@@ -61,6 +61,8 @@ import matplotlib as mpl
61
61
from mpl_toolkits.mplot3d import Axes3D
62
62
from matplotlib.animation import FuncAnimation
63
63
from IPython.display import HTML
64
+ from matplotlib.patches import Polygon
65
+ from mpl_toolkits.mplot3d.art3d import Poly3DCollection
64
66
```
65
67
66
68
## Definitions and examples
@@ -743,6 +745,11 @@ This is, in some sense, a steady state probability of unemployment.
743
745
744
746
Not surprisingly it tends to zero as $\beta \to 0$, and to one as $\alpha \to 0$.
745
747
748
+
749
+
750
+
751
+
752
+
746
753
### Calculating stationary distributions
747
754
748
755
A stable algorithm for computing stationary distributions is implemented in [ QuantEcon.py] ( http://quantecon.org/quantecon-py ) .
@@ -757,6 +764,11 @@ mc = qe.MarkovChain(P)
757
764
mc.stationary_distributions # Show all stationary distributions
758
765
```
759
766
767
+
768
+
769
+
770
+
771
+
760
772
### Asymptotic stationarity
761
773
762
774
Consider an everywhere positive stochastic matrix with unique stationary distribution $\psi^* $.
@@ -767,17 +779,24 @@ For example, we have the following result
767
779
768
780
(strict_stationary)=
769
781
``` {prf:theorem}
782
+ :label: mc_gs_thm
783
+
770
784
If there exists an integer $m$ such that all entries of $P^m$ are
771
- strictly positive, with unique stationary distribution $\psi^*$, then
785
+ strictly positive, then
772
786
773
787
$$
774
788
\psi_0 P^t \to \psi^*
775
789
\quad \text{ as } t \to \infty
776
790
$$
791
+
792
+ where $\psi^*$ is the unique stationary distribution.
777
793
```
778
794
795
+ This situation is often referred to as ** asymptotic stationarity** or ** global stability** .
796
+
797
+ A proof of the theorem can be found in Chapter 4 of {cite}` sargent2023economic ` , as well as many other sources.
798
+
779
799
780
- See, for example, {cite}` sargent2023economic ` Chapter 4.
781
800
782
801
783
802
@@ -848,103 +867,96 @@ Here
848
867
You might like to try experimenting with different initial conditions.
849
868
850
869
851
- #### An alternative illustration
852
870
853
- We can show this in a slightly different way by focusing on the probability that $\psi_t$ puts on each state.
854
871
855
- First, we write a function to draw initial distributions $\psi_0$ of size ` num_distributions `
856
-
857
- ``` {code-cell} ipython3
858
- def generate_initial_values(num_distributions):
859
- n = len(P)
860
-
861
- draws = np.random.randint(1, 10_000_000, size=(num_distributions,n))
862
- ψ_0s = draws/draws.sum(axis=1)[:, None]
863
-
864
- return ψ_0s
865
- ```
872
+ #### Example: failure of convergence
866
873
867
- We then write a function to plot the dynamics of $(\psi_0 P^t)(i)$ as $t$ gets large, for each state $i$ with different initial distributions
868
874
869
- ``` {code-cell} ipython3
870
- def plot_distribution(P, ts_length, num_distributions):
875
+ Consider the periodic chain with stochastic matrix
871
876
872
- # Get parameters of transition matrix
873
- n = len(P)
874
- mc = qe.MarkovChain(P)
875
- ψ_star = mc.stationary_distributions[0]
877
+ $$
878
+ P =
879
+ \begin{bmatrix}
880
+ 0 & 1 \\
881
+ 1 & 0 \\
882
+ \end{bmatrix}
883
+ $$
876
884
877
- ## Draw the plot
878
- fig, axes = plt.subplots(nrows=1, ncols=n, figsize=[11, 5])
879
- plt.subplots_adjust(wspace=0.35)
885
+ This matrix does not satisfy the conditions of
886
+ {ref}` strict_stationary ` because, as you can readily check,
880
887
881
- ψ_0s = generate_initial_values(num_distributions)
888
+ * $P^m = P$ when $m$ is odd and
889
+ * $P^m = I$, the identity matrix, when $m$ is even.
882
890
883
- # Get the path for each starting value
884
- for ψ_0 in ψ_0s:
885
- ψ_t = iterate_ψ(ψ_0, P, ts_length)
891
+ Hence there is no $m$ such that all elements of $P^m$ are strictly positive.
886
892
887
- # Obtain and plot distributions at each state
888
- for i in range(n):
889
- axes[i].plot(range(0, ts_length), ψ_t[:,i], alpha=0.3)
893
+ Moreover, we can see that global stability does not hold.
890
894
891
- # Add labels
892
- for i in range(n):
893
- axes[i].axhline(ψ_star[i], linestyle='dashed', lw=2, color = 'black',
894
- label = fr'$\psi^*({i})$')
895
- axes[i].set_xlabel('t')
896
- axes[i].set_ylabel(fr'$\psi_t({i})$')
897
- axes[i].legend()
895
+ For instance, if we start at $\psi_0 = (1,0)$, then $\psi_m = \psi_0 P^m$ is $(1, 0)$ when $m$ is even and $(0,1)$ when $m$ is odd.
898
896
899
- plt.show()
900
- ```
897
+ We can see similar phenomena in higher dimensions.
901
898
902
- The following figure shows
899
+ The next figure illustrates this for a periodic Markov chain with three states.
