Abstract
The probabilistic model introduced by Ernst Zermelo (1929) provides a fundamental framework for inferring player strengths from tournament results. In this article we study a linear approximation of the Zermelo equilibrium equations near the symmetric point where players have comparable strength. The resulting system takes the form of a Laplacian equation on the tournament graph.
This formulation naturally leads to an iterative solution in which information propagates along the network of games. In the case of Swiss tournaments, where all players play the same number of rounds, the first iteration produces the Buchholz score, while higher powers of the propagation matrix correspond to longer chains of opponents.
This interpretation shows that classical tie-break procedures can be understood as truncated approximations of a diffusion process on the tournament graph, connecting practical tournament rules with probabilistic models and graph-theoretic structures.
The Zermelo model
The probabilistic model proposed by Ernst Zermelo (1929)
Die Berechnung der Turnier-Ergebnisse als ein Maximumproblem der Wahrscheinlichkeitsrechnung
is one of the earliest attempts to infer player strengths from tournament results. In this model every player \(i\) is assigned a positive strength parameter
\[ u_i>0. \]
The probability that player \(i\) defeats player \(j\) is
\[ P_{ij}=\frac{u_i}{u_i+u_j}. \]
Thus stronger players have higher winning probability, but outcomes remain probabilistic. The goal of the model is to determine the strengths \(u_i\) that best explain the observed tournament results.
Likelihood Formulation
Suppose that between players \(i\) and \(j\)
- \(w_{ij}\) = number of wins of (i) against (j)
- \(w_{ji}\) = number of wins of (j) against (i)
and \[ n_{ij}=w_{ij}+w_{ji} \]
is the total number of games between them.
Ignoring draws for the moment, the probability of observing exactly these results is
\[ L_{ij}= \left(\frac{u_i}{u_i+u_j}\right)^{w_{ij}} \left(\frac{u_j}{u_i+u_j}\right)^{w_{ji}} . \]
The likelihood of the entire tournament is the product over all pairs
\[ L(u_1,\dots,u_n)= \prod_{i<j} \left(\frac{u_i}{u_i+u_j}\right)^{w_{ij}} \left(\frac{u_j}{u_i+u_j}\right)^{w_{ji}} . \]
Taking logarithms gives the log-likelihood
\[ \ell(u)= \sum_{i<j} \left[ w_{ij}\log u_i + w_{ji}\log u_j - n_{ij}\log(u_i+u_j) \right]. \]
The strengths \(u_i\) are obtained by maximizing this function.
Maximum Likelihood Equations
Differentiating with respect to \(u_i\):
\[ \frac{\partial \ell}{\partial u_i} = \sum_{j\ne i} \left( \frac{w_{ij}}{u_i} - \frac{n_{ij}}{u_i+u_j} \right). \]
At the maximum likelihood point
\[ \frac{\partial \ell}{\partial u_i}=0. \]
Multiplying by \(u_i\) gives
\[ \sum_{j\ne i} w_{ij} = \sum_{j\ne i} n_{ij}\frac{u_i}{u_i+u_j}. \]
Define
\[ W_i=\sum_{j\ne i} w_{ij} \]
which is the total number of wins of player \(i\).
We obtain the equilibrium equations
\[ W_i= \sum_{j\ne i} n_{ij}\frac{u_i}{u_i+u_j}. \]
The interpretation is simple:
observed wins = expected wins predicted by the model
Thus the model balances observed results with probabilistic expectations.
Remark on Draws
Chess games may end in draws.
In rating systems such as Elo, a draw is counted as half a point for each player. In the likelihood formulation this is equivalent to treating a draw as half a win and half a loss. Therefore the equations remain valid if we replace the number of wins (W_i) with the total score
\[ S_i = \text{wins} + \tfrac12 \text{draws}. \]
From now on we use \(S_i\), since chess tournaments are scored in points.
Linearization of the Zermelo Model
The equations above are nonlinear, which makes them difficult to solve analytically.
However, important insight can be obtained by studying the model near an equilibrium where players have similar strength.
Assume
\[ u_i = u_0 + \varepsilon_i \]
where the deviations \( \varepsilon_i \) are small.
The winning probability becomes
\[ P_{ij}= \frac{u_0+\varepsilon_i}{2u_0+\varepsilon_i+\varepsilon_j}. \]
Expanding to first order gives
\[ P_{ij} \approx \frac12 + \frac{\varepsilon_i-\varepsilon_j}{4u_0}. \]
Thus the probability depends linearly on the difference of strength deviations.
