Macro-dynamics of the energy transition : robustness,

money velocity and public policies

Edouard Dossetto

To cite this version:

Edouard Dossetto. Macro-dynamics of the energy transition : robustness, money velocity and public
policies. Economics and Finance. Université Panthéon-Sorbonne - Paris I, 2022. English. �NNT :
2022PA01E004�. �tel-03725727�

HAL Id: tel-03725727

https://theses.hal.science/tel-03725727
Submitted on 18 Jul 2022

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 UNIVERSITÉ PARIS 1 PANTHÉON-SORBONNE
UFR02 d’Économie
Laboratoire de rattachement : Centre d’Économie de la Sorbonne
UMR 8174

THÈSE
Pour l’obtention du titre de Docteur en Économie
Présentée et soutenue publiquement
le 28 Février 2022 par
Edouard Dossetto

Macro-dynamics of the energy transition:

robustness, money velocity and public policies

Sous la direction de
• M. Christophe Chorro
Maı̂tre de conférence en Mathématiques appliquées - HDR, Paris 1 Panthéon-Sorbonne
• et M. Gaël Giraud
Directeur de recherche au CNRS — Centre d’Économie de la Sorbonne

Composition du jury

• M. Jean-Bernard Châtelain, Professeur des Universités, Paris 1 Panthéon-Sorbonne, PSE

• M. Matheus R. Grasselli, Professeur des Universités, McMaster University Rapporteur
• M. Stefano Demichelis, Professeur des Universités, Università Di Pavia Rapporteur
• M. Adrien Nguyen-Huu, Maı̂tre de conférence, Université de Montpellier
• M. Joseph E. Stiglitz, Professeur des Universités, Columbia University
“L’Université Paris 1 Panthéon-Sorbonne n’entend donner aucune approbation, ni
improbation aux opinions émises dans cette thèse ; elles doivent être considérées
comme propres à leur auteur”

1
 À mes grands-parents

2
Acknowledgments

This PhD thesis is the result of my five-year research in the field of macro-economic modeling in
order to provide decision makers with a robust tool to capture “climate-driven Minsky moments”
and identify policies avoiding it. It was carried out between September 2016 and February 2022
at the Centre d’Économie de la Sorbonne (Université Paris 1 PanthéonSorbonne). This work was
supported by the Commission for Doctoral Formation of my State body : the Corps des Ingénieurs
des Ponts, des Eaux et Forêts.

I deeply thank my supervisors Christophe Chorro and Gaël Giraud for giving me the chance
to research on such an interesting topic and to get insight into the research work of two di↵erent
places. I feel grateful for that, as well for all the time and energy they invested in me. Without
their supervisions, I would not have been able to complete this work.

My expression of gratitude goes also to my rapporteurs, Prof. Stefano Demichelis and Prof.
Matheus R. Grasselli, who have done me the honour of reading and assessing my PhD thesis. I
sincerely thank the member of this jury for their interest in my work and their presence at the
defense : Prof. Nicolas Bouleau, Prof. Jean-Bernard Châtelain, Asst. Prof. Adrien Nguyen-Huu
and Prof. Joseph E. Stiglitz.

Thanks to the Alliance Doctoral Mobility Grant, I had the opportunity to work in an amazing
environment during a three-month stay in Columbia University, following the invitation of Prof.
Joseph E. Stiglitz. I would like to use the opportunity to thank him and his assistants, Mrs
Susanna Asher De Martino and Mr Caleb A. Oldham, as well as the many administrative members
of the Alliance Program.

Furthermore, I want especially thank my co-worker in Columbia, Mrs Anna McClelland-Enger,

and in France, Mrs Fatma Rostom, for their feedbacks and work support, as well as the many
people that helped me to write this thesis, e.g. by inspiring scientific conversations and inputs,
amazing programming support, great cooperations, constant assistance, or kind encouragements:
Dr Matthew Arran, Mr Stanislas Augier, Dr Emmanuel Bovari, Mr Thomas Deshays, Mrs Andrée
Kahlmorgan, Dr Florent McIsaac, Mr Clovis Pareiko and Dr Antonin Pottier.

Eventually, I would like to thank all my family and especially my parents, Geneviève and Michel
- maybe they do not understand my research, but they understand me, support me and love me -,
my brother, Olivier, for his excellent ideas in modeling, my uncle, Marc, for his support and Marie
for her constant encouragement and her love.

3
Résumé Substantiel

Cette thèse prend son origine dans le discours fondateur de Mark Carney (gouverneur de la
Banque d’Angleterre) sur les risques financiers induits par le climat.

La probabilité d’un “climate-driven Minsky moment” [Carney, 2019] ou moment de Minsky (in-
duit par le climat) - terme utilisé pour désigner un e↵ondrement soudain du prix des actifs et
pour dépeindre les crises financières survenues en 2001 et 2007-2009 - combinée à l’incapacité d’une
grande majorité des modèles de l’époque à prévoir de telles crises nous invite à nous interroger sur
la fiabilité des modèles actuels pour l’orientation des décideurs. Si l’on ne peut ignorer l’ampleur
des impacts socio-économiques de la récente crise sanitaire et la remise en question de nos politiques
économiques qu’elle a engendrée, cette thèse se concentre sur un autre défi majeur de notre époque:
la nécessaire transition de nos systèmes économiques pour assurer la résilience de nos sociétés face
à cette autre menace fondamentale qu’est le changement climatique.

Le changement climatique constitue en e↵et une menace considérable pour nos sociétés en raison
de son ampleur et de son impact dans tous les secteurs et toutes les régions du monde,1 son caractère
prévisible,2 son irréversibilité 3 et sa dépendance vis-à-vis des actions à court terme.4 Avec le
sommet de Rio, le protocole de Kyoto et maintenant les règles de mise en œuvre de l’accord de
Paris (finalisé en 2021), les États se sont positionnés en pierres angulaires de la lutte contre le
changement climatique. Lors de la COP21, la sphère politique mondiale a confirmé sa volonté de
faciliter la transition vers une société bas carbone.5
1
Le rapport du Network for Greening the Financial System (NGFS) indique que le changement climatique a↵ectera
tous les agents économiques (ménages, entreprises, gouvernements). Les risques seront probablement corrélés à des
points de basculement et potentiellement aggravés par ceux-ci, de manière non linéaire. Cela signifie que les impacts
pourraient être beaucoup plus importants, plus étendus et plus variés que ceux d’autres changements structurels.
2
Bien que les résultats exacts, l’horizon temporel et la trajectoire future soient incertains, il existe un degré élevé
de certitude pour qu’une certaine combinaison de risques physiques et de transition se matérialise à l’avenir. [Carney,
2019]
3
Comme le montrent les travaux du GIEC, la concentration des émissions de gaz à e↵et de serre (GES) dans
l’atmosphère détermine l’impact du changement climatique et aucune technologie mature n’a encore une portée
suffisante pour inverser le processus. Au-delà d’un certain seuil, elles présentent un fort potentiel pour imposer des
conséquences irréversibles sur une multitude d’écosystèmes, même si l’horizon temporel et le degré de gravité restent
incertains [IPCC, 2014]. De nombreux changements dus aux émissions passées et futures de gaz à e↵et de serre seront
irréversibles, en particulier vis-à-vis des modifications des océans, des calottes glaciaires et du niveau mondial des
mers. [IPCC, 2021]
4
Les actions menées aujourd’hui détermineront la nature et l’ampleur des catastrophes liées au climat. Elles exigent
donc des stratégies fiables et globales de long terme. Par exemple : des réductions fortes, rapides et durables des
émissions de CH4 limiteraient également l’e↵et de réchau↵ement résultant de la diminution de la pollution par les
aérosols et amélioreraient la qualité de l’air. [IPCC, 2021]
5
Par la suite, la loi française sur la transition énergétique, adoptée en juillet 2015, exige des entreprises et des
institutions financières qu’elles rendent compte des risques climatiques et qu’elles limitent la hausse mondiale des
températures à moins de 2 degrés Celsius, et de préférence aussi près que possible de 1,5 degré Celsius. Le ‘Green

4
 La demande d’actions immédiates par le Groupe d’experts intergouvernemental sur l’évolution
du climat ( [IPCC, 2014], [IPCC, 2018] et plus récemment dans le rapport de son premier groupe
de travail [IPCC, 2021]) afin d’atténuer les impacts à court terme et l’irréversibilité à grande échelle
du changement climatique, nécessite la construction de modèles fiables pour garantir la transition
vers une économie à faible émission de carbone (ou même une économie neutre du point de vue
climatique). Le développement de mon travail vise à étendre un modèle de transition macro-
économique existant, stock-flux cohérent, avec une dynamique non linéaire et à en évaluer la fiabilité.
La dynamique de ce type de modèle conduit en e↵et à des équilibres multiples incorporant la
possibilité d’un e↵ondrement. L’utilisation d’un tel modèle fournit donc un cadre dans lequel
les politiques publiques capables d’éviter ce que Marc Carney appelle un “climate-driven Minsky
moment” [Carney, 2019], peuvent être analysées.

Ce terme fait référence à l’hypothèse d’instabilité financière (FIH) de Minsky, qui a analysé dans
les années 1970 la probabilité d’une instabilité intrinsèque du système financier et la manière dont
elle peut impliquer une nouvelle crise comparable en terme d’ampleur à celle de 1929. Comme
expliqué dans [McIsaac, 2016], le principal avantage de ce cadre est sa capacité à reproduire une
crise financière, telle que la crise des subprimes, de manière endogène, alors que la grande majorité
des Dynamic Stochastic General Equilibrium (DSGE), très étudiés dans la littérature, ne peuvent
guère considérer ces situations autrement que comme des “cygnes noirs” ou des risques importants
ayant une très faible probabilité de se matérialiser.

Il existe donc, d’une part, un besoin évident de disposer de modèles de transition vers une société
neutre en carbone, capable de prédire le risque d’un “moment de Minsky induit par le climat”, afin
d’orienter les décisions de politique publique ; et d’autre part, la conscience que la production d’un
tel modèle présente de grandes difficultés analytiques, comme le souligne [Carney, 2019], aggravées
par l’incapacité des modèles macro-économiques standard à prévoir les crises financières des années
2000.

Cet e↵et cocktail motive une analyse de ces modèles à la frontière de la modélisation mathématique,
des hypothèses économiques (ou financières), et des objectifs et contraintes écologiques. Cette anal-
yse en introduction de mon travail se fonde sur le rôle ambigu des modèles, dû à l’incertitude et à
la nature non-bijective de la modélisation (Section 1.2), et la crise des modèles macro-économiques,
dont certaines caractéristiques conduisent aux incapacités indiquées ci-dessus (Section 1.3).

L’ambiguı̈té des modèles, source d’incertitudes pour l’orientation des décideurs.

Il est en e↵et essentiel de tenir compte de la polyvalence et de l’ambiguı̈té des modèles pour orienter
au mieux nos décisions publiques. Les modèles mathématiques et les simulations informatiques sont
utilisés dans une grande variété de domaines et d’institutions. Les écologistes les développent par
exemple pour étudier les écosystèmes et la dynamique des populations. Les institutions du monde
entier les utilisent pour étudier les économies mondiales et nationales. En ce qui concerne le change-
ment climatique, deux grands types de modèles sont utilisés : a) les modèles physiques qui évaluent
l’évolution du climat et b) les modèles macro-économiques qui évaluent l’impact économique ou
Deal” de l’UE et son paquet “Fit-for-55” comprennent une législation contraignante pour garantir que l’UE atteigne
ses objectifs climatiques et énergétiques de 2030. Lors de la COP26, la présidence britannique a réussi à placer la
neutralité carbone au centre des discussions.

5
financier du changement climatique. Dans certains cas, les modèles physiques permettent de corro-
borer ou de falsifier des hypothèses sur les systèmes écologiques, et les modèles macro-économiques
permettent d’orienter les politiques publiques. Par exemple, en France, le Commissariat général
du développement durable (du Ministère de la transition écologique) évalue l’impact de la stratégie
nationale bas carbone avec le modèle Three-ME2. Ses résultats prévoient une augmentation du PIB
de 1,6 d’ici 2035, alors que le ministère français des Finances prévoit un e↵et de contraction, avec
son propre modèle Mésange (voir [Grandjean and Giraud, 2017]).

Les modèles physiques (modèles climatiques ou météorologiques) utilisés pour prédire l’évolution
de l’environnement peuvent être évalués par leur capacité à simuler les conditions climatiques
actuelles et passées, et à reproduire les variations climatiques à chaque échelle de temps. Ces
modèles ont donc une capacité prédictive qui permet d’identifier une politique publique efficace
avec fiabilité (robustesse) [Grandjean and Giraud, 2017].

Pour les modèles macro-économiques, atteindre des résultats robustes est un sujet beaucoup plus
sensible en raison a) des échecs d’une grande majorité de modèles à prédire les crises des années
2000,6 et b) de la sensibilité (manque de robustesse) des modèles actuels.7 Il est clair aujourd’hui
que cette sensibilité a conduit à de grandes sous-estimations de l’impact économique du climat.8
C’est pourquoi une analyse de sensibilité globale des modèles macro-économiques est nécessaire
pour éclairer les discours sur la fiabilité de leurs résultats, et ainsi contribuer à tempérer la crise
qu’ils traversent.

La capacité du modèle à représenter fidèlement la réalité dépend en e↵et, à la fois d’études

empiriques et de considérations théoriques. Par exemple, les fonctions comportementales (comme
la courbe de Phillips ou la fonction d’investissement en économie) sont construites pour s’adapter à
des données qui impliquent un grand nombre de calibrations ou d’estimations (basées sur des études
empiriques [Grandjean and Giraud, 2017]), alors que la fonction de production (comme [Cobb and
Douglas, 1928], [Leontief, 1951], CES [Solow, 1956] ou Putty-clay au sens de [Akerlof and Stiglitz,
1969]) est un concept clé de la théorie de la croissance économique, et peut donc faire apparaı̂tre
des divergences entre les lois de production réelles et leur représentation fonctionnelle (voir [Stiglitz,
2018]).

La variété des entrées d’un modèle sou↵re, elle aussi, d’incertitudes sur ses valeurs réelles. Toute
prévision doit donc tenir compte de ces décalages possibles afin de fournir des résultats fiables (ro-
bustes). Cette considération critique est l’objet de l’analyse d’incertitude (UA) et de l’analyse de
sensibilité (SA). En particulier dans les sciences sociales, ou dans des environnements économiques
complexes, où ces calibrations et estimations de paramètres conduisent à des résultats très sensi-
bles aux paramètres, l’impact des incertitudes des entrées ne peut être considéré comme négligeable.
Ainsi, la remise en question de la robustesse des modèles macro-économiques est de plus en plus con-
sidérée comme essentielle pour mener des politiques publiques dans ces domaines [European Com-
mision, 2009].
6
Voir Section 1.3.2.
7
Voir une revue de certains de ces modèles dans la Section 2.4.
8
Comme exposé dans le livre de [Pottier, 2014].

6
 Les modèles macro-économiques écologiques tentent de relier les environnements climatiques et
économiques. Certains d’entre eux visent à évaluer l’impact d’une politique publique,9 avec une
analyse coûts-bénéfices ou coût-efficacité, à l’échelle nationale, régionale ou mondiale. Inspirés
de [La↵argue, 2012] et [Grandjean and Giraud, 2017], ces modèles macro-économiques sont classés
comme suit :

• Modèles néo-keynésiens : avec des ajustements de court terme, ces modèles sont pilotés par
la demande globale (par exemple, Mésange). Des déséquilibres temporaires sont possibles en
raison de l’inertie des ajustements de quantités ;

• Modèles d’équilibre général (CGE) : basés sur une représentation walrasienne de l’économie,
où les prix sont parfaitement flexibles et où les préférences d’un agent rationnel optimisant
une fonction d’utilité sont explicitement décrites. Divers types de modèles existent dans cette
catégorie (tels que les modèles d’évaluation intégrée - Integrated assessment model comme
celui de [Nordhaus, 2008]). En général, les institutions utilisent à la fois un modèle DSGE
estimé (Global Integrated Monetary and Fiscal Model - GIMF, New Area-Wide Model of The
Euro area - NAWM, ...) et un modèle néo-keynésien pour leurs prévisions ;

• Modèles de déséquilibre (comme ThreeMe, un modèle post-keynésien) : avec une dynamique

des prix spécifique, indépendante de l’ajustement de l’o↵re et de la demande, une production
multisectorielle et une dynamique des salaires à travers une nouvelle courbe de Phillips à
court terme. Le modèle GEMMES (General Monetary and Multisectoral Macrodynamics for
the Ecological Shift) est basé sur les travaux de [Grasselli and Costa Lima, 2012] avec une
dynamique non linéaire hors équilibre inspirée des travaux de [Goodwin, 1967], [Akerlof and
Stiglitz, 1969], [Keen, 1995]. Ces équations sont basées sur la dynamique proie-prédateur de
Lotka-Volterra, qui prend en compte la dette privée et/ou publique. Ces types de modèles sont
capables de représenter l’hypothèse d’instabilité financière de Minsky (FIH) et l’éventualité
de krachs. La dynamique stochastique permet l’émergence d’indicateurs agrégés, mais ils
sont fortement critiqués en raison de leur forte dépendance aux paramètres, de leur caractère
myope (sans anticipations) et de leur croissance exogène. Cependant, ils semblent être les plus
à même de représenter les trajectoires que nous voulons explorer, si nous pouvons compenser
ces critiques. Une version estimée du modèle existe avec une boucle de rétroaction climatique
développée dans [Bovari et al., 2018] et une première approche dédiée à l’économie brésilienne
a été mise en œuvre par [Bastidas et al., 2017].

Pour faire face au risque d’un moment de Minsky lié au climat, en utilisant un modèle robuste pour
orienter les politiques publiques afin de garantir une transition vers une société neutre en carbone,
ces modèles de déséquilibre semblent les plus convaincants car capables de surmonter certaines des
défaillances des autres modèles macro-économiques.10
9
Mésange (Bercy), Three-Me (OFCE / ADEME), Imaclim (CIRED), Nemesis (Erasme), Gemini (Lausanne), GEM
E3 (Commission européenne), GEMMES (AFD), etc.
10
Un aperçu plus large de ces critiques est disponible dans les chapitres de la thèse, mais nous pouvons en afficher
certaines ici, par exemple :
• La plupart des nouveaux modèles classiques n’ont pas d’énergie ou de matière dans leur fonction de production
(exception : lorsqu’on utilise la fonction de production KLEM avec Capital (K), Travail (L), Énergie (E),
Matière (M), par exemple, [Stiglitz, 1974a]). Ces modèles considèrent généralement que, l’élasticité du PIB à
l’énergie est égale à la quantité nominale d’énergie dans le PIB, ce qui est négligeable. En fait, [Giraud and
Kahraman, 2014] a montré que cette part en volume représente en moyenne 60% (6 fois plus que ce que les
nouveaux économistes classiques admettent habituellement).
• La monnaie est toujours considérée comme neutre lorsqu’elle est prise en compte, alors qu’elle est démontrée

7
 Selon l’analyse de [Grandjean and Giraud, 2017], on peut en e↵et s’interroger sur la capacité des
modèles macro-économiques à prédire des trajectoires fiables, notamment en raison de la définition
de leurs entrées,11 l’ambiguı̈té de leurs indicateurs de production (PIB12 ; Coût des politiques
publiques ; Taux de chômage13 ), leur manque de back-testing, leur sensibilité potentiellement élevée
aux paramètres calibrés, leur incapacité à représenter les trajectoires hors équilibre à équilibres
multiples, notamment en économie du changement climatique.

Dans le cas spécifique de l’économie du changement climatique, les modélisateurs sont en ef-
fet confrontés à deux domaines principaux : l’économie et l’environnement. Dans le domaine
de l’environnement, les chercheurs modélisent des interactions complexes entre des composants
dépendant d’observables (éventuellement probabilistes). En général, ces comportements peuvent
être décrits par des lois et des équations physiques. Ces équations fondamentales, telles que la
conservation de l’énergie, le transfert de chaleur, etc., ont déjà été établies et testées empirique-
ment. La résolution de ces équations, par le biais d’une méthode numérique et d’un échantillonnage
choisis, peut introduire certaines approximations. Ces équations font intervenir divers paramètres,
parmi lesquels certaines constantes fondamentales ne sont pas discutables (comme la constante de
Planck ou de Boltzman). Les incertitudes des modèles ne proviennent donc pas principalement des
équations elles-mêmes, mais plutôt de la manière dont elles sont mises en œuvre et résolues afin de
représenter le système réel.

En macroéconomie en revanche, les modélisateurs doivent représenter des interactions complexes

entre des unités de décision complexes (les humains), dont la description comportementale par des
équations mathématiques n’est pas indiscutable [Grandjean and Giraud, 2017]. Outre les doutes
légitimes que l’on peut avoir sur la pertinence de certains indicateurs et les biais possibles dans
la méthodologie de leur construction, les données sous-jacentes de ces modèles macro-économiques
sont difficiles d’accès en raison de leur nature ‘socialement construite”. Leur construction dépend
donc d’institutions disposant de techniques et de capacités di↵érentes pour évaluer finement ces
données d’entrée. D’autres paramètres d’entrée sont par ailleurs supposés constants, comme le taux
de dépréciation du capital, ce qui est tout sauf certain.
comme endogène dans le livre Illusion Financière de [Giraud, 2014].
• Ils n’admettent pas de dynamique globale hors équilibre (voir par exemple les modèles CGE) ;
• Ils sont principalement linéaires dans la dynamique présumée, même s’ils représentent des phénomènes très
complexes ;
• Ils ne prennent pas en compte la dette privée (comme le montre les travaux de [Keen, 2013] sur le lien dette-
déflation. Si un endettement excessif est si susceptible de générer une déflation, il est pour le moins surprenant
que le ratio d’endettement ne fasse pas partie d’un tel modèle — mentionnons une exception notable, [Krugman
and Eggertsson, 2013], qui est cependant une approche sans monnaie...) ni le secteur bancaire (uniquement
considéré comme des intermédiaires financiers). Voir par exemple Debt-Deflation versus the Liquidity Trap : the
Dilemma of Nonconventional Monetary Policy de [Giraud and Pottier, 2016] et Illusion Financière, [Giraud,
2014].
• En raison des hypothèses anticipations rationnelles et de demande globale, ils ne peuvent pas prédire un
équilibre avec un sous-emploi important (du moins lorsque le marché du travail est parfaitement flexible) et
peuvent à peine considérer les crises financières comme des cygnes noirs.

11
Par exemple, la controverse sur le capital de Cambridge à la fin des années 1950 exprime l’écart entre la définition
de la nature et du rôle des biens d’équipement. [Stiglitz, 1974b].
12
Voir une revue des critiques dans [Jany-Catrice and Méda, 2015] ou la proposition de [Piketty, 2009].
13
Les experts de la BCE estiment le taux d’atonie du marché du travail à environ 18% dans la zone euro, soit
presque le double du taux officiel d’Eurostat qui se situe autour de 9,5%. Cela montre le paradoxe actuel concernant
le calcul de cet indicateur [European Central Bank, 2017]

8
 Les incertitudes de ces entrées et la difficulté d’accès aux données nuisent à la fiabilité des résultats
ou à la qualité des recommandations en termes de politique publique (surtout lorsque le résultat
lui-même est un indicateur ambigu). Les modèles macro-économiques sont en e↵et susceptibles
de dépendre de nombreux paramètres. Si certains d’entre eux sont estimés, un grand nombre de
paramètres ne sont que calibrés. Les méthodes d’estimation peuvent conduire à des décalages entre
le résultat et la réalité. Ces types d’inadéquation peuvent être réduits en testant rétrospectivement
(ce que l’on nomme back-testing) le modèle sur des données historiques.

Ces estimations et calibrations des paramètres peuvent avoir des impacts importants sur les
résultats. La finalité de l’analyse de sensibilité est de décrire le lien entre l’impact sur les sorties et
les incertitudes sur les entrées. Ce type d’analyse est détaillé dans l’annexe A.2. Dans un système
très complexe, tel que les sciences de l’environnement ou la macroéconomie, avec de nombreuses non-
linéarités et paramètres, nous pouvons reconnaı̂tre le fait qu’une incertitude mineure peut entraı̂ner
une grande variation de la sortie.

Parmi les di↵érentes analyses de sensibilité, la méthode bayésienne est la plus fréquemment
utilisée pour calibrer les modèles DSGE, mais cette méthode comporte de nombreux biais a pri-
ori. L’inférence bayésienne utilisée pour construire l’analyse de sensibilité des modèles DSGE
est donc fortement dépendante des choix a priori des modélisateurs. Dans un domaine où les
faits et les valeurs sont difficiles à distinguer [Pottier, 2014], on peut comprendre comment une
méthode bayésienne pour calibrer un modèle DSGE pourrait potentiellement confirmer les biais
originaux des modélisateurs. D’autres analyses de sensibilité (dites locales) impliquent la sélection
d’un point nominal autour duquel tester la sensibilité du modèle.14 Le point nominal sélectionné
dans l’espace d’entrée (quelle que soit la méthode choisie) est critique pour l’analyse de sensibilité,
car le modélisateur l’explore dans un voisinage proche de cette valeur particulière.

Dans un domaine où les incertitudes sur les paramètres sont importantes, en raison de données
difficilement accessibles et socialement construites, on peut comprendre la nécessité de “back-tester”
le modèle et d’éprouver sa robustesse à ces incertitudes par l’analyse de sensibilité, afin de fournir
le résultat le plus fiable, même si la sortie reste discutable.

Sélection d’une famille de modèle capable de décrirer les risques induits par un moment
de Minksy dû au climat

Le modèle historique de Solow est un modèle dynamique, mais souvent utilisé pour analyser un
état stationnaire à long terme. Cette habitude est tout sauf scientifique, car elle présuppose que
notre économie réelle a atteint son équilibre. Les modèles d’équilibre général tentent de calculer
de nombreux équilibres statiques à chaque pas de temps, en tenant compte des éventuels chocs
exogènes pour représenter la grande volatilité des indicateurs économiques réels. Cependant, ces
modèles considèrent la cinématique du système à travers le prisme d’une succession d’équilibres.
Les modèles DSGE quant à eux, supposent que l’équilibre de Solow est atteint, mais que des chocs
exogènes en éloignent le système. Leurs analyses consistent à calculer le retour à l’état d’équilibre
de long terme, quelle que soit la nature du choc. Par essence, ces modèles ne paraissent donc pas
en mesure de prédire les trajectoires vers un e↵ondrement ou un moment de Minsky induit par le
climat. La dynamique hors équilibre d’un DSGE est par construction toujours transitoire.
14
Ces méthodes sont également présentées en annexe A.2.4.

9
 Pour respecter un objectif climatique, nous devons intégrer les contraintes climatiques dans le
modèle macro-économique afin de décrire les trajectoires de transition (comme cela a été remarqué
dans [Canfin et al., 2016]). L’objectif est de garantir une meilleure cohérence entre l’analyse à
court terme et les objectifs de décarbonation à long terme. Pour cette raison, selon [Grandjean and
Giraud, 2017], les modèles macro-économiques doivent :

• Être “Stock-Flow Consistent” (une condition de base pour garantir la fiabilité du modèle en
termes de comptabilité) ;

• Être “Physic-Consistent” (et en particulier avec les lois physiques, telles que la 2nd loi de la
thermodynamique) ;

• Décrire explicitement la circulation de la monnaie, et ne pas la supposer neutre a priori ;

• Être “back-testé” et analysé en termes de sensibilité aux paramètres ;

• Éviter les variables cachées ou infalsifiables.

Une typologie de modèles plus intéressante que les autres pour notre étude sont ainsi les modèles de
déséquilbre mentionnés précédemment (tels que Three-Me et son extension IMACLIM-3ME [Ghersi,
2020], [Akerlof and Stiglitz, 1969] ou [Grasselli and Costa Lima, 2012]). Ces modèles évaluent par
construction les trajectoires vers des équilibres multiples et sont capables de décrire la dynamique
globale hors équilibre.

Contrairement aux modèles macro-économiques habituels, le modèle [Grasselli and Costa Lima,
2012] est capable d’éviter certains de leurs défauts et de capturer les caractéristiques spécifiques dont
nous avons besoin pour prédire un éventuel e↵ondrement (provoqué par exemple, par une quantité
explosive de dettes). Ce modèle est basé sur les travaux de Goodwin et introduit la dette [Goodwin,
1967] et [Keen, 1995]. La nature et la stabilité des équilibres de dette ont été analysées dans [Grasselli
and Costa Lima, 2012] et [McIsaac, 2016] pour les extensions de Van Der Ploeg.15 Le modèle est
basé sur la logique prédateur-proie de Lotka-Volterra, avec une équation classique d’accumulation
du capital et deux fonctions comportementales (sur, d’une part, l’investissement dépendant des
profits, et, d’autre part, le lien entre le taux d’emploi et la part des salaires, à travers une courbe de
Phillips à court terme), la variation de la dette privée (comme la di↵érence entre l’investissement et
les profits) et un taux de croissance exogène de la population et de la productivité du travail. Les
défauts du modèle résident dans son modèle de croissance exogène, l’utilisation de la loi de Say (qui
peut être assouplie par l’introduction d’inventaires [Grasselli and Nguyen-Huu, 2018]), sa nature
myope (sans anticipations), et sa forte dépendance aux paramètres.16

Dans le cadre de notre travail, trois axes de recherche sont explorés pour évaluer la capacité de ce
modèle à répondre à l’objectif de notre thèse : la construction d’un modèle de transition robuste qui
doit être en mesure de fournir des informations sur les politiques publiques permettant d’éviter soit
un équilibre d’attraction avec du capital brun uniquement, soit un e↵ondrement dû à un “moment
de Minsky induit par le climat”. Cette construction est basée sur une structure stock-flux cohérente
avec une dynamique non linéaire conduisant à des équilibres multiples, qui peut intégrer ce moment
de Minsky. Ces axes sont les suivants :
15
Introduisant une fonction de production CES au lieu de la fonction de production Leontief [Van der Ploeg, 1985].
16
Voir aussi l’étude de [Pottier and Nguyen-Huu, 2017] sur la dépendance de la dynamique du modèle à la fonction
d’investissement choisie.

10
 • Quantifier la robustesse d’un modèle et comparer sa sensibilité à celle d’autres modèles macro-
économiques ;

• Étendre un modèle macro-économique existant, [Grasselli and Costa Lima, 2012], en intro-
duisant les inventaires, la vitesse de la monnaie, l’intervention du gouvernement afin d’aider
l’orientation des politiques publiques en conjonction avec les modèles actuels appliqués ;

• Évaluer l’impact des politiques publiques afin de garantir une transition entre les di↵érentes
natures de capital dans ce modèle.

Pour répondre aux critiques formulées à l’encontre du modèle [Grasselli and Costa Lima, 2012]
concernant sa forte dépendance vis-à-vis des paramètres, le chapitre 2 explore la robustesse du
modèle avec deux fonctions de production (Leontief et CES), à travers un analyse de sensibilité
globale avec des indices de Sobol’ estimés par la méthode [Saltelli, 2002].17 Le chapitre passe
également en revue certaines analyses sur des modèles DSGE, de la théorie des cycle réels (RBC) et
des modèles d’évaluation intégrée (IAM). Le chapitre 3 vise à améliorer notre modèle en introduisant
la vélocité de la monnaie (selon la théorie quantitative de la monnaie) dans un cadre assouplissant la
loi de Say par l’introduction des inventaires (inspirée par les travaux de [Grasselli and Nguyen-Huu,
2018]. Finalement, le modèle de transition est construit avec deux natures de capital et peut être
généralisé dans un cadre “Putty-Clay” (voir [Akerlof and Stiglitz, 1969], [Cass and Stiglitz, 1969]).
Le modèle de transition le plus simple possible, avec seulement deux natures de capital, fournit
déjà un critère de stabilité et des leviers possibles par l’intervention du gouvernement (au sens
de [Costa Lima and Grasselli, 2014]). Ce chapitre 4 étudie les équilibres multiples à long terme et
les critères de stabilité du modèle de transition avec di↵érents types d’intervention gouvernementale.

Quantifier la robustesse d’un modèle et comparer sa sensibilité à celle d’autres modèles

macro-économiques

Dans le chapitre 2, l’objectif est de tester la robustesse du modèle [Grasselli and Costa Lima,
2012] en termes de sensibilité aux paramètres. Ce travail a été mené avec le soutien du programme
Alliance (Alliance Doctoral Mobility Grant) et la Columbia University Graduate School of Arts and
Sciences, sur l’invitation du Prof. Joseph E. Stigltiz.