903
900
904
901
``` {code-cell} ipython3
905
- # Define the number of iterations
906
- # and initial distributions
907
- ts_length = 50
908
- num_distributions = 25
909
-
910
- P = np.array([[0.971, 0.029, 0.000],
911
- [0.145, 0.778, 0.077],
912
- [0.000, 0.508, 0.492]])
913
-
914
- plot_distribution(P, ts_length, num_distributions)
915
- ```
916
-
917
- The convergence to $\psi^* $ holds for different initial distributions.
902
+ ψ_1 = (0.0, 0.0, 1.0)
903
+ ψ_2 = (0.5, 0.5, 0.0)
904
+ ψ_3 = (0.25, 0.25, 0.5)
905
+ ψ_4 = (1/3, 1/3, 1/3)
918
906
907
+ P = np.array([[0.0, 1.0, 0.0],
908
+ [0.0, 0.0, 1.0],
909
+ [1.0, 0.0, 0.0]])
919
910
911
+ fig = plt.figure()
912
+ ax = fig.add_subplot(projection='3d')
913
+ colors = ['red','yellow', 'green', 'blue'] # Different colors for each initial point
920
914
921
- #### Example: failure of convergence
922
-
915
+ # Define the vertices of the unit simplex
916
+ v = np.array([[1, 0, 0], [0, 1, 0], [0, 0, 1], [0, 0, 0]])
923
917
924
- In the case of a periodic chain, with
918
+ # Define the faces of the unit simplex
919
+ faces = [
920
+ [v[0], v[1], v[2]],
921
+ [v[0], v[1], v[3]],
922
+ [v[0], v[2], v[3]],
923
+ [v[1], v[2], v[3]]
924
+ ]
925
925
926
- $$
927
- P =
928
- \begin{bmatrix}
929
- 0 & 1 \\
930
- 1 & 0 \\
931
- \end{bmatrix}
932
- $$
926
+ def update(n):
927
+ ax.clear()
928
+ ax.set_xlim([0, 1])
929
+ ax.set_ylim([0, 1])
930
+ ax.set_zlim([0, 1])
931
+ ax.view_init(45, 45)
932
+
933
+ # Plot the 3D unit simplex as planes
934
+ simplex = Poly3DCollection(faces,alpha=0.05)
935
+ ax.add_collection3d(simplex)
936
+
937
+ for idx, ψ_0 in enumerate([ψ_1, ψ_2, ψ_3, ψ_4]):
938
+ ψ_t = iterate_ψ(ψ_0, P, n+1)
939
+
940
+ point = ψ_t[-1]
941
+ ax.scatter(point[0], point[1], point[2], color=colors[idx], s=60)
942
+ points = np.array(ψ_t)
943
+ ax.plot(points[:, 0], points[:, 1], points[:, 2], color=colors[idx],linewidth=0.75)
944
+
945
+ return fig,
933
946
934
- we find the distribution oscillates
947
+ anim = FuncAnimation(fig, update, frames=range(20), blit=False, repeat=False)
948
+ plt.close()
949
+ HTML(anim.to_jshtml())
950
+ ```
951
+ This animation demonstrates the behavior of an irreducible and periodic stochastic matrix.
935
952
936
- ``` {code-cell} ipython3
937
- P = np.array([[0, 1],
938
- [1, 0]])
953
+ The red, yellow, and green dots represent different initial probability distributions.
939
954
940
- ts_length = 20
941
- num_distributions = 30
955
+ The blue dot represents the unique stationary distribution.
942
956
943
- plot_distribution(P, ts_length, num_distributions)
944
- ```
957
+ Unlike Hamilton’s Markov chain, these initial distributions do not converge to the unique stationary distribution.
945
958
946
- Indeed, this $P$ fails our asymptotic stationarity condition, since, as you can
947
- verify, $P^t$ is not everywhere positive for any $t$.
959
+ Instead, they cycle periodically around the probability simplex, illustrating that asymptotic stability fails.
948
960
949
961
950
962
(finite_mc_expec)=
@@ -1105,42 +1117,6 @@ mc = qe.MarkovChain(P)
1105
1117
ψ_star
1106
1118
```
1107
1119
1108
- Solution 3:
1109
-
1110
- We find the distribution $\psi$ converges to the stationary distribution more quickly compared to the {ref}` hamilton's chain <hamilton> ` .
1111
-
1112
- ``` {code-cell} ipython3
1113
- ts_length = 10
1114
- num_distributions = 25
1115
- plot_distribution(P, ts_length, num_distributions)
1116
- ```
1117
-
1118
- In fact, the rate of convergence is governed by {ref}` eigenvalues<eigen> ` {cite}` sargent2023economic ` .
1119
-
1120
- ``` {code-cell} ipython3
1121
- P_eigenvals = np.linalg.eigvals(P)
1122
- P_eigenvals
1123
- ```
1124
-
1125
- ``` {code-cell} ipython3
1126
- P_hamilton = np.array([[0.971, 0.029, 0.000],
1127
- [0.145, 0.778, 0.077],
1128
- [0.000, 0.508, 0.492]])
1129
-
1130
- hamilton_eigenvals = np.linalg.eigvals(P_hamilton)
1131
- hamilton_eigenvals
1132
- ```
1133
-
1134
- More specifically, it is governed by the spectral gap, the difference between the largest and the second largest eigenvalue.
1135
-
1136
- ``` {code-cell} ipython3
1137
- sp_gap_P = P_eigenvals[0] - np.diff(P_eigenvals)[0]
1138
- sp_gap_hamilton = hamilton_eigenvals[0] - np.diff(hamilton_eigenvals)[0]
1139
-
1140
- sp_gap_P > sp_gap_hamilton
1141
- ```
1142
-
1143
- We will come back to this when we discuss {ref}` spectral theory<spec_markov> ` .
1144
1120
1145
1121
``` {solution-end}
1146
1122
```
0 commit comments