Not Strassen's theorem
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: I'm Ruth.
Warren: And I'm Warren, here with Iris.
Ruth: We're looking at the paper: "Not Strassen's theorem".
Warren: Here's the gist of the paper. Suppose there is an uncertain state of the world, drawn from a CDF G.
Ruth: It is well known that the random posterior mean induced (via Bayes’s rule) by a signal must have a distribution F that is a mean-preserving contraction of G, and conversely that any mean-preserving contraction is induced by some signal The Core Distinction between Theorems The paper distinguishes between Strassen’s theorem and Theorem 1, highlighting that they concern different relationships between distributions and random variables.
Warren: Strassen’s theorem relates a distribution G to the set of distributions F that can be induced by a martingale structure where X ∼ F and Y ∼ G.
Ruth: In contrast, Theorem 1 fixes a probability space P and a random variable Y on P, asking for which distributions F there exists an X such that (X, Y) is a martingale with X ∼ F.
Paper summary: Ruth: So we're looking at this paper, "Not Strassen's theorem," and it sets up a comparison between two different ideas about how random states of the world and signals interact. The main point here is that the random posterior mean you get from Bayes’ rule must have a distribution that is a mean-preserving contraction of the original state distribution.
Warren: That sounds like some kind of mathematical constraint on how much uncertainty can be reduced by getting more information about something. So, the paper claims this relationship between distributions and signals is not what Strassen's theorem deals with. It's a distinct result that has versions alluded to by other authors like Hardy, Littlewood and Pólya, Blackwell and Kolotilin.
Iris: That distinction really matters because it separates how we look at the setup. Strassen’s theorem relates a distribution G to the set of distributions F that can be induced by a martingale structure where you have X from F and Y from G.
Ruth: Right, so that's about having both things—the state and the signal—present in one setup. But Theorem one fixes a probability space P and a random variable Y on that space, and it asks for which distributions F there exists an X such that X from F makes (X, Y) a martingale <ref:2611.15527#pg1>.
Warren: So Theorem one is about finding the distribution of the signal given the state's distribution, whereas Strassen’s theorem seems to be about a more direct relationship between two pre-existing distributions without fixing a specific underlying probability space P for both <ref:2611.15527#pg1>. That's where things get tricky.
Iris: The authors point out that while Theorem one implies the non-trivial direction of Strassen’s theorem, the reverse isn't necessarily true <ref:2611.15527#pg1>. This means you can have a mean-preserving contraction that doesn't necessarily come from a simple martingale structure as defined in Strassen's version.
Ruth: And what they emphasize is the role of richness here. Richness is crucial because it guarantees that for any distribution conditional on the state Y, there actually exists some random variable X with that exact conditional distribution.
Warren: So richness is like saying every possible way the world could look given a state Y can actually be represented by some observable X. That's a strong assumption in economics, but they use it to make their equivalence between Theorem one and mean-preserving contractions hold <ref:2611.15527#pg1>.
Iris: They connect this to feasibility. Corollary one shows that a distribution F is feasible given Y if and only if it belongs to the set of mean-preserving contractions of G <ref:2611.15527#pg1,Y if and only if>. This is the core characterization we need here for understanding posterior means.
Paper summary: Ruth: To unpack that, they show you can construct a signal Z such that the expectation of Y given Z results in a distribution F, which then satisfies that martingale condition. And if F is already in the set of mean-preserving contractions of G, Theorem one guarantees there exists an X from F such that the conditional expectation E(Y given X) equals X almost surely <ref:2611.15527#pg1>.
Warren: So the characterization boils down to this: feasibility for a posterior mean distribution F is exactly equivalent to F being a mean-preserving contraction of G. That's what makes sense for understanding what kind of beliefs we can form after seeing some data about an uncertain state.
Iris: They also look at the case where you only have the distribution G and not the specific random variable Y, which they call Theorem one in Corollary two <ref:2611.15527#pg1>. In that under-specified model, feasibility still maps directly onto being in MPC(G).