Les modèles macro-économiques actuellement utilisés sont confrontés à diverses défaillances (voir
[Stiglitz, 2018]), notamment dans un environnement économique complexe impliquant un grand
nombre de paramètres calibrés et estimés, ce qui conduit à des résultats très sensibles aux paramètres
(voir [Grandjean and Giraud, 2017]). Par conséquent, la construction d’un modèle, notamment en
sciences sociales ou en macroéconomie, nécessite de considérer trois types d’incertitudes :

• Incertitude sur la précision entre le modèle et la réalité (Sensibilité aux représentations) ;

• Incertitudes numériques correspondant aux choix d’échantillonnage ;

• Incertitudes liées aux paramètres (Sensibilité aux paramètres).

17
Le choix de la méthode est basé sur un état de l’art dans ce domaine (voir annexe A.2). Notons que la dépendance
aux valeurs initiales a été explorée à travers l’étude du bassin d’attraction par [McIsaac, 2016]. Par conséquent, nous
pouvons concentrer notre étude sur les paramètres uniquement.

11
Après une étude des di↵érentes méthodes d’analyse de sensibilité (voir l’annexe A.2.5), une analyse
de sensibilité globale (GSA) avec des indices de Sobol estimés par la méthode [Saltelli, 2002] a été
choisie en fonction des caractéristiques de notre modèle : présence de non-linéarités, petit nombre
d’entrées, et coûts de calcul modestes.

Inspiré par les travaux du chimiste Cukier, I.M. Sobol’ a en e↵et proposé une analyse de sensibilité
globale basée sur la décomposition de la variance par le biais d’un échantillonnage de Monte-Carlo.
Étant donné le modèle suivant Y = f (X), cette analyse de la variance (ANOVA) vise à définir la
part de variabilité de la sortie due à ses entrées, ce que l’on appelle les “indices de Sobol’ ”.

Dans le cadre de notre étude, nous nous intéressons à une hiérarchisation des facteurs, afin de
qualifier l’influence relative de chaque facteur d’entrée sur la sortie. Le principe d’application de
l’ANOVA avec indices de Sobol est décrit ci-dessous, mais l’estimation de ces indices reste difficile
(en particulier pour l’e↵et de Sobol de second ordre, représentant les e↵ets conjoints entre les sorties)
et est développée en annexe A.2.6. Plusieurs études fructueuses ont déjà été menées en économie
sur les modèles DSGE (par exemple, [Ratto, 2008]).

Dans son article fondateur [Sobol, 1967], Sobol a proposé d’exprimer f 2 L1 ([0, 1]p , dx) en sommes
de dimension croissante:
p
X p X
X p
f (x1 , ..., xp ) = f0 + fi (xi ) + fi,j (xi , xj ) + ... + f1,...,p (x1 , ..., xp ). (1)
i=1 i=1 j>i

En imposant que 8s 2 {1, ..., p}, 8 i1 < i2 < ... < is 2 {1, ..., p}s et 8 k 2 {i1 , ..., is },
Z 1
fi1 ,...,is (xi1 , ..., xis )dxk = 0
0
ce développement de f en 2n sommes de di↵érentes dimensions existe et est unique, et les termes
associés sont orthogonaux deux à deux dans L1 ([0, 1]p , dx).
Supposons maintenant que l’on puisse exprimer une variable aléatoire de carré intégrable, Y , sous
la forme Y = f (X) où X = (X1 , ..., Xp ) est un vecteur de p variables aléatoires indépendantes
uniformément distribuées sur [0, 1], nous avons 18

p
X p X
X p
Y = f0 + fi (Xi ) + fi,j (Xi , Xj ) + ... + f1,...,p (X1 , ..., Xp ). (2)
i=1 i=1 j>i

Par conséquent, en utilisant l’orthogonalité par paire dans la décomposition de Sobol’, nous pouvons
obtenir facilement une expression récursive de chaque terme en utilisant les espérances condition-
18
Ici l’indépendance des variables aléatoires (X1 , ..., Xp ) est une hypothèse forte nécessaire pour obtenir la
décomposition de la variance sous-jacente alors que nous pouvons évidemment relâcher l’hypothèse sur la distri-
bution en nous rappelant que pour une variable aléatoire réelle Z, FZ (U ) et Z sont équidistribués lorsque U est
une variable aléatoire uniforme sur [0, 1] et FZ le pseudo-inverse de la fonction de distribution de Z. Dans le cas de
variables dépendantes, on peut s’opposer à la difficile généralisation des expressions récursives ci-dessus de chaque
terme en utilisant les espérances conditionnelles. Voir la note de bas de page 23 pour plus d’explications.

12
nelles
f0 = E(Y )
fi (Xi ) = E(Y |Xi ) E(Y )
fi,j (Xi , Xj ) = E(Y |Xi , Xj ) fi (Xi ) fj (Xj ) E(Y ) (3)
..
.
et déduire la décomposition dite ANOVA-HDMR pour la variance de Y .
p
X p X
X p
V (Y ) = Vi + Vi,j + ... + V1,...,p
i=1 i=1 j>i

m
p p p (4)
X Vi X X Vi,j V1,...,p
1= + + ... +
V (Y ) V (Y ) V (Y )
i=1 i=1 j>i

où

Vi = V (E(Y |Xi ))
Vi,j = V (fi,j (Xi , Xj )) = V [E(Y |Xi , Xj )] V [E(Y |Xi )] V [E(Y |Xj )] (5)
..
.
Maintenant nous pouvons définir les indices de Sobol : 8 s 2 {1, ..., p}, 8 i1 < i2 < ... < is 2
{1, ..., p}s .

Vi1 ,...,is
Si1 ,...,is = (6)
V (Y )

où Si est une mesure de l’impact de Xi sur la variance totale, Si,j est une mesure de l’impact
conjoint de (Xi , Xj ) sur la variance totale, etc....

Dans cet esprit, pour mesurer l’impact total (somme des e↵ets uniques et des e↵ets conjoints liés)
de Xi sur la variance de Y nous définissons19

E(V (Y |Xĩ )) V (E(Y |Xĩ ))

ST i = =1 (7)
V (Y ) V (Y )

où Xĩ = (X1 , ..., Xi 1 , Xi+1 , ..., Xp ). Par exemple pour un modèle à 3 entrées, nous avons

ST1 = S1 + S1,2 + S1,3 + S1,2,3 = 1 S2,3 . (8)

19
Les propriétés des indices de Sobol’ sont bien connues dans la littérature, et peuvent être résumées comme suit :
Pp P
1. Si1 ,...,is = 1
s=1 1i1 <i2 <...<is p

2. 8i 2 {1, ..., p}, Si , STi 2 [0, 1].

Ces deux propriétés correspondent à l’interprétation des indices de Sobol’ comme contributions relatives à la variance
de la sortie V (Y ).

13
 Le processus pour estimer les indices est basé sur la méthode [Saltelli, 2002] : deux matrices
d’échantillonnage initiales sont générées par des Latin Hypercube Sampling (LHS). Une revue de
la littérature montre que des GSA ont également été réalisées sur des modèles DSGE, RBC et le
modèle DICE [Nordhaus, 2008].

Avec la méthode des indices de Sobol’ estimés selon [Saltelli, 2002], la sensibilité d’une sortie
du modèle [Grasselli and Costa Lima, 2012] à ses entrées est plus équilibrée que celle d’un modèle
DSGE, RBC ou IAM (DICE), dépendant principalement d’un unique paramètre souvent très difficile
à estimer. Notre étude montre aussi qu’il est possible d’améliorer légèrement la robustesse du
modèle en endogénéisant les paramètres principalement responsables de cette sensibilité. Ainsi, le
modèle [Grasselli and Costa Lima, 2012] avec une fonction de production CES est légèrement plus
robuste qu’avec une fonction de production Leontief.

Nous avons également étudié le modèle original [Goodwin, 1967] (voir annexe A.5) et nous avons
remarqué que la sensibilité du modèle original dépend des mêmes paramètres que son extension
et que ses performances en termes de robustesse sont légèrement meilleures que celles du modèle
[Grasselli and Costa Lima, 2012], ce qui pourrait être dû au plus petit nombre de non-linéarités.
Ces résultats sont assez rassurants quant au fait que l’extension du modèle ne semble pas a↵ecter
la nature profonde de sa sensibilité. Ils nous amènent à nous interroger sur un éventuel compromis
entre la robustesse et le réalisme d’un modèle pour trouver un équilibre entre un modèle très robuste
incapable de s’approcher de la réalité (tel que le modèle [Goodwin, 1967]), et un modèle très fin
et très sensible (comme le DSGE de [Ratto, 2008]). De ce point de vue, le modèle [Grasselli and
Costa Lima, 2012] semble être un compromis intéressant, avec une description plus fine de l’économie
et une sensibilité équilibrée entre plusieurs entrées.

Étendre un modèle macro-économique existant, [Grasselli and Costa Lima, 2012], en

introduisant les inventaires et la vitesse de la monnaie

Jusqu’à présent, l’analyse de la monnaie se heurte à un certain nombre de paradoxes et d’impossibilités

qui a↵ecte notre compréhension de l’interaction entre la monnaie, sa vélocité et la macro-dynamique.
Dans le chapitre 3, nous proposons une dynamique stock-flux cohérente avec concurrence imparfaite
où ces paradoxes peuvent être levés.

Une version préliminaire, sans monnaie, de notre dynamique non linéaire et stock-flux cohérente
est réalisée dans un premier temps. En plus de fixer la notation pour le modèle complet, nous
nous écartons de la littérature précédente en nous concentrant sur une économie avec une fonction
de production ces, une tarification oligopolistique et une dynamique des inventaires. Ce modèle
présente quatre équilibres de long terme, économiquement significatifs, et localement stables pour
cette dynamique, dont l’un est caractérisé par un ratio d’endettement fini et un emploi strictement
positif — l’équilibre solovien — et un autre par un endettement infini et un emploi nul.

Dans un second temps, nous introduisons la monnaie à travers la célèbre ‘équation d’échange”, où
T est le volume des transactions. Dans ce cadre, s’il y a une o↵re excédentaire, Y Yd , le volume
e↵ectif de transaction est Yd . Si, au contraire, la demande excède l’o↵re, Y  Yd , les entreprises
vendront des stocks pour répondre à la demande, et par conséquent, T est toujours égal à Yd , tant
que le stock d’inventaire reste non négatif. Si ce stock devient nul, le secteur de la production ne

14
peut plus répondre à la demande chaque fois que Y  Yd : un certain rationnement doit avoir
lieu et le volume de transaction doit être égal à la production : T = Y . Nous laissons cette
dernière situation pour une étude plus approfondie,20 et nous concentrerons donc notre analyse sur
les solutions “intérieures” du système di↵érentiel représentant le modèle avec monnaie où V 0
tout au long du processus. Par conséquent, l’équation d’échange se lit comme suit

M v = pYd , (9)

où M doit être compris comme la somme des comptes bancaires détenus par des organismes non
gouvernementaux (ménages et entreprises). Il s’agit d’un stock endogène de monnaie intérieure,
généré en reconnaissance de dette.21 La variable, v, définie par (9), désigne la vitesse de la monnaie,
à savoir la vitesse à laquelle les transactions réelles se produisent dans l’économie.

Mais d’où vient cet argent ? Dans notre économie de crédit, il ne peut être créé que par un
crédit bancaire supplémentaire. Inversement, lorsqu’une dette est remboursée (que ce soit une dette
des firmes, Df , ou des ménages, Dh ), la quantité de monnaie correspondante retourne au bilan du
système bancaire, et est donc ‘détruite”. En conséquence,

Ṁ := D˙f + Ḋh . (10)

Remarquez que, chaque fois que D˙f + Ḋh = 0, aucune monnaie supplémentaire ne circule dans
l’économie : l’un des deux secteurs prête à l’autre l’incrément de dette nécessaire. Avec les équations
dynamiques de la dette, l’équation (10) conduit à la dynamique endogène suivante de la monnaie :

Ṁ = cV̇ + p K. (11)

où V est le volume d’inventaires, valorisé au coût unitaire de production, c tandis que est le taux
de dépréciation du capital, K, en termes réels, valorisé au prix p.

Malgré sa simplicité, l’équation (11) fournit quelques informations utiles. En e↵et, si l’on mainte-
nait la ‘loi” de Say et que l’on négligeait le coût irrécupérable induit par la dépréciation du capital,
la quantité de monnaie M serait nécessairement constante, et deviendrait donc, là encore, non per-
tinente. On déduirait également de (11) que Ḋf = Ḋh . Le fameux problème de la poule et de
l’œuf — qui cause l’autre, l’épargne ou le crédit ? — devient à nouveau inextricable. Cela signifie
que, pour résoudre cette énigme et pouvoir observer la création endogène de crédit dans un modèle
macro-économique, l’abandon de la “loi” de Say est une condition nécessaire.

L’introduction de la monnaie permet de résoudre les premiers paradoxes auxquels nous venons de
faire allusion : il n’est pas nécessaire de s’appuyer sur l’hypothèse d’anticipations rationnelles pour
que l’économie atteigne un équilibre de long terme avec un taux d’inflation stable. Nous montrons
aussi que la monnaie, telle qu’elle est ici introduite, s’avère ne pas être neutre à la fois à court
et à long terme. Sa création est ici déterminée de manière endogène par le crédit bancaire. La
cohérence stock-flux permet de suivre la création monétaire et son déplacement le long du circuit
de l’économie agrégée.
20
Considérant que ce type de “problème limite” conduit à une dynamique non di↵érentiable, qui sera étudiée dans
un travail ultérieur.
21
C’est-à-dire que nous négligeons ici la monnaie extérieure qui serait libre de dette, voir [Gurley and Shaw, 1960].

15
 Dans ce modèle, l’évolution de sa vitesse de circulation peut être observée. En particulier, la baisse
de la vitesse de circulation de la monnaie observée dans plusieurs pays à la suite d’une crise financière
peut être analysée. Notre modèle apporte aussi un éclairage nouveau sur les liens entre la vitesse
de circulation de la monnaie et la trajectoire de la dette. La faible inflation, une faible croissance
du revenu réel, une augmentation de la dette privée et une diminution de la vélocité des échanges
s’avèrent être les caractéristiques des trajectoires macro-économiques menant à un endettement de
longue durée. Nous fournissons les conditions pour lesquelles, cependant, une politique monétaire
à faible taux d’intérêt peut stimuler l’activité économique afin d’échapper à la trappe à liquidité.

Par exemple l’intervention de la banque centrale peut avoir un e↵et positif évident pour prévenir
une crise caractérisée par l’e↵ondrement des taux d’emploi. La politique monétaire prévient ce
résultat en conditionnant le taux d’intérêt nominal à court terme au ratio dette privée/résultat : plus
ce dernier est élevé, plus le coût de l’argent doit être faible. Les résultats en matière de persistance
sont beaucoup plus forts que, par exemple, l’éventuelle instabilité locale et asymptotique d’états
stationnaires indésirables pour des valeurs de paramètres typiques du modèle : l’intervention de la
banque centrale, sous la forme d’une politique monétaire suffisamment réactive, empêche l’économie
de rester en permanence à des niveaux d’emploi arbitrairement bas, quelles que soient les conditions
initiales du système. Pour faire écho à [Costa Lima and Grasselli, 2014], il se peut que la stabilisation
d’une économie monétaire instable soit un ordre trop élevé pour la banque centrale, mais que la
déstabilisation d’une crise stable de dette-déflation soit possible.

Il y a, bien sûr, un certain nombre d’aspects institutionnels que nous avons négligés dans ce
chapitre : réserves obligatoires, exigences en matière d’adéquation des fonds propres (comme le
cadre de Bâle iii/crr), réglementation prudentielle... [Bovari et al., 2020] ont introduit quelques
raffinements dans la modélisation du système bancaire. dans une économie “réelle”. Ajoutés au
cadre actuel, ces raffinements pourraient nous permettre d’aborder des questions importantes autour
d’une réglementation des banques basée sur l’e↵et de levier contracyclique.

Peut-être est-il tout aussi important de considérer les économies ouvertes et l’interaction entre la
création de crédit non neutre à partir de rien, les taux de change et les emprunts internationaux. La
thèse principale défendue par [Werner, 2014b] devrait alors être examinée afin de vérifier dans quelles
conditions les pays en développement auraient dû s’appuyer sur la création monétaire nationale
plutôt que d’emprunter de l’argent à l’étranger, au risque de se retrouver “englués dans une spirale
de dettes en devises étrangères, alors qu’en réalité les banques étrangères ont simplement créé
l’argent à partir de rien, ce que le pays en développement aurait pu arranger par le biais de leurs
propres banques nationales” ( [Werner, 2016]). Comme l’indique clairement cet auteur, le fait de
considérer la création monétaire endogène “a également des implications sur la question de savoir
qui doit payer pour le sauvetage des banques, en déplaçant le pendule des contribuables vers les
banquiers centraux”. Nous laissons aux recherches futures le soin d’étudier ces questions fascinantes.

Évaluer des leviers de politiques publiques pour garantir une transition entre di↵érentes
natures de capital dans ce modèle.

Après avoir étudié la robustesse du modèle et sa capacité à intégrer la vitesse de circulation de

la monnaie et à relâcher la loi de Say, l’objectif de ce chapitre 4 est de proposer un modèle simple
de transition énergétique des combustibles fossiles aux énergies renouvelables.

16
 Nous considérons donc deux types de capital, un capital “brun” (ou sale) qui a besoin d’énergie
fossile pour fonctionner, et un capital ‘vert” (ou propre), qui est moins productif mais ne dépend pas
de l’énergie fossile. Le prix de l’énergie fossile doit être compris comme le prix relatif des énergies
fossiles et renouvelables. L’investissement privé est orienté vers le type de capital le plus rentable, en
fonction du prix (exogène) de l’énergie fossile. Les deux types de capital sont de type “putty-clay”
( [Johansen, 1959]) par rapport à l’énergie : une fois qu’un bien d’équipement est créé, son rapport
de consommation de capital à l’énergie est fixé une fois pour toutes, et les deux types de capital
ne sont pas substituables. Cette structure de production s’inscrit dans la même macrodynamique
proie-prédateur que précédemment, stock-flux cohérente [Goodwin, 1967] où l’investissement privé
peut être financé par la dette [Grasselli and Costa Lima, 2012]. La transition vers les énergies
renouvelables a lieu si l’économie fonctionne asymptotiquement avec du capital vert uniquement.
Deux facteurs principaux sont mis en évidence dans ce chapitre 4 : le ratio production/capital
du capital brun par rapport au capital vert et le prix relatif des énergies fossiles par rapport aux
énergies renouvelables.

Dans ce travail, l’investissement est orienté vers le type de capital le plus rentable, en fonction
de sa productivité et du prix relatif de l’énergie fossile, mais ne s’ajuste pas instantanément (bien
que plus l’écart entre les taux de profit est élevé, plus l’ajustement est rapide). Cette inertie de
l’investissement peut être considérée comme reflétant les e↵ets de complémentarité entre les deux
types de capital (qui sont par ailleurs de parfaits substituts en tant que facteurs de production),
la spécialisation de la main-d’œuvre ou les externalités liées aux infrastructures. Elle s’écarte du
passage instantané du capital brun au capital vert, souvent postulé dans la littérature.22

Suite à [Grasselli and Nguyen-Huu, 2018], et comme dans le chapitre précédent, la loi de Say est
abandonnée ici et la politique publique est introduite par le biais des dépenses publiques et des taxes
d’une manière similaire à [Costa Lima and Grasselli, 2014]. L’interaction entre la dynamique des
inventaires et la politique publique conduit à une dynamique similaire à celle bidimensionnelle de
type Aubry-Mather sur le tore avec un petit cycle (lié à la dynamique proie-prédateur sous-jacente
entre l’emploi et les salaires) entrelacé avec un plus grand cycle induit par la réponse publique
contracyclique. Avec la dynamique de la dette privée, de la part des salaires et du sous-emploi, cela
conduit à un système dynamique non linéaire à 10 dimensions. Nous montrons que ce système peut
être réduit à un système à 5 dimensions. Ce système n’est cependant pas autonome car il dépend du
prix de l’énergie fossile, donné de manière exogène, qui détermine l’allocation des investissements.

Pour cette étude, nous considérons donc une économie fermée à quatre secteurs comprenant des
ménages, un secteur bancaire, un secteur public et des entreprises privées produisant un seul bien
de consommation à capital brun ou vert. Par souci de simplicité, nous nous limitons au cas extrême
de deux types de capital : le capital brun d’une part, qui a besoin d’un flux constant d’énergie
fossile pour produire ; et le capital vert d’autre part, qui fonctionne sans énergie fossile mais est
moins productif que le capital sale.

La production, Y, peut être produite de deux manières : soit par rapport au stock, K1 , de capital
brun avec un flux, E, d’énergie fossile et de travail, L1 , soit en exploitant un stock de capital propre,
22
Voir, par exemple, le modèle DICE de Nordhaus [Nordhaus, 2014] ; pour un modèle de changement climatique
partageant notre cadre théorique, voir [Bovari et al., 2018].

17
K2 , avec un flux, L2 , de travail seul.23 Pour chaque type k = 1, 2, le ratio capital/production ⌫k est
constant. Les biens d’équipement installés, quel que soit leur type, ne peuvent pas être convertis
les uns dans les autres : si les investisseurs veulent passer d’un type à un autre, ils doivent investir
dans de nouveaux biens d’équipement. En ce qui concerne la production de biens, les deux types
de biens, Y1 et Y2 , sont de parfaits substituts :
K1 E K2
Y = Y1 + Y2 avec Y1 = = = a1 L1 et Y2 = = a2 L2 , (12)
⌫1 ⌫1 ⌫2
où ⌫1 , ⌫2 , a1 , a2 > 0 avec ak le taux de productivité du travail pour chaque type de capital k = 1, 2. Le
plein emploi du capital est supposé tout au long de l’équation. Toutes les quantités sont exprimées en
termes réels et sont des quantités nettes, ce qui signifie que les recettes et les dépenses intermédiaires
sont déduites de la production annuelle finale.

On montre que le système n’admet que deux états stationnaires correspond respectivement à
un équilibres fonctionnant et investissant uniquement en capital vert et un autre avec uniquement
du capital brun. On peut montrer que, sous des conditions tout à fait raisonnables, détaillées
dans l’annexe C.1.2, dans chaque cas, huit points d’équilibre émergent. Par conséquent, le modèle
présente 16 états stables à long terme.

Comme dans la littérature consacrée à ce type de macrodynamique (voir [Grasselli and Costa Lima,
2012] et [Giraud and Grasselli, 2019]), nous montrons que, indépendamment de la nature du capital
prévalant à long terme, il existe un état stationnaire souhaitable correspondant à une part de salaire
et un taux d’emploi positifs, et un niveau fini de ratios d’endettement, de subventions et de taxes.
Suivant une remarque de [Giraud and Grasselli, 2019], c’est l’analogue d’un équilibre solovien parce
que le taux de croissance réel, à cet état d’équilibre, est égal à la croissance de la productivité du
travail (progrès technique) plus la croissance de la population. Si nous utilisions une fonction de
production ces, le modèle de Solow deviendrait un cas particulier du présent modèle.

Le deuxième type d’équilibre à long terme est un état stationnaire déflationniste associé à une
montée en flèche des ratios d’endettement tandis que les salaires et l’emploi tendent vers zéro.
Sous des spécifications raisonnables de la courbe de Phillips à court terme, ainsi que des fonctions
d’investissement et de consommation, nous montrons que ces équilibres déflationnistes par la dette,
qu’ils soient sales ou propres, sont localement instables. Ceci contraste fortement avec la littérature
déjà mentionnée. La force qui déstabilise les équilibres déflationnistes s’avère être l’introduction
de l’énergie dans le processus de production et l’inertie avec laquelle l’investissement agrégé passe
d’un type de capital à un autre. Cela soulève une question intéressante pour les recherches futures
: quelles sont les forces qui poussent une économie vers la déflation par la dette dans un monde
où plusieurs types de capitaux coexistent et où l’investissement présente une certaine inertie ?
[Costa Lima and Grasselli, 2014] avaient fourni un mécanisme complètement di↵érent qui déstabilise
la crise dette-déflation, basé sur l’intervention publique.

Deux autres équilibres correspondent à des états déflationnistes avec un niveau de dette fini :
l’un avec un salaire nul, l’autre avec un salaire positif. Leur stabilité asymptotique locale ne peut
être jugée a priori et doit être vérifiée au cas par cas. Il en va de même pour l’équilibre dit de
23
Pour simplifier les notations, nous n’introduisons pas explicitement les énergies renouvelables, qui pourraient être
utilisées par les deux types de capital. un tel ajout pourrait facilement être fait sans améliorer significativement nos
résultats, ni leur clarté.

18
l’esclavage (sans salaire mais avec emploi). Tous, cependant, sont structurellement instables et
donc non pertinents dans notre étude.

Le principal moteur de la transition vers la durabilité de l’économie s’avère être la di↵érence entre
le prix relatif de l’énergie fossile et l’écart de productivité entre les deux types de capital. Si le prix
de l’énergie fossile reste trop faible par rapport à l’écart de productivité entre les deux types de
capital, l’économie fonctionne asymptotiquement exclusivement avec du capital brun. Le coût de
l’énergie fossile est compensé à long terme par une diminution de la part des salaires, grâce à la
productivité plus élevée du capital. Inversement, lorsque le prix de l’énergie fossile est suffisamment
dissuasif, l’investissement se concentre finalement sur le capital propre. Des simulations numériques
montrent les e↵ets de changements dans les paramètres et dans le prix relatif de l’énergie fossile.

Afin de satisfaire ce critère de transition que notre étude a mis en lumière, E ⌫ > 0, nous
pouvons introduire une taxe carbone, c (selon, par exemple, le corridor de valeurs établi par [Stiglitz
et al., 2017]), afin d’augmenter le prix des énergies fossiles, E ou réduire le ratio de capital vert
par rapport à la production, ⌫2 . En e↵et, comme ⌫ = ⌫11 1
⌫2 , le critère de transition sera plus
facilement satisfait. La première option risque de réduire les profits privés et donc de favoriser
l’endettement privé. La deuxième option présente le risque d’augmenter la dette publique. Nous
identifions les conditions sous lesquelles la transition vers le capital vert se fait sans conduire à un
surendettement, qu’il soit privé ou public.

Bien que ces deux outils politiques aient été largement préconisés dans la littérature, leur effi-
cacité à long terme n’a pas été clairement démontrée jusqu’à présent. Lorsque seul l’outil de la taxe
carbone est utilisé, nous montrons que les critères de transition et de stabilité locale sont identiques
: E+ c ⌫ > 0. Lorsqu’elle est suffisamment élevée, une taxe sur le carbone incite fortement
à passer des combustibles fossiles aux énergies renouvelables, mais au prix d’une pénalisation des
entreprises. L’état stationnaire vert de type solovien est localement asymptotiquement stable, et
si l’économie converge, elle doit devenir asymptotiquement verte car les équilibres sales s’avèrent
asymptotiquement instables. Inversement, si la taxe carbone reste trop faible, l’économie ne peut
s’empêcher, si elle converge, de le faire vers l’un des nombreux équilibres à forte intensité de com-
bustibles fossiles, car les états d’équilibre verts deviennent instables.

La deuxième option est réalisée en introduisant des subventions publiques à l’innovation verte,
dans l’espoir de pouvoir garantir que la transition aura lieu, quel que soit le point de départ de
l’économie. Inspirés par l’équation classique d’accumulation du capital, nous proposons une dy-
namique similaire pour le stock de connaissances qui a un impact sur la productivité du travail et
donc sur le ratio capital/production, ⌫2 , du capital vert, qui influence à son tour la productivité du
travail associée au capital vert.

Sous les mêmes conditions raisonnables que précédemment, le modèle présente les mêmes types
d’états stationnaires - qu’ils soient bruns ou verts. Le principal écart par rapport à la section
précédente provient toutefois de l’introduction de la variable ⌫2 : de nouveaux équilibres appa-
raissent, moins manichéens que précédemment, puisqu’ils sont compatibles avec la coexistence du
capital vert et du capital brun à long terme. Mais ils soulèvent, bien sûr, de nouvelles questions
: dans quelles conditions l’économie convergera-t-elle vers un état durable à zéro carbone, par
opposition à une situation à faible carbone où, même à long terme, le capital sale continuerait à
fonctionner, bien que de manière modérée ?

19
 Pour chacune des variantes propres de ces équilibres, nous avons toujours des équilibres localement
stables sous les hypothèses standard : la situation souhaitable avec des salaires et des emplois posi-
tifs, et un niveau fini de ratios d’endettement, des subventions et des taxes. Comme précédemment,
l’état d’équilibre déflationniste associé à une montée en flèche des ratios d’endettement alors que les
salaires et les emplois se réduisent à zéro est localement instable. Ce qui est nouveau, cependant,
c’est que, le régime permanent de type solovien brun s’avère également localement instable sous
réserve que le rapport du stock de connaissance sur le capital vert est tel que sa dérivée par rapport
au temps reste bornée supérieurement chaque fois que le volume de capital vert tend vers zéro. Cette
restriction revient à dire que, chaque fois que le capital propre se réduit à zéro (en raison du manque
d’investissements verts et de la dépréciation exponentielle), le savoir-faire écologique encapsulé dans
R se réduit encore plus rapidement. Cela peut être réalisé, par exemple, en imposant que le taux
de décroissance des connaissances soit supérieur au taux de dépréciation du capital vert.

Dans ce chapitre, nous avons aussi présenté une paramétrisation particulière conduisant à une
transition avec un faible prix de l’énergie fossile. Dans cet exemple, le financement public et privé
de la R&D verte permet d’assurer la transition, même si l’économie suivait initialement une voie
carbonée. L’introduction d’une taxe carbone accélère la transition mais ne la garantit pas. Comme
dans [Acemoglu et al., 2012], tout retard dans la mise en œuvre des politiques publiques de transition
peut être très coûteux en temps et en argent. Cependant, contrairement au résultat obtenu dans
cet article, ici, aucune transition ne peut se produire en cas de laisser-faire.

De plus, nous avons montré qu’en général, de nouveaux types d’états stables à long terme
émergent, où le capital brun et le capital vert coexistaient. Dans notre exemple spécifique, cela
est dû au faible niveau d’investissements publics et privés dans la R&D verte. Bien que moins
simplistes que nos précédents équilibres manichéens, ces états stables hybrides suggèrent que, non
seulement le laissez-faire est insuffisant, mais qu’une intervention publique trop timide peut échouer
à conduire l’économie vers une situation de neutralité carbone à long terme.

Plusieurs améliorations de cet article peuvent être envisagées. La dynamique des prix devrait
évidemment intégrer le coût de l’énergie et être au moins partiellement orientée par la demande.
Cela permettrait aux prix de la consommation de refléter l’augmentation du coût des combustibles
fossiles chaque fois qu’une taxe sur le carbone est mise en pratique. L’impact sur la demande
e↵ective serait vraisemblablement non négligeable. La principale question serait alors de savoir si
une demande déprimée serait capable d’empêcher le secteur privé d’investir suffisamment dans les
infrastructures vertes pour atteindre un équilibre propre de type solovien. La même question serait
probablement renforcée si les ménages consommaient également de l’énergie (et pas seulement le
secteur de la production) et supportaient donc directement une partie du coût supplémentaire induit
par une taxe carbone.

D’autre part, nous nous sommes exclusivement concentrés, ici, sur l’analyse à long terme. Une
analyse à court terme permettrait également de comprendre les avantages et les inconvénients des
leviers politiques considérés dans ce papier et qui ont été très débattus aux États-Unis dans la
première partie du mandat Biden. En particulier, l’instabilité asymptotique de l’état stationnaire
de type solovien “sale” devrait être mise en perspective avec l’urgence temporelle induite par le
changement climatique.

20
 Une autre amélioration de ce papier est laissée pour un travail ultérieur, qui consiste à endogénéiser
le prix, E , de l’énergie fossile. Enfin, en se basant sur le présent travail, on pourrait introduire
une variété de di↵érents types de capitaux dans un secteur de production “putty-clay” similaire à
celui introduit par [Akerlof and Stiglitz, 1969] afin d’obtenir un cadre plus réaliste que le cas à deux
capitaux envisagé ici.

