, the same label denotes a reparameterized Schrödinger Bridge training objective in which forward and backward transition models are trained by squared-error regression to conditional means of adjacent states, thereby making bridge training resemble standard diffusion-model mean prediction (Tang et al., 25 Aug 2025). Across these usages, the common structure is iterative alternation over simpler subproblems, while “mean-matching” refers either to exact linear moment constraints or to regression toward conditional means.

1. Variational definition and core mathematical structure

In the information-projection formulation, IPMM is a special case of iterative proportional fitting procedures (IPFP) for entropy minimization. The basic problem is to find, for a Polish space $X$, a probability measure $\pi$ closest to a reference measure $\theta \in \mathcal P(X)$ in Kullback–Leibler divergence subject to constraints: $$\pi^*=\arg\min_{\pi\in\mathcal Q} D_{\mathrm{KL}}(\pi\|\theta), \qquad \mathcal Q=\bigcap_{i=1}^N \mathcal Q_i .$$ Here

$D_{\mathrm{KL}}(\pi\|\theta)= \begin{cases} \displaystyle \int_X \log\!\left(\frac{d\pi}{d\theta}\right)\,d\pi, & \pi\ll\theta,\[4pt] +\infty, & \text{otherwise}. \end{cases}$

When each constraint set is determined by linear moment conditions,

$$\mathcal Q_i = \left\{ \pi\in\mathcal P(X): \int h_i\,d\pi=\int h_i\,d\mu \ \text{for all } h_i\in H_i \right\},$$

with $H_i\subset L^\infty(X)$ and $\mu$ equivalent to $\theta$, the resulting algorithm is naturally described as IPMM: each step updates the current distribution so that all moments in one function space $H_i$ are matched while remaining KL-close to the previous iterate (Eckstein et al., 27 Feb 2025).

The iterative scheme starts from $\pi$0. For outer iteration $\pi$1 and block $\pi$2,

$\pi$3

with the cycle closed by $\pi$4. Each update is a partial information projection. In the mean-matching interpretation, the procedure successively enforces the expectations associated with $\pi$5, and, under the regularity conditions stated in the convergence theory, converges to the unique KL projection onto $\pi$6 (Eckstein et al., 27 Feb 2025).

This formulation places IPMM inside the exponential-family geometry of entropy minimization. Projecting onto one $\pi$7 is equivalent to multiplicatively tilting the current measure so as to satisfy the corresponding moment constraints. A finite-dimensional instance arises when $\pi$8 and $\pi$9: the two alternating projections enforce the moments of the $\theta \in \mathcal P(X)$0 and the $\theta \in \mathcal P(X)$1 in turn, and the intermediate densities have exponential-family form with Lagrange multipliers chosen to satisfy those expectations (Eckstein et al., 27 Feb 2025).

2. Exponential convergence and the geometry of constraint spaces

A central result for IPMM in the moment-constrained setting is exponential convergence of the KL objective under a set of structural assumptions. The conditions singled out are: strong duality for the information-projection problem; coincidence of primal and dual IPFP; existence of an optimizer $\theta \in \mathcal P(X)$2; uniform boundedness of the log-densities of all iterates and the optimizer relative to $\theta \in \mathcal P(X)$3; and closedness of the sum of the closed constraint spaces,

$\theta \in \mathcal P(X)$4

Under these hypotheses, there exist constants $\theta \in \mathcal P(X)$5 and $\theta \in \mathcal P(X)$6 such that

$\theta \in \mathcal P(X)$7

so the KL error decreases geometrically with the number of full cycles (Eckstein et al., 27 Feb 2025).