Expected Score
The expected score of player \(i\) is
\[ E_i=\sum_{j\ne i} n_{ij}P_{ij}. \]
Using the linear approximation
\[ E_i \approx \sum_{j\ne i} n_{ij} \left( \frac12 + \frac{\varepsilon_i-\varepsilon_j}{4u_0} \right). \]
Define
\[ m_i=\sum_{j\ne i}n_{ij} \]
which is the number of games played by player \(i\). Then
\[ E_i \approx \frac{m_i}{2} + \frac{m_i}{4u_0}\varepsilon_i - \frac{1}{4u_0} \sum_{j\ne i} n_{ij}\varepsilon_j. \]
Linear System for Strength Deviations
At equilibrium the observed score equals the expected score:
\[ S_i=E_i. \]
Substituting the approximation yields
\[ m_i\varepsilon_i - \sum_{j\ne i} n_{ij}\varepsilon_j = 4u_0 \left( S_i-\frac{m_i}{2} \right). \]
This is a linear system. The matrix
\[ L = D-A \]
appearing here is the Laplacian of the game graph, where
- \(D=\text{diag}(m_i)\)
- \(A=(n_{ij})\)
Graph Interpretation
Each player is a node of a graph and every game creates a connection between two nodes. The Laplacian equation
\[ L\varepsilon = b, \]
where \(b = 4u_0 \left( S_i-\frac{m_i}{2} \right)\), means that player strength is determined by a balance between
- his own score
- the strengths of the opponents he faced.
The constant factor \(4u_0\) only rescales all strength deviations by the same positive constant. Since rankings depend only on the relative ordering of players, this factor can be absorbed into the definition of the variables. For simplicity we therefore rewrite the system as
\[ L\varepsilon = S - \frac{m_i}{2}. \]
Iterative Solution
The linear system obtained from the linearization is
\[ m_i\varepsilon_i - \sum_j n_{ij}\varepsilon_j = S_i-\frac{m_i}{2}. \]
Solving for \( \varepsilon_i \) gives
\[ \varepsilon_i = \left(\frac{S_i}{m_i}-\frac12\right) + \frac{1}{m_i}\sum_j n_{ij}\varepsilon_j . \]
Define
\[ c_i=\frac{S_i}{m_i}-\frac12 \]
and
\[ W_{ij}=\frac{n_{ij}}{m_i}. \]
The coefficient \(W_{ij}\) represents the fraction of games played by player \(i\) against player \(j\). The system can therefore be written compactly as
\[ \varepsilon = c + W\varepsilon . \]
This naturally suggests the fixed-point iteration
\[ \varepsilon^{(t+1)} = c + W\varepsilon^{(t)}. \]
Interpretation:
player strength = performance relative to 50% + average strength of opponents.
The matrix \(W\) therefore describes how performance information propagates through the network of games.
The Swiss Tournament Case
In a Swiss tournament all players normally play the same number of rounds. Thus
\[ m_i = m \]
for all players. Although this assumption is not mathematically necessary, it simplifies the formulas considerably and makes the structure of the iteration clearer. Now
\[ W_{ij}=\frac{n_{ij}}{m} \]
and since every player plays \(m\) games
\[ \sum_j W_{ij}=1 . \]
Thus \(W\) simply computes the average strength of the opponents of player (i). The iteration becomes
\[ \varepsilon^{(t+1)} = c + W\varepsilon^{(t)} . \]
where \(c = \left(\frac{S}{m}-\frac12\right)\).
Propagation of Information
To understand the meaning of the iteration, expand the solution. Repeated substitution gives
\[ \varepsilon^{(t)} = c + Wc + W^2c + \dots + W^{t}c . \]
If the process converges,
\[ \varepsilon = (I-W)^{-1}c = c + Wc + W^2c + W^3c + \dots \]
which represents a diffusion of information along longer and longer paths in the tournament graph. Each power of \(W\) corresponds to an additional step in the network of opponents.
Iteration 0
Take the simplest initial estimate
\[ \varepsilon^{(0)}_i = S_i . \]
Players are ranked purely by their score.
| criterion | information used |
|---|---|
| Score | direct results |
This is the primary ranking criterion in Swiss tournaments.