Pour conclure, ces cinq années de doctorat ont permis de tester la robustesse d’un modèle de
déséquilibre, capable de simuler un moment de Minsky. Les extensions de ce modèle ne semblant
pas fondamentalement altérer sa robustesse, des améliorations lui ont été apportées pour répondre
aux principales critiques faites à cette famille de modèles. Ces améliorations concernent notamment
le relâchement de la “loi” de Say par l’introduction des inventaires, qui se révèlent indispensable
pour pouvoir étudier la création monétaire. La vélocité de la monnaie et la création monétaire
par les banques peut être suivies dans cette version étendue du modèle d’origine [Grasselli and
Costa Lima, 2012]. Enfin, notre travail a consisté à construire un modèle de transition simple, avec
seulement deux types de capitaux, basé sur [Grasselli and Costa Lima, 2012] et à en étudier les
multiples équilibres et les leviers possibles pour garantir la transition de l’un vers l’autre capital.

21
Contents

1 Introduction 27
1.1 “The world is on an unsustainable path” [BP, 2019] . . . . . . . . . . . . . . . . . . . 28
1.1.1 A long-time known trajectory . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.1.2 A “climate-driven Minsky moment” [Carney, 2019] . . . . . . . . . . . . . . . 30
1.2 The ambiguous role of models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.2.1 Uncertainty: the very essence of a model . . . . . . . . . . . . . . . . . . . . . 32
1.2.2 The sensitivity analysis to build reliable models . . . . . . . . . . . . . . . . . 34
1.3 The crisis of macro-economic models . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
1.3.1 An historical typology of macro-economic models . . . . . . . . . . . . . . . . 37
1.3.2 A largely documented diagnostic of a crisis . . . . . . . . . . . . . . . . . . . 39
1.3.3 Macro-economic modeling: map of the problematic . . . . . . . . . . . . . . . 42
1.3.4 A mix of science and ethical values in Climate Change Economics . . . . . . 45
1.4 Research questions and executive summary . . . . . . . . . . . . . . . . . . . . . . . 47
1.4.1 Research questions and axes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
1.4.2 Thesis organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

2 Building a global sensitivity analysis to quantify the robustness of macro-economic

models 51
2.1 General structure of the [Grasselli and Costa Lima, 2012] model . . . . . . . . . . . 53
2.1.1 Some classical assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.1.2 Stock-flow consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.1.3 Aggregate behavioural rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.2 Motivation of our uncertainties analysis (UA) and sensitivity analysis (SA) methods 56
2.2.1 Definition and selection of uncertainty and sensitivity analysis . . . . . . . . 57
2.2.2 A global sensitivity analysis with Sobol’ indices . . . . . . . . . . . . . . . . . 59
2.2.3 Construction of the Sobol’ indices . . . . . . . . . . . . . . . . . . . . . . . . 60
2.3 Global sensitivity analysis of the [Grasselli and Costa Lima, 2012] model . . . . . . . 63
2.3.1 Identification of the inputs/outputs of our study . . . . . . . . . . . . . . . . 63
2.3.2 Sampling method: Latin Hypercube . . . . . . . . . . . . . . . . . . . . . . . 64
2.3.3 Sobol’ indices computation for the [Grasselli and Costa Lima, 2012] inputs . 65
2.3.4 Convergence of the Sobol’ indices of the model . . . . . . . . . . . . . . . . . 68
2.4 Comparison with other macro-economic models . . . . . . . . . . . . . . . . . . . . . 71
2.4.1 Review of the [Ratto, 2008] GSA about the [Lubik and Schorfheide, 2007] model 71
2.4.2 Review of the [Harenberg et al., 2017] GSA about a canonical RBC model . . 73
2.4.3 Review of the [Miftakhova, 2019], [Butler et al., 2014] and [Anderson et al.,
2014] GSA about DICE [Nordhaus, 2008] . . . . . . . . . . . . . . . . . . . . 74
2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

22
CONTENTS

3 Money Velocity in an Imperfectly Competitive, Stock-Flow Consistent Dynam-

ics 79
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.1.1 Paradoxes associated with money . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.1.2 Monetary policy and persistence . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.2 The moneyless economy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.2.1 Stock-flow consistent inventories . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.2.2 Expected profits and debts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.2.3 Some accounting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
3.2.4 Aggregate behavioural rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3.2.5 Oligopoly pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
3.3 Long-run “real” perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
3.3.1 The moneyless vector field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
3.3.2 Long-term real steady states . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.3.3 Numerical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.4 Introducing money . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
3.4.1 The non-neutral dynamics of money . . . . . . . . . . . . . . . . . . . . . . . 96
3.4.2 Endogeneous money creation . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
3.5 Long-run non-neutrality and monetary policy . . . . . . . . . . . . . . . . . . . . . . 99
3.5.1 Monetary steady states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3.5.2 Monetary persistence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

4 Directed Technical Change in an inventory Stock-flow Consistent Dynamics with

Government: from dirty to clean Capital 105
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.2 Inventory dynamics with government intervention . . . . . . . . . . . . . . . . . . . 107
4.2.1 Dirty vs clean capital . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.2.2 Expected profits, taxes and subsidies . . . . . . . . . . . . . . . . . . . . . . . 108
4.2.3 Some accounting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
4.2.4 Aggregate behavioural rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.3 Long-run stability study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4.4 Introducing government levers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
4.4.1 Introduction of a carbon tax . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
4.4.2 Innovation and directed technical change . . . . . . . . . . . . . . . . . . . . 122
4.5 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

5 Conclusion 132

A Building a global sensitivity analysis to quantify the robustness of macro-economic

models 142
A.1 Overview of macro-economic modeling: Features and criticisms of the [Grasselli and
Costa Lima, 2012] model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
A.1.1 Features and criticisms of the [Grasselli and Costa Lima, 2012] model . . . . 143
A.2 Some reminder on sensitivity analysis (SA) . . . . . . . . . . . . . . . . . . . . . . . 154
A.2.1 Uncertainties Analysis (UA): exploring the space of the input factors . . . . . 154
A.2.2 Sensitivity Analysis (SA): from uncertainties in the input to uncertainty in
the output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

23
CONTENTS

A.2.3 Screening Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

A.2.4 Local Sensitivity Analysis (LSA) . . . . . . . . . . . . . . . . . . . . . . . . . 160
A.2.5 Global Sensitivity Analysis (GSA) . . . . . . . . . . . . . . . . . . . . . . . . 163
A.2.6 Estimation of the indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
A.2.7 Synthesis and selection of the GSA: Sobol’ indices estimated by Mc Kay’s
method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
A.3 Second-order Sobol’ indices of the [Grasselli and Costa Lima, 2012] model with 100
simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
A.4 Convergence of the Sobol’ indices of the [Grasselli and Costa Lima, 2012] model with
a Leontief production function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
A.5 Sobol’ indices of the [Goodwin, 1967] model . . . . . . . . . . . . . . . . . . . . . . . 175
A.5.1 Sobol’ indices of the [Grasselli and Costa Lima, 2012] model with CES pro-
duction function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176

B Money Velocity in an Imperfectly Competitive, Stock-Flow Consistent Dynam-

ics 181
B.1 Long-run properties of the moneyless model . . . . . . . . . . . . . . . . . . . . . . . 181
B.1.1 The reduced dynamical system . . . . . . . . . . . . . . . . . . . . . . . . . . 181
B.1.2 Existence and stability of long-run equilibria . . . . . . . . . . . . . . . . . . 182
B.2 The monetary dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
B.2.1 The reduced monetary dynamical system . . . . . . . . . . . . . . . . . . . . 188
B.2.2 Long-run equilibria with money . . . . . . . . . . . . . . . . . . . . . . . . . . 189
B.3 Normalization of the Philips’ Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
B.3.1 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197

C Directed Technical Change in an inventory Stock-flow Consistent Dynamics with

Government: from dirty to clean Capital 200
C.1 Stability of equilibria in the transition model with government . . . . . . . . . . . . 200
C.1.1 The reduced dynamical system . . . . . . . . . . . . . . . . . . . . . . . . . . 200
C.1.2 Existence of long-run equilibria . . . . . . . . . . . . . . . . . . . . . . . . . . 203
C.1.3 Analysis of the di↵erent equilibria . . . . . . . . . . . . . . . . . . . . . . . . 204
C.1.4 Stability study of equilibria with brown or green capital . . . . . . . . . . . . 205
C.2 Stability of equilibria with subsidies and carbon tax . . . . . . . . . . . . . . . . . . 214
C.2.1 The reduced dynamical system . . . . . . . . . . . . . . . . . . . . . . . . . . 214
C.2.2 Analysis of the di↵erent equilibria . . . . . . . . . . . . . . . . . . . . . . . . 218

24
List of Abbreviations

ANOVA ANalysis Of the VAriance

BP Beyond Petroleum (formerly British Petroleum)
CES Constant Elasticity of Substitution
CGE Computable General Equilibrium
COP Conference Of the Parties
CPE Chaos Polynomial Expansion
CPI Consumer Price Index
DICE Dynamic Integrated model of Climate and the Economy
DoE Design of Experiment
DSGE Dynamic Stochastic General Equilibrium
DSGM Derivative-based Global Sensitivity Measure
E3 Economy-Energy-Environment
EC European Commission
ECB European Central Bank
EU European Union
FAST Fourier Amplitude Sensitivity Test
FF Full or Fractional Factorial
FIH Financial Instability Hypothesis
FIPECO FInance Publique et ECOnomie
GAMS General Algebraic Modeling System
GDP Gross Domestic Product
GEMMES General Monetary and Multisectoral Macrodynamics for the Ecological Shift
GSA Global Sensitivity Analysis
HANDY Human And Nature DYnamics
HDMR High Dimensional Model Representation
IAM Integrated Assessment Model
IES Intertemporal Elasticity of Substitution

25
CONTENTS

IMF International Monetary Fund

IPCC Intergovernmental Panel on Climate Change
LED Light Emitting Diode
LHS Latin Hypercube Sampling
LSA Local Sensitivity Analysis
MC Monte-Carlo
MCA Monte-Carlo Analysis
NFGS Network for Greening the Financial System
OAT One-at-A-Time
OECD Organisation for Economic Co-operation and Development
OLH Optimal Latin Hypercube
OLHS Optimal Latin Hypercube Sampling
PCC Partial Correlation Coefficient
PI Proportional-Integral
PPP Purchasing Power Parity
PRRC Partial Rank Correlation Coefficient
QoI Quantity of Interest
RBC Real Business-Cycle
R&D Research & Development
RE Renewable Energies
SA Sensitivity Analysis
SCC Social Cost of Carbon
SDR State Dependent Regression
SRC Standardized Regression Coefficient
SRRC Standardized Regression Rank Coefficient
TFP Total Factor Productivity
UA Uncertainties Analysis
UN United Nations
VAR Vector AutoRegression
WB World Bank

26
Chapter 1

Introduction

The origin of this thesis lies in Mark Carney’s (Governor of the Bank of England) foundational
speech about climate-induced financial risks. In 2015, he stated that the climate change challenge
creates three types of potential risks for financial institutions: a) Physical climate risks associated
with the physical damages to assets,1 b) liability risks associated with the impacts that could
arise tomorrow if parties who have su↵ered loss or damage from the e↵ects of climate change seek
compensation from those they hold responsible,2 and c) Carbon risks associated with the financial
viability of a necessary transition to a low-carbon economy [Carney, 2015].

Three years later, in their October 2018 progress report, the members of the Network for Greening
the Financial System (NGFS) recognised that the challenges we face are unprecedented, urgent and
analytically difficult. Indeed, while dealing with climate change, we are at the same time facing a
deep crisis in financial and macro-economic models.

The likelihood of a climate-driven Minsky moment [Carney, 2019] - the term used to refer to a
sudden collapse in asset prices and to depict the financial crises that occurred in 2001 and 2007-
2009 - combined with the inability of a large majority of models at that time to predict such crises
invites us to question the reliability of today’s models for the orientation of decision makers. While
we cannot ignore the magnitude of the socio-economic impacts of the recent health crisis and the
questioning of our economic policies that it has engendered, this thesis focuses on another major
challenge of our time: the necessary transition of our economic systems to ensure the resilience of
our societies in the face of this other fundamental threat that is climate change.
1
These changes in turn may translate into adaptation costs and economic loss of value.
2
At the time, one might have thought that it would be decades before such claims were made, and that carbon
extractors and emitters - and, if covered by liability insurance, their insurers - would be the hardest hit [Carney, 2015].
But in 2021, at COP26, loss and damages were at the centre of the discussion and result in a cover decision [on Climate
Change 1/CP.26, 2021]

27
CHAPTER 1. INTRODUCTION

The demand for immediate actions by the Intergovernmental Panel on Climate Change, [IPCC,
2014], [IPCC, 2018] and more recently in the report of its first Working Group [IPCC, 2021], in
order to mitigate the short-term impacts and far-reaching irreversibility of climate change, requires
the building of reliable models to guarantee transition to a low-carbon economy (or even a climate
neutral economy). The development of my work aims at evaluating such reliability and o↵ering
extensions to an existing stock-flow consistent macro-economic transition model with a non-linear
dynamic. This dynamic leads to multiple equilibria incorporating the possibility of a collapse.
The use of such a model provides a framework where public policies capable of avoiding it can be
analysed.

1.1 “The world is on an unsustainable path” [BP, 2019]

Climate change is undoubtedly a tremendous threat to our societies due to its magnitude and
far-reaching impact across all sectors and geographies,3 its foreseeable nature,4 its irreversibility5
and its dependency on short-term actions.6 With the Rio Summit, the Kyoto protocol and now
the Paris Agreement Rulebook (finalized in 2021), States positioned themselves as cornerstones in
the fight against climate change. During the COP21, the global political sphere reconfirmed its
willingness to facilitate the transition to a low-carbon society. France’s Energy Transition Law,
adopted in July of 2015, requires companies and financial institutions to report on climate risk and
limit the global rise of temperatures to less than 2 C, and preferably as close to 1.5 C as possible.
The EU’s Green Deal and its “Fit-for-55” package include binding legislation to ensure that the
EU meets its 2030 climate and energy targets. At COP26, the UK Presidency succeeded in putting
carbon neutrality at the centre of discussions.

1.1.1 A long-time known trajectory

The extreme e↵ects of climate change are already easily observed around the world. From op-
pressive heatwaves in North America to typhoons in south-east Asia and severe droughts in Africa
and Australia, climate-related catastrophes are damaging infrastructure, degrading public health,
decreasing productivity, and diminishing our collective wellbeing [Carney, 2019]. As exemplified by
the foreign electronic or automotive industry after flooding in Thailand,7 and the five-fold increase
in insurance losses over the past three decades, these repercussions are increasingly severe and global
in reach. In the energy sector, data compiled in the BP Statistical Review of World Energy suggest
that:
3
NFGS report states: climate change will a↵ect all agents in the economy (households, businesses, govern-
ments).The risks will likely be correlated with and potentially aggravated by tipping points, in a non-linear fashion.
This means the impacts could be much larger, and more widespread and diverse than those of other structural changes.
4
While the exact outcomes, time-horizon and future pathway are uncertain, there is a high degree of certainty that
some combination of physical and transition risks will materialise in the future [Carney, 2019].
5
As shown in the work reported by the IPCC group, the concentration of greenhouse gas (GHG) emissions in the
atmosphere determines the impact of climate change and no mature technology is yet far-reaching enough to reverse
the process. Above a certain threshold, they demonstrate great potential to impose irreversible consequences on a
multitude of ecosystems, even if the time-horizon and degree of severity remain uncertain [IPCC, 2014]. Many changes
due to past and future greenhouse gas emissions are irreversible for centuries to millennia, especially changes in the
ocean, ice sheets and global sea level. [IPCC, 2021]
6
Actions today will determine the nature and magnitude of climate-related disasters. Thus, they demand reliable
and comprehensive long-term strategies. E.g.: strong, rapid and sustained reductions in CH4 emissions would also
limit the warming e↵ect resulting from declining aerosol pollution and would improve air quality. [IPCC, 2021]
7
In 2011, Western Digital and Nissan su↵ered losses of 235M and 67M , respectively (see [Chongvilaivan, 2012]).

28
CHAPTER 1. INTRODUCTION

“In 2018, global energy demand and carbon emissions from energy use grew at their
fastest rate since 2010/11, moving even further away from the accelerated transition
envisaged by the Paris climate goals.”

According to the authors, the growth in energy consumption in 2019 is clearly climate-related
because families and businesses around the world increased their demand for cooling and heating
in response to an unusually large number of hot and cold days. This rise in energy consumption
led to an acceleration in carbon emissions. Despite the many lock-downs around the world and the
shutdown of many economic activities leading to an unprecedented drop of 5.4 per cent in 2020,
global carbon dioxide emissions bounced back to pre-COVID levels, and concentrations of GHGs in
the atmosphere continue to rise.8 More generally, experts conclude that the global pace of progress
is “inconsistent with the Paris climate goals”. They stated:

“The world is on an unsustainable path: the longer carbon emissions continue to rise,
the harder and more costly will be the eventual adjustment to net-zero carbon emissions.”
[BP, 2019]

In their highly influential book The limits to growth 9 and 1974 study about economic growth
and resource depletion, Meadows and colleagues accurately illustrated the societal risks associated
with surpassing sustainable thresholds10 nearly fifty years ago [Meadows et al., 1972] and [Meadows
et al., 1974] for our societies. Most of the scenarios the researchers depicted ultimately led to
global economic collapse. Back-testing the model forty years later [Turner, 2008], [Turner, 2012]
and [Turner, 2014] updated a version of the model demonstrating that the collapse is even more
probable.11 However, the researchers’ conclusions have not been taken into serious consideration
due to their basic core economic model.

In [McIsaac, 2016] and [Bovari et al., 2018], the authors used a much finer economic model and
the United Nations (UN) population growth series to explore the question of whether or not global
warming could induce a similar breakdown of the world economy. They are reaching the same
conclusions as their counterparts, with even more precision about the process of climate change on
economic fundamentals.12

Motesharrei and colleagues develop the Human And Nature Dynamics (HANDY) model to il-
lustrate the factors which led to the collapse or sustainability of advanced, complex, and powerful
societies throughout history [Motesharrei et al., 2014]. The authors identified two important factors
8
The Emissions Gap Report 2021 shows that new national climate pledges combined with other mitigation measures
put the world on track for a global temperature rise of 2.7 C by the end of the century. [UNEP, 2021]
9
Updated in [Meadows and Randers, 2012].
10
Also warned in [Barnosky et al., 2012].
11
Turner’s model forecasts a collapse in living standards due to resource scarcity and carbon saturation in the
twenty-first century.
12
Their findings are threefold: 1) the +2 C target is already out of reach, absent negative emissions. A result that
is confirmed by the IPCC report WG I [IPCC, 2021] and the UNFCCC synthesis report in Oct 2021 of nationally
determined contributions which assesses the temperature rise at 2.7 C. [UNFCCC, 2021] ; 2) the long-run (resp.
short-run) consequences of the process of climate change on economic fundamentals may lead to severe economic
consequences absent (resp. too rapid) proactive climate policies. Global warming (resp. too fast transition) forces the
private sector to leverage in order to compensate for output and capital losses (resp. to lower carbon emissions), thus
endangering financial stability; 3) Implementing an adequate carbon price trajectory, as well as increasing the wage
share, fostering employment, and reducing private debt make it easier to avoid unintended degrowth and to reach a
+2.7 C target.

29
CHAPTER 1. INTRODUCTION

which appeared frequently throughout the study: 1) the stretching of resources due to the strain
placed on the ecological carrying capacity,13 and 2) the economic stratification of society into Elites
and Masses (or “Commoners”).14 They conclude that the model produces two di↵erent scenarios
of collapse, with one being due to a scarcity of labor (which follows an inequality-induced famine)
and the other to a scarcity of Nature (caused by a depletion of natural resources).

“In both of these scenarios, the Elites, due to their wealth, do not su↵er the detrimental
e↵ects of the environmental collapse until much later than the Commoners. This bu↵er
of wealth allows Elites to continue “business as usual” despite the impending catastrophe.
It is likely that this is an important mechanism that would help explain how historical
collapses were allowed to occur by elites who appear to be oblivious to the catastrophic
trajectory”.

1.1.2 A “climate-driven Minsky moment” [Carney, 2019]

Although the trajectories to a resource-driven collapse are well-modeled for a long time, the last
decade has witnessed the emergence of threat due to the financialization 15 of societies combined
with these climate-related risks. Indeed, as detailed in the work of the 2 Investing Initiative group16
[Cleveland et al., 2015], a systemic risk to financial stability associated with the transition to a
low-carbon economy has not been carefully investigated to date. Such an assessment of systemic
risk would require a better understanding of climate and carbon risks in the realm of financial
institutions. Whereas no such analysis exists yet, the topic is now gaining recognition and priority
on the agendas of macro-prudential authorities.

To avoid being “oblivious to the catastrophic trajectory” 17 [Motesharrei et al., 2014], thirty-four
central banks and supervisors across five continents, which account for half of all global green-
house gas emissions, created the Network for Greening the Financial System (NFGS) in 2017. The
coalition’s first report delivers recommendations to ensure a smooth transition to a low-carbon
economy, which include integrating the monitoring of climate-related financial risks into day-to-
day supervisory work, adding sustainability criterion into portfolio management, collaborating to
share knowledge about climate-risks data, and developing the capacity for in-house climate-risk
assessments. However the report recognizes the analytically difficult character of this approach.

At the same time, the shift towards greater electrification seems more and more necessary with
the increasing demand for power. Consequently, the necessary energy transition can only take place
if accompanied by a corresponding decarbonization program [BP, 2019]. For the financial sector,
this multifaceted transition demands new resolutions and actions by operators (disclosure) and
investors (divestment and engagement) [Cleveland et al., 2015] to mitigate the risk of what Mark
Carney calls a “climate-driven Minsky moment” [Carney, 2019].
13
[Abel, 2013], [Catton, 1982], [Kammen et al., 1994], [Ladurie, 1987], [Ponting, 1993], [Redman, 2004], [Wood,
1998] and [Wright, 2004].
14
[Brenner, 1976], [Turchin and Nefedov, 2009], [Diamond and Van Dijk, 2008] and [Goldstone, 1991].
15
Financialization is defined as “the growing importance of finance in daily activities, both at the national and global
levels”. At the heart of the problem of financialization is a growing inequality of income and wealth, as Nobel Laureate
Joseph E. Stiglitz shrewdly observed [Sheng, 2013].
16
A multi-stakeholder think-tank working to align the finance sector with 2 C climate goals.
17
Also written in [Carney, 2019] as “the obvious risks before our eyes”.

30
CHAPTER 1. INTRODUCTION

This term refers to the Minsky’s Financial Instability Hypothesis (FIH), which analyzed in the
1970s the likelihood of an intrinsic instability of the financial system and how it can imply a new
financial crisis comparable in terms of magnitude to that of 1929. As explained in [McIsaac, 2016],
the primary advantage of this framework is its ability to reproduce a financial crisis, such as the
subprime mortgage crisis, endogenously, while the large majority of Dynamic Stochastic General
Equilibrium (DSGE) models, which are well developed in literature, can hardly consider these
situations other than as “black swans” or large risks having a very small probability to materialize.

On the one hand, there is a clear need for a transition model to a carbon-neutral society, that can
predict the risk of a “climate-driven Minsky moment”, in order to orientate public policy decisions.
On the other hand, there is great analytical difficulty in producing such a model, as pointed out
by [Carney, 2019], which is compounded by the inability of standard macro-economic models to
predict the financial crises of the 2000s.

This cocktail e↵ect motivates an analysis of the nature of these models at the intersection of math-
ematical modeling, economic (or financial) assumptions, and ecological objectives and constraints.
This analysis is the object of the following sections about: the ambiguous role of models, due to
uncertainty and the non-bijective nature of modeling (Section 1.2), and the crisis of macro-economic
models, of which some features leads to this inability indicated above (Section 1.3).

1.2 The ambiguous role of models

Considering the versatility and ambiguity of models is critical. For example, mathematical models
and computer simulations are used in a wide variety of fields and institutions. Ecologists at all levels
of experience and scope develop them to study ecosystems and population dynamics. Institutions
around the world use them to study global and national economies. In regard to climate change,
two main types of models are used: a) physical models which assess the evolution of climate and b)
macro-economic models which assess the economic or financial impact of climate change. In some
instances, physical models help to corroborate or falsify hypotheses about the ecological systems,
and macro-economic models help to guide public policies.18

As noted by [Cariboni et al., 2007], models are often utilized in order to answer specific questions
about the current or future behavior of the system under analysis (e.g., is extinction within a fixed
time horizon a concrete threat?) or to investigate interactions between model components [Zaldıvar
et al., 1998]. More specifically, such system and interactions may be represented by mathematical
models and computer simulations such as di↵erential equations, interacting particles, or algebraic
equations. The models used to design structures, or to help policy decisions, influence in turn the
development of new behaviors in many spheres of society.

However, models themselves are built under uncertainties in their representation of reality, with
samplings of continuous phenomena, and the possible errors in the values of their own factors (e.g., a
population growth rate) what is called the parameterization of the model. These uncertain factors
18
E.g., in France, the Commissioner-General for Sustainable Development (of the Ministry of Environment) assesses
the impact of the National Low-Carbon Strategy with the Three-ME2 model. Their results forecast a GDP increase
of 1.6 by 2035, whereas the French Ministry of Finance predicts a contractionary e↵ect, with its own Mésange model
(see [Grandjean and Giraud, 2017]).

31
CHAPTER 1. INTRODUCTION

come in a variety of forms.19 While seeking scientific insight, designing the model, or making
decisions, the user of these models is thus limited by such uncertainties surrounding their input
variables and parameters (both are called factors). The reliability of the results produced by the
model, called robustness, is defined as its ability to retain its qualitative properties independently
of the errors that may arise from these uncertainties. A sensitivity analysis of the model’s input
variables and parameters allows for a quantification of its robustness of uncertainty.

1.2.1 Uncertainty: the very essence of a model

In their major book about modeling, [Saltelli et al., 2000] illustrate the role of models in a
scientific process in the sense of [Rosen, 1991], summarized in Figure 1.1. The non-bijective nature
of modeling generates di↵erent ambiguities surrounding the role of models in various fields. This
nature is due to the variety of uncertainties one may face in the building of a model.

Figure 1.1: Modeling after [Rosen, 1991]

The “world”, or “natural system” on the left of the Rosen’s diagram, is the object of this inves-
tigation. The model operates on the assumption that there are rules governing its behavior, be it
natural (as in the case of the diagram) or artificial. The object of the model is to represent such
rules. To this end, we “encode” the world into a system of formal relationships such as hypotheses,
di↵erential equations, etc. The decoding of this set of structures allows us to predict or understand
the “world’s” behavior, because both the “world” system and model are “entailed” with these rules.
The Rosen’s intuition postulates that there are no rules dictating how to encode or decode either
the “world” or the model, which would imply that the relationship between the “world” and its
representation is not a bijective function.20 In his words:

“While the world and the model are each internally “entailed”, nothing entails the world
with the model”.

As explained by [Saltelli et al., 2008a], among the reasons for this paradox is the fact that
the portion of the world captured by the model is an arbitrary “enclosure” of an otherwise open,
19
E.g., chemical, material, or mechanical properties, electrical capacitance or resistance, or atmospheric conditions,
behavioral economic function, etc.
20
In mathematics, a bijection, bijective function, or one-to-one correspondence is a function between the elements
of two sets, where each element of one set is paired with exactly one element of the other set, and each element of the
other set is paired with exactly one element of the first set. There are no unpaired elements. [Nikulin, 2001].

32
CHAPTER 1. INTRODUCTION

interconnected system.21 This idea might be quite disturbing for a student accustomed to the
beauty and apparent self-evidence of physical laws, but, as the authors say:

“Practitioners of modeling have come to live with the rather unpleasant reality that more
than one model may be compatible with the same set of data or evidence. Some have
gone so far as to coin a word for this paradox: equifinality22 [Beven, 1993], [Beven and
Beven, 2001] and [Saltelli et al., 2004] meaning that di↵erent models can lead to the
same end.”

Besides this partial “enclosure” of the model, which leads to equifinality, another ambiguity of
modeling resides in the “non-injective” nature of modeling, by which various realities can be mod-
eled by the same formal system. The scientific inquiry and modeling strategy might consequently
be rather arduous as a result of the ambiguity between model and reality. For instance, in macroe-
conomics, Pottier’s criticism of the [Acemoglu et al., 2012] model illustrates this “second layer”
of uncertainty in modeling, in which the model’s mathematical formalism could represent either
innovation, or nudges to managers, or the learning-by-doing abilities of the system - all with the
same modeling prediction (decoding) of very di↵erent public policies [Pottier, 2014].

Moreover, when modeling continuous phenomena with di↵erential equations, mathematical reso-
lutions or computing simulations involve sampling techniques that lead to mismatches between the
continuous reality and its discrete representation.23 Thus, when we define a model, we must even-
tually define its inputs and outputs according to its nature and purpose. As summarized in [Saltelli
et al., 2008a], a model can be:

• Leading to diagnostic or prognostic: a distinction between models used to understand a

law or to predict the behavior of a system ;

• Data-driven or law-driven: a law-driven model tries to put together accepted laws which
have been attributed to the system, in order to predict its behaviour. For example, we use
Darcy’s and Ficks’ laws to understand the motion of a solute in water flowing through a porous
medium. A data-driven model tries to treat the solute as a signal and to derive its properties
statistically.24

In every case, the model’s ability to faithfully represent the reality depends both on empirical
studies and theoretical considerations. For example, behavioral functions (such as the Phillips curve
or the investment behavioral function in economics) are built to fit data that involve a large number
of calibrations or estimations (based on empirical studies [Grandjean and Giraud, 2017]), whereas
the production function (such as [Cobb and Douglas, 1928], [Leontief, 1951], CES [Solow, 1956]
or Putty-clay in the sense of [Akerlof and Stiglitz, 1969]) is a key concept of economic growth
theory, and thus can present discrepancies between the actual production laws and their functional
representation (see [Stiglitz, 2018]).
21
Especially when dealing with the study of nonobservable parts of a system. e.g.,when astrophysicists introduce
dark matter into equations to explain the expansion of the Universe.
22
Also called model indeterminacy.
23
In the field of digital signal processing, the Nyquist-Shannon signal sampling theorem gives a criterion about the
minimal sampling frequency according to the maximal signal frequency to avoid distorsion of the information [Shannon,
1949].
24
Many other categorizations of models are possible: [Bell, 1988] distinguishes between formal (axiomatic), descrip-
tive and normative models (rules an agent should follow to reach a target).

33
CHAPTER 1. INTRODUCTION

A model’s variety of inputs su↵er the uncertainty of their actual values. Any forecast should thus
take into account these possible mismatches in order to provide reliable (robust) results. This critical
consideration is the object of the uncertainty (UA) and sensitivity analysis (SA). Especially in the
social sciences, or in complex economic environments, where these calibrations and estimations of
parameters lead to highly parameter-sensitive results, the impact of inputs’ uncertainties cannot
be considered negligible. Thus, questioning the robustness of macro-economic models comes to be
more and more seen as essential to conduct public policies in these fields [European Commision,
2009].

1.2.2 The sensitivity analysis to build reliable models

A possible definition of sensitivity analysis is given by [Saltelli et al., 2004]: “The study of how
uncertainty in the output of a model (numerical or otherwise) can be apportioned to di↵erent sources
of uncertainty in the model inputs”.

While building a model or interpreting numerical simulations, Sensitivity Analysis (SA) methods
are thus powerful tools, providing insight on the uncertainty in a model’s output [Iooss and Lemaı̂tre,
2015] and its relative dependence on di↵erent sources of uncertainty in the model’s input [Saltelli
et al., 2000] in order to quantify the relative importance of these input variables on the output.
More specifically, the purpose of SA is to answer the questions:

• “Which of the uncertain input factors are most influential in determining the variability af-
fecting the inference?”

• “If the uncertainty of one of the inputs could be eliminated, which one should be chosen in
order to reduce to the minimum the variance of the output of interest?”

• “Are there factors whose e↵ect on the output is so low that they can be confidently fixed
anywhere in their ranges of variation without a↵ecting the results?”