Ruth: I wonder if this means we don't need to fix an underlying probability space P to talk about these posterior means when the state of the world is just given by its distribution G. That seems like a big relief for modeling real-world uncertainty.
Warren: It does, but they caution against using Strassen’s theorem directly in that context because it forces you to choose both the signal and the state simultaneously, which isn't practical when you only have the state distribution G as an input.
Iris: The analogy they use is comparing this to weak versus strong solutions of stochastic differential equations. A strong solution requires a fixed probability space P, while a weak solution involves defining it on another space with an exogenous forcing term B prime.
Ruth: That comparison helps frame the problem—the structure of the signal and state interaction is analogous to how we define our mathematical models in finance or physics. The weak solution being more convenient but lacking economic meaning because you don't know the forcing term beforehand, right?
Warren: Exactly. If you're just given the distribution G, you're essentially dealing with a situation where the forcing term is exogenous to your model setup, which is a classic weak solution scenario.
Iris: So if we only have G and want to talk about feasible posterior means, the result points us directly back to Corollary one and Theorem one <ref:2611.15527#pg1>. It’s not about finding some hidden structure imposed by Strassen's theorem that we can't see when the state is already fixed by its distribution.
Paper summary: Ruth: So, what does this mean for people actually running these models? If I have a distribution G of my market state, I know that any posterior mean I derive from Bayes’ rule must fall into this MPC(G) set to be considered feasible.
Warren: It means you don't need to worry about whether the specific mathematical machinery used to get that posterior mean satisfies the complex martingale structure of Strassen's theorem. You just need to check if it meets the contraction requirement defined by MPC(G).
Iris: The limitation they state is that they are focusing on this relationship within a fixed bounded interval, specifically using the unit interval zero one for all random variables involved <ref:2611.15527#pg1,the unit interval 0, 1>. They aren't immediately extending this to more complex spaces like continuous states with infinite dimensions.
Ruth: That's a fair limitation. The result as stated is very concrete on that bounded space, which is where we usually start in modeling things like belief updates in simple systems. But the structure they found still seems robust for the underlying idea of signal-state interaction.
Warren: It confirms that the mathematical object we care about—the posterior mean distribution F—is fundamentally constrained by its relationship to G through this contraction property, regardless of whether we use Strassen's theorem or Theorem one to describe it <ref:2611.15527#pg1>.
Iris: So, to summarize this paper "Not Strassen's theorem," the main contribution is establishing a precise equivalence between feasibility for posterior means and belonging to the set of mean-preserving contractions of the state distribution G. This result is framed around Theorem one which deals with finding a signal X given an observed random variable Y <ref:2611.15527#pg1>.
Ruth: It seems this paper provides a solid mathematical foundation for understanding exactly what kind of beliefs are physically or economically possible after updating our knowledge about an uncertain state. It clarifies the relationship between different ways of modeling that update process.
Warren: I think it helps us avoid getting stuck trying to force every posterior mean into the framework of Strassen's theorem when we only have the state distribution G defined upfront. It streamlines the analysis by giving us a clear benchmark: MPC(G).
Iris: So, for anyone looking at modeling uncertainty, whether in markets or physics, this paper tells you that your resulting belief distribution must respect this contraction property relative to your initial state distribution. That's the practical implication here.
Conclusion: Ruth: So we’ve been looking at how distributions of signals and states interact, specifically this paper called "Not Strassen's theorem."
Warren: Yeah, it boils down to this idea that there’s a clear line between what we can mathematically achieve with a martingale structure and what’s actually feasible in the real world.
Iris: Basically, they show that for any posterior mean distribution you get from Bayes rule, it has to fit inside this set called mean-preserving contractions of the original state distribution.
Ruth: And that means if you want a belief update to be possible, it has to obey this contraction property relative to the initial uncertainty. It’s not just any arbitrary mathematical shape for your new belief distribution.
Warren: That's interesting because it gives us a concrete benchmark instead of just throwing around abstract theorems about martingale structures. It’s saying, if you can't get there through that contraction path, then the update isn't possible under these conditions.