The rate is controlled by both dual smoothness/strong convexity and the geometry of the constraint subspaces. The contraction factor is given in terms of

$\theta \in \mathcal P(X)$8

where $\theta \in \mathcal P(X)$9 and $$\pi^*=\arg\min_{\pi\in\mathcal Q} D_{\mathrm{KL}}(\pi\|\theta), \qquad \mathcal Q=\bigcap_{i=1}^N \mathcal Q_i .$$0 are the strong convexity constant and Lipschitz gradient constant of the dual penalization on the region where log-densities are bounded by $$\pi^*=\arg\min_{\pi\in\mathcal Q} D_{\mathrm{KL}}(\pi\|\theta), \qquad \mathcal Q=\bigcap_{i=1}^N \mathcal Q_i .$$1, and $$\pi^*=\arg\min_{\pi\in\mathcal Q} D_{\mathrm{KL}}(\pi\|\theta), \qquad \mathcal Q=\bigcap_{i=1}^N \mathcal Q_i .$$2 is the sum operator modulo its kernel. The quantity $$\pi^*=\arg\min_{\pi\in\mathcal Q} D_{\mathrm{KL}}(\pi\|\theta), \qquad \mathcal Q=\bigcap_{i=1}^N \mathcal Q_i .$$3 is the condition number of the sum operator and encodes how well conditioned the decomposition into constraint blocks is (Eckstein et al., 27 Feb 2025).

For two subspaces, this conditioning admits an explicit expression through the Friedrichs angle. If $$\pi^*=\arg\min_{\pi\in\mathcal Q} D_{\mathrm{KL}}(\pi\|\theta), \qquad \mathcal Q=\bigcap_{i=1}^N \mathcal Q_i .$$4 is the cosine of the Friedrichs angle between $$\pi^*=\arg\min_{\pi\in\mathcal Q} D_{\mathrm{KL}}(\pi\|\theta), \qquad \mathcal Q=\bigcap_{i=1}^N \mathcal Q_i .$$5 and $$\pi^*=\arg\min_{\pi\in\mathcal Q} D_{\mathrm{KL}}(\pi\|\theta), \qquad \mathcal Q=\bigcap_{i=1}^N \mathcal Q_i .$$6, then

$$\pi^*=\arg\min_{\pi\in\mathcal Q} D_{\mathrm{KL}}(\pi\|\theta), \qquad \mathcal Q=\bigcap_{i=1}^N \mathcal Q_i .$$7

A larger angle, equivalently a smaller cosine, yields a better condition number and therefore faster contraction; a small angle indicates near-linear dependence of the moment families and correspondingly slower convergence (Eckstein et al., 27 Feb 2025). The paper also recalls the Deutsch criterion that $$\pi^*=\arg\min_{\pi\in\mathcal Q} D_{\mathrm{KL}}(\pi\|\theta), \qquad \mathcal Q=\bigcap_{i=1}^N \mathcal Q_i .$$8 is closed if and only if the Friedrichs cosine $$\pi^*=\arg\min_{\pi\in\mathcal Q} D_{\mathrm{KL}}(\pi\|\theta), \qquad \mathcal Q=\bigcap_{i=1}^N \mathcal Q_i .$$9, so positivity of the angle is exactly the closed-sum condition in the two-block case.

This gives a precise geometric interpretation of IPMM. Each $D_{\mathrm{KL}}(\pi\|\theta)= \begin{cases} \displaystyle \int_X \log\!\left(\frac{d\pi}{d\theta}\right)\,d\pi, & \pi\ll\theta,\[4pt] +\infty, & \text{otherwise}. \end{cases}$0 is a family of moment functions, and the algorithm alternates between correcting the expectations of these families. If two families overlap strongly in $D_{\mathrm{KL}}(\pi\|\theta)= \begin{cases} \displaystyle \int_X \log\!\left(\frac{d\pi}{d\theta}\right)\,d\pi, & \pi\ll\theta,\[4pt] +\infty, & \text{otherwise}. \end{cases}$1, enforcing one tends to undo or redundantly reinforce the other. If they are nearly orthogonal, each update contributes largely independent information. This suggests that block design, reparameterization, or choice of reference measure $D_{\mathrm{KL}}(\pi\|\theta)= \begin{cases} \displaystyle \int_X \log\!\left(\frac{d\pi}{d\theta}\right)\,d\pi, & \pi\ll\theta,\[4pt] +\infty, & \text{otherwise}. \end{cases}$2 can materially affect practical convergence by changing the effective angles between the $D_{\mathrm{KL}}(\pi\|\theta)= \begin{cases} \displaystyle \int_X \log\!\left(\frac{d\pi}{d\theta}\right)\,d\pi, & \pi\ll\theta,\[4pt] +\infty, & \text{otherwise}. \end{cases}$3.

3. Relation to Sinkhorn, optimal transport, and dual block updates

Classical Sinkhorn scaling is a canonical special case of IPMM’s ambient IPFP framework. In entropic optimal transport with fixed marginals, the admissible set is the intersection of two marginal-constraint sets, one for the first marginal and one for the second. These can be written through the function spaces

$D_{\mathrm{KL}}(\pi\|\theta)= \begin{cases} \displaystyle \int_X \log\!\left(\frac{d\pi}{d\theta}\right)\,d\pi, & \pi\ll\theta,\[4pt] +\infty, & \text{otherwise}. \end{cases}$4

Alternating KL projections onto these two sets is precisely Sinkhorn’s row/column rescaling algorithm. In the product reference geometry, the two spaces are orthogonal in $D_{\mathrm{KL}}(\pi\|\theta)= \begin{cases} \displaystyle \int_X \log\!\left(\frac{d\pi}{d\theta}\right)\,d\pi, & \pi\ll\theta,\[4pt] +\infty, & \text{otherwise}. \end{cases}$5, so the Friedrichs angle is $D_{\mathrm{KL}}(\pi\|\theta)= \begin{cases} \displaystyle \int_X \log\!\left(\frac{d\pi}{d\theta}\right)\,d\pi, & \pi\ll\theta,\[4pt] +\infty, & \text{otherwise}. \end{cases}$6, the cosine is $D_{\mathrm{KL}}(\pi\|\theta)= \begin{cases} \displaystyle \int_X \log\!\left(\frac{d\pi}{d\theta}\right)\,d\pi, & \pi\ll\theta,\[4pt] +\infty, & \text{otherwise}. \end{cases}$7, and the condition number is $D_{\mathrm{KL}}(\pi\|\theta)= \begin{cases} \displaystyle \int_X \log\!\left(\frac{d\pi}{d\theta}\right)\,d\pi, & \pi\ll\theta,\[4pt] +\infty, & \text{otherwise}. \end{cases}$8, which yields the best possible conditioning in the general rate formula (Eckstein et al., 27 Feb 2025).

The same framework extends beyond classical two-marginal transport. The convergence result is presented as unifying and extending recent results from multi-marginal, adapted, and martingale optimal transport, because all of these can be encoded as information projections onto intersections of linear-constraint sets. In that sense, IPMM is a “Sinkhorn-type” procedure for arbitrary linear expectation constraints, not merely for marginals (Eckstein et al., 27 Feb 2025).

The dual viewpoint clarifies why alternating mean-matching can contract exponentially. Dual variables $D_{\mathrm{KL}}(\pi\|\theta)= \begin{cases} \displaystyle \int_X \log\!\left(\frac{d\pi}{d\theta}\right)\,d\pi, & \pi\ll\theta,\[4pt] +\infty, & \text{otherwise}. \end{cases}$9, with $$\mathcal Q_i = \left\{ \pi\in\mathcal P(X): \int h_i\,d\pi=\int h_i\,d\mu \ \text{for all } h_i\in H_i \right\},$$0, parameterize exponential tilts of the reference measure. The primal optimizer has density proportional to

$$\mathcal Q_i = \left\{ \pi\in\mathcal P(X): \int h_i\,d\pi=\int h_i\,d\mu \ \text{for all } h_i\in H_i \right\},$$1

The dual functional is strongly convex and smooth on bounded log-density regions, and each IPFP step corresponds to updating only one block $$\mathcal Q_i = \left\{ \pi\in\mathcal P(X): \int h_i\,d\pi=\int h_i\,d\mu \ \text{for all } h_i\in H_i \right\},$$2 while the others are fixed. This is analogous to a block-coordinate or proximal update on the dual objective, with the conditioning of the block decomposition governed by the same operator $$\mathcal Q_i = \left\{ \pi\in\mathcal P(X): \int h_i\,d\pi=\int h_i\,d\mu \ \text{for all } h_i\in H_i \right\},$$3 that appears in the contraction rate (Eckstein et al., 27 Feb 2025). A plausible implication is that IPMM is best viewed not merely as an alternating projection method, but as a geometrically preconditioned block optimization scheme in the dual exponential-family coordinates.

4. IPMM as a Schrödinger Bridge training objective

A distinct but closely related use of IPMM appears in discrete-time Schrödinger Bridge (SB) modeling. In that setting, one considers a time grid $$\mathcal Q_i = \left\{ \pi\in\mathcal P(X): \int h_i\,d\pi=\int h_i\,d\mu \ \text{for all } h_i\in H_i \right\},$$4, a data distribution $$\mathcal Q_i = \left\{ \pi\in\mathcal P(X): \int h_i\,d\pi=\int h_i\,d\mu \ \text{for all } h_i\in H_i \right\},$$5 at time $$\mathcal Q_i = \left\{ \pi\in\mathcal P(X): \int h_i\,d\pi=\int h_i\,d\mu \ \text{for all } h_i\in H_i \right\},$$6, a prior $$\mathcal Q_i = \left\{ \pi\in\mathcal P(X): \int h_i\,d\pi=\int h_i\,d\mu \ \text{for all } h_i\in H_i \right\},$$7 at time $$\mathcal Q_i = \left\{ \pi\in\mathcal P(X): \int h_i\,d\pi=\int h_i\,d\mu \ \text{for all } h_i\in H_i \right\},$$8, and a reference path measure $$\mathcal Q_i = \left\{ \pi\in\mathcal P(X): \int h_i\,d\pi=\int h_i\,d\mu \ \text{for all } h_i\in H_i \right\},$$9 generated by a forward Markov chain

$H_i\subset L^\infty(X)$0

with Gaussian transitions

$H_i\subset L^\infty(X)$1

The discrete SB problem is

$H_i\subset L^\infty(X)$2

and classical iterative proportional fitting solves it by alternating KL projections on the full joint path measure, first to enforce the terminal marginal and then to enforce the initial marginal (Tang et al., 25 Aug 2025).

That full-path IPF is theoretically sound but impractical. Diffusion Schrödinger Bridge (DSB) therefore works with Gaussian conditional transitions and trains forward and backward neural networks $H_i\subset L^\infty(X)$3 and $H_i\subset L^\infty(X)$4 for the conditional means. The original DSB loss uses targets involving both successive states and current network evaluations, which the paper identifies as problematic for three reasons: it doubles the number of forward passes per training sample; the target is offset by another network that is itself changing, which leads to instability; and the functional form does not match standard score-based generative model (SGM) objectives, which obstructs direct use of pre-trained SGMs as initialization (Tang et al., 25 Aug 2025).

IPMM is introduced as a reparameterized objective that replaces those original losses by direct mean-prediction losses: $H_i\subset L^\infty(X)$5 The backward network is thus trained to match the conditional mean of $H_i\subset L^\infty(X)$6 given $H_i\subset L^\infty(X)$7, and the forward network to match the conditional mean of $H_i\subset L^\infty(X)$8 given $H_i\subset L^\infty(X)$9. The Gaussian transitions are parameterized as

$\mu$0

$\mu$1

with $\mu$2 and $\mu$3 (Tang et al., 25 Aug 2025).

In this usage, “mean-matching” is local and conditional rather than global and moment-constrained: the algorithm alternates between fitting forward and backward one-step conditional means along bridge trajectories. The “iterative proportional” aspect remains, because the method alternates between two directions of bridge refinement in an IPF-like manner.

5. Equivalences, algorithmic structure, and empirical behavior

The training procedure alternates over epochs. In a backward epoch, one samples trajectories from the prior by iterating the current backward transition model and then minimizes $\mu$4. In a forward epoch, one samples trajectories from the data by iterating the current forward transition model and then minimizes $\mu$5. This alternation defines a discrete-time IPF-like scheme at the level of Gaussian conditionals rather than full path measures (Tang et al., 25 Aug 2025).

Two theoretical propositions motivate the reparameterization. First, under the Gaussian conditional parameterization with fixed variance $\mu$6, the simplified IPMM losses are approximately equivalent to the original DSB losses when the step sizes are small and successive marginals are close, which are the standard SGM assumptions. Second, when $\mu$7, IPMM, Iterative Proportional Terminus-Matching (IPTM), and Iterative Proportional Flow-Matching (IPFM) induce the same Gaussian bridge conditionals under the stated approximations, so they are theoretically equivalent at the level of the SB solution and differ primarily by parameterization and training convenience (Tang et al., 25 Aug 2025).

This equivalence explains IPMM’s role in unifying SB methods with diffusion-model practice. Standard SGMs are trained by MSE losses that predict a next-state mean, noise, or data variable; IPMM uses the same mean-prediction form for both bridge directions. The paper therefore argues that pre-trained SGMs can be used as drop-in initializations for SB training. It gives, for example, an initialization of the backward model

$\mu$8

where $\mu$9 is a pre-trained SGM predicting a related denoising quantity, and states that such alignment is valid for IPMM, IPTM, and IPFM but not for the original DSB loss because DSB’s targets have misaligned time indices and nonlocal combinations of network evaluations (Tang et al., 25 Aug 2025).

Empirically, the paper reports Gaussian SB experiments in dimensions $\theta$0 using the averaged metric $\theta$1 over time. IPMM trained from scratch consistently yields lower KL than DSB, IPML, DSBM, BM$\theta$2, and D-IMF, and SGMs initialization further decreases IPMM’s KL; the paper notes that DSB can even worsen under SGM initialization because of target misalignment. On image tasks, IPMM is one of the three reparameterized models evaluated by FID for unconditional generation, conditional generation, and unpaired translation. IPTM and IPFM tend to obtain the best FID on those image benchmarks, but IPMM remains competitive and shares the same training framework (Tang et al., 25 Aug 2025).

The implementation details reinforce the diffusion-model interpretation. For images, the models use ADM or LDM-style U-Net architectures. Training uses Adam with learning rate $\theta$3, $\theta$4, no weight decay, and batch size $\theta$5, with one direction trained until approximate convergence before switching to the other. This suggests that IPMM is not a separate numerical ecosystem so much as a bridge-compatible reparameterization of existing diffusion-model infrastructure.

6. Broader uses, neighboring literatures, and terminological caution

Outside these two explicit definitions, the label IPMM is used more loosely as an interpretive umbrella for iterative proportional procedures that match means or covariate averages. In the social-science discussion of IPF and the NM-method, “Iterative Proportional Mean-Matching” is not directly defined; rather, the paper sketches what such a method would look like if one replaced marginal constraints by mean constraints, and argues that a KL-based iterative proportional method would be appropriate for calibration or completion problems but not necessarily for counterfactual constructions, where the authors instead favor rank-based constraints and cumulative Kullback–Leibler logic (Naszodi, 2023). In causal inference, multimarginal unbalanced optimal transport is computed by a generalized Sinkhorn/IPFP algorithm and yields matching weights that balance covariate distributions across treatment arms, which has been interpreted as an IPMM-like procedure because balancing whole distributions implies balance of means and higher moments (Gunsilius et al., 2021). In group-matching for subjects and items, the term is again not explicit, but the algorithms iteratively eliminate units so that retained groups become statistically similar with respect to average values and other criteria for multiple covariates (Kiss et al., 2021).

This suggests an important terminological caution: IPMM is not a single universally standardized algorithm across all fields. At least two explicit meanings are established in the literature represented here. One is the information-geometric meaning: IPFP specialized to linear mean or moment constraints, with convergence governed by strong convexity and the Friedrichs-angle geometry of constraint spaces (Eckstein et al., 27 Feb 2025). The other is the Schrödinger Bridge meaning: a reparameterized objective in which forward and backward Gaussian transition models are trained by MSE regression to adjacent-state conditional means, enabling direct reuse of diffusion-model architectures and pre-trained SGMs (Tang et al., 25 Aug 2025). The broader matching, calibration, and counterfactual literatures show that the phrase is also useful descriptively for iterative proportional schemes targeting covariate means, but they do not furnish a single canonical formal definition.

Taken together, these strands place IPMM at the intersection of entropy minimization, Sinkhorn-type alternating projections, bridge-based generative modeling, and statistical balancing. The precise meaning depends on whether “mean-matching” refers to linear moment constraints in a KL projection problem or to conditional-mean regression inside a Gaussian transition model. The common principle is iterative enforcement of mean-related structure by proportional updates.