Iteration 1
Applying the iteration once gives
\[ \varepsilon_i^{(1)} = c_i + \sum_j W_{ij}S_j . \]
Multiplying by \(m\) to remove normalization,
\[ m\varepsilon_i^{(1)} = (S_i-\tfrac{m}{2}) + \sum_j n_{ij}S_j . \]
The dominant term is
\[ \sum_j n_{ij}S_j . \]
But this is exactly the Buchholz score
\[ BH_i = \sum_{j\in Opp(i)} S_j . \]
Thus Buchholz corresponds to the first propagation step \(W S\). It measures the strength of the opponents faced by player \(i\).
Iteration 2
Applying the iteration again gives
\[ \varepsilon^{(1)} = c + WS. \]
\[ \varepsilon^{(2)} = c + Wc + W^2S . \]
The term \(W^2S\) propagates information two steps in the tournament graph. The second iteration produces the sum of the Buchholz scores of the opponents (a “Buchholz of Buchholz”), corresponding to two-step propagation in the tournament graph.
Global Interpretation
The successive approximations therefore correspond to increasing levels of information about the tournament.
| iteration | information used |
|---|---|
| 0 | direct results |
| 1 | opponents (Buchholz) |
| 2 | opponents of opponents |
| ... | longer chains |
| ∞ | global Zermelo equilibrium |
Classical tie-breaks can therefore be interpreted as
early truncations of the diffusion process that leads to the global Zermelo ranking.
They approximate the full probabilistic model using only local information from the tournament graph. The first iteration naturally produces the quantity
\[ S_i + BH_i, \]
which combines the player’s score with the strength of the opponents faced. In the usual Swiss system these two components are separated into a primary score and a Buchholz tie-break, which may produce paradoxical situations where the tournament winner has a significantly weaker opposition. The linearized Zermelo model naturally combines both components into a single performance measure.
Final Remark
A Neutral Fictitious Opponent
The Laplacian system obtained from the linearization,
\[ L\varepsilon = b, \]
has a well-known structural property: the Laplacian satisfies
\[ L\mathbf{1}=0. \]
Therefore the solution is determined only up to an additive constant. This reflects the fact that the model depends only on differences of strength: adding the same constant to all players does not change any predicted result.
A convenient way to remove this indeterminacy is to introduce a fictitious neutral opponent. Suppose that every player plays two additional games against this opponent, winning one and losing one.
For each player (i) this gives
\[ m_i' = m_i + 2, \qquad S_i' = S_i + 1 . \]
However,
\[ S_i' - \frac{m_i'}{2} = (S_i+1) - \frac{m_i+2}{2} = S_i - \frac{m_i}{2}, \]
so the right-hand side of the linear system remains unchanged. The Laplacian matrix, on the other hand, becomes
\[ L' = L + 2I . \]
The system therefore becomes
\[ (L+2I)\varepsilon = b . \]
This modification has several useful consequences:
- the matrix becomes invertible, so the solution is unique;
- the fictitious opponent connects all players through a common node, ensuring that the tournament graph is connected;
- the relative differences \( \varepsilon_i - \varepsilon_j \) remain unchanged, so the ranking structure is preserved.
Interestingly, the matrix (L+2I) coincides with the matrix used in the ranking method of Wesley N. Colley.
Random Walk Interpretation
Another illuminating interpretation emerges from the matrix
\[ W_{ij}=\frac{n_{ij}}{m_i}. \]
Since
\[ \sum_j W_{ij}=1, \]
the matrix (W) is row-stochastic and can be interpreted as the transition matrix of a random walk on the tournament graph.
Starting from player (i), one step of the process consists of moving to one of his opponents chosen uniformly at random.
In this interpretation
\[ (W^k)_{ij} \]
represents the probability that a random walk starting from player (i) reaches player (j) after (k) steps.
The iterative solution
\[ \varepsilon = c + Wc + W^2c + W^3c + \dots \]
can therefore be viewed as a diffusion of performance information through the tournament graph, where the influence of results propagates along chains of opponents of increasing length.
Conclusion
The analysis presented here shows that classical Swiss tie-breaks are not arbitrary constructions but can be interpreted as early truncations of a diffusion process on the tournament graph arising from the linearized Zermelo model. In this perspective, the Laplacian structure provides a natural mathematical framework linking probabilistic models of competition, graph theory, and practical ranking procedures used in tournaments.
version 1: 2026.03.15