An overview of SA methodologies can be found in Chapter 2 and its appendix. Some rele-
vant applications of SA techniques to ecological and environmental science include, atmospheric
chemistry [Campolongo et al., 1999], transport emissions [Kioutsioukis et al., 2005], geographic
information systems [Crosetto and Tarantola, 2001], environmental management [EPA, 2003] and
population dynamics [Zaldivar et al., 2000].

According to Saltelli and colleagues,25 the role of UA and SA is crucial in ecological risk as-
sessment, which utilizes analytical models to estimate the impact of human actions on natural
resources and interpret the significance of that impact in light of the uncertainties identified in each
component of the evaluation process.

Some e↵ort has been put into understanding the role of SA from an environmental regulatory
point of view. Both the report on Good Practice Guidance and Uncertainty Management in Na-
tional Greenhouse Gas Inventories [EPA, 2003] and the Draft Guidance on the Development, Eval-
uation and Application of Regulatory Environmental Models [IPCC, 2014] provide information
about model use and model science in environmental settings. The EPA’s (Environmental Pro-
tection Agency) report also contains recommendations on good practices for UA and SA [EPA,
25
[Saltelli et al., 2000], [Saltelli et al., 2004] and [Saltelli et al., 2008a].

34
CHAPTER 1. INTRODUCTION

2003]. In Europe, sensitivity analysis is mentioned in the guidelines for impact assessment [Euro-
pean Commision, 2009]. Among them, the Global Sensitivity Analysis (GSA) appears to fulfill the
macro-prudential authorities’ need for robust models.

According to the model’s goals, nature and properties, types of uncertain inputs, and outputs
(also called quantities of interest – QoI), various GSA can be used. In [Iooss and Lemaı̂tre, 2015],
the authors give many application examples ; for instance, [Makowski et al., 2006] analyze, for a
crop model prediction, the contribution of 13 genetic parameters on the variance of two outputs.
Additionally, Lefebvre and colleagues use GSA to determine the most influential input among a
large number of inputs (around thirty) [Lefebvre et al., 2010].

According to [Iooss and Lemaı̂tre, 2015], the objectives of GSA may include:

• identify and prioritize the most influential inputs,

• identify non-influential inputs in order to fix them to nominal values,

• map the output behavior in function of the inputs by focusing on a specific domain of inputs
if necessary,

• calibrate some model inputs using some available information (real output observations, con-
straints, etc.).

When the model is non-linear and non-monotonic, with a small number of parameters and a
short running time for each simulation, the De Rocquigny’s decision diagram suggests using the
GSA method with the so-called Sobol’ indices [De Rocquigny, 2008]. The method, developed by
the Russian mathematician Sobol in [Sobol, 1967], is based on the decomposition of the output
variance: for a given function f 2 L1 ([0, 1]p , dx) Sobol considers a decomposition into terms of
increasing dimensions
p
X p X
X p
f (x1 , ..., xp ) = f0 + fi (xi ) + fi,j (xi , xj ) + ... + f1,...,p (x1 , ..., xp ) (1)
i=1 i=1 j>i

and imposes that 8s 2 {1, ..., p}, 8i1 < i2 < ... < is 2 {1, ..., p}s and 8k 2 {i1 , ..., is },
Z 1
fi1 ,...,is (xi1 , ..., xis )dxk = 0 .
0

Consequently this expansion of f into 2n summands of di↵erent dimensions exists, is unique and the
related terms are pairwise orthogonal in L1 ([0, 1]p , dx). Assuming now a square integrable random
variable Y as Y = f (X) where X = (X1 , ..., Xp ) is a vector of p independent random variables
uniformly distributed on [0, 1], we have:26
p
X p X
X p
Y = f0 + fi (Xi ) + fi,j (Xi , Xj ) + ... + f1,...,p (X1 , ..., Xp ). (2)
i=1 i=1 j>i

26
Here one can see that the variance decomposition strongly depends on the independence of the random variables
(X1 , ..., Xp ).

35
CHAPTER 1. INTRODUCTION

which implies27
p
X X p
p X p
X p
X X Vi,j p
Vi V1,...,p
V (Y ) = Vi + Vi,j + ... + V1,...,p , 1 = + + ... +
V (Y ) V (Y ) V (Y ) (4)
i=1 i=1 j>i i=1 i=1 j>i

V
where Si1 ,...,is = iV1 ,...,i s
(Y ) . Si is a measure of the single impact of Xi on the total variance, Si,j is a
measure of the joint impact of (Xi , Xj ) on the total variance, etc... In this spirit, to quantity the
total impact (sum of single and related joint e↵ects) of Xi on the variance of Y we define

E(V (Y |Xĩ )) V (E(Y |Xĩ ))

ST i = =1 (5)
V (Y ) V (Y )

where Xĩ = (X1 , ..., Xi 1 , Xi+1 , ..., Xp ).

For example, in a model with 3 inputs we have

ST1 = S1 + S1,2 + S1,3 + S1,2,3 = 1 S2,3 . (6)

Monte Carlo sampling-based methods have been developed to estimate these Sobol’ indices (see
[Saltelli, 2002]), although McKay, Iman and colleagues demonstrate the efficiency of other types
of stratified sampling with a pick freeze method (such as Latin Hypercube Sampling - see [McKay
et al., 2000] and [Iman, 1999]). These methods (GSA with Sobol’ indices estimated by the [Saltelli,
2002] method) are specifically designed for non-linear models with interactions between parameters.
In general, Global Sensitivity Analysis can be conducted with variance-based, frequency-based or
meta-models methods. Their results are quite similar [Saltelli and Bolado, 1998], but latter two
methods require the a priori choice of an upper bound (in the degree of the polynomial or the
harmonic), which can introduce biases in the study to assess actual values of the sensitivity indices.

Physical models (climate or meteorological models) used to predict the evolution of the environ-
ment can be evaluated through their ability to: a) simulate present and past climate conditions and
b) reproduce climate variations at each time-scale. These models thus have a predictive ability that
allows for the identification of an e↵ective public policy with reliability (robustness) [Grandjean and
Giraud, 2017].

For macro-economic models, the challenge of achieving this reliability is a much more sensitive
subject because of a) the failures of a large majority of models to predict the 2000s crises,28 and
b) the sensitivity (lack of robustness) of today’s models.29 It is clear today that this sensitivity has
27
Using the pairwise orthogonality in the Sobol’ decomposition, we can obtain easily a recursive expression of each
term using conditional expectations
f0 = E(Y )
fi (Xi ) = E(Y |Xi ) E(Y )
fi,j (Xi , Xj ) = E(Y |Xi , Xj ) fi (Xi ) fj (Xj ) E(Y ) (3)
..
.

and deduce the so-called ANOVA-HDMR decomposition for the variance of Y .

28
See Section 1.3.2 and 1.3.3.
29
See a review of some of these models in Section 2.4.

36
CHAPTER 1. INTRODUCTION

led to large underestimations of the economic impact of climate.30 Therefore, a global sensitivity
analysis of macro-economic models is necessary to inform discourses around the reliability of their
results, and thus to help make up for the crisis they are facing.

1.3 The crisis of macro-economic models

Macroeconomics has historically been a controversial sphere, despite the existence of a mainstream
theory. The mainstream theory failed to predict the crises of the 2000’s in spite of the multitude
of contemporary models, vivifying the old fatalistic diagnosis of the crisis of macro-economic mod-
els. We should interpret this diagnosis as a demand for the selection of a model that is able to
accommodate our scientific expectations in climate change economics.

1.3.1 An historical typology of macro-economic models

Various types of macro-economics models can be described:

• Statistical models, such as Vector AutoRegression (VAR) models, link linear statistical rela-
tionships to macro-economic variables of multiple time series. They are a generalization of the
univariate autoregressive model. According to [Bjornland, 2000], even though VAR models
may not satisfy Lucas’ criteria for policy intervention, they are still useful for anticipating
the impact of policy actions that fall within the realm of historical experience. In particular,
shifts in policy rules can be somewhat subsumed under stable policy rules ;

• Theoretical models: since the publication of The General Theory of Employment, Interest
and Money [Keynes, 1936], economists have been divided into micro and macroeconomics.
Micro-economic models are interested in individuals’ and firms’ behavior and solving utility
or profit maximization problems under certain constraints. Macro-economic models study ag-
gregated indicators which consider the whole economy. The theory is based on a mathematical
interpretation of the Keynes’ and Hicks’ work about the Great Depression [McIsaac, 2016].
In the 1960s, Lucas and his contemporaries of the new classical macroeconomics developed
micro-foundations (criticized by the Sonnenschein-Mantel-Debreu Theorem) and a complex
mathematical basis known as Dynamic Stochastic General Equilibrium (DSGE). Under the
assumption of an absence of rigidities in market mechanisms,31 the DSGE models add im-
perfections and exogenous shocks that alter the equilibrium and impinging economic growth.
These models overpowered macroeconomics and are widespread among national and interna-
tional institutions.32 Generally, they are founded on the following assumptions [Annicchiarico,
30
Exposed in the book of [Pottier, 2014].
31
First models called Real Business Cycle models postulated a representative consumer who operates in perfectly
competitive markets.
32
Example of DSGE models across the world [Grandjean and Giraud, 2017]:
– Europe: Germany: [Pytlarczyk, 2005] ; France: Mésange [Bardaji et al., 2017] ; UK: BEQM [Harrison et al.,
2005] ;
– Euro Zone: BCE, NAWM [Christiano et al., 2010] ; European Commission: QUEST III [Ratto, 2008] and
DYNARE [Adjemian et al., 2011] ;
– Americas: Brasil: SAMBA ; Canada: ToTEM [Fenton and Murchison, 2006] ; United States: SIGMA, [Erceg
et al., 2006] ;
– International Institutions, e.g., IMF: GIMF [Laxton et al., 2010] ;
– World Bank: Integrated macroeconomic model for poverty analysis (IMMPA) is a dynamic CGE.

37
CHAPTER 1. INTRODUCTION

2010]:

– Perfect competition in all markets ;

– All prices adjust instantaneously ;
– Rational expectations ;
– No asymmetric information ;
– The competitive equilibrium is Pareto optimal ;
– Firms are identical and price takers ;
– Infinitely lived identical price-taking households.

The models’ general equilibrium nature is presumed to capture the interaction between policy
actions and agents’ behavior, while the models specify assumptions about the stochastic shocks
that give rise to economic fluctuations. Hence, the models are presumed to “trace more clearly
the shocks’ transmission to the economy” [Sbordone et al., 2010].

An other classification of these models can be done by way of their di↵erent purposes:

• Cyclical forecasts models (e.g., Mésange, Egée and Opale for the French Ministry of Economy)
where the future values of the main indicators are computed with very short-tem horizon
(around one month up to two years) ;

• Integrated Assessment Models (IAM) which aim to assess economic impacts of climate change.
They are subdivided into a) Process-based IAM [Sathaye and Shukla, 2013], also called E3
(Economy-Energy-Environment), usually with a cost-e↵ectiveness analysis such as Imaclim-R
(CIRED), and b) those based on a cost-benefit analysis (such as the famous DICE [Nordhaus,
2008], PAGES, FUND, etc.). IAM are usually based on a Computable General Equilibrium
model (CGE) with a climate feedback loop (e.g. a damage function) ;

• Theoretical macro-economic models which aim to explore a particular phenomenon without

necessarily an empirical basis, to identify stylized facts both corroborated by the internal
logic of the model and the data. However, as detailed in [Pottier, 2014], this juxtaposition of
the model and the data, without an actual confrontation, might lead to poor analyses called
just-so-stories.

Ecological macro-economic models attempt to link climate and economic environments. Some of
them aim to assess the impact of a public policy,33 with cost-benefits or cost-e↵ectiveness analysis,
at a national, regional or global scale. Inspired by [La↵argue, 2012] and [Grandjean and Giraud,
2017], these macro-economic models are classified as:

• Neo-Keynesian based models: with quantity adjustment in the short-run, these models are
driven by aggregate demand (e.g.,Mésange). Temporary disequilibria are possible due to the
inertia of quantities adjustments ;

• Computable General Equilibrium models: based on a Walrasian representation of the economy,

where prices are perfectly flexible and preferences of a rational agent optimizing a utility
function are explicitly described. Various types of models exist in this category (such as
33
Mésange (Bercy), Three-Me (OFCE / ADEME), Imaclim (CIRED), Nemesis (Erasme), Gemini (Lausanne), GEM
E3 (European Commission), GEMMES (AFD), etc.

38
CHAPTER 1. INTRODUCTION

Integrated Assessment Models). Generally, institutions use both an estimated DSGE (Global
Integrated Monetary and Fiscal model - GIMF, New Area-Wide Model of The Euro area -
NAWM, ...) and a neo-keynesian model for their forecast ;

• Disequilibrium models (such as ThreeMe, a post-Keynesian model): with a specific prices’

dynamic, independent of the adjustment in supply / demand, a multisectoral production and
a wages’ dynamic through a new short-term Phillips curve. GEMMES is based on the work
of [Grasselli and Costa Lima, 2012] with a non-linear out-of-equilibrium dynamic inspired by
the work of [Goodwin, 1967], [Akerlof and Stiglitz, 1969], [Keen, 1995]. These equations are
based on the Predator-Prey Logic of Lotka-Volterra, which takes into account private and/or
public debt. These types of models are able to represent the Minsky’s Financial Instability
Hypothesis (FIH) and predict crashes. The stochastic dynamic allows the emergence in aggre-
gated indicators, but they are heavily criticized due to their high-dependency on parameters,
their myopic nature (without expectations) and exogenous growth. However, they seem to
be the most able to represent the trajectories we want to explore, if we can compensate for
these criticisms. An estimated version of the model exists with a Climate Feedback Loop
developed in [Bovari et al., 2018] and a first approach dedicated to the Brazilian economy was
implemented by [Bastidas et al., 2017].

For the purpose of addressing the risk of a “climate-driven Minsky moment”, using a robust
model to orientate public policies towards a smooth transition to a carbon-neutral society, these
disequilibrium models seem indeed the most compelling because they are able to overcome some of
the failures of other macro-economic models.

1.3.2 A largely documented diagnostic of a crisis

The crisis of economics and economic models is an ancient topic that was reignited by the inability
of major macro-economic models to predict the financial crises of the 2000s. This inability reveals
a failure both in the profession, as expressed in [Galbraith, 2009], [Boyer, 2012] and [Guesnerie,
2013], and in the core theory of the model at that time. Prof. Blanchard34 who claimed in his 2009
paper that “the state of macro is good” [Blanchard, 2009],35 stated in 2016 [Blanchard, 2016]:

“I see the current DSGE model as seriously flawed [...] They are based on unappealing
assumptions. Not just simplifying assumptions, as any model must, but assumptions
profoundly at odds with what we know about consumers and firms.”

In [Christiaens, 2018], the author wrote the Kocherlakota’s36 words:

“...we do not have a settled successful theory of the macro-economy. The choice made
25-40 years ago - made then for a number of excellent reasons - should not be treated as
written in stone or even in a pen”
34
Chief economist at the International Monetary Fund between 2008 - 2015.
35
While Robert Lucas, in his presidential address to the American Economic Association at the University of
Chicago in 2003, was declaring that “central problem of depression-prevention has been solved”.
36
President of the Federal Reserve Bank of Minneapolis 2009.

39
CHAPTER 1. INTRODUCTION

More recently, the Nobel Laureate, Prof. Joseph E. Stiglitz,37 wrote that DSGE “failed to in-
corporate key aspects of economic behavior, e.g.,incorporating insights from information economics
and behavioral economics. Inadequate modelling of the financial sector meant they were ill-suited for
predicting or responding to a financial crisis” in [Stiglitz, 2018], while Prof. Paul Romer,38 speaking
about DSGE, declared also in The Trouble with Macroeconomics [Romer, 2016]: “For more than
three decades, macroeconomics has gone backwards [...] The trouble is not so much that macroe-
conomists say things that are inconsistent with the facts. The real trouble is that other economists
do not care that the macroeconomists do not care about the facts.”

In France, the FIPECO39 presented a comparison between the forecast from the French Ministry
of Economy and reality, and concluded by: “In economic matters, the observed reality is never
compatible with predictions. Statistically, they noticed that the modeling does not perform better
than a simple persistence model. These criticisms about DSGE are not new. Indeed in his book,
Debunking Economics:The Naked Emperor Dethroned?, [Keen, 2011] developed a series of theoretical
criticisms of mainstream macroeconomics. A broader overview of these criticisms is available in the
next section, but we can have some of them displayed here, such as:

• Most of new classical models have no energy or matter in their production function (exception:
when using KLEM production function with Capital (K), Labor (L), Energy (E), Matter(M),
e.g., [Stiglitz, 1974a]);40

• Money is always considered neutral when it is taken into account;41

• They do not admit global out-of-equilibrium dynamics (see e.g., CGE models);

• They are mainly linear in the presumed dynamics, even if they represent very complex phe-
nomenon;

• They do not take into account private debt42 (exception: [Krugman and Eggertsson, 2013]
but no money...) nor the banking sector (only considered as financial intermediaries);43

• Due to rational expectations and aggregate demande assumptions, they cannot predict an
equilibrium with large underemployment (at least when the labor market is perfectly flexible)
and can barely consider financial crises as black swans.

Perhaps the most compelling criticism of classical macroeconomic models is that they fail to
consider private debt, which generates an inability to predict trajectories to collapse other than
black swans. For example, the oil-debt link is well-illustrated by Prof. Joseph E. Stiglitz in his
book The Great Divide [Stiglitz, 2015], in which he illustrates how the Federal Reserve contributes
37
Nobel Laureate in 2001 and World Bank Chief Economist between 1997-2000.
38
Nobel Laureate in 2018 and World Bank Chief Economist between 2016 - 2018.
39
Finance Publique et Economie, independent and free-access association analyzing public finance and french econ-
omy. Source http://www.fipeco.fr/pdf/0.48968300%201464608610.pdf
40
These models generally consider that, according to the cost-share theorem, the GDP-elasticity to energy equals
the nominal amount of energy in GDP, which is negligible. In fact, [Giraud and Kahraman, 2014] showed that this
share in volume accounts in average for 60% (6 times greater than what new classical economists usually admit).
41
Whereas it is demonstrated as endogenous in the book Illusion Financière of [Giraud, 2014].
42
See the work of [Keen, 2013] about the link debt-deflation. If excessive debt is so likely to generate a deflation, it
is at least surprising the debt ratio was not part of such a model.
43
See Debt-Deflation versus the Liquidity Trap: the Dilemma of Nonconventional Monetary Policy of [Giraud and
Pottier, 2016] and Illusion Financière, [Giraud, 2014].

40
CHAPTER 1. INTRODUCTION

to the aggravating factors of the financial crisis by misunderstanding the attenuating power of debt
to limit oil price inflation:

“The war in Iraq made matters worse, because it led to soaring oil prices. With America
so dependent on oil imports, we had to spend several hundred billion more to purchase
oil [...] Normally this would have led to an economic slowdown as it had in the 1970s.
But the Fed met the challenge in the most myopic way imaginable. The flood of liquidity
made money readily available in mortgage markets even to those who would normally
not be able to borrow. And, yes, this succeeded in forestalling an economic downturn ;
America’s household saving rate plummeted to zero. But it should have been clear that
we were living on borrowed money and borrowed time.”

Indeed, the very low interest rates of the Fed (from 0.98% to 5.3 % between 2002 and 2006) allow
every American to live on credit, thus circumventing inflation in oil prices [McIsaac, 2016]. The
book’s analysis illustrates how the financial crisis is, above all, a crisis of private debt. It does so
by explaining that DSGE models at the time were incapable of considering such a crisis probable.
This incapability motivates a revival of Minsky’s idea of the Financial Instability Hypothesis (FIH)
about the intrinsic instability of capitalism.44

According to the analysis of [Grandjean and Giraud, 2017], one can also question the ability of
macro-economic models to predict reliable trajectories, especially because of their input definition,45
the ambiguity of their output indicators (GDP46 ; Cost of public policy ; Unemployment rate47 ),
their lack of back-testing, their potentially high sensitivity to calibrated parameters, their inabil-
ity to represent out-of-equilibrium trajectories to multiple equilibria, especially in climate change
economics.
44
The Minsky’s idea was to consider a time in which economic stability incites investors to take risks (with optimism).
This risk-taking reflects innovations transforming production and society. However, they also perturb the economic
environment. With its sheep-like behavior, the capitalist system switches from a time of constant growth to a time of
euphoria where expectations about the future become increasingly optimistic. Financial levers are easier. Investments
are encouraged by risk-taking, and they begin to exceed profits, leading to an increase in private debt. The financing
demand skyrockets. Subsequently, interest rates increase so much that financial viability may not be satisfied. At the
peek of this euphoric period, when financial-market operators feel a too big mismatch between valuation and actual
values, they sell their assets, igniting a credit collapse. Because of the increasing debt and high interest rates, profits
are fewer and fewer, which will make investments subsequently plummet. The euphoric period thus leads to a recession
in which profit rates come back to more moderate values. This story is at the heart of the naturalist novel L’Argent by
Emile Zola, in which the author depicts money and debt as living entities and describes the premisses of the collapse
of the Napoleonic 2nd Empire. In [Krugman and Eggertsson, 2013], the authors illustrate this phenomenon with the
coyote and the road runner cartoons, in which the coyote continues to run even when he is mid-air and about to
fall from a cli↵. When optimism comes back, the same scheme will appear, but this time, the debt amount at the
beginning of the euphoric period will be higher. For this reason, [Krugman and Eggertsson, 2013] try to put the
debt at the heart of new classical DSGE models, but without money. Despite their attempts, the large majority of
new-classical models still does not take into account private debt in their dynamic.
45
E.g.,the Cambridge capital controversy in the late 1950s expresses the gap between definition of the nature and
role of capital goods. [Stiglitz, 1974b].
46
See a review of the criticisms in [Jany-Catrice and Méda, 2015] or the proposition of [Piketty, 2009].
47
The ECB experts estimate the labour market slack around 18% in the Euro Zone, almost twice the official
rate from Eurostat around 9.5%. This shows the current paradox about the calculation of this indicator [European
Central Bank, 2017]

41
CHAPTER 1. INTRODUCTION

1.3.3 Macro-economic modeling: map of the problematic

In climate change economics, modelers are facing two main fields: Economics and Environment.
In environmental science, researchers are modeling complex interactions between components de-
pending on (possibly probabilistic) observables. However these behaviors can usually be described
through physical laws and equations. These fundamental equations, such as the conservation of
energy, heat transfer, etc., have already been established and empirically tested. Solving these
equations, through a selected numerical method and sampling, can introduce some approximations.
These equations involve various parameters, among which some fundamental constants are not ques-
tionable (such as the Planck’s or Boltzman’s constant). The uncertainties in the models are thus
not primarily coming from the equations themselves, but rather from the way they are implemented
and solved in order to represent the actual system.

In Macroeconomics, modelers have to represent complex interactions between complex decision

units (humans), of which the behavioral description through mathematical equations is not some-
thing unquestionable [Grandjean and Giraud, 2017]. Furthermore, whereas all environmental mod-
els are governed by the laws of thermodynamics,48 a majority of macro-economic models still ignore
them, despite the words of Sir [Eddington, 2012]:49

“The law that entropy always increases holds, I think, the supreme position among the
laws of Nature. If someone points out to you that your pet theory of the universe is
in disagreement with Maxwell’s equations, then so much the worse for Maxwell’s equa-
tions. If it is found to be contradicted by observation, well, these experimentalists do
bungle things sometimes. But if your theory is found to be against the second law of
thermodynamics I can give you no hope; there is nothing for it but to collapse in deepest
humiliation.”

Some works of economists, such as [Georgescu-Roegen, 1993] or [Ayres and Warr, 2005], are
pioneers in taking into account physical laws and constraints into macroeconomic theory, but this
kind of link is still rare in the field. Moreover, an important number of models are facing other
structural problems in the definition of their inputs / outputs (indicators), the back-testing of their
results, and their high-dependency on parameters, sampling choices, and representations.

The ambiguity of indicators based on hard-to-access data

Usual macro-economic models deliver predictive trajectories about some indicators (or outputs,
also-called Quantity of Interest in Sensitivity Analysis). The relevance of these indicators is ques-
tionnable because of their ambiguous definition, their difficult methodology of construction and the
hard-to-access data they are based on.

For instance, the main indicator of these models, Gross Domestic Product (GDP), is criticized
heavily through a large literature (see [Jany-Catrice and Méda, 2015] for a survey, or the [Piketty,
2009] proposition to switch the indicator to Net National Income). According to the [Kaya, 2014]’s
48
More specifically, the second law of thermodynamics which states that the total entropy of an isolated system can
never decrease over time. It means that a natural process runs only in one sense, and is not reversible. For example,
heat always flows spontaneously from hotter to colder bodies, and never the reverse, unless external work is performed
on the system. This implies the impossibility of a perpetual motion machine, or the Carnot’s theorem that limits the
maximum efficiency for any possible engine.
49
British astrophysicist and philosopher of science.

42
CHAPTER 1. INTRODUCTION

equation, about the link between GDP growth and CO2 emissions growth, it seems indeed dubious
to consider it as a relevant indicator to the task of building a climate policy. [Giraud and Kahraman,
2014] stated the link between energy consumption growth and GDP growth, evidenced in the [BP,
2019] report, questioning the integrity of pursuing GDP growth if that growth is so strongly linked
to increases in our (mostly fossil) energy consumption.

According to [OECD, 1996], the unemployment rate is also an ambiguous indicator, because:
(1) many classifications of unemployment may lead to various possible results due to accounting
manipulations; (2) in France, the methodology di↵ers according to the institutions and thus deliver
di↵erent results; and (3) the indicator is based on a strong conventional component that can be
caused by many sources (discouraged workers; unregistered; ...).

The (expected) cost of a public policy is often used to orientate a policy maker’s decision. This is
the case, for example, in Integrated Assessment Models (IAM) such as DICE, which relies on a cost-
benefit analysis. This analysis is generally conducted through the maximization of the aggregated
utility function.50 The aggregation methodology is debatable, as detailed in the works of many
economists.51 One of the best example is the Sonnenschein-Mantel-Debreu Theorem which proves
that microeconomic rationality assumptions have no equivalent macroeconomic implications, leading
to the conclusion that aggregate demand can be any polynomial [Rizvi, 2006]. Furthermore, the
utility’s maximization is generally based on consumption, without consideration for a possible limit
to this consumption.

Besides the dubious relevance of these indicators, and the possible biases in their building method-
ology, the underlying inputs or data of these macro-economic models are hard-to-access due to their
socially-built nature. Their construction thus depends on institutions with di↵erent techniques and
abilities to finely assess the input. Other input parameters are assumed to be constant, such as the
depreciation rate of capital, which is anything but certain.

A need for back-testing and assessing sensitivity of models

These inputs’ uncertainties and the difficulty to access to data impinge the reliability of the results
or the recommendation in terms of public policy (especially when the result itself is an ambiguous
indicator). The macro-economic models are indeed likely dependent on many parameters. While
some of them are estimated, a large number of parameters are only calibrated. The estimation
methods can lead to mismatches between the result and the reality. These types of mismatches can
be reduced through back-testing the model on historical data.

These estimations and calibrations of parameters can have important impacts on the outputs.
This is the end of Sensitivity Analysis: to describe the link between impact on the outputs and
uncertainties on the inputs. This type of analysis is detailed in appendix A.2. In a very complex
system, such as environmental science, or macroeconomics, with numerous nonlinearities and pa-
rameters, we can acknowledge the fact that a minor uncertainty might lead large variation of the
output.
50
Inspired from the Ramsey-Cass-Koopsmans model, maximizing levels of consumption over successive generations
[Collard, 2012].
51
E.g., [Arrow, 1999], [Arrow et al., 2004].

43
CHAPTER 1. INTRODUCTION

Among the various sensitivity analyses, the bayesian method is the most frequently used to
calibrate DSGE models, but this method involves many a priori biases. The bayesian inference
used to build the sensitivity analysis of the DSGE models is consequently highly dependent on the
a priori choices of the modelers. In a field where facts and values are hard to distinguish [Pottier,
2014], one can understand how a bayesian method to calibrate a DSGE model could potentially
confirm the original biases of the modelers. Other SA (so-called local) implies the selection of a
nominal point around which testing the sensitivity of the model.52 The selected nominal point in
the input’s space (whatever the method selection is) is critical for the sensitivity analysis, because
the modeler is exploring it in a close neighborhood of this particular value.

In a field with important uncertainties on parameters, due to hard-to-access and socially-built

data, one can understand the need for back-testing the model and experiencing its robustness to
these uncertainties through sensitivity analysis, in order to deliver the most reliable result, even if
the output remains questionable.

The questionable validity of functional representations and dynamics

Among the di↵erent representations of the production function in Macroeconomics, one of the
most famous is the Cobb-Douglas production function, which is stilled used in various models and
was at the center of the Cambridge Capital Controversy.53 The theory itself of a production function
is empirically hard to describe (even at a micro-level) and covers many theoretical challenges about
the possible substitution between factors, the elasticity and inertia of these substitutions, etc.

Many other functional representations in economics54 are unfalsifiable in the sense of [Popper,
1973],55 who proposed that “the criterion of the scientific status of a theory is its falsifiability, or
refutability, or testability” [Popper, 1963]. This argument helps us to understand that we need
to take into account the limitation of these models and not dismiss all functional representations,
following the aphorism of [Box, 1976]: “All models are wrong, some are useful.”

In addition to these functional representations, the very di↵erent dynamics between the di↵er-
ent types of models can be questioned. For this reason, major institutions are generally using
several models with di↵erent dynamics to compute a range of scenarios covering these theoretical
divergences in the functional representations.

The historical Solow’s model is dynamic but often used to analyse a steady-state in the long-run.
This habit is anything but scientific, because it presupposes that our real economy has reached
its equilibrium, and nothing could be less certain. Computable General Equilibrium models try
to calculate many static equilibrium at each time step, taking into account possible exogenous
shocks to represent the high volatility of actual economic indicators. However these models view
the kinematics of the system through the lens of a succession of equilibria. DSGE models assume
the Solow’s equilibrium is reached, but exogenous shocks move the system away. Their analyses
52
These methods are also presented in appendix A.2.4.
53
See the work of [Samuelson, 1966], and the more recent book of [Felipe and McCombie, 2013], which is summarized
in [Guerrien and Gun, 2014].
54
E.g.,the utility function, which can not be observed [Hodgson, 2012].
55
A statement, hypothesis, or theory has falsifiability (or is falsifiable) if it is contradicted by a basic statement,
which, in an eventual successful or failed falsification, must respectively correspond to a true or hypothetical obser-
vation.

44
CHAPTER 1. INTRODUCTION

consist in computing the return to the long-run steady-state, whatever the nature of the shock is.
By essence, these models are not able to predict trajectories to collapse or a climate-driven Minsky
moment. The out-of-equilibrium dynamic of a DSGE is by construction always transitional.

A more interesting typology of models for our study are the disequilibrium models mentioned
above (such as Three-Me and its extension IMACLIM-3ME [Ghersi, 2020], [Akerlof and Stiglitz,
1969] or [Grasselli and Costa Lima, 2012]). These models assess by construction the trajectories
to multiple equilibria and are able to describe global out-of-equilibrium dynamics. The GEMMES
model56 is an application of this type of model about various geography [Bastidas et al., 2017]
and production functions [McIsaac, 2016], with climate feedback loop [Bovari et al., 2018] and
various features (Government intervention [Costa Lima and Grasselli, 2014], Inventories [Grasselli
and Nguyen Huu, 2016], Inequalities [Giraud and Grasselli, 2017], etc.).

One conclusion drawn from this line of thought is that we cannot use DSGE models to represent
the trajectories we seek to explore in order to avoid a “climate-driven Minsky moment”. For this
reason, we used a disequilibrium model, as an alternative to the previously described defaults.
As detailed in appendix A.1.1, the primarily necessary features we aim to o↵er are a non-linear,
global out-of-equilibrium dynamic, and a consideration of money and private debt, in which mass
unemployment and financial crashes can occur and the climate feedback loop can be analyzed.

The selected model is based on the work of Goodwin and introduces debt [Goodwin, 1967]
and [Keen, 1995]. The nature and stability of the debt equilibria were analyzed in [Grasselli and
Costa Lima, 2012] and [McIsaac, 2016] for the Van Der Ploeg’s Extensions.57 The model is based
on the Lotka-Volterra Predator-Prey Logic, with a classical capital’s accumulation equation and
two behavioral functions (about, on the one hand, investment depending on profits, and, on the
other hand, the link between employment rate and wage share, through a short-term Phillips curve),
private debt variation (as the di↵erence between investment and profits) and exogenous population
and labor productivity growth rate. The flaws of the model lie in its exogenous growth model, the
use of the Say’s law (which can be relaxed by the introduction of inventories [Grasselli and Nguyen-
Huu, 2018]), its myopic nature (without expectations), and its high parameter-dependency58 and
its criticism stems from these flaws.

In order to contribute e↵ectively to the improvements of financial and macro-economic models

built to assess public policies for preventing climate-driven collapse, a robust transition model have
to be provided in order to display transition trajectories, while bearing in mind the obstacles of the
climate change economics.

1.3.4 A mix of science and ethical values in Climate Change Economics

The standard evaluation of the economic impacts of climate change highlights specific issues, be
they technical (about the right indicator) or more philosophical. This evaluation is indeed contained
within a big-picture portrayal of human history as a relentless and inevitable trend “toward greater
levels of social complexity, political organization, and economic specialization, with the development
of more complex and capable technologies supporting ever-growing population, all sustained by the
56
Developed at the French Development Agency under the direction of Gaël Giraud.
57
Introducing a CES production function instead of the Leontief production function [Van der Ploeg, 1985].
58
See also the study of [Pottier and Nguyen-Huu, 2017] about the dependency of the model dynamics to the selected
investment function.

45
CHAPTER 1. INTRODUCTION

mobilization of ever-increasing quantities of material, energy, and information” [Motesharrei et al.,

2014] on the one hand.

In this story, human society is well-satisfied with its achievements. Wealth has never been higher,
a new liberalism leads us to a capitalism without any complexes, and there is no problem that
innovation, or a right price, cannot solve, because our knowledge is in a constant state of progression.
Any problems remaining in our society do not really matter because they can be solved by way
of political will, or more preferably, by letting the market mechanisms guide us to the harmony.
From an economic point of view, the ecological problem is rather simple. Ecological damages are
externalities. The obvious solution consists of putting a price on the externality so that it can be
used optimally. The price adjustment would make up the cost that the externality may impose
upon society.59

However, on the other hand, in the same field (climate change economics), a very di↵erent vision
exists, as described, for example, in [Pottier, 2014]. In this story, climate threats are not mere
accidents, but rather consequences of a specific way of life that liberalism and globalisation fostered
around the world. This lifestyle is reliant on the consumption and exploitation of natural resources,
and an “out of sight, out of mind” attitude toward waste, which ultimately ends up in natural
habitats. This lifestyle inevitably leads to a mass depletion of resources and biodiversity. The
widespread combustion of fossil fuels releases greenhouse gases into the atmosphere at unprecedented
levels and provokes climate disruption. Overfishing and ocean acidification destroy marine species.
This ecological crisis is the result of a dysfunctional society.

These antinomic visions co-exist in the same discipline, with one being inspired by biology and the
other by economics. The gap between the two visions can be attributed to a deficiency in our tools
for understanding climate change from the economic perspective, and also to an epistemological
mixing of values, as is shown in [Pottier, 2014]. Generally speaking, three main tools can be utilised
to reframe the ecological crisis according to its economic impacts: the discount rate, a damage
function of the cost-benefit analysis, and the cost-e↵ectiveness analysis.

• Discount rate: in most macro-economic models addressing climate, the discount rate repre-
sents the preference for the present, covers a wide range of possible values,60 and is a very
influential parameter on the output (see the controversy about the [Stiglitz et al., 2017] re-
port in [Godard, 2007]). Hidden behind the technical debate around the actual values of
the discount is an ethical question: should we value people living today more than future
generations [Beckerman et al., 2007]? For example, [Mertens and Rubinchik, 2012] state that
discount rate must equal the growth rate of real per capita income whereas [Roemer, 2011]
shows that “a very small discount rate (large discount factor) should be used”;

• Damage functions: they aim at assessing the economic impact of climate change through a
function depending usually on increases in temperature, as they impact production [Nordhaus,
2008] or labor productivity [Burke et al., 2015], but very large uncertainties remain about
59
This position is defended by important economists claiming that ecological issues can be addressed through a
relevant price. See, for example, the op-ed article of Tirole and Gollier who advocates for a global carbon price [Tirole
and Gollier, 2015], or the famous reports of [Stiglitz et al., 2017] or [Canfin et al., 2016].
60
As noticed by [Grandjean and Giraud, 2017], the Ramsey’s equation states that the discount rate must equal the
sum of the time-preference rate and the product between the growth rate of the economy by the risk-averse rate. But
the debate is still open about the values of these elements.

46
CHAPTER 1. INTRODUCTION

the parameter and the form of these functions, which leads to important variations between
scenarios such as those from [Nordhaus, 2008], [Weitzman, 2012] or [Dietz and Stern, 2015];
• Cost-e↵ectiveness analysis: according to [Pindyck, 2017], cost-benefit analyses, and especially
damage functions,“create a perception of knowledge and precision, but that perception is illu-
sory and misleading”. For this reason, numerous models are now involved in cost-e↵ectiveness
analysis through a disaggregation of economic sectors in order to avoid criticism on aggregated
production, and model in a finer way the relationship between production and climate.

Beyond the technical debate, [Pottier, 2014] shows how the canonical epistemology (mainly in-
spired by Milton Friedman) of economic analysis built a positive and normative economics on
the separation between facts and values. [Ferraton, 2008] contests this separation, and shows that
economists act as if they are reaching objective, positive results and conclusions when, in fact, they
are always tainted with philosophical elements. With the absence of axiological neutrality [Ferraton,
2011], inspired from the Weberian ideal of research motivation, one must be very careful with their
personal values, and not just their quantitative ones, when conducting research.

To follow these guidelines and contribute to research in economics in the fairest way, I will try to
be explicit, at each step of my work, about my assumptions and the values which underlie them.
My own perspective on climate change economics is that natural habitats are degraded because
of human activity, on the one hand. On the other hand, the society’s financialization introduces
strong biases into our understanding of our own system, making its reactions highly non-linear.
From the perspective of an economist, I agree with Minsky’s idea of an intrinsic instability of
capitalism. Moreover, I do believe, as the experiment showed us, that our mainstream models are
not able to capture the magnitude of the problem - both economically and ecologically. As a former
student of mechanical engineering, I also consider that our macro-economic models su↵er from their
parameter-sensitivity, and thus that their results are highly questionable.

Consequently, I will direct my research to a family of models able to take into account a “climate-
driven Minsky moment”: disequilibrium models. Furthermore, I will try to assess and improve their
robustness. I am deeply convinced of our need for robust macro-economic models not only to assess
the economic impacts of climate change, but more importantly, to identify the public policies best
able to redirect our system’s trajectory to a more sustainable path.

1.4 Research questions and executive summary

1.4.1 Research questions and axes
This work seeks to address three research questions with the objective of building a robust tran-
sition model that can take into account a “climate-driven Minsky moment”. These questions are:
what are the primary necessary features of a model to assess the macrodynamics of the transition?
How should we evaluate or model these features quantitatively (from an economic viewpoint)?
Which public policies are likely to guarantee a transition in this model with these features?

In this framework, these questions correspond to three research axes, each of them leading to the
publication of an article (working papers). These axes are the following:
• Quantifying the robustness of a model and comparing its sensitivity with those of other
macroeconomic models ;

47
CHAPTER 1. INTRODUCTION

• Extending an existing macro-economic model, [Grasselli and Costa Lima, 2012], by introducing
inventories, money velocity, government intervention in order to help the orientation of public
policies in conjuction with applied current models ;

• Assessing the impact of public policies in order to guarantee a transition between di↵erent
natures of capital in this model.

1.4.2 Thesis organization

This robust transition model must be able to provide information about public policies avoiding
either an attracting equilibria with brown capital only, or a collapse due to a “climate-driven Minsky
moment”. This building is based on a stock-flow consistent model with a non-linear dynamic leading
to multiple equilibria, that can integrate this Minsky moment.

To comply with the climate objective, we have to integrate climatic constraints into the macro-
economic model in order to describe transition trajectories (as noticed in [Canfin et al., 2016]). The
aim is to guarantee the best consistency between short-run analysis and long-term decarbonation
objectives. For this reason, according to [Grandjean and Giraud, 2017], macro-economic models
have to:

• Be Stock-Flow Consistent (a basic condition to guarantee the reliability of the model in terms
of accounting);

• Be Physic-Consistent (and especially with the physical laws, such as the 2nd law of thermo-
dynamics);

• Explicitly describe the money circulation, and not assume a priori a neutral money;

• Be back-tested and analyzed in terms of parameter-sensitivity;

• Avoid hidden or unfalsifiable variables.

In contrast with usual macro-economic models, the [Grasselli and Costa Lima, 2012] model is able
to avoid some of their multiple failures and to capture specific features we need in order to predict
possible collapse driven (e.g., by an explosive amount of debt). This [Grasselli and Costa Lima,
2012] model has been backtested and extended through e.g. the work of [McIsaac, 2016] and [Bovari
et al., 2018] with a climate feedback loop.

But one can highlight some criticisms to the model such as its exogenous growth model, its as-
sumptions about the Say’s Law,61 or its myopic nature (no expectations are introduced in the model)
or its input-dependency (parameter and initial values of state variables): all stability criterion are
satisfied under certain conditions on parameters.
61
Which equals output and demand, and consequently drives the expression of the investment behavioral function
according to the adage: “all profits are reinvested and all wages are consumed”. There is no inventories or bu↵er
between output and demand.

48
CHAPTER 1. INTRODUCTION

To respond to the criticism made against the original [Grasselli and Costa Lima, 2012] model
about its high parameter-dependency, Chapter 2 explores the robustness of the model with two pro-
duction functions (Leontief and CES), through a GSA with Sobol’ indices estimated by the [Saltelli,
2002] method.62 It also reviews some global SA about Dynamic Stochastic General Equilibrium
(DSGE), Real Business Cycle (RBC) models and Integrated Assessment Models (IAM). Chapter 3
extends the model by introducing money velocity (according to the Quantitative Theory of Money)
in a framework relaxing the Say’s law by the introduction of inventories (inspired by the work
of [Grasselli and Nguyen Huu, 2016]. Eventually, the transition model is built with two natures
of capital and can be generalized through the Putty-Clay framework (see [Akerlof and Stiglitz,
1969], [Cass and Stiglitz, 1969]). The simplest transition model, with two natures of capital only,
delivers yet stability criterion and possible levers through government intervention (in the sense
of [Costa Lima and Grasselli, 2014]). This Chapter 4 studies the long-run multiple equilibria and
stability criteria of the transition model with di↵erent types of government intervention.

Chapter 2: Building a global sensitivity analysis to quantify the robustness of macro-

economic models

In Chapter 2, the robustness of the [Grasselli and Costa Lima, 2012] model was tested in terms
of sensitivity to parameters. After a study of the various sensitivity analysis methods (see also
the appendix of Chapter 2), a Global Sensitivity Analysis (GSA) with Sobol’ indices estimated
by the [Saltelli, 2002] method was selected according to the features of the models: nonlinearities,
a small number of inputs, and modest computational costs. The process is based on two initial
sampling matrices generated by Latin Hypercube Sampling (LHS). A review of literature shows
that GSA has been also conducted on the DSGE, RBC, and DICE models [Nordhaus, 2008]. This
work was done with the support of the Alliance Program (Alliance Doctoral Mobility Grant) in
Columbia University Graduate School of Arts and Sciences, under the invitation of Prof. Joseph E.
Stigltiz.

With the Sobol’ indices estimated by the [Saltelli, 2002] method, the sensitivity of an output in
the [Grasselli and Costa Lima, 2012] model to its inputs is more balanced than the one in a DSGE,
RBC or IAM (DICE)63 model – mainly depending on a single parameter hard-to-assess. One can
slightly improve the model’s robustness by endogeneizing the main influential parameter as in the
case of a CES production function.

Eventually we also studied the original [Goodwin, 1967] model (see appendix A.5) and noticed
that the sensitivity of the original model depends on the same parameters as its extension. We
showed that its performance in terms of robustness are slightly better than the [Grasselli and
Costa Lima, 2012] model, which might be due to the smaller number of nonlinearities.

These results are quite reassuring that refining the model does not seem to a↵ect the deep nature
of its sensitivity. They might make us wonder about a possible compromise between the robustness
and the realism of a model to find an equilibrium between a very robust model unable to approximate
62
The selection of the method is based on a State of Art in this field (see appendix A.2). Note that the dependency
to initial values has been explored through the study of basin of attraction by [McIsaac, 2016]. Consequently we can
focus our study on parameters only.
63
See respectively [Ratto, 2008], [Harenberg et al., 2017], [Nordhaus, 2008], [Miftakhova, 2019], [Butler et al., 2014]
and [Anderson et al., 2014].

49
CHAPTER 1. INTRODUCTION

reality (such as the [Goodwin, 1967] model), and a very fitted and very sensitive model (such as
the DSGE of [Ratto, 2008]). From this point of view, the [Grasselli and Costa Lima, 2012] model
seems to be an interesting compromise, with a finer description of the economy and a more robust
structure.

Chapter 3: Money Velocity in an Imperfectly Competitive, Stock-Flow Consistent Dy-

namics

So far, the analytics of money has been cluttered by a number of paradoxes and impossibilities
obstructing access to a proper understanding of the interaction between money, its velocity and
macro-dynamics. We propose a stock-flow consistent dynamics in continuous time with imperfect
competition where these paradoxes can be solved. Money turns out to be non-neutral both in
the short- and the long-run. Its creation by credit banking allows its quantity circulating in the
economy, as well as the speed at which it circulates, to be endogenously determined by the need
to finance investment and consumption. In particular, the decline in money velocity observed in
several countries in the aftermath of the Great financial crisis can be analyzed.

Our set-up sheds new light on the links between money velocity and the debt-deflationary path
on which several countries seem to have been progressing despite unconventional monetary policies.
Low inflation, low real income growth, increasing private debt and a declining velocity of trades
turn out to be the hallmark of the macro-economic trajectories leading to a debt-deflationary long-
run crisis. We provide conditions under which, however, a low-interest rate monetary policy can
stimulate economic activity so as to escape from the liquidity trap.

Chapter 4: Directed Technical Change in a Stock-flow Consistent Dynamics with Gov-

ernement: from Brown to Green Capital

Now we explore the robustness of the model and its ability to integer money velocity and to
relax Say’s law, the aim of this chapter is to propose a simple model of energy transition from
fossil fuels to renewable energies. We consider here the extreme case of two natures of capital, one
brown (that needs fossil energy to operate), the other, green (free of fossil energy), which, however,
is less productive than the dirty one. This dual capital structure is embedded into a stock-flow
consistent macrodynamics where investment can be financed by debt and the productive sector
adjusts shifts in demand thanks to inventories. The transition is said to take place if the economy
does asymptotically operate with green capital only.

Absent any public intervention, the economy converges to a purely brown production sector:
unsurprisingly, renewable energies are kicked out as long as green capital remains less productive
than the brown one. The inertia a↵ecting the transition of aggregate investment from one type
of capital to the other, however, turns out to destabilize the debt-deflationary equilibria of the
dynamics. Finally, we pinpoint two main levers for a government intervention: increasing the energy
price via a carbon tax or the output-to-capital ratio of green capital via publicly subsidized directed
technical innovation. The first option runs the risk of reducing private profits, hence fostering
private debts. The second runs the dual risk of boosting public debt. We identify conditions under
which the transition to green capital takes place without leading to an overhang of debt, be it
private or public. Numerical simulations illustrate the properties of each equilibrium.

50
Chapter 2

Building a global sensitivity analysis

to quantify the robustness of
macro-economic models

Currently used macro-economic models are facing various defaults (see [Stiglitz, 2018]), especially
in a complex economic environment involving a large number of calibrated / estimated parameters
which leads to highly parameter-sensitive results (see [Grandjean and Giraud, 2017]). Moreover,
in social sciences, where most data are approximations, the impact of inputs’ uncertainties cannot
be considered negligible, and questioning the robustness of our macro-economic models becomes
more and more essential to conduct public policies in these fields [European Commision, 2009].
Consequently, building a model, especially in social sciences or macroeconomics, requires considering
three types of uncertainties:

• Uncertainty on the accuracy between model and reality (Sensitivity to representations);

• Numerical uncertainties corresponding to sampling choices;

• Parameter uncertainties (Sensitivity to parameters).

In the frame of the sensitivity to parameters, inputs’ uncertainties are usually modeled by prob-
ability density functions. We estimate the probability density of the input and try to quantify its
propagation through the model. Unlike local sensitivity analysis (LSA), global sensitivity analysis
(GSA) methods evaluate the e↵ect of this propagation while all other factors are varied as well
and thus they account for interactions between variables and do not depend on the stipulation of a
benchmark. Among the large number of available approaches1 , the method of global sensitivity in-
dices developed by Sobol’ (1993, 2001) is based on the ANOVA decomposition of a high dimensional
model representation [Kucherenko and Shah, 2007].
1
Such that, for example:
• Momentum Methods, where the output’s average and variance are compared to the probability density of
inputs (hard in highly non-linear model) ;
• Sampling Methods (which can be conducted either with a Monte-Carlo or a Latin Hyper Cube sampling).
These methods are very time-consuming and can miss some non-linearity phenomena ;
• Response Surface Methods (also called Emulators or Meta-Models), which are less time consuming and ensure
to capture combined parameters’ phenomenon.

51
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

This Sobol’ GSA is superior to other SA methods, such as those based on joint e↵ects or regression
coefficients, because it doesn’t require linearity assumptions. Moreover, it captures both individual
and interaction e↵ects of the input parameters on the output variance, overcoming drawbacks such
as the “curse of dimensionality” [Bellman, 1957]. Various efficient practical implementations of these
indices have been proposed in the literature (see [Saltelli, 2002] for a survey). In macroeconomics, a
number of SA already exists for DSGE models (see [Ratto, 2008], and [Harenberg et al., 2017] for a
survey of the methods or the SA made by Nordhaus about the canonical model [Nordhaus, 2008]).

Our sensitivity analysis is applied to a specific type of macro-economic models such as the [Gras-
selli and Costa Lima, 2012] model, where investment can be financed by debt. The main features
of this approach when compared to competing macro-economic models are its stock-flow consis-
tent methodology leading to multiple equilibria and its 3D non-linear dynamics describing the time
evolution of the wage share, employment rate, and private debt of the system.

This type of models, also called disequilibrium models (such as Three-Me, or [Akerlof and Stiglitz,
1969], or [Grasselli and Costa Lima, 2012] based on a Lotka-Volterra approach between wages
and employment rate2 ), can answer to some of the criticisms of a large majority of usual models
because of their global out-of-equilibrium non-linear dynamics. For a more detailed motivation of
the selection of this family of models, we refer the reader to the introduction.3

With this family of models, we can indeed study trajectories towards various equilibria, be they
“good” or “bad” (e.g. in the case of a crisis), their basin of attractions [McIsaac, 2016] and manage
public policies to stay in it [Costa Lima and Grasselli, 2014]. Many extensions are available to
cope with fundamendal economic aspects: money velocity can be introduced to overcome neutral
or exogenous money, we can consider di↵erent trajectories for private and public debts, integrate
climate feedback loops [Bovari et al., 2018] and associated transition mechanism or observe the
non-linear inventory dynamics of the model without Say’s Law [Grasselli and Nguyen-Huu, 2018].

However, the output values of the model (wage share, employment rate and debt ratio of the
system), the stability criterion of the considered equilibra and the dimensions of their basins of
attraction are obviously dependent on certain parameters of the system. This motivates a dedicated
sensitivity analysis to understand the possibility for a decision-maker to influence certain inputs in
order to drastically transform the trajectory of the system.

The rest of this study is organized as follows. Section 2.1 is dedicated to a brief review of the
features of the model [Grasselli and Costa Lima, 2012]. Then an overview of the various SA methods
is conducted in Section 2.2 in order to motivate the selected method - a global sensitivity analysis
(GSA) with Sobol’ indices - according to the characteristics of the model. Section 2.3 contains our
application of the GSA to the model and concludes about its macroeconomic consequences and
the possible improvements we should make, i.e. endogenizing growth or relaxing the assumption
of constant capital-to-output ratio. In Section 2.4, we propose a qualitative comparison with other
GSA conducted on various types of macro-economic models such as Dynamic Stochastic General
Equilibrium (DSGE) model [Ratto, 2008], Real Business Cycle (RBC) model [Harenberg et al.,
2017] or Integrated Assessment Model: DICE [Nordhaus, 2008].
2
Inspired from [Goodwin, 1967] and developed by [Keen, 1995].
3
See also Appendix A.1 for the justification of our assumption on the model.

52
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

2.1 General structure of the [Grasselli and Costa Lima, 2012]

model
The model is a 3D non-linear dynamical system describing the time evolution of the wage share,
employment rate, and private debt of the system. It proposes a stock-flow consistent approach
leading to multiple equilibria where investments can be financed by private debt. Two behavioral
functions are assumed between wages and employment on the one hand, and, on the other hand,
about investment depending on the net profit. A Leontief’s production function is assumed involving
a fixed capital-to-output ratio. Consumption, here, is only an adjustment variable. We refer the
reader to see appendix A.1.1 and the above-mentioned article for comprehensive information about
the properties and construction of the model. In this section, we only present a summary of the
essential elements for the continuation of our study.

2.1.1 Some classical assumptions

First, a classical capital’s accumulation equation is assumed, i.e. that K̇, the change in capital
stock K, is depending on the investment I and a fixed depreciation rate :

K̇ = I K . (1)

Assuming the Say’s Law, total sales demand Yd is equal to the total yearly output Y (so consumption
C acts only as an adjustement variable):

C + I = Yd = Y , (2)

where ⇧ is the net profit for firms after paying wages, W .

As in [Grasselli and Costa Lima, 2012], we consider the following Leontief production function:
⇢
K
Y = min ; aL , (3)
⌫

which means, when assuming full capital utilization: Y = K⌫ = aL, where ⌫ is a constant capital-to-
output ratio, L is the number of employed workers and a is the labour productivity, that is to say,
the number of units of output per worker per year. All quantities are assumed to be quoted in real
rather nominal terms, thereby already incorporating the e↵ects of inflation, and are net quantities,
meaning that intermediate revenues and expenditures are deducted from the final yearly output.
Let the labor productivity be given by a(t) = a0 e↵t and the total labour force by: N (t) = N0 e t ,
L(t)
we then define the employment rate: = N (t)

2.1.2 Stock-flow consistency

2.1.3 Aggregate behavioural rules
The rate of change in real wages w is an increasing function of the employment rate :

ẇ = ( )w (4)

where (·) is known as an increasing short-run Philips curve (see, e.g., [Gordon, 2011] for a historical
survey, [Gordon, 2013], [Mankiw, 2001] and [Mankiw, 2016], as well as [Grasselli and Nguyen Huu,
2016]). Moreover, we define the wage share: ! := W wL
Y = aL = a .
w

53
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

Inspired from [Minsky, 1982], and formalized by [Keen, 1995], the aggregate investment is ex-
pressed as a function, (·) of the net profit share:

I = (⇡)Y, (5)

where the net profit share, ⇡ := Y⇧ , with ⇧ the net profit after paying wages and interests on debt
(i.e. ⇧ := Y W rD, where D is the amount of debt in real term and r a constant interest rate).
This implies that the growth rate for the economy in this model, Ŷ , is:
Ẏ K̇ I K (⇡)
Ŷ := = = = . (6)
Y K K ⌫

As in Minsky’s and Kindleberger’s theories, current cash-flows validate past liabilities and form
the basis for investment decisions.4 The change in real debt, D, is thus :

Ḋ := I ⇧. (7)

The dynamical variables in this model are so the wage share !, the employment rate and the
debt ratio d := DY . Thus the di↵erential system is:
8
> !˙ ẇ ȧ
> ! = w a = ( ) ↵ , !˙ = ! ( ( ) ↵)
>
>
>
>
>
>
>
>
>
> ✓ ◆
<˙ Ẏ ȧ Ṅ (⇡) (⇡)
= = ↵ , ˙ = ↵ (8)
>
> Y a N ⌫ ⌫
>
>
>
>
>
> ✓ ◆ ✓ ◆
>
>
>
> d˙ Ḋ Ẏ (⇡)Y ⇡Y (⇡) ˙ (⇡)
: = = , d = (⇡) ⇡ d
d D Y D ⌫ ⌫

Under the quite reasonable conditions presented in [Grasselli and Costa Lima, 2012], the dynam-
ical system admits only three equilibria:
• a so-called Good Equilibrium – corresponding to a desirable situation with positive equilibrium
final values of wages and employment, and a finite level of debt ratio, (!1 ; 1 ; d1 ), – which is
locally stable under certain conditions on the set of parameters ;
• a Slavery Equilibrium with employment but no wages, which is structurally unstable, and so
will not be studied further ;
• an Explosive-debt Equilibrium (also called the Bad Equilibrium) associated with a skyrock-
eting level of debt ratios while wages and employment shrink to zero. This long-term steady
state is also locally stable for a broad family of parameters.

Figures 2.1a and 2.1b are the trajectories of the two locally stable equilibria of the system con-
sidering a short-run Phillips curve of the form, ( ) = (1 1 )2 0 , and the following investment
function: (⇡) = k0 + k1 ek2 ⇡e , as in [Grasselli and Costa Lima, 2012] and with parameters specified
below (see Table 2.1).5
4
For simplicity again, the model assumes that non-financial firms injects all their profits, ⇧, for investment purposes
(no redistribution of their expected profit to their shareholders).
5
As exposed in [McIsaac, 2016] based on World Bank Data, Penn World Table and BIS Statistics Explorer.

54
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

Symbol & Definition Value Symbol & Definition Value

0.04
↵ labor productivity growth rate 0.025 0 Phillips curve ( ) parameter 1 0.042
0.043
population growth rate 0.02 1 Phillips curve ( ) parameter 1 0.04
depreciation rate 0.01 k0 Investment function (⇡) constant -0.0065
⌫ capital-to-output ratio 3 k1 Investment function (⇡) parameter e 5
r short-run nominal interest rate 3% k2 (⇡) exponential parameter 20

Table 2.1: Canonical parameterization of [Grasselli and Costa Lima, 2012].

(a) Trajectory from initial values: !0 = 0.8, 0 = 0.9 (b) Trajectory from initial values: !0 = 0.7, 0 = 0.7
and d0 = 0.1 to the Good Equilibrium with final values: and d0 = 0.1 to the Explosive-debt Equilibrium with
(!1 , 1 , d1 ) = (0.8361, 0.9686, 0.0702) final values (0, 0, 9 · 10101 )

Figure 2.1: Trajectories leading to the locally stable equilibria of system of the [Grasselli and
Costa Lima, 2012] model with the set of parameters defined in Table 2.1.

[McIsaac, 2016] studied the basins of attraction6 of these equilibria. The dimension of the basins
are depending on the set of parameters, as they are modifying the condition of stability (Routh-
Hurwitz criterion). Figure 2.2 represents the basin of attraction of the Good Equilibrium7 .8

6
I.e. all initial values of input’s variables leading to this equilibrium according to our choice of parameters.
7
Figure displayed with the .R code of [Augier, 2018]
8
One can notice that the initial value of (!, ) leading to the Good Equilibrium (green points of the Figure 2.2) are
in a neighborhood of this equilibrium (the equilibrium values are detailed in Fig. 2.1a: (!1 , 1 ) = (0.8361, 0.9686).
Note that, for the debt ratio the range of possible initial values leading to the Good Equilibrium is quite large compared
with the one of ! or .

55
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

Figure 2.2: Basin of attraction of the Good Equilibrium for the set of parameters defined in Table
2.1 ; each green point is a set of initial values for the inputs (!; ; d) leading to the Good Equilibrium
; the rest of the space leads to the Explosive Debt Equilbrium

2.2 Motivation of our uncertainties analysis (UA) and sensitivity

analysis (SA) methods
Numerical models are now widely used to study complex systems involving large numbers of
parameters, be it in engineering, computation science, environmental science, macroeconomics or
finance. Their ability to faithfully represent the reality depends both on empirical studies and
theoretical considerations.

As an example, behavioral functions9 are built to fit data, involving a large number of calibration
or estimation (based on empirical studies) which leads to highly parameter-sensitive results [Grand-
jean and Giraud, 2017], whereas production function10 is a key concept of the economic growth
theory, and thus can present discrepancies between the actual production laws and their functional
representation (see [Stiglitz, 2018]).

These di↵erent types of model’s inputs su↵er from various nature of uncertainties and any pre-
vision should be undertaken into a probabilistic frame to take them into account.11 This is the
9
Such as the Philips’ curve or the investment behavioral function in [Grasselli and Costa Lima, 2012].
10
Such as [Cobb and Douglas, 1928], Putty-clay in the sense of [Akerlof and Stiglitz, 1969], CES [Solow, 1956]
or [Leontief, 1951] as used in [Grasselli and Costa Lima, 2012].
11
To provide high-quality controls of model-based previsions, SA has been confirmed by [European Commision,
2009] in a working document (The EC handbook for extended impact assessment which states that a good sensitivity
analysis should conduct analyses over the full range of plausible values of key parameters and their interactions, to
assess how impacts change in response to changes in key parameters.) and the journal Science Online reviewed
by [Saltelli et al., 2008a] highlights the importance of SA and UA in falsifying or corrobating a simulation model.

56
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

principle of an uncertainty and sensitivity analysis.12

Uncertainties analysis (UA) studies the propagation of uncertainties in the model by assessing
the uncertainty on the outputs due to uncertainties on the inputs, whereas sensitivity analysis is
designed to assess (qualitatively or quantitatively) the respective contribution of each inputs to the
output uncertainties.13

We refer the interested reader to appendix A.2.1 and A.2.2 for more details on UA et SA, including
a survey of classical tools and a comprehensive justification of the approach used in this chapter to
analyse 3D non-linear dynamical systems.

2.2.1 Definition and selection of uncertainty and sensitivity analysis

When building a simulation model, many sources of uncertainty can indeed arise. To improve the
reliability of the process, it is necessary to understand these risks. According to [Macdonald et al.,
1999] and [De Wit, 2001], the sources of uncertainty can be classified as follows:

• Uncertainty on the accuracy between model and reality: This uncertainty appears
when building the mathematical representation of the real economic phenomenon. Mathemat-
ics are an approximation of the economic reality based on restrictive assumptions, such as, for
example, rational agent, Say’s Law, or the production function used in our macro-economic
models.

• Numerical uncertainty corresponding to sampling choices: In macroeconomics, as

in several other fields, mathematical models are often described by partial derivative equa-
tions, without explicit closed-form solutions. Numerical resolution of these equations involves
an approximation of the continuous model by finite di↵erence schemes. With smaller dis-
cretizations,14 one can reduce the impact of this uncertainty, but we can not totally discard
it.

• Parameter uncertainty: macro-economic models are usually built with a high number of
calibrated or estimated parameters, involving uncertainties about their actual values. The
knowledge of these parameter uncertainties and the quantification of their impact on the
outputs of a model is at the core of the so-called “propagation of uncertainties” [Saltelli et al.,
2008a] and [Bouleau, 2022].15
12
See appendix A.2 for a definition and appendix A.2.1 and A.2.2 for a review of these various techniques
13
See e.g. [Lam and Hui, 1995] , [Macdonald et al., 1999], [Helton et al., 2006], [Saltelli et al., 2008a]. The di↵erence
between the definition of uncertainty analysis (UA) and sensitivity analysis (SA) is stated by [Helton et al., 2006] as
that: ”Specially, uncertainty analysis refers to the determination of the uncertainty in analysis results that derives from
uncertainty in analysis inputs, and sensitivity analysis refers to the determination of the contributions of individual
uncertain inputs to the uncertainty in analysis results.” UA and SA are most often run together, where the SA is used
to rank the uncertainty sources (identified by UA), according to their influences on outputs [Saltelli et al., 2008a].
Another usual definition of a SA method is the following: The study of how uncertainty in the output of a model
(numerical or otherwise) can be apportioned to di↵erent sources of uncertainty in the model input [Saltelli et al.,
2008b].
14
In the field of digital signal processing, the Nyquist-Shannon signal sampling theorem gives a criterion about the
minimal sampling frequency according to the maximal signal frequency to avoid distorsion of the information [Shannon,
1949].
15
See also [Bouleau, 2008] and [Bouleau, 2019].