Iris: The authors focus on Theorem one here, which links this feasibility directly to finding a specific random variable X and a signal. They also handle the tricky case where you don't know the exact random variable Y beforehand.
Ruth: It feels like they’re trying to strip away some of that overly complex machinery, showing us what really matters for modeling uncertainty in economics or any field dealing with state estimation.
Warren: The implication is that when we model something uncertain, we should focus on these contraction constraints rather than getting bogged down in whether a specific mathematical structure perfectly matches an older theorem about martingales.
Iris: It’s a big simplification if you can trust this equivalence holds across different setups, even when the initial state distribution G is all you have.
Ruth: So the paper suggests that for practical modeling, the necessary and sufficient condition for a posterior mean to be feasible is simply belonging to that set of mean-preserving contractions of G.
Warren: It sets up a solid test. If your model generates a posterior mean distribution outside those bounds, you know it’s not going to work under these assumptions.
Iris: And they do point out the limitations, though—the main result is tied to this bounded interval of random variables they used for the proof.
Ruth: Exactly. So while this gives us a very strong characterization for simple models, we gotta keep in mind that extending it to more complicated spaces is where the real work lies.
Warren: That’s what we’ll be digging into next—how does this relate to those stochastic differential equations they mentioned?
Ludvig Sinander
University of Oxford
econ.TH
Submitted: 2026-11-11
Updated: 2026-11-11
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 78/100
The gist: Suppose there is an uncertain state of the world, drawn from a CDF G.
Key concepts
- MPC(G)
- This set contains distributions F that are mean-preserving contractions of the initial state distribution G. It is crucial for characterizing feasible posterior means, ensuring that the resulting belief distribution has a specific relationship with the original state uncertainty.
- Strassen's Theorem
- This theorem relates a distribution G to distributions F induced by a martingale structure involving two random variables, X and Y. It concerns the relationship between distributions arising from different probabilistic setups, specifically when both the signal and the state are chosen.
- Theorem 1
- This theorem provides an equivalence: F belongs to MPC(G) if and only if there exists a random variable X with distribution F such that E(Y|X) = X almost surely. It is the key result for characterizing feasible posterior-mean distributions in the under-specified model.
Terminology
Summary
Suppose there is an uncertain state of the world, drawn from a CDF G. p. 0 It is well known that the random posterior mean induced (via Bayes’s rule) by a signal must have a distribution F that is a mean-preserving contraction of G, and conversely that any mean-preserving contraction is induced by some signal
The Core Distinction between Theorems
The paper distinguishes between Strassen’s theorem and Theorem 1, highlighting that they concern different relationships between distributions and random variables. Strassen’s theorem relates a distribution G to the set of distributions F that can be induced by a martingale structure where X ∼ F and Y ∼ G. p. 2 In contrast, Theorem 1 fixes a probability space P and a random variable Y on P, asking for which distributions F there exists an X such that (X, Y) is a martingale with X ∼ F. p. 2 The authors note that the non-trivial direction of Theorem 1 implies the non-trivial direction of Strassen’s theorem, but not vice versa. p. 2
The Role of Richness and Feasibility
Theorem 1 establishes an equivalence between a distribution F belonging to MPC(G) and the existence of a random variable X such that X ∼ F and E(Y X) = X a.s. The concept of richness is crucial here, as it ensures that for any imaginable distribution conditional on the state Y, there exists a random variable X which has that conditional distribution. p. 3 Corollary 1 further connects feasibility—where F is the posterior-mean distribution induced by some signal given Y—to MPC(G) via Theorem 1. p. 0
Characterization of Feasible Distributions
Corollary 1 provides a precise characterization for distributions feasible given Y, stating that F is feasible given Y if and only if F belongs to MPC(G). The proof shows this by constructing a signal Z such that E(Y Z) ∼ F, which implies X = E(Y Z) satisfies E(Y X) = X a. p. 4s., thus showing F belongs to MPC(G). p. 1 Conversely, if F ∈ MPC(G), Theorem 1 guarantees the existence of an X ∼ F such that E(Y X) = X a. p. 2s, and setting Z = X yields a posterior mean M ∼ F, proving feasibility. p. 4