57
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

To assess the impact of uncertainties on outputs, we have to propagate the uncertainties through
the model. We so choose a technique to explore the numerical model. These inputs’ uncertainties are
usually modeled by a probability density functions, used widely since the 70s [Sobol, 1967] [McKay
et al., 2000].

Among the various methods to conduct an uncertainty analysis (UA),16 we selected a stratified
sampling with a Latin Hypercube Sampling (LHS) method in order to maximize the inputs space
coverage17 and because of its features when dealing with very non-linear model, without linear
approximations, and its ability to be optimized - as detailed in [McKay et al., 2000] or [Saltelli,
2002].

For sensitivity analysis, a wide range of techniques is also available, but these techniques are
commonly grouped into local and global methods [Lomas and Eppel, 1992], which distinctions are
listed by [Hoes and De Vann, 2005].18 Local SA are considered to be computationally faster and
easier, but less accurate when compared to the sophisticated global SA methods. According to the
literature, it also has been stated that the global SA should be used, when input variables are in
a non-linear model and from di↵erent magnitudes of uncertainty sources [Cukier et al., 1973].19 .
To allievate the lacks of local SA, when studying very non-linear models, or exploring large inputs
spaces, one can conduct a global sensitivity analysis.20

Global sensitivity analysis (GSA) are indeed used for many di↵erent purposes, such as identifying
critical regions in the space of inputs, establishing importance of inputs factors in regard with their
impact on ouputs, simplifying models, etc. Most of these GSA in litterature are based on derivatives
@Yi
@Xi of an output Yi versus an input Xi , which can be thought as a mathematical definition of the
sensitivity of Yi versus Xi . The derivative based approach is very efficient in computer time, but
are unwarranted when model inputs are uncertain or when the model is of unknown linearity.

In ordre to help the practitioner to choose the most appropriate method for its problem and its
model, some authors have proposed decision trees. Figure 2.3 reproduces the flowchart of [De Roc-
quigny, 2008]. Although useful to fix some ideas, such diagrams are rather simple and should be
used with caution.

Among the various GSA methods21 and according to the decision tree above (Figure 2.3), we
selected a Sobol’s indices method (with the [Saltelli, 2002] estimation method22 ) in order to study
the [Grasselli and Costa Lima, 2012] macroeconomic (apriori non monotonic) model, with a non-
linear dynamics, partial derivative equations and around 10 inputs (parameters of the model),
16
See appendix A.2.1 for a description of these methods.
17
LHS are indeed generally used to improve the inputs space coverage, especially when processing is expensive due
to a high number of parameters or wide ranges of uncertainty.
18
See also in Table A.1 of appendix A.2.2.
19
LSA indeed focuses on the local impact of input variables on model outputs. It is usually carried out by assuming
the linear or monotonous model, or computing partial derivatives of the outputs, with respect to a small interval
fractional variation of an input variable around its normal value [Saltelli et al., 2008a].
20
See appendix A.2.2 for an overview of the di↵erent available methods.
21
See appendix A.2.5 for an overview of these methods and their features.
22
See appendix A.2.6 for an exhaustive description of this method and appendix A.2.6 for the motivation to use it
in order to estimate Sobol’ indices.

58
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

Figure 2.3: Decision diagram for the choixe of a SA method (from [De Rocquigny, 2008]

2.2.2 A global sensitivity analysis with Sobol’ indices

As recommended in [Saltelli et al., 2008b], the relative influence of the inputs variance on output
variance allows better results about the global sensitivity measure of a very non-linear system. This
is in order to exploit the statistical properties of variance [...] to investigate how factors contribute
to the variance. Interesting features of variance-based methods are:

• Model independence: the sensitivity measure is model-free ;

• Capacity to capture the influence of the full range of variation of each input factor ;

• Appreciation of interaction e↵ects among input factors;

• Capacity to tackle groups of input factors: uncertain factors might pertain to di↵erent logical
types, and it might be desirable to decompose the uncertainty according to these types.

The main drawbacks of variance-based measures are their computational cost, which is not a big
issue for our model (less than ten inputs) and the high number of possible methods, leading to
di↵erent rankings of factors according to their objectives. To avoid this kind of confusing results
we have to define a Setting. A setting is a way of framing the sensitivity quest in such a way that
the answer can be confidently entrusted to a well-identified measure [Saltelli et al., 2008b].

Inspired by the work of the chemists Cukier, I.M. Sobol’ proposed a Variance-based Global Sensi-
tivity Analysis through a straightforward Monte-Carlo implementation of the concept of computing

59
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

sensitivity for arbitrary group of factors. Given the following model Y = f (X): this ANalysis Of
the VAriance (ANOVA) aims at defining the share of variability of the output due to its inputs,
what is called Sobol’ indices.

In the frame of our study, we are interested in a Factor Prioritization Setting, to qualify the relative
influence of each input factors on the output. The application principle of the ANOVA with Sobol’
indices is described below, but the estimation of these indices remains challenging (especially for
second-order Sobol’ e↵ect, representing joint e↵ects between output) and is developed in appendix
A.2.6. However various successful studies were conducted about DSGE models (e.g. [Ratto, 2008]).

2.2.3 Construction of the Sobol’ indices

In the case of a linear model Y with p 2 N⇤ independent inputs (X1 , ... , Xp ), we have
p
X
Y = ai Xi , ai 2 R. (9)
i=1

Assuming the independence of the inputs,23 the variance of Y is given by

p
X
V (Y ) = a2i V (Xi ) (10)
i=1

and, in this case, the sensitivity of Y to Xi is usually quantified measuring the impact of Xi on the
total variance by the so-called Standardized Regression Coefficient:

V (Xi )
SRCi = a2i (11)
V (Y )
p
P
fulfilling by construction SRCi = 1.
i=1
What is more, in this particular case, we don’t have interaction e↵ects in the sense that the im-
pact of two or more inputs is just the sum of their single e↵ects. For a non-linear model, where
we can not apriori make hypothesis on the structure, we need to define sensitivity indices from a
well-chosen decomposition of the output variance.

23
The issue of GSA with dependent variables has been the object of intense recent research but is not considered in
our study as we make the classical assumptions of apriori independent inputs at a first step of the study. A prospect of
this work can be to check the independence of these inputs. We review here the di↵erent improvements made to take
into account dependent inputs. About Sobol’ indices, [Xu and Gertner, 2008] exhibits a Sobol’ indices decomposition
into dependent/correlated and independent/uncorrelated components for linear models.With this decomposition, [Li
et al., 2010] build the Sobol’ indices for a general model. An other technique is proposed by [Mara and Tarantola,
2011], using the Gram-Schmidt process to decorrelate the inputs variables. They also propose to define new indices
through the Sobol’ indices of the decorrelated problem. [Chastaing et al., 2013] provide a theoretical framework to
generalize the ANOVA decomposition to problems with dependent variables, based on the work of [Hooker, 2007]. In
contrast to the other works which focus on generalizing the ANOVA decomposition, [Kucherenko et al., 2012] develop
the Sobol’ indices via the law of total variance. Recent work of [Tarantola and Mara, 2017] considered estimating
Sobol’ indices with dependent variables using the Fourier Amplitude Sensitivity Test. In our study, inputs of the
model are the exogenous parameters of [Grasselli and Costa Lima, 2012], assumed to be independent. Thus we can
only consider the core method of the Sobol’ indices to assess its sensitivity to its inputs. But these extensions can be
useful when using e.g. a Taylor’s rule to endogeneize the interest rate r.

60
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

In his seminal paper [Sobol, 1967], Sobol proposed to expand f 2 L1 ([0, 1]p , dx) into summands of
increasing dimensions

p
X p X
X p
f (x1 , ..., xp ) = f0 + fi (xi ) + fi,j (xi , xj ) + ... + f1,...,p (x1 , ..., xp ). (12)
i=1 i=1 j>i

Imposing that 8s 2 {1, ..., p}, 8i1 < i2 < ... < is 2 {1, ..., p}s and 8k 2 {i1 , ..., is },
Z 1
fi1 ,...,is (xi1 , ..., xis )dxk = 0
0
this expansion of f into 2n summands of di↵erent dimensions exists, is unique and the related terms
are pairwise orthogonal in L1 ([0, 1]p , dx).

Suppose now that we can express a square integrable random variable Y as Y = f (X) where
X = (X1 , ..., Xp ) is a vector of p independent random variables uniformly distributed on [0, 1], we
have24

p
X p X
X p
Y = f0 + fi (Xi ) + fi,j (Xi , Xj ) + ... + f1,...,p (X1 , ..., Xp ). (13)
i=1 i=1 j>i

Consequently, using the pairwise orthogonality in the Sobol’ decomposition, we can obtain easily a
recursive expression of each term using conditional expectations

f0 = E(Y )
fi (Xi ) = E(Y |Xi ) E(Y )
fi,j (Xi , Xj ) = E(Y |Xi , Xj ) fi (Xi ) fj (Xj ) E(Y ) (14)
..
.

and deduce the so-called ANOVA-HDMR decomposition for the variance of Y

p
X p X
X p
V (Y ) = Vi + Vi,j + ... + V1,...,p
i=1 i=1 j>i

m
p p p (15)
X Vi X X Vi,j V1,...,p
1= + + ... +
V (Y ) V (Y ) V (Y )
i=1 i=1 j>i

where
24
Here the independence of the random variables (X1 , ..., Xp ) is a strong hypothesis necessary to obtain the under-
lying variance decomposition while we can obviously relax the assumption on the distribution remembering that for
a real random variable Z, FZ (U ) and Z are equidistributed when U is an uniform random variable on [0, 1] and FZ
the pseudo-inverse of the distribution function of Z. In the case of dependent variables, one can oppose the difficult
generalisation of the above recursive expressions of each term using conditionnal expectations. See the footnote 23
for more explanations.

61
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

Vi = V (E(Y |Xi ))
Vi,j = V (fi,j (Xi , Xj )) = V [E(Y |Xi , Xj )] V [E(Y |Xi )] V [E(Y |Xj )] (16)
..
.
Now, inspired by the linear case, we can define the Sobol indices: 8s 2 {1, ..., p}, 8i1 < i2 < ... <
is 2 {1, ..., p}s

Vi1 ,...,is
Si1 ,...,is = (17)
V (Y )

where Si is a measure of the single impact of Xi on the total variance, Si,j is a measure of the joint
impact of (Xi , Xj ) on the total variance, etc...

In this spirit, to quantity the total impact (sum of single and related joint e↵ects) of Xi on the
variance of Y we define

E(V (Y |Xĩ )) V (E(Y |Xĩ ))

ST i = =1 (18)
V (Y ) V (Y )

where Xĩ = (X1 , ..., Xi 1 , Xi+1 , ..., Xp ).

For example, in a model with 3 inputs we have

ST1 = S1 + S1,2 + S1,3 + S1,2,3 = 1 S2,3 . (19)

Properties of the Sobol’ indices are well known in the literature, and can be summarized as:
p
P P
1. Si1 ,...,is = 1
s=1 1i1 <i2 <...<is p

2. 8i 2 {1, ..., p}, Si , STi 2 [0, 1]

3. 8i 2 {1, ..., p}, Si  STi

The first two properties corresponds to the interpretation of the Sobol’ indices as relative contri-
butions to the variance of the output V (Y ). The third property expresses that Si as the contribution
of Xi is contained in STi as the contribution of Xi and all of its interactions.

Conclusion To summarize this section, our study will be based on a global sensitivity analysis
with a Latin Hypercube Sampling method to identify the inputs that are the most contributors
to the variance of the output (by comparing their total impact Sobol’ indices, STi ). We will then
conduct a global sensitivity analysis (GSA) on these main influent inputs in order to quantify their
contribution to the outputs uncertainty. This GSA will be conducted by comparing first-order and
total-e↵ect Sobol’ indices estimated with the [Saltelli, 2002] method.

62
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

2.3 Global sensitivity analysis of the [Grasselli and Costa Lima,

2012] model
The aim of this section is to test the robustness of the [Grasselli and Costa Lima, 2012] model, i.e.
here to analyze the global sensitivity of its outputs to uncertainties on its inputs. Consequently, we
first have to state the inputs/outputs of our study. We give details about the range of the inputs,
the sampling and the design of the experiment. From this point, we can estimate the Sobol’s indices
with a strong basis of various combinations of our inputs.

This GSA with the Sobol’ indices estimated by the [Saltelli, 2002] method will be conducted with:
1. A qualitative analysis with an Optimal Latin Hypercube Sampling to identify the main in-
fluential parameters of the [Grasselli and Costa Lima, 2012] model on the outputs - i.e. the
Good Equilibrium values of the wage share, the employment rate and the debt-to-output ratio.
This step investigates all the parameters and their joint e↵ect. The identification is done by
comparing the relative total-e↵ect Sobol’ indices of each input in subsection 2.3.3.

2. The refining of our meshing about the main influential parameters in order to capture a
quantitative assessment of the first-order, total-e↵ect and second-order (joint e↵ect) Sobol’
indices of these inputs. Refining is done with 200, 300, and 500 simulations which was sufficient
to guarantee a maximal standard deviation less than 0.05 (see subsection 2.3.4) .

2.3.1 Identification of the inputs/outputs of our study

The inputs of the [Grasselli and Costa Lima, 2012] model are both input’s variables (such as wage
share, employment rate or debt ratio) and ex-ante fixed parameters (such as the labor productivity
growth rate or the capital-to-output ratio). As the impact of input’s variables has been extensively
studied25 in the [McIsaac, 2016], [Bovari et al., 2018] and [Augier, 2018], especially through basins
of attraction, we will not studied it further here.

The inputs of our study are consequently the parameters of the model. At a first stage, we consider
the impact of all the parameters. The range of their possible values are chosen to guarantee the
convergence to the Good Equilibrium (see Table 2.2). These ranges might seem small, but wider
ranges does not guarantee a sufficient number of combinations leading to the Good Equilibrium, due
to its stability criterion. As one can show it, when running the sensitivity analysis with wider ranges,
the number of combinations avoiding the large basin of attraction of the Explosive Debt Equilibrium
shrinks drastically and therefore is not more relevant than smaller ranges in a neighborhood of the
Good Equilibrium we want to study.

In this work, we are only looking at the values of the Good Equilibrium (!1 , 1 , d1 )26 as outputs
of our study and how they can be influenced by uncertainties on the inputs. One can also be
interested in observing the impact of inputs on the Explosive Debt Equilibrium values, but as
(!, , d) ! (0, 0, +1), the examination of their variance is less meaningful, and is highly dependent
on the time horizon we consider. As the Slavery Equilibrium is structurally unstable, we do not
study it further.
25
But these studies take as assumptions certain values of parameters that guarantee the stability of the studied
equilibria. One can wonder about the uncertainties of these parameters and their potential impact on the outputs of
the model.
26
See equation (26) of the appendix A.1.1 for the expression of these final values at the Good Equilibrium.

63
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

Symbol Definition Min Value Max Value Range Nom. Value

↵ labor productivity growth rate 0 0.03 0.03 0.025
population growth rate 0.015 0.03 0.015 0.02
depreciation rate 0.005 0.03 0.025 0.01
⌫ capital-to-output ratio 2 4.5 2.5 3
0.04
0 Philips Curve ( ) parameter 0.038 0.042 0.004 1 0.042
0.043
1 Philips Curve ( ) parameter 6.333 E-5 7 E-5 6.67 E-6 1 0.04
k0 Investment function (⇡) constant -0.007 0.0004 0.074 -0.0065
k1 Investment function (⇡) parameter 0.0064 0.007 0.0006 e 5
k2 (⇡) exponential parameter 19.8 20.4 0.6 20
r nominal interest rate 1% 8% 0.07 3%

Table 2.2: Definition and Range of the inputs of our study

2.3.2 Sampling method: Latin Hypercube

One of the main issues of this type of study lies in the sampling choices: a too precise sampling
leads to very long rendering due to the skyrocketing number of simulations to launch27 ; and a too
wide one can miss some important phenomenon due to the nonlinearity of our model.

We compare here two types of Latin Hypercube Sampling: the so-called Random one and the
Optimal Latin Hypercube Sampling. Comparison here is made in regard with a Full Factorial
Sampling - when every possible combination of levels of each inputs is tested.

• The (Random) Latin Hypercube Sampling (LHS) is a Design of Experiment (DoE)

ensuring a maximum inputs’ space coverage [Iman, 1999]. It is a stratified statistical method
generating a near-random sample of parameter values from a multidimensional distribution.
According to [McKay et al., 2000], it confers to LHS a bigger precision than other stratified
sampling methods;

• The Optimal Latin Hypercube Sampling (OLHS) can be implement to deliver even
better coverage ensuring to capture the maximum of information with the minimum of com-
binations28 . As a comparison with more usual methods such as One-At-A-Time samplings
for every inputs, also called Fractional Factorial Sampling (or Full Factorial Sampling - FF),
the OLHS is known to be a much more efficient technique. These di↵erent techniques are
27
As an example, in the [Saltelli, 2002] method (see appendix A.2.6), with p inputs, and n combinations of the
input values to test, we need to build first two samples of the n-combinations, then run n ⇤ p combinations for the
! of the first-order Sobol’ indices, n ⇤ p combinations for the estimation of the total-e↵ect Sobol’ indices, and
estimation
p
n⇤ combinations for the estimation of the joint e↵ect. E.g. with 10 inputs and 100 combinations, it leads to
2
!
10
100 ⇤ 10 ⇤ 2 + 100 ⇤ 10 + 100 ⇤ 10 + 100 ⇤ = 8500 combinations. With no-parallelization, the running time of one
2
simulation is around 10s, i.e. a total time of 23h30.
28
These types of matrix are not reproducible (unless using the same random seed), because their construction is
based on a randomization, followed by a stochastic optimization process which aims at spreading points as evenly as
possible within the design space (according to an optimality criteria such maxi-min distance criteria).

64
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

detailed in [Burman and Plackett, 1946], improved by [Box and Behnken, 1960] and Taguchi
methods [Karna et al., 2012].

The concept of these DoE can be illustrated by Figure 2.4 for two parameters X1, X2. With the
same number of combinations, the FF explores only 3 levels (with redundant levels of a parameter),
missing thus some areas of the inputs space, whereas Random LHS guarantees that no combination
contains the same value for a given parameter and OHLS satisfies an evenly spreading.29

Figure 2.4: Comparaison of samplings with only three possible levels of two inputs

For our study with the OLHS, we choose a sampling of 100 levels for each parameter as it guar-
antees an inputs’ space coverage 95% (see Figure A.4b from [Iman, 1999]) and is more easily
reproductible than a Random LHS.

2.3.3 Sobol’ indices computation for the [Grasselli and Costa Lima, 2012] inputs
In this part, we present our results regarding the relative importance of the inputs according
to their first-order and total-e↵ect Sobol’ indices in order to capture joint e↵ects. This leads to
the identification of four main influential parameters: the labor productivity growth rate ↵, the
population growth rate , the depreciation rate and the capital-to-output ratio ⌫. One can notice
that the nominal interest rate r is very influential at first-order, but does not have many joint
e↵ects, and so is not a key parameter in terms of total-e↵ect Sobol’ indices.

Complying the [Saltelli, 2002] method to estimate the Sobol’ indices of each input of the [Grasselli
and Costa Lima, 2012] model, we conduct the calculation with the following steps:

1. We build two sample matrix X(1) and X(2) with the Optimal Latin Hypercube sampling
(100 combinations of inputs, ensuring an inputs’ space coverage 95%) ;

2. We build the dedicated matrix for the first-order, second-order and total e↵ect Sobol’ indices
with the [McKay et al., 2000] method (See appendix A.2.6).30 ;
29
In our study, considering, for example, (4 inputs ⇥ 7 levels =) 2041 simulations to run with FF Sampling leads
to almost the same standard deviations of the Sobol’ indices than 100 simulations of OLHS with ten replicates and is
consequently twice slower.
30
First-order Matrix contains 10 columns (10 inputs) and 100 lines (100 combinations of inputs), Total-e↵ect
! Matrix
10
contains also 10 columns and 100 lines, whereas Second-order Matrix contains 10 inputs and 100 ⇤ = 4500 lines.
2

65
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

3. We run the [Grasselli and Costa Lima, 2012] model with every combination of parameters (each
line of these matrix) and record the Good Equilibrium values: wage share !1 , employment
rate 1 and debt ratio d1 ;
4. We compute the Sobol’ indices by comparing the variance of this output vectors to the ex-
pected variance of these vectors with no perturbations, and we replicate the whole process 10
times in order to display the standard deviation of our results.

Our results are gathered in Table 2.3.31 We rank the parameters according to their total-e↵ect
Sobol’ indices, in order to take into account the possible joint e↵ects between them. This ranking is
of course di↵erent from the one at first-order (without joint e↵ects).32 According to the total-e↵ect
Sobol’ indices, the most influential parameters for all outputs are:
• The labor productivity growth rate ↵ with a mean total-e↵ect on all outputs about 0.861,
• The population growth rate : 0.788 ,
• The depreciation rate : 0.894,
• And the capital-to-output ratio ⌫: 0.8995.
Here it is important to note that this ranking is possible only because it is very stable regardless of
the chosen output. If this were not the case no overall conclusion would be possible
One can notice that the nominal interest rate which has an important first-order Sobol’ indices on
!1 , is not very influential at a global scale because its interaction with other parameters does not
lead to important variation of the outputs.

An interesting question here is to wonder if the most influential parameters we highlighted are
hard to estimate or calibrate. Indeed, a non-sensitive parameter that is difficult to estimate does
not represent a big problem for the robustness of the model, or a very sensitive parameter that is
easy to estimate or calibrate gives confidence in the results of the model, and conversely.

As detailed in [McIsaac, 2016], demographic data from United Nations time series are available
and can easily permit approximations through logistic function (but more rarely with an exponential
as in the [Grasselli and Costa Lima, 2012] - or only when considering periods around 50 years -
which can be the case in this family of models). This gives some confidence about the ability to
estimate the population growth rate, with a certain confidence.
31
Results about second-order Sobol’ indices (joint e↵ects between factors) are displayed in appendix A.3.
32
Indeed, as one can see in Table 2.3, the analysis of the most influential parameters on each output at first-order
gives dissonant information because it does not capture the joint e↵ects between parameters which are undoubtedly
important in this non-linear model. On the wage share, !1 , the most influential parameters are the capital-to-
output ratio ⌫, with a first-order Sobol’ indices about 0.243, the nominal interest rate r: 0.215, the population growth
rate : 0.118, and the constant from the investment function k0 : 0.095. This result can be surprising because one
can expect the labor productivity growth rate ↵ to be, one of the main first-order influential parameter. The reason
why it is not the case might be due to the feedback loops of the Di↵erential System containing proportional-integral
(PI) controls. Indeed, the [Grasselli and Costa Lima, 2012] model reacts as a Closed Loop Transfer Function with
proportional-integral control in its direct chain. In the field of mechanical engineering (and especially in process
control) PI controls are known to constrain static perturbation on the inputs (such the ones of ↵, 0 , or 1 ). On
the employment rate, 1 , the most influential parameters are the labor productivity growth rate ↵: 0.146, the
investment function constant k0 : 0.139, the parameter of its exponential shape k2 : 0.125 and the depreciation rate
: 0.122. On the debt ratio, d1 , the most influential ones are again the capital-to-output ratio, ⌫: 0.138, the
population growth rate, : 0.119, the labor productivity growth rate, ↵: 0.106, and the investment function constant
k0 : 0.094.

66
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

Infl.
First-order Si Total e↵ect SiT
Rank
Input Description !1 1 d1 !1 1 d1 #
labor
0.092 0.146 0.106 0.827 0.8995 0.855
↵ productivity 3
(0.0548) (0.0966) (0.0831) (0.1498) (0.0938) (0.0962)
growth rate
population 0.118 0.076 0.119 0.805 0.763 0.798
4
growth rate (0.0533) (0.0529) (0.0947) (0.1347) (0.1640) (0.1817)
depreciation 0.079 0.122 0.076 0.900 0.864 0.917
2
rate (0.0571) (0.0592) (0.0579) (0.0592) (0.1376) (0.0424)
capital-to- 0.243 0.094 0.138 0.925 0.864 0.910
⌫ 1
output ratio (0.0870) (0.0745) (0.0628) (0.0548) (0.1110) (0.0617)
Philips Curve
0.00706 0.032 0.0064 0.1464 0.156 0.1458
0 ( ) 10
(0.0048) (0.0440) (0.0064) (0.003) (0.0875) (0.003)
parameter
Philips Curve
0.00707 0.041 0.0057 0.1466 0.225 0.1458
1 ( ) 9
(0.0048) (0.0428) (0.0053) (0.003) (0.2030) (0.003)
parameter
Investment
0.095 0.139 0.094 0.252 0.254 0.231
k0 function (⇡) 8
(0.0699) (0.0702) (0.0597) (0.3132) (0.2733) (0.2783)
constant
(⇡) 0.063 0.066 0.077 0.325 0.347 0.369
k1 6
parameter (0.0432) (0.0631) (0.0581) (0.3472) (0.3466) (0.3404)
(⇡)
0.078 0.125 0.040 0.427 0.461 0.428
k2 exponential 5
(0.0386) (0.0929) (0.0480) (0.3952) (0.3776) (0.3824)
parameter
nominal 0.215 0.102 0.068 0.308 0.374 0.146
r 7
interest rate (0.0811) (0.0710) (0.0500) (0.0493) (0.2266) (0.0030)

Table 2.3: Sobol’ indices and standard deviations (with 10 bootstrap replicates) of all inputs (ranked
from #1 the main influent parameter to #10 the less influent one) of the [Grasselli and Costa Lima,
2012] model on the outputs: the Good Equilibrium values of wage share !1 , employment rate 1
and debt ratio d1

As also exposed in their paper, by averaging the time series of the labour productivity growth rate
(from 1990 to 2011 - World Bank and Penn World Data), a better approximation is obtained than
with the OLS regression. This can give some confidence in the possibility to use a fixed parameter
to estimate this parameter.

On the contrary, the capital-to-output ratio is hard to define and is undoubtedly variable (between
3 and 3.5 in the period 1990 - 2011 according the World Bank Data and Penn World Table.) and
the depreciation rate, , is often calibrated as an educated guess [Smets and Wouters, 2007]. One
can imagine that such difficulty to estimate them can impede the robustness of the model.

In the next subsection 2.3.4, we will quantify more precisely the influence of these parameters and
we will see in the next section 2.4, that the sensitivity of the model to these difficult-to-estimate

67
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

parameters is quite low compared to other more traditional models.

2.3.4 Convergence of the Sobol’ indices of the model

After identifying the most influential parameters we need to refine our meshing to quantitatively
assess their contribution to the variance of the outputs. Consequently we are exploring the most
influential parameters,33 by refining the meshing of this input’s space following the method exposed
in [Harenberg et al., 2017] to give a better assessment of the Sobol’ indices. As in [Harenberg et al.,
2017], we choose to stop this exploration when the standard deviation of every indices is less than
0.05.

This exploration is conducted with an increasing number of simulations. We need to go up to

500 simulations to ensure that the standard deviation of each indice is less than 0.05. The indices
and standard deviations for 500 simulations are displayed in Table 2.4.34 Figure 2.5 displays the
value of each Sobol’ indice and its standard deviation for each exploration of 200, 300, 400 and 500
simulations respectively.

N = 500 simulations First-order Si Total e↵ect SiT

Input Description !1 1 d1 !1 1 d1
labor productivity 0.088 0.133 0.105 0.948 0.933 0.950
↵
growth rate (0.0470) (0.0482) (0.0344) (0.0321) (0.0125) (0.0315)
0.077 0.070 0.090 0.835 0.869 0.875
population growth rate
(0.0421) (0.0260) (0.0376) (0.0463) (0.0454) (0.0431)
0.053 0.100 0.073 0.919 0.929 0.938
depreciation rate
(0.0227) (0.0248) (0.0447) (0.0421) (0.0322) (0.0171)
0.189 0.060 0.122 0.935 0.938 0.942
⌫ capital-to-output ratio
(0.0498) (0.0457) (0.0403) (0.0377) (0.0281) (0.0304)

Table 2.4: Sobol’ indices and standard deviations (with 10 bootstrap replicates and 500 simulations)
of the main influent inputs of the [Grasselli and Costa Lima, 2012] model on the outputs: the Good
Equilibrium values of wage share !1 , employment rate 1 and debt ratio d1

Figure 2.5a displays the convergence of the four most influential parameter on the three outputs
(!1 , 1 , d1 ). One can notice that the highest first-order Sobol’ indices is obtained for the influence of
the capital-to-output ratio, ⌫, to the wage share, !1 (see also Figure 2.6) with an indice around 0.189
(with a standard deviation around 0.0498). Figure 2.5b, 2.5c and 2.5d are an extraction of Figure
2.5a and displays the results for the wage share, the employment rate and debt ratio respectively.

This result is important for the rest of the study to understand if this model is qualitatively
more robust than others classical models (by comparing their highest first-order Sobol’ indice for
example), but also to see how we can improve the robustness of the [Grasselli and Costa Lima,
2012] model (by “endogenizing” the most influential parameter).
33
Labor productivity growth rate ↵; population growth rate ; depreciation rate and capital-to-output ration ⌫.
34
See in appendix A.4 for the results of each exploration.

68
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

(a) First-order indices of (↵ ; ; ; ⌫) on (!1 ; 1 ; d1 ) (b) First-order indices of (↵ ; ; ; ⌫) on !1 for

for various number of simulations various number of simulations

(c) First-order indices of (↵ ; ; ; ⌫) on 1 for (d) First-order indices of (↵ ; ; ; ⌫) on d1 for

various number of simulations various number of simulations

Figure 2.5: Convergence of the Sobol’ indices of the main influent inputs (↵ ; ; ; ⌫) on (!1 ; ; d1 )

Another remarkable result can be identified by studying the original [Goodwin, 1967] model:35 in
this simpler model, the most influential parameters are still the labor productivity, ↵, the population
growth rate, , the depreciation rate, , and the capital-to-output ratio, ⌫. This result seems to
confirm the idea that refining this model(e.g. by introducing a debt ratio as in [Grasselli and
Costa Lima, 2012]) does not fundamentally a↵ect the nature of its sensitivity.

Since the maximum first-order Sobol’ indice of the model depends on the uncertainty on the
capital-to-output ratio, one can decide to improve the robustness of the model by changing its
production function from Leontief to CES in order to “endogenise” this capital-to-output ratio.36
The results of this analysis shows that at first-order and for 100 simulations (10 bootstrap), the
[Grasselli and Costa Lima, 2012] model with CES production is slightly more robust to its inputs
than the same model with a Leontief production function. At a total-e↵ect order, the quadruplet
35
See appendix A.5 for the whole study.
36
See appendix A.5.1 for an exhaustive analysis.

69
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

(a) First-order Sobol’ indices and standard deviation of (↵ ; ; ; ⌫) on (!1 ; 1 ; d1 )

(b) Total-e↵ect Sobol’ indices and standard deviation of (↵ ; ; ; ⌫) on !1

Figure 2.6: Sobol’ indices of [Grasselli and Costa Lima, 2012] for the most influential parameters
and their standard deviation.

(↵, , , ⌫) is replaced by (↵, , , C) which seems coherent with the sensitivity of [Grasselli and
Costa Lima, 2012] be it with a Leontief or a CES production function.

70
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

2.4 Comparison with other macro-economic models

In the economic literature, a number of sensitivity or uncertainties analysis are conducted with
various methods (mainly Local or Bayesian approaches), but a very little number of them are Global
Sensitivity Analysis (GSA). The European Commission in its [European Commision, 2009] report
highlighted however its importance in order to conduct public policies. [Ratto, 2008], for the Joint
Research Center of the European Commission, delivers a GSA with the DYNARE code [Adjemian
et al., 2011] of a core DSGE model: the [Lubik and Schorfheide, 2007] model.

[Harenberg et al., 2017] and [Harenberg et al., 2019] are successful studies comparing di↵erent
techniques to analyze sensitivity of a canonical Real Business Cycle (RBC) model, and especially
through GSA. [Miftakhova, 2019] and [Miftakhova, 2021] applied the same GSA method (with Chaos
Polynomial Expansion) about the most commonly employed Integrated Assessment Model (IAM),
DICE [Nordhaus, 2008]. The author states that, when applied to a selection of parameters, the
method produces results consistent with the conclusions of [Nordhaus, 2008]. Full analysis however
changes drastically the ranking and the relative impact of the inputs with the most influential
parameters being among those omitted from the aforementioned restricted setting (confirmed by
[Anderson et al., 2014], [Butler et al., 2014]).

In this part, we qualitatively compare the first-order Sobol’ indices of [Grasselli and Costa Lima,
2012] with these models. This comparison should be taken with caution as the quantities of interest
of each model are di↵erent.

2.4.1 Review of the [Ratto, 2008] GSA about the [Lubik and Schorfheide, 2007]
model
The [Lubik and Schorfheide, 2007] model is a a small-scale, structural general equilibrium model of
a small open economy (SOE), in the variables aggregated output (yt ), CPI (Consumer Price Index)
inflation rate (⇡t ), nominal interest rate (Rt ), terms of trade (qt ), exogenous world output (yt⇤ ),
world inflation shock (⇡t⇤ ) and nominal exchange rate (et ). It is a simplified version of the [Galiánd
Monacelli, 2005] model. As explained in [Lubik and Schorfheide, 2007], the model consists of a
forward-looking (open economy) IS (Investment-Savings)-curve and a Phillips curve. Monetary
policy is described by an interest rate rule, while the exchange rate is introduced via the definition
of the CPI and under the assumption of Purchasing power parity (PPP). Specifically, the evolution
of the SOE is determined by the following equations. The consumption Euler equation is rewritten
as an open economy IS-curve:

yt = Et yt+1 [⌧ + ↵(2 ↵)(1 ⌧ )](Rt Et ⇡t+1 )

1 ⌧ ⇤
(20)
↵[⌧ + ↵(2 ↵)(1 ⌧ )]Et qt+1 ↵(2 ↵) yt+1 Et zt+1
⌧
where ↵ 2 [0; 1] is the import share, and ⌧ the inter-temporal substitution elasticity, zt is the growth
rate of an underlying non-stationary world technology process At .37
37
In order to guarantee the stationarity of the model, all real variables are therefore expressed in terms of percentage
deviations from At . See [Lubik and Schorfheide, 2005] for further discussion of such a specification. Note also that
optimal price setting of domestic firms leads to the open economy Phillips curve:
k
⇡t = Et ⇡t+1 + ↵ Et qt+1 ↵ qt + (yt ȳt )
⌧ + ↵(2 ↵)(1 ⌧) (21)
and ⇡t = et + (1 ↵) qt + ⇡t⇤

71
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

Assuming monetary policy is described by an interest rate rule, the authors includes nominal
exchange rate depreciation et in the policy rule:

Rt = ⇢R Rt 1 + (1 ⇢R )[ 1 ⇡t + 2 yt + 3 e t ] + "R
t (22)

where 1 , 2 , 3 > 0 are policy coefficients, ⇢R 2 [0, 1] a smoothing persistence term and "R
t an
exogenous policy shock.

The GSA method applied in [Ratto, 2008] is a High Dimensional Model Representation (HDMR)
[M. Sobol, 1990] of the DSGE model described as:

Et (g(yt+1 ; yt ; yt 1 ; ut ; X)) =0 (23)

where yt is the vector of endogenous variables, ut the vector of exogenous shocks, and X =
(X1 ; ...; Xp ), the array of structural parameters they study. They estimate the state dependent
regression coefficients of their truncated HDMR, by a procedure of recursive filtering and smooth-
ing estimation (Kalman filter) based on a Monte-Carlo sampling of their data.

The study shows the use of the HDMR to characterize the relationship between the reduced form
coefficients of this rational expectation model and the structural coefficients. They explain that
“the polynomial parameterization of the SDR estimation of the HDMR provides the analyst with a
direct, albeit approximated, analytic representation of the solution procedure of rational expectation
models”.

They analyse in particular the reduced form coefficients describing the relationships between ⇡t
vs. Rt 1 . The conclusions of their study are that their SDR estimation explains more than 99% of
the total variation of the log-transformed coefficient (within sample).

The largest impact is given by ⇢R (smoothing persistence term of the nominal interest rate),
with a first-order Sobol’ indices Si = 0.896, followed by k (a slope coefficient treated as structural
parameter even if it is defined as a function depending on underlying parameters capturing the
degree of prices stickiness, labor supply and demand elasticities) with a Si = 0.06 and 1 , a policy
parameter, with Si = 0.04. All the remaining parameters have an irrelevant e↵ect.

As a conclusion, we can underline that a model with a first-order Sobol’ indices equal to 0.9 is
highly likely to be sensitive to uncertainties on the input responsible for this variation. If the alluded
input is moreover a smoothing persistence parameter hard to estimate empirically, one can wonder
about the global robustness of such a model and its recommendations, especially when working to
orientate public policies of the European Commission. Of course, we can hardly compare our results
to this study, as the quantity of interest (⇡t vs. Rt1 ) is very di↵erent from ours.38
where ȳt = ↵(2 ↵)(1 ⌧ )/⌧ yt⇤ is potential output in the absence of nominal rigidities, and k > 0 a slope coefficient,
depending on underlying structural parameters, such as labor supply and demand elasticities and parameters capturing
the degree of price stickiness. In order to study exchange rate policies, the authors introduce the nominal exchange
rate et via the definition of the CPI and assumed that relative PPP holds, leading the previous second equation,
where ⇡t⇤ is a world inflation shock, treated as an unobservable.
38
Sobol’ indices of the [Grasselli and Costa Lima, 2012] model are less than 0.2 for each quantity of interest.

72
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

2.4.2 Review of the [Harenberg et al., 2017] GSA about a canonical RBC model
The aim of this article is to transfer the recent advances in sensitivity analysis methods from
the engineering and mathematic fields to economics. It gathers and compares di↵erent methods
and especially the GSA with Sobol’ indices, applied to a canonical Real Business Cycle model with
capital adjustment cost. The model was widely studied (see e.g. [Aruoba et al., 2006] or [Den Haan
et al., 2011]) and focus, following the economic literature, on two quantities of interest: average
production and the ratio of the variance of log production in the model over its empirical counterpart.
The estimation of the indices relays on the chaos polynomial expansion (CPE) approximation to
identify interactions and calculate statistical moments as well as the probability density of the
quantities of interest. The allocation problem is described by a dynamical optimization:
1
X 1
t (ct (1 lt ) 1 )1 ⌧
V (kt ; at ) = max 1 E0 1 (24)
{ct ;lt ;it }0
t=0
1 ⌧

where ct is the consumption, 1 lt is leisure in each period, the discount factor, ⌧ the inter-
temporal elasticity of substitution (IES), and the share parameter in the composite commodity.
The decision variables are consumption ct , labor lt , and investment it . The aggregate resource
constraint is given by
✓ ◆2
it
qt = c t + it + k t (25)
2 kt
with qt the quantity of produced goods, kt the capital stock, and its depreciation rate. The
production technology is assumed to be: qt = eat kt↵ lt1 ↵ depending on at the productivity of the
factors, with a technical substitution rate ↵.39

The inputs of the study are the seven parameters of the system: the discount factor , the
IES ⌧ , the share parameter in the composite commodity , the technical substitution rate ↵, the
depreciation rate , the auto joint e↵ects coefficient ⇢, strength of adjustment cost , the standard
deviation of the normally distributed shock .

Two quantities of interest are observed as outputs: average production, Q1 = E[qt ], and the
variance of log production in the model over its empirical counterpart Q2 = 12 V ar[log(qt )] with q2
q
the empirical variance of log production in the data (commonly set at 0.019 in the literature). The
study displays the first-order and total-e↵ect Sobol’ indices of each input in relation with the two
quantities of interest (see [Harenberg et al., 2017] and the appendix of [Harenberg et al., 2019]).

In [Harenberg et al., 2017], the main influential parameter on the average production, E[qt ], is the
capital share ↵, with a first-order Sobol’ indice about to 0.75 and a total-e↵ect Sobol’ indice about
0.85. Here again, it is a high sensitivity for a model to a single parameter. One can also notice that
there is no big di↵erences between total-e↵ect and first-order Sobol’ indices in this model, which
means that it is not very non-linear.
39
Capital adjustment cost are modelled as in [Den Haan et al., 2011] with governing the size of cost. The capital
transition and stochastic productivity process are given by
kt+1 = it + (1 )kt
(26)
at+1 = ⇢at + "t+1

where ⇢ is an auto joint e↵ects coefficient and "t a normally distributed shock (std. deviation: ).

73
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

2.4.3 Review of the [Miftakhova, 2019], [Butler et al., 2014] and [Anderson
et al., 2014] GSA about DICE [Nordhaus, 2008]
Parameter uncertainty in Integrated Assessment Models (IAM) is highly debated but only par-
tially explored in the literature: expert opinions on the importance of particular inputs (let alone
their probability distributions) often diverge, hampering policy assessment and the decision-making
process ( [Tol, 1965], [Heal and Millner, 2014], [Athanassoglou, 2014]).

In this study, the authors implement a GSA with Sobol’ indices or density sensitivity indices
about the most commonly employed IAM: DICE [Nordhaus, 2008]. They compared their results to
the GSA developed in [Nordhaus, 2008]. The authors demonstrate the application of the method,
when applied to the same selection of parameters, produces results consistent with the conclusions of
[Nordhaus, 2008]. Full analysis, however, changes the ranking dramatically, with the most influential
parameters being among those omitted from the aforementioned restricted setting. We also show
that caution is needed when analyzing the full set of parameters because the credibility of the results
relies on the independence assumption.

DICE is a neoclassical growth model that resolves a trade-o↵ between consumption, investment,
and emissions reduction. Given a discount rate ⇢, the model maximizes total social welfare over its
time span,
TX
m ax

W = (1 + ⇢) t U (Ct , Lt ) (27)
t=1

with a constant-elasticity ↵ of the utility function of consumption Ct , and labor Lt .

(Ct /Lt )1 ↵ 1
U (Ct , Lt ) = Lt (28)
1 ↵
Production function takes the standard Cobb-Douglas form with Hicks neutral total factor produc-
tivity (TFP) At , labor Lt , capital Kt , as input factors. Total output Yt is reduced by the cost of
mitigation, a power function of emission reduction policy µ, and by the damage induced by the
climate change, ⌦t ,
Yt = ⌦t (1 ✓1 µ✓2 )At L1t Kt (29)
where the reduction in economic output caused by climate change is a function of the atmospheric
temperature anomaly TAt (measured as the increase in the atmospheric temperature from the
preindustrial level),
1
⌦t = (30)
1 + a1 TAt + a2 TAa3t
Given a depreciation rate , and investment It , capital accumulates according to:
Kt = (1 )Kt1 + It 1 (31)
The production of goods a↵ects the climate system by introducing anthropogenic CO2 emissions
at a rate determined by the emissions intensity t and reduced by the chosen rate of mitigation µt ,
the total emissions levels is the sum this anthropogenic and a natural source Etreet
Et = ECO2 t + Etreet
(32)
ECO2 t = t (1 µt )At L1t Kt

74
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

The core of the climate module is its carbon cycle with a three-layer model (atmospheric, upper
layer and lower ocean of CO2 concentrations) with transitions among layers:

MATt+1 = Et + b11 MATt + b21 MUt

MUt+1 = b12 MATt + b22 MUt + b32 MLt (33)
MLt+1 = b23 MUt + b3 3MLt
The net change in atmospheric CO2 concentrations indices the net energy flow to the atmosphere
(called radiative forcing, Ft ),
✓ ◆
MATt + MATt+1
Ft = ⌘ log2 + FEXt (34)
2MAT1750
where FEXt is the exogenous radiative forcing. This energy causes a change in the atmospheric
temperature, TAt , and, subsequently, the ocean temperature, TOt ,

TAt+1 = TAt (35)

[Butler et al., 2014] displayed how all inputs can be represented in a synthetic figure. Respecting
the selection made by [Nordhaus, 2008], [Miftakhova, 2019] and [Miftakhova, 2021]preselects the
same eight input parameters that presumably a↵ect the inference from the model the most, and
assume that they are normally distributed, as summarized in Table 2.5.

Parameter Definition Mean Std

a2 Damage function coefficient 0.0028 0.0013
b12 Carbon cycle transition 0.189 0.017
gA0 Initial growth rate of TFP 0.092 0.040
g0 Initial change of decarbonization -0.07 0.02
CCum Maximum extraction of fossil fuels 6,000 1,200
pBACK Cost of backstop technology 1.17 0.468
LASY M Asymptotic population 8,600 1,892
Equilibrium climate sensitivity 3.00 1.11

Table 2.5: Parameters selected for the analysis and their distributions (from [Nordhaus, 2008]).

For these eight parameters, the results of first-order and total-e↵ect Sobol’ indices are consistent
with [Nordhaus, 2008]: the damage function coefficient, a2 , contributes to the variance of the output
the most, followed by climate sensitivity, , and the initial growth rate of TFP, gA0 . Asymptotic
population size, LASY M , plays a far less significant role, while the e↵ect of the uncertainty in the
rest of the parameters is negligible.

As explained by [Miftakhova, 2019], the inference above relies on the preliminary subjective
selection of a set of parameters that potentially have a strong e↵ect on the SCC (Social Cost of
Carbon). Performing a generalized experiment by treating all parameters equally and assuming that
we do not have any information on their probability distributions, leads to very di↵erent results.
The computed Sobol’ indices reveal a notable di↵erence to the previous, restricted setting. The
biggest e↵ects are indeed attributed to the damage function exponent, a3 (first-order Sobol’ indice
about 0.3), capital elasticity, (about 0.16), and elasticity of marginal utility of consumption, ↵,
(about 0.12) all three missing in the subset selected for the previous analysis.

75
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

The resulting ranking therefore suggests that it can be highly misleading to rely on a prior
subjective judgment about the importance of input parameters. To infer a model’s sensitivity to
the uncertainty in its input factors it is necessary to consider their complete set and discriminate
among the parameters based on a full quantitative decomposition of uncertainty in the model. The
results of [Miftakhova, 2019] are in agreement with those of [Anderson et al., 2014].

We can not directly compare these various studies in terms of Sobol’ indices, as the outputs are
not the same, but one can notice that the outputs of some of these models are highly sensitive
(between 0.55 to 0.9 for the main quantities of interest of DSGE and RBC) to their inputs.40

A more precise study of the DICE model reveals that not only the ranking, but also the values
of the GSA of [Nordhaus, 2008] are biased, due to a prior selection of parameters [Butler et al.,
2014], [Anderson et al., 2014], [Miftakhova, 2019]. These new studies deliver values of maximal
Sobol’ indices about 0.3 when studying the output: Social Cost of Carbon 2015. But here again,
the main influential input is arbitrary imposed (damage function exponent a3 in [Nordhaus, 2008])
which can be drastically di↵erent (see [Bovari et al., 2018] for a comparison of various damage
functions).

2.5 Conclusion
In this study, we developed a sensitivity analysis of the parameters of a macro-economic model.
We study the relative influence of uncertainties on all inputs’ parameters of our model: the [Gras-
selli and Costa Lima, 2012] with a Leontief production function. We selected a variance-based
Global Sensitivity Analysis (GSA), because this model contains a large number of nonlinearities
and apriori independent inputs. Due to the small number of inputs of the model (and consequently
its low computational cost), we chose the Sobol’ indices method to conduct our GSA according to
the decision tree of [De Rocquigny, 2008] . The calculation of the Sobol’ indices is based on combi-
nations of inputs values obtained from a Design of Experiment (DoE): an Optimal Latin Hypercube
Sampling, according to the [Saltelli, 2002] method of estimation.

At a first step, we regarded all the possible inputs of the model and observed their impact on three
outputs: the Good Equilibrium value of the wage share, the employment rate and the debt ratio of
the economy. The most influential parameters on these outputs are the labor productivity growth
rate, the population growth rate, the depreciation rate and the capital-to-output ratio. The last
one has the maximal first-order Sobol’ indice (less than 0.2). These values are not very dependent
on the number of simulations as we show it by successively increasing the number of simulations
until the satisfaction of a certain criterion about the standard deviation of our results (less than
0.05).

One can notice here that two of these most influential parameters can be assessed with a certain
confidence as detailed in [McIsaac, 2016] (from demographic data from United Nations time series
for the population growth rate and by averaging the time series of the labour productivity growth
rate). On the contrary, the capital-to-output ratio and the depreciation rate are hard to assess that
can impede the robustness of the model.
40
By the way, these inputs (discount factor , capital share ↵, auto joint e↵ects coefficient ⇢ in [Harenberg et al.,
2017] or smoothing persistence of the interest rate ⇢R in [Ratto, 2008]) are generally hard to estimate, what can lead
consequently to questionnable results about the prediction of these models.

76
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

Then we displayed various similar GSA about DSGE, RBC and IAM model, and their maximal
first-order Sobol’ indices. We can not directly compare these results as the outputs and inputs of
each study are di↵erent, but we can question the reliability of the information delivered by these
models highly sensitive on an arbitrarily fixed parameter.

Indeed, our work has identified the most influential parameters of the model [Grasselli and
Costa Lima, 2012] to help modelers refine their work when building predictions with this model. On
the contrary of DSGE or RBC model of [Ratto, 2008] and [Harenberg et al., 2017] (which are very
dependent on a single parameter), we show that the sensitivity of the [Grasselli and Costa Lima,
2012] model is more balanced between four parameters (the capital-to-output ratio, the depreciation
rate, the population growth rate and the labor productivity growth rate).

We can make no quantitative conclusions about the relative robustness of a model compared
with an other one, because their respective robustness depends on the uncertainty on the most
influential parameters (the smoothing persistence for [Ratto, 2008], the capital share for [Harenberg
et al., 2017] and the four parameters of [Grasselli and Costa Lima, 2012]).

However, we can notice that the model observed in [Ratto, 2008] and [Harenberg et al., 2017] are
highly sensitive to a single parameter that is generally hard to assess. Modelers can now choose
to focus their e↵orts on refining the value of this parameter hard-to-assess when working with
these type of models or they can decide to work with [Grasselli and Costa Lima, 2012] where large
uncertainties in each of the four influential parameters in [Grasselli and Costa Lima, 2012] will be
less likely and have an equal impact on the variance of the output (the maximum share of the
impact is 0.19).

A prospect of the study can be to improve this balanced sensitivity of our model. One can
think about endogeneizing the most influential parameters (e.g. capital-to-output ratio and labor
productivity growth rate) either by using di↵erent production function, such as a Constant Elasticity
Substitution (CES) function (see [McIsaac, 2016] for the use of the Van der Ploeg’s extension) or a
Putty-Clay production function (see [Akerlof and Stiglitz, 1969]) instead of our Leontief production
function ; or by developing the model with an endogenous growth.

The study of the [Grasselli and Costa Lima, 2012] model with a CES production function is
provided in appendix A.5.1. At a total-e↵ect level (i.e. with all joint e↵ects between inputs), the
main influent parameters are the same as before. At a first-order level, the main influent parameter
on the wage share is now the nominal interest rate. An improvement of these performances is
possible by introducing a Taylor’ rule to endogeneize it.

Eventually we also studied the original [Goodwin, 1967] model (see appendix A.5) and noticed
that the sensitivity of the original model depends on the same parameters as its extension. This is
quite reassuring that refining the model does not seem to a↵ect the deep nature of its sensitivity as
we can improve it by multiple features about money, inventory or di↵erent natures of capital.

We showed that the performance of the original [Goodwin, 1967] in terms of Sobol’ indices are
slightly better than the [Grasselli and Costa Lima, 2012] model, which might be due to the smaller
number of nonlinearities. These results might make us wonder about a possible compromise between
the robustness and the realism of a model to find an equilibrium between a very robust model unable

77
CHAPTER 2. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

to approximate reality (such as the [Goodwin, 1967] model), and a very fitted and very sensitive
model (such as the DSGE of [Ratto, 2008]). From this point of view, the [Grasselli and Costa Lima,
2012] model seems to be an interesting compromise, with a finer description of the economy and a
more balanced sensitivity.

78
Chapter 3

Money Velocity in an Imperfectly

Competitive, Stock-Flow Consistent
Dynamics

Abstract
So far, the analytics of money has been cluttered by a number of paradoxes and impos-
sibility results obstructing access to a proper understanding of the interaction between
money, its velocity and macro-dynamics.
We propose a stock-flow consistent dynamics in continuous time with imperfect compe-
tition where these paradoxes can be solved. Money turns out to be non-neutral both in
the short- and the long-run. Its creation by credit banking allows its quantity circulating
in the economy, as well as the speed at which it circulates, to be endogenously deter-
mined by the need to finance investment and consumption. In particular, the decline
in money velocity observed in several countries in the aftermath of the Great financial
crisis can be analyzed.
Our set-up sheds new light on the links between money velocity and the debt-deflationary
path on which several countries seem to have been progressing despite unconventional
monetary policies. Low inflation, low real income growth, increasing private debt and a
declining velocity of trades turn out to be the hallmark of the macro-economic trajecto-
ries leading to a debt-deflationary long-run crisis. We provide conditions under which,
however, a low-interest rate monetary policy can stimulate economic activity so as to
escape from the liquidity trap.

3.1 Introduction
From 2008:Q1 to 2019:Q3, the M 1-velocity of trades in the US economy, measured as the number
of times a US Dollar has changed hand within one year through some trade in the real economy1 ,
has shrunk from more than 10 to less than 6. This decline is well approximated by an exponential
decreasing at the constant rate -1.4% (see Figure 3.1). Another example of a declining velocity is
provided by the Czech Koruna over the period 2008-2018, which fell from 2.5 to 1.5, hence losing
40% (see Figure 2).
1
Income velocity is measured, here, as the ratio between nominal gdp and the monetary aggregate, M1 .

79
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

Figure 3.1: Income velocity in US Dollars, 2008-2019.

More generally, during the ten years following the Great Financial Crisis, the following OECD
currencies have experienced an average slow-down of -1.3% in the speed at which they circulate in
their respective real economy: the US Dollar, the Euro, the Australian and the Canadian Dollar,
the Swiss Franc, the Swedish Krona, the Chilean Peso, the Korean Won, the Turkish Lira, the
Czech Koruna and the Zloty. Significant exceptions are provided by the Iceland Krona, the Indian
Rupee, the Brazilian Real, the Norwegian Krone and the Russian Ruble.

Figure 3.2: Income velocity in Czech Koruna, 2008-2018.

This illustrates the first of three stylized facts that seem to have characterized several oecd
countries in the aftermath of the Great Financial Crisis:

1. A sharp decline in money velocity,

2. A strong growth of the ratio private debt/gdp, and

3. A stagnation of real gdp itself — or, at best, a slow real growth rate compared to the paces
observed up to 2007.

Indeed, most of these countries experienced an significant increase of the private debt-to-gdp ratio:
Australia, Canada, Poland, Sweden, Switzerland, Turkey, Chile and 9 Euro Zone countries (Austria,

80
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

Belgium, France, Finland, Greece, Ireland, Luxembourg, Netherlands, Slovakia) have gone through
an exponential increase of this ratio at the average rate of 2.8% between Q1 2008 and Q1 2014. To
give just one example: in Canada, this ratio jumped from about 175% to more than 250% within
6 years (see Figure 3).

Figure 3.3: Private debt/GDP Canada, 2008-2014.

In order to put these magnitudes into perspective, we present a macro-dynamics which can be
viewed as a monetary extension of the Solow model (see [Van der Ploeg, 1985] and [Bastidas et al.,
2019]).2 At variance with the celebrated workshorse, here, the phase space admits at least two long-
run equilibria: the first one is but the embedding within our setting of the unique Solow equilibrium;
the second steady state is a debt-deflationary one à la [Fisher, 1933] where an overhang of private
debt leads to a sharp decline of money velocity and ultimately a fall in gdp. [?] had provided a
full characterization of debt-deflation and the liquidity trap induced by an excess of private debt
but their analysis remained entirely confined to a static equilibrium framework and money was
not viewed as being endogenously created by banking credit. Here, by contrast, we provide a
narrative of the path ultimately leading to a liquidity trap, along which several countries may have
been traveling in the recent past — and, as we shall shortly see, money exclusively arises from
endogenously determined banking credit.

A simple calibration of our model leads to evolution rates of both money velocity and debt ratios
that are quite close to the ones recently observed in the countries mentioned above.3 If confirmed,
it would mean that these countries are converging towards a debt-deflationary equilibrium.

3.1.1 Paradoxes associated with money

The theory surrounding money, however, is notoriously paved with various paradoxes which, so
far, have not been solved.

A. Banking credit creation

2
See also [Grasselli and Costa Lima, 2012], [Costa Lima and Grasselli, 2014], [Grasselli and Nguyen-Huu, 2018],
[Giraud and Grasselli, 2019] and [McIsaac, 2016].
3
Of course, there are exceptions e.g. Hungary, Island, Israël, Japan, Russia New Zealand and 5 Euro Zone countries
(Germany, Portugal, Latvia, Slovenia, Spain) but this seems to provide a pattern.

81
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

First, as shown by [Werner, 2016],4 during the past century, three di↵erent theories of banking
were dominant at di↵erent times: (1) The currently prevalent financial intermediation theory of
banking says that banks collect deposits and then lend these out, just like other non-bank financial
intermediaries. (2) The older fractional reserve theory of banking says that each individual bank is a
financial intermediary without the power to create money, but the banking system collectively is able
to create money through the process of ?multiple deposit expansion’ (the ?money multiplier’) fueled
by the money created by the central bank. (3) The credit creation theory of banking, predominant
a century ago, argues that each individual bank creates credit and money newly when granting a
bank loan. Since according to the dominant financial intermediation theory, banks are virtually
identical with other non-bank financial intermediaries, they are usually neglected in the models
used in economics or by central bankers. “Should this theory not be correct, as writes [Werner,
2016], currently prevailing economics modelling and policy-making would be without empirical
foundation.” Confirming what had already been assessed by [Werner, 2014a], [Werner, 2016] proves
that the financial intermediation and the fractional reserve theories of banking are rejected by the
evidence.
As concluded by [Werner, 2016], “one of the implications of this study is that it does not make
much sense to build economic theories of the financial sector, if these are not based on institutional
(and accounting) realities. The role of accounting and law in economics should be increased, both in
research and in the teaching of economics. This includes the role of national income accounting and
flow of funds information [...], which have to be reconciled with those records of the banks. These
are not only the ?central settlement bureau, a kind of clearing house or bookkeeping centre for the
economic system? ( [Schumpeter and Nichol, 1934], p. 124), but also the creators and allocators
of the money supply. The reflection of empirical bank reality within theories and textbooks surely
must become the ?new normal’ in finance and economics.”
This paper aims at paving the road towards such a renewal of economic modelling by endorsing the
credit creation theory of banking within a stock-flow consistent setting.5

B. Instability of inflation

Second, monetarist theories assume that money supply can anchor prices through the famous equa-
tion of exchange first expressed by ( [Fisher, 1912]):

M v = pT, (1)
where M is money, p is the aggregate price index, T is real output, and v stands for money velocity,
measuring how rapidly the value of output pT “turns over” with respect to the money supply, M .
Most monetarist models assume from the outset that v is an increasing function of the inflation rate,
i. At first glance, this seems all the more natural as the return to holding money is i. However,
as emphasized, e.g., by [Taylor, 2009], pp. 70 sq, this leads to a seemingly insurmountable puzzle.
Indeed, di↵erentiating (1) and rearranging, one gets

di v ⇣ ⌘
= i+g M̂ , (2)
dt dv/di
where a hat signifies growth rate. Suppose that T and M are predetermined (or determined by
market forces that are independent from v and p), then imposing dv/di > 0 induces an unstable
dynamics of inflation: a higher value of i increases di/dt, which feeds back into a further increase
4
See also [Werner, 2014a], [Werner, 2014b], [Werner, 2014c].
5
See [Godley and Lavoie, 2006] for stock-flow consistent macro-modelling.

82
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

in i. In other words, apart in some degenerate situations, this viewpoint necessarily leads to the
conclusion that hyperinflation is unavoidable. Models with perfect foresight therefore try to avoid
this stumbling block by imposing from the outset that asset holders will rationally anticipate an
inflation path along which v is constant and di/dt = 0, so that (2) reduces to

i = M̂ g.
This argument, however, has to face yet another paradox, namely the result provided by [?] showing
that such a “perfect foresight” inflation path is not locally stable: any mistake by any asset holder
in the formation of expectations along this path will lead to another inflation path from which
no perfect foresight expectations can jump back to the initial trajectory. Generically, therefore,
economic actors will fail to coordinate on this “perfect foresight” inflation path.6 As a consequence,
the economy will sooner or later leave this ”rational” trajectory — and shall never return to it,
being unavoidably trapped by some hyperinflation spiral, courtesy of Eq. (2).
Next, money velocity is not observed to be wildly unbounded: even in chronically inflationary
Brazil, the ratio of GDP to money supply reached a level of around 65 in the mid-1990s, which
may have been the upper-bound ever recorded in history. This empirical observation is yet another
characteristic needing explanation. Indeed, by contrast, Eq. (1) and (2) do not provide any upper-
bound on v. Is there any mechanism, hidden behind these two equations, that would give an implicit
cap on the speed of money circulation?

C. Money neutrality?

Finally, monetary analysis has been plagued for decades by an ongoing debate on the alleged
neutrality of money. Money is said to be neutral when a once-and-for-all change in the money
supply or money demand has no real e↵ects. Money is super-neutral when a change in the growth
rate of its supply (or demand) has no real e↵ect. And money is non-neutral when a change in its
supply or demand does have real e↵ects. Current mainstream wisdom claims that7
1/ Money is neutral in the long run.
2/ Money is approximately super-neutral in the long run, but not exactly so. This is
especially true if there are other distortions such as taxes on nominal investment income.
3/ Money is strongly non-neutral in the short run, as monetary shocks a↵ected real
wages, real output, employment, real interest rates, debt defaults, and many other real
variables. The short run can last for years.
Statement 3/ is confirmed by empirical investigations.8 The difficulty with 1/ and 2/, of course, is
that, if money is neutral in the long-run, it implies that any real short-term e↵ects are ultimately
reversed. To be convincing, a sensible monetary theory should be able to describe how this reversal
happens across time. How long is the long-run? We are not aware of any theoretical framework
compatible with the credit creation theory of banking that would highlight a mechanism allegedly
responsible for turning short-term non-neutral money into being long-run neutral. In this paper, we
show that within our framework, 3/ is confirmed but neither 1/ nor 2/ hold: money is non-neutral
at every time scale — and even with perfectly flexible prices.9
6
See also [Guesnerie, 2005].
7
See, e.g., for a good synthesis of this standpoint.
8
For classical references, see [Friedman M., 1963], the Volcker disinflation, and [Mussa, 1986].
9
More often than not, short-run money non-neutrality obtains thanks to some price stikiness or wage rigidity (see,
e.g., [Mankiw and Romer, 1991] for a textbook illustration). Here, such an expedient is superfluous: both prices and
wages are perfectly flexible.

83
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

Throughout the paper, we have tried to cast the discussion within the simplest possible setting.
Nonetheless, among the not completely standard features of our model is the fact that we need to
drop Say’s law so as to be able to consider the endogenous dynamics of inventories. The rationale
for this is the following. As observed by [Pottier and Nguyen-Huu, 2017], when combined with
usual accounting identities, Say’s law makes the monetary interpretation of any dynamics akin to
the one studied here quite controversial: everything goes as if firms would borrow “money” out of
households’ savings (or vice versa), so that “money” could equally well be replaced by a consumption
good or by some numéraire. In other words, despite the formal invocation of “money”, one would
still deal with the modelling of a “real” (moneyless) economy. As we shall see, decoupling e↵ective
demand from the supply forces turns out to be a necessary condition for the genuine emergence of
money: at each point of time, the stock of inventories is then modified proportionately to the gap
between saving and debt, which allows for money creation (or destruction).

3.1.2 Monetary policy and persistence

In the years following the Great Recession, the United States, Japan as well as several European
countries have turned to unconventional monetary policies to stimulate economic activity. After
two decades of slow growth, the Bank of Japan decided to employ a zero interest rate policy (zirp)
to combat deflation and promote economic recovery [Kuroda, 2016]. A similar policy has been
implemented by the US and UK.10 Under zirp, however, the central bank can no longer reduce
interest rates, rendering conventional monetary policy ine↵ective — so goes the argument. On
the other hand, as seen in the Eurozone, over-extending a zero interest rate policy can also result
in negative interest rates.11 Thus, several authors have challenged the value of zero interest rate
policies, pointing to liquidity traps amongst several other pitfalls.12
Our analysis confirms the primary benefit of low interest rates to be their ability to stimulate
economic activity. We identify conditions under which, near-zero interest rates lower the cost of
borrowing, hence spur spending on business capital and investments. Businesses’ increased capital
spending then creates jobs and consumption opportunities. Under the identified conditions, there is
no need for the central bank to set an implied negative nominal interest rate, where loans actually
receive interest.
To prove this result, we adopt the persistence viewpoint. Persistence theory (see [Smith and Thieme,
2011] and [Costa Lima and Grasselli, 2014]) studies the long term behaviour of dynamical systems,
in particular the possibility that one or more variables remain bounded away from zero. It has been
successfully applied by [] in order to assess the impact of public spending in a moneyless economy.
Here, we show that a near-zero interest policy can guarantee that the employment rate does not
remain indefinitely trapped at arbitrarily small values. This is in sharp contrast with what happens
in the model without monetary policy intervention, where the employment rate is guaranteed to
converge to zero and remain there forever if the initial conditions are in the basin of attraction of the
debt-deflationary long-run steady ztate. Furthermore, as with any persistence result, Proposition
1 is a global one: no matter how disastrous the initial conditions are, a sufficiently responsive
monetary policy can bring the economy back from a state of crisis associated with zero employment
rates.
10
See Bank of England, “Monetary Policy Report” [Bank of England, 2020] and Board of Governors of the Federal
Reserve System. “Open Market Operations.” [Federal Reserve, 2021b].
11
European central bank. ”The ECB‘s Negative Interest Rate.” [European Central Bank, 2021]
12
Federal Reserve Bank of St. Louis. ”The Liquidity Trap: An Alternative Explanation for Today’s Low Inflation.”
[Federal Reserve, 2021a].

84
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

The rest of the paper is organized as follows. We set the stage in section 2 with a preliminary,
moneyless version of our non-linear, stock-flow consistent dynamics. Besides fixing the notation for
the full-blown model, we depart from the previous literature by focusing on an economy with a ces
production function, oligopoly pricing, and inventory dynamics. Section 3 exhibit four economically
meaningful and locally stable long-run steady states for this dynamics, among which one is char-
acterized by a finite debt ratio and strictly positive employment — the Solovian equilibrium— and
another one by infinite debt and zero employment. The next section is dedicated to the introduction
of money. Section 5 shows that it solves the first two paradoxes just alluded to: there is no need
to rely on the rational expectation hypothesis for the economy to reach a long-run steady state
with a stable inflation rate. We show money to be non-neutral both in the short- and long-term.
Its creation is endogenously determined by banking credit. Stock-flow consistency then proves to
enable tracking money creation and its travel along the circuit of the aggregate economy. The
monetarist interpretation of the equation of exchange holds at the monetary Solovian equilibrium,
but only there. Moreover, near-zero monetary policy is shown to enable avoiding the liquidity trap
associated wit the debt-deflationary steady-states. A last section concludes.

3.2 The moneyless economy

This section introduces our main notations and the underlying “real” (i.e., moneyless) economy.
Let us consider a three-sector closed economy, E, with no public sector, comprising households, a
banking sector and private non-financial firms producing a single consumption good. Labor forces,
N , grow at some deterministic rate N̂ := Ṅ /N = n(t) 0, possibly time-dependent. We denote
the stock of capital in real terms by K, the total output by Y , and the number of employed workers
by L. For the sake of simplicity, we assume that the economy is endowed with a ces production
function
h i 1/⌘
⌘
Y := A bK + (1 b)e↵t L ⌘
,

where A > 0 is a (constant) total factor productivity, b 2 [0, 1] reflects the capital intensity of
production, ↵ > 0 is the (constant) growth rate of Harrod-neutral technical progress, and the short-
1 13
run elasticity between capital and labor is captured by ⇣ := 1+⌘ . As usually, changes in capital
stock are given by

K̇ := I K, (3)
where > 0 is a constant depreciation rate and I stands for the real investment flow.

3.2.1 Stock-flow consistent inventories

Let us denote total real demand by Yd :

Yd := C + I, (4)
13
Following [Bastidas et al., 2019], we shall restrict ourselves to the economically more relevant case where ⌘ 0,
so that ⇣ 2 [0, 1]. Even though the empirics of this family of endogenous growth cycle models needs to be further
developed, we chose a CES production function since, when backtested on the U.S. economy at least, it seems to
outperform the Leontie↵ specification, [McIsaac, 2021]. On the other hand, the Leontie↵ approach implies that ⌫ be
constant across time —which is obviously not the case in any country where data is available— while the Leontie↵
viewpoint allows for an endogenous capital-to-output ratio.

85
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

where C denotes aggregate consumption. The di↵erence between output and demand determines
changes in the level of inventories, V , held by firms:

V̇ := Y Yd . (5)

Let p 0 refer to a price deflator. Nominal sales are given by pYd = pC + pI while nominal
output is

Yn := pYd + cV̇ , (6)

where c := wL/Y , is the unit cost of production, with w being the unitary money wage and
W := wL the wage bill. Denoting ! := wL/pY the wage share, one gets c = p!. In other words,
we follow [Godley and Lavoie, 2006] (eq. (8.24) and (8.25)), [Grasselli and Nguyen-Huu, 2018] by
valuing inventory changes at cost c, consistently with current accounting practices: cV is what
is incurred by the acquisition department of a firm14 in order to purchase unsold goods from the
production department.15

Let r 0 be a constant, nominal short-run interest rate16 while Df denotes the nominal stock
of firms’ debts. The pre-dividend net profit of firms after paying wages, interest on debt, and
accounting for depreciation (i.e., consumption of fixed capital) is given by17

⇧f := Yn W rDf p K. (7)

3.2.2 Expected profits and debts

As in Minsky’s and Kindleberger’s theories, current cash-flows validate past liabilities and form the
basis for investment decisions. At variance with these authors, however, we introduce here a small
twist by letting investment depend upon expected sales, pY , and not Yn . This enables to circumvent
any circularity in the definition of profits (since Yn itself depends upon current investments) and,
quite realistically, make investment depend upon (a reduced-form of) entrepreneurs’ expectations.
Non-financial firms redistribute only a (constant) fraction, 2 (0, 1), of their expected profit,
f := ⇧ e , to their shareholders. The retained profits (or savings), ⇧f ⇧e , are reinjected for
investment purposes. It follows from firms’ budget constraint that

Ḋf := pI Sf = pI ⇧f + ⇧e . (8)

Observe that, whenever Yn < pY , firms’ expected profit turns out to have been over-optimistic:
⇧f < ⇧e . The fraction (⇧e ⇧f ) that is distributed to shareholders is therefore financed not with
cash currently on hand, but through new debt financing, embodied in (10).18
Next, as in [Giraud and Grasselli, 2019], the household budget constraint implies that, whenever
nominal household consumption exceeds their disposable income, the di↵erence needs to be financed
14
Represented in the capital account of Table 4.1, see below.
15
Therefore, Yn = pY if, and only if, either p = c or V̇ = 0.
16
Section 5 will consider time-varying short-run interest rate policies.
17
As pointed out by [Grasselli and Nguyen-Huu, 2018], even though changes in inventory add to profits, sales
constitute the only way for firms to have positive gross (i.e., before interest payment and depreciation of capital)
profit, since pC + pI + cV̇ W = pYd + c(V̇ Y ) = (p c)Yd .
18
Debt-financed dividends, known as dividend recap, have become an almost standard practice in the U.S. in the
last decade. According to Moody’s Investors Service, in the first three quarters of 2012 there were over $ 16.5 billion
in debt-financed dividend payments (see the Wall-Street journal, https://on.wsj.com/3jqQvb5).

86
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

by an increase, Ḋh , in debt. Conversely, if households’ disposable income exceeds nominal household
consumption, household saving is used to reimburse (part of) their net aggregate debt. Firms and
banks are privately owned by households.19 Let us denote by f and b the dividends received
by households from firms and banks respectively. The households’ disposable income is W + f +
b rDh . The banks’ current profit is ⇧b := r(Df + Dh ). For simplicity again, we assume that,
facing negligible operating costs, banks redistribute all their profits, b := r(Df + Dh ).

On the supply side, expected profit, ⇧e , is defined as the profit that would be accrued were
aggregate demand to absorb the output, Y :20

⇧e := pY W rDf p K.
The dynamics of households’ debt can now be written

Ḋh := pC (W + f + b rDh ) = pC W rDf ⇧e

= Yn cV̇ pI W rDf ⇧e (9)
= D˙f cV̇ + p K.
As will be clear shortly, the latter equation is key for the introduction of money.

3.2.3 Some accounting

The balance sheet, transactions, and flow of funds for this economy are described in Table [4.1].
Unless otherwise stated, all quantities are functions of time. Households’ assets consist thus of firms’
equity, Ef , bank equity, Eb , and current money deposits, Mh . Their liabilities consist of loans, Lh ,
resulting in net worth (i.e., wealth) Xh := E + Mh Lh , with E := Ef + Eb . Firms’ assets are the
capital stock, pK, plus inventories, cV , and current money deposits, Mf , whereas their liabilities
are their loans, Lf , and equity Ef := pK + cV + Mf Lf .21 Similarly, loans, L := Lh + Lf are
the assets of the banking sector. Its liabilities are the deposits, M := Mh + Mf whereas bank
equity is Eb = L M . Hence, the net worth of firms and banks remains Xf = Xb = 0. Again
for simplicity, we assume that firms do not keep positive deposits, preferring to use any balances
to repay their loans instead. As already said, all their net profits minus dividends are reinvested:
⇧f f = ⇧f ⇧e = Sf . Savings of the banking sector are given by22

Sb = rL rM b = r(Df + Dh ) b = 0. (10)

Finally, the financial balance row on Table [4.1] obeys the following ex post accounting identity
between total nominal savings and investment in the economy E:

S := Sh + Sf + Sb = p(I K) + cV̇ . (11)

We eventually get an expression of the total wealth, X, in E:
19
For simplicity, at variance with [Giraud and Grasselli, 2019], we don’t draw any distinction between wage-earners
and “rentiers” earning dividends.
20
Expected profit could arguably be thought of as depending upon current sales, and not supply capacity. This,
however, would make the analysis more complex without impairing qualitatively our main results.
21
Here, equity is treated as a balancing variable so that the net worth of non-financial firms is always Xf = 0.
22
To save notations, we assume that banks have no consumption. Several alternative specifications of banks’
behavior would be conceivable. Details can be found in [Giraud and Grasselli, 2019] .

87
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

Households Firms Banks Row Sum

Balance sheet
Capital stock +pK pK
Inventory +cV cV
Deposits +Mh +Mf M 0
Loans Lh Lf L 0
Equity E Ef Eb 0
Column sum (Net worth) Xh Xf = 0 Xb = 0 X
Transactions Current Capital
Consumption pC +pC 0
Investment +pI pI 0
Change in Inventory +cV̇ cV̇ 0
Depreciation p K +p K 0
Wages +wL wL 0
Interest on loans rLh rLf +rL 0
Interest on deposits +rMh +rMf rM 0
Pre-dividend net profits ⇧f +⇧f + ⇧b 0
Dividends f + b f b 0
Column sum (balances) Sh 0 Sf p(I K) cV̇ Sb = 0 0
Flows of funds
Change in capital stock pK̇ pK̇
Change in inventory cV̇ cV̇
Change in deposits +Ṁh +Ṁf Ṁ 0
Change in loans L̇h L̇f +L̇ 0
Column sum (savings) Sh Sf Sb pK̇ + cV̇
Change in firm equity +Ėf (Sf + ṗK + ċV ) 0
Change in bank equity +Ėb Sb 0
Change in net worth Ė + Sh ˙ + (cV
(pK) ˙ )

Table 3.1: Balance sheet, transactions and flow of funds for the three-sector, closed
economy with inventories and prices

X = X h = E f + E b + Mh Lh
= pK + cV.

In words, total wealth equals the wealth of the households, which in turn is equivalent to the
productive assets of the corporate sector.

3.2.4 Aggregate behavioural rules

Aggregate behavioural rules will now add some flesh to the skeleton provided by Table [4.1].
Similarly to [Grasselli and Nguyen-Huu, 2018], the aggregate investment function of firms is given

88
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

by
(⇡e )
I := K, (12)
⌫(!)
where ⌫(!) := K/Y is the endogenous capital-to-output ratio, and (·) is a continuous, increasing
function of the normalized expected profit:

⇧e
⇡e := =1 ! rdf ⌫(!). (13)
pY
The rate of change in real wages is a function of the current employment rate and inflation:
ẇ ṗ
:= ( ) + , (14)
w p
where (·) is an increasing short-run Phillips curve, and 2 [0, 1] is a measure of money illusion —
with = 1 corresponding to no monetary illusion.23
Aggregate consumption is given by a function, ch (·), of disposable income (cf. (11)):

C := ch (! + rdf + ⇡e )Y. (15)

Finally, the short-run price-wage dynamics is given by:24
ṗ
i := i(!) := = ⌘p (µ! 1). (16)
p
In words, prices relax at speed 1/⌘p > 0 around the unit cost of production, c = p!, inflated by some
markup, µ 1, whose magnitude depends upon the imperfection of competition on the commodity
market. The markup epitomizes the gap between the average cost of production, wL/Y = c, and
the long-run price.25

3.2.5 Oligopoly pricing

The production sector is assumed to behave like a large oligopoly which, given some stock of
capital K, adjusts the quantity, L, of hired labor so as to maximise its current, nominal net profit,
⇧f . The Marshallian first-order condition of this optimization programme is w@L/@Y = p or,
equivalently, @Y /@L = w/p. However, due to imperfect competition, the ruling price, p, will usually
be higher than marginal cost. More precisely, a monopoly’s decision problem is usually thought of
as choosing the quantity, Y , so as to maximise its profit:

max p(Y )Y c(Y ), (17)

Y 0

where p(·) stands for the inverse aggregate demand function, and c(·) for the monopoly’s cost
function. As is well-known, under standard smoothness assumptions, a non-zero solution, Y ⇤ > 0,
to (17) is given by
23
See, e.g., [Gordon, 2011] for a historical survey, [Gordon, 2013], [Mankiw, 2001] and [Mankiw, 2016], as well
as [Grasselli and Nguyen Huu, 2016]).
24
See for example [Godard, 2007], Section 8.3.
25
This is a cost-pushed inflation dynamics. There are good arguments showing that, in certain countries and for
certain time periods, inflation can also be demand-pulled. We could have let inflation in (21) depend upon the
dynamics of inventories at the cost, however, of a more complex analysis involving Hopf bifurcations leading to steady
states being replaced by limit-cycles. This extension is left for further research.

89
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

p0 (Y ⇤ )Y ⇤ + p(Y ⇤ ) = c0 (Y ⇤ ). (18)

Since, usually, p(·) is decreasing, we get indeed that p(Y ⇤ ) > c0 (Y ⇤ ). Moreover, it is equally well-
known that, in general, the average cost of production, c(Y ⇤ )/Y ⇤ = wL⇤ /Y ⇤ = ! ⇤ at equilibrium
is located somewhere above the marginal cost, c0 (Y ⇤ ) = w@L/@Y , and below the monopoly price,
p(Y ⇤ ). In our setting, this means that there exists some µ 1 such that p = µ! and

@L
p= w
, (19)
@Y
Of course, when = µ = 1, we recover the textbook (perfectly competitive) case. On the other
hand, the price dynamics (21) is precisely saying that prices are relaxing towards the markup
augmented average cost. Now, some manipulations analogous to the ones performed by [Van der
Ploeg, 1985] lead to the endogenous capital-to-output ratio as a function of ! and :

K 1 ⇣1 !⌘ 1/⌘
⌫(!) := = , (20)
Y A b
(⇡e ) !˙
g := g(!, , df ) := Ŷ = . (21)
⌫(!) (1 !)⌘
As an immediate consequence, whenever ! is increasing along the endogenous business cycle, real
growth is a decreasing function of , hence will tend to be lower if the oligopolistic power of firms
increases. Since, however, µ should increase as well, inflation should accelerate: the resulting
outcome on nominal growth remains ambiguous. On the contrary, along the declining phase of !,
a tougher competition will depress both real and nominal growth. However, at a long-run steady
state where, say, !˙ = 0, g is unambiguously a decreasing function of the oligopolistic power of firms
as long as wages are non-zero.

Y
Labor productivity a := L obtains in a similar way:
✓ ◆1
↵t !(t) ⌘
a(!(t)) = Ae . (22)
1 b
Therefore, since ⌘ > 0 (see footnote 1), labor productivity will tend to rise if competition deteriorates
on the commodity market while salary negotiations on the labour market remain untouched. This
should not come as a surprise: an increase in works exactly as a decrease in the workers’ real
wage. More generally, (19) is equivalent to

w 1 @Y
= .
p @L
Taking > 1 is a way to reflect, e.g., the productivity-pay gap ( [Stansbury and Summers, 2018]).26
26
There is a deeper justification for > 1, provided by the criticism formulated in [Kucherenko and Shah, 2007]
and [?] on the textbook first-order condition of profit-maximization. Once it is taken into account that the total
di↵erential of profit should intervene in the first-order characterization of profit-optimization (and not just a partial
derivative) the standard equality between marginal revenue and marginal cost leads to a condition analogous to (19)
— even within a Marshallian “atomistic” set-up.

90
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

3.3 Long-run “real” perspectives

This section provides a brief account of the phase space of our basic “real” (i.e., moneyless) economy.
A numerical example illustrates the main emerging properties.

3.3.1 The moneyless vector field

The employment rate dynamics can now be expressed as:

ˆ = Ŷ (⇡e ) !˙
â N̂ = (↵ + n + ). (23)
⌫(!) ⌘!(1 !)

The wage share dynamics also obtains as

!˙ ⌘ ⇣ ⌘
= ( ) ↵ (1 ) i(!) . (24)
! ⌘+1

Df Dh
The debt ratios are defined by df := pY , and dh := pY . Writing ch := C/Y , their dynamics
follow from (11), (10) and (21):

(⇡e ) (1
!)yd + ⇡e + ⌫(!)
dˆf = +r g i
df
ch (! + rdf + ⇡e ) ! rdf ⇡e
and dˆh = g i.
dh
where the normalized e↵ective demand is given by:

yd := Yd /Y = ch (! + rdf + ⇡e ) + (⇡e ).27 (25)

The overall dynamics therefore reduces to a 4-dimensional, non-linear autonomous system:28

8 ⌘! h i
>
> !˙ = ( ) ↵ (1 )i(!)
>
> ⌘+1
>
> h (⇡ ) i
>
< ˙ e !˙
= (↵ + n + ) (26)
> ⌫(!) ⌘!(1 !)
>
> ⇥ ⇤
>
>
> d˙f = (⇡e ) (1 !)(yd ) + (1 ) ⌫(!) df g(!, , df )
(1 )r + i(!)
>
:˙ ⇥ ⇤
dh = ch (! + rdf + ⇡e ) ! rdf ⇡e dh g(!, , df ) + i(!) .
where expected profit, ⇡e , is defined by (13) as a function of ! and df .
27
This relationship between the demand-to-output ratio, yd , the investment behavioural function, (.), and ag-
gregate consumption, ch (.), comes directly from Yd = C + I. It eases the study of (26) by reducing the number
of di↵erential equations: households’ debt turns out to be entirely determined by the other variables, as noticed in
Appendix B.1.1.
28
The capital-to-output ratio, ⌫(!), can be treated as an auxiliary variable. Indeed, ⌫⌫˙ = ⌘(1 !˙ !) so that ⌫(!)
˙ =0
if, and only if, !˙ = 0 which is by definition true at any equilibrium point.

91
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

Normalising our state variables with Yn (instead of pY ), one would recover the alternative share
of wages, corporate debt ratio and profit rate obtained by using the expressions:

Yn pYd + c(Y Yd ) (p c)Yd c

= = + = (1 !)yd + ! ,
pY pY pY p

8
>
> W pY !
>
> !
˜= =
>
> pY Y n (1 !)y d+! (27)
>
< Df pY df
so that d˜f = =
>
> pY Yn (1 !)yd + !
>
>
>
> ⌫(!) (1 !)yd rdf ⌫(!)
>
:⇡˜f = 1 !n rdnf = .
(1 !)yd + ! (1 !)yd + !

3.3.2 Long-term real steady states

Under reasonable conditions detailed in Appendix C.1.2,29 the vector field (26) exhibits four types
of meaningful steady points, analogous to those described in [Grasselli and Nguyen Huu, 2016]. The
first one is a Solovian equilibrium corresponding to a desirable situation with positive wages and
employment, and a finite level of debt ratios, dh and df .30 Following [Giraud and Grasselli, 2019]
we call it Solovian because the growth rate, g, at this steady state equals labor productivity growth
↵ (technical progress) plus population growth, n. Moreover it verifies the Golden rule. Solow’s
workhorse model being a particular case of the present one,31 this equilibrium can be interpreted
as the embedding of the unique equilibrium exhibited by the standard Solow model within our
broader setting. This steady state, however, is no more globally stable as in the Solow model, but
turns out to be locally stable under standard assumptions. Fig. 3.3.3 provides an illustration of the
way inflation and the demand-to-output ratio, yd , evolve in the neighborhood of such a desirable
equilibrium.

The second long-run steady state of (26) is a debt-deflationary situation associated with a sky-
rocketing level of debt ratios while wages and employment shrink to zero.32 This long-term collapse
is also locally stable for a broad family of parameters. Fig. 3.3.3 shows how oscillations of inflation
and the demand-to-output ratio, yd , damp as the economy converges towards such an undesir-
able steady state. The next two steady states correspond to deflationary states with a finite debt
level — one with zero wage, the other one with positive wages. Their local asymptotic stability
cannot be judged a priori and must be checked on a case-by-case basis. Eventually, there are
two other steady states corresponding respectively to a trivial (!, , df ) = (0, 0, 0) and a slavery
(!, , df ) = (0, 5 , df5 ) long-run equilibrium. Both, however, are structurally unstable and therefore
irrelevant.
29
See Appendix C.1.2 for a complete analysis of the existence and stability of steady states.
30
See Figure 3.4a for the trajectory towards this steady state.
31
It can be recovered by putting r = = 0, = µ = 1, ch (x) ⌘ x, (⇡e ) ⌘ ⇡e , and ' ⌘ 0.
32
See Figure 3.4b for the trajectory towards this debt-deflationary equilibrium.

92
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

Symbol & Definition Value Symbol & Definition Value

0.04
↵ labor productivity growth rate 0.025 0 Phillips curve ( ) parameter 1 0.042
0.043
n population growth rate 0.02 1 Phillips curve ( ) parameter 1 0.04
depreciation rate 0.01 k0 Investment function constant -0.0065
A Factor productivity 1/3 k1 Investment function parameter e 5
b Share of capital 0.135 k2 (·) exponential parameter 20
⌘ Elasticity of substitution 50 r Short-run nominal interest rate 3%
Degree of monetary illusion 0.8 µ Mark-up 1.8
c min. value of the consumption ratio 0.01 c+ max. value of the consumption ratio 0.99
⌘p Price relaxation 0.03 Share of profits saved by firms 0.1

Table 3.2: Parameters

3.3.3 Numerical Example

The baseline parameters for our simulations are provided in Table [3.2].33 We rely on a short-run
Phillips curve of the form:

1
( )= 0, (28)
(1 )2

with parameters specified below. The investment function is taken from [Grasselli and Costa Lima,
2012]: (⇡e ) = k0 + k1 ek2 ⇡e . Following [Giraud and Grasselli, 2019], the consumption ratio is given
by:
n o
ch (! + rdf + ⇡e ) = min c ; c+ tanh(! + rdf + ⇡e ) . (29)

By analogy with the Goodwin cycles empirically identified for the US economy (see [Grasselli and
Maheshwari, 2017], [McIsaac, 2021]), one cycle at the beginning of a trajectory corresponds to 10-15
years, so that one time unit in our discrete-time simulation must correspond to approximately one
week.
Let us wrap together the various elements of this preliminary (moneyless) model. Dropping
Say’s law, allowing for some substituability between capital and labour (or, equivalently, endogeniz-
ing the capital-to-output ratio), and for imperfect competition on the commodity market does not
qualitatively change the phase space of a closed economy, E. We recover endogenous business cycles
which, depending upon the basin of attraction to which their starting point belongs, will converge
to one of a finite number of long-run steady states, broadly similar to the ones already found in the
above cited literature. As already suggested in [Bastidas et al., 2019], the CES production function
seems to induce a smaller magnitude of endogenous cycles and a more rapid convergence to the
Solovian equilibrium than, say, a Leontief technology. Notice as well that, in all the examples we
have gone through, the debt ratio, df , never decreases along any trajectory leading to one of the
two stable steady states (this need not be true with a Leontief production structure). Finally, the
basin of attraction of the Solovian equilibrium shrinks as competition becomes more oligopolistic
(see Figure 3.6) making it more difficult to reach this desirable steady state.
33
See [Grasselli and Costa Lima, 2012], [McIsaac, 2016], [Grasselli and Nguyen Huu, 2016] and [Giraud and Grasselli,
2019] for exhaustive justifications of these values.

93
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

(a) Trajectories from initial values: !0 = 0.8, 0 = (b) Trajectories from initial values: !0 = 0.3,
0.9 and d0 = 0.1, to the Solovian equilibrium with 0 = 0.5 and d0 = 1 to the debt-deflationary equi-
final value: !1 = 0.79 , 1 = 0.97 , df1 = 0.602 librium: !p2 = 0, 2 = 0 , df2 = 9.10101

Figure 3.4: Trajectories leading to the locally stable steady states of (26).

3.4 Introducing money

In this section, we explicitly introduce money. So far, indeed, E could be interpreted as a purely
”real” (moneyless) economy. Indeed, prices given by (21) can be viewed as being only relative
prices. The system (26) is 0-homogeneous with respect to (p, w): multiplying both the consumption
price index and the unitary wage by the same > 0 would leave the dynamics of E unchanged.
Similarly, the dynamics of debts (11) can be interpreted as follows: any additional credit received
by either households or firms is lent to them by firms, taking either from the stock of inventories or
from the sunk cost linked to capital depreciation.34 In other words, there is no nominal anchor for
prices and wages. As we now show, however, money can be introduced so as to provide this gauge.

Remember the celebrated “equation of exchange” (1), where T is the volume of transactions. In
our setting, if there is oversupply, Y Yd , the e↵ective volume of transaction is Yd . If, on the
contrary, demand exceeds supply, Y  Yd , firms will sell inventories to meet demand, and as a
consequence, T is still equal to Yd , as long as the stock of inventories remains non-negative. If
this stock becomes null, the production sector can no more meet demand whenever Y  Yd : some
rationing must occur and the volume of transaction should equal production: T = Y . We leave
the latter situation for further investigation,35 and shall therefore focus our analysis on “interior”
solutions to (26) where V 0 throughout.36 As a consequence, the equation of exchange reads

M v = pYd , (30)

where M is to be understood as the sum of bank accounts held by non-governmental bodies (house-
holds and firms). It is an endogenous stock of inside money, generated as acknowledgement of
debt.37 The variable, v, defined by (30), denotes income velocity, namely the speed at which real
34
A similar interpretation — leading to the same conclusion that E is moneyless — has been advocated by [Pottier
and Nguyen-Huu, 2017], in a simpler framework.
35
Considering this type of “boundary problem” leads to a non-di↵erentiable dynamics, to be studied in a subsequent
work.
36
This restriction was actually already at work in [Grasselli and Nguyen Huu, 2016] as well as in the previous
sections of the present paper.
37
That is, we neglect here outside money that would be free of debt, see [Gurley and Shaw, 1960].

94
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

(a) Convergence of inflation, i, to its Solovian (b) Convergence of inflation, i, to its debt-
asymptotic value: i1 = 1.2% deflationary asymptotic value i = 3%

(c) Convergence of the demand-to-output ratio, yd , (d) Convergence of the demand-to-output ratio,
along a path to the Solovian equilibrium, with final yd , to its upper-bound, c+ + k0 , along a path to
value: yd1 = 0.83 the debt-deflationary equilibrium

Figure 3.5: Inflation and demand-to-output ratio

(a) Basin of attraction of the Solovian equilibrium (b) Basin of attraction of the Solovian equilibrium
with perfect competition: = µ = 1. with imperfect competition: > µ > 1.

Figure 3.6: Influence of imperfect competition on the basins of attraction.

95
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

transactions occur in the economy.In a similar vein, equation (2) must be interpreted as

di v ⇣ ⌘
= i + gd M̂ ,
dt dv/di
with gd := Ẏd /Yd .

Where does money come from? In our credit-economy, it can only be created through some
additional banking credit. Conversely, when a debt is reimbursed, the corresponding quantity of
money flows back to the balance sheet of the banking system, and is therefore “destroyed”. As a
consequence,

Ṁ := D˙f + Ḋh . (31)

Notice that, whenever D˙f + Ḋh = 0, no additional money flows into the economy: one of either
sector is lending to the other the increment of debt required. Together with the debt dynamics in
(11) and (10), Eq. (31) leads to the following endogenous dynamics of money:

Ṁ = cV̇ + p K. (32)

Despite its simplicity, Eq. (32) delivers some useful insights. Indeed, were we to maintain Say’s
“law” and to neglect the sunk cost induced by capital depreciation, the quantity of money M would
necessarily be constant, and would therefore become, again, irrelevant. One would also deduce from
(32) that Ḋf = Ḋh . The famous chicken-and-egg problem — who causes the other, savings or
loans?— would again become inextricable. This means that, in order to solve this conundrum and
to be able to observe endogenous credit creation in a macro-econmic model, giving up Say’s “law”
is a necessary condition.

3.4.1 The non-neutral dynamics of money

Defining the money-to-output ratio by m := M/pY , one gets
⇥ ⇤
ṁ = !(1 yd ) + ⌫(!) m g + i , (33)

which we add to the dynamical system (26). The new vector field therefore involves the following
state variables (!, , df , dh , m). Income velocity, v, can be written v = ymd , and is, therefore, a
function of !, , df , dh , and m. Its dynamics is given by

v̂ = ŷd m̂. (34)

Monetarist theories often assume that v is an increasing function of the inflation rate: the return
to holding money being equal to i, as inflation runs faster, it erodes the real value of the money
stock more rapidly so that households and firms should flee to other assets. In our framework, this
property is broadly implied by (1).
From a slightly di↵erent viewpoint, one could argue that the speed at which prices adjust should also
be an increasing function of v: the more frequently trades take place, the more often price changes
can materialize. For instance, this would certainly be true in (double) auction mechanisms.38 Here,
this could be captured by introducing income velocity into the di↵erential equation that governs
inflation (21) as follows:
38
Cf. e.g. [?].

96
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

i(!, df , dh ) = ⌘p v(µ! 1). (35)

In words, (35) means that the speed, ⌘p v, at which prices relax to changes in their long-run value
(given by µ!) is an increasing function of v.
Whether we adopt (35) or keep (16) in order to account for price dynamics, notice that, being now
dependent upon !˙ (see (21)), the real growth rate, g, depends upon v, and hence m. Therefore, at
least in the short-and medium-run, money is non-neutral.
Notice, moreover, that the instability problem, linked to the conventional interpretation of the
equation of exchange alluded to in the Introduction, no more shows up. Indeed, (2)

di v ⇣ ⌘
= i + gd M̂
dt dv/di
does not necessarily lead to unstable inflation since, from now on, both g, i and m̂ depend upon v.
As a matter of fact, courtesy of (24), the dynamics of inflation can equivalently be written:

di ⌘p m⌘! h i
= ⌘p m!˙ = ( ) ↵ (1 )i .
dt ⌘+1
Since 2 (0, 1), a higher value of i decreases di/dt, which now feeds back into a decrease in i.
The overall dynamics boils down to the following 4-dimensional system:
8
> ⌘! h i
>
> !˙ = ( ) ↵ (1 )⌘ p v(µ! 1)
>
> ⌘+1
>
> h i
< ˙ (⇡e ) !˙
= (↵ + n + ) (36)
> ⌫(!) ⌘!(1 !)
>
> ⇥ ⇤
>
> d˙f = (⇡e ) (1 !)(yd ) + (1 ) ⌫(!) df g (1 )r + ⌘p v(µ! 1)
>
>
:
ṁ = !(1 yd ) + ⌫(!