Implications for Under-Specified Models
Corollary 2 addresses the case where only the distribution G of the state of the world is given, showing that F is feasible given Y (where Y ∼ G) if and only if F belongs to MPC(G). The authors caution against directly applying Strassen’s theorem in this context because it implies choosing both the signal and the state, which is not economically meaningful when the state is exogenously given. p. 5 The relevant characterization for feasible posterior-mean distributions in this under-specified model remains MPC(G). p. 5
Analogy to Stochastic Processes
The distinction between Strassen’s theorem and Theorem 1 is compared to the difference between weak and strong solutions of stochastic differential equations (SDEs). A strong solution requires a given filtered probability space P, whereas a weak solution involves defining the process on another filtered probability space P' with a different exogenous forcing term B'. Weak solutions are noted as convenient but lack economic meaning because the forcing term is exogenously given. p. 6
Feasible Posterior-Mean Distributions
The paper defines a signal as (M, π), where M is a non-empty compact metric space and π is a Borel-measurable map [0, 1] → ∆(M). The interpretation of this signal structure is that M represents messages and π(x) is the distribution of messages sent if the state is x ∈ [0, 1]. p. 3 This structure allows for defining a posterior belief via Bayes’s rule, whose expectation is called the posterior mean. p. 3
The role of richness is to ensure that for any imaginable distribution conditional on the state Y, there exists a random variable X which has that conditional distribution. This richness provisos are ubiquitous in economics, but usually left implicit. p. 3
The authors state that the fundamental issue is that the model is under-specified (Y and the probability space are not given: only G is). The more reasonable interpretation of this under-specified model is that there is a fixed exogenous probability space and state of the world Y (a random variable whose distribution is G), even if this was not explicitly mentioned. p. 5 p. 5 p. 5 Then feasible posterior-mean distributions are characterised by Corollary 1 in the previous section. p. 5 p. 5 p. 5
The fact that this model is under-specified is ultimately harmless, because one way or another, the answer to ‘which posterior-mean distributions are feasible?’ is ‘those in MPC(G)’. p. 1 The relevant result is Theorem 1. p. 5 p. 5 p. 5
The authors conclude that it does not make economic sense to attribute the ‘MPC(G)’ characterisation of feasible posterior-mean distributions to Strassen’s theorem, which is relevant only if the state of the world is chosen (subject to a distribution constraint). Instead, Theorem 1 is the relevant result. p. 5
The paper concludes that the fundamental issue is that the model is under-specified (Y and the probability space are not given: only G is). The more reasonable interpretation of this under-specified model is that there is a fixed exogenous probability space and state of the world Y (a random variable whose distribution is G), even if this was not explicitly mentioned. Then feasible posterior-mean distributions are characterised by Corollary 1 in the previous section. The relevant result is Theorem 1.
The authors state that the fundamental issue is that the model is under-specified (Y and the probability space are not given: only G is). The more reasonable interpretation of this under-specified model is that there is a fixed exogenous probability space and state of the world Y (a random variable whose distribution is G), even if this was not explicitly mentioned. Then feasible posterior-mean distributions are characterised by Corollary 1 in the previous section. The relevant result is Theorem 1.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Ability to characterize feasible posterior-mean distributions based on mean-preserving contractions of a given distribution G using Theorem 1, which provides an equivalent condition for feasibility:
F is feasible given Y if and only if F belongs to MPC(G).
-
Developing algorithms that find a random variable X such that E(Y X) equals X almost surely, where Y has distribution G and F belongs to MPC(G), by leveraging the connection between Theorem 1 and Strassen's theorem.
-
Creating models for decision-making under uncertainty where the signal realization induces a posterior mean distribution that is a mean-preserving contraction of the prior state distribution, allowing for inference about signals without needing to specify the underlying probability space P or random variable Y explicitly, as suggested by Corollary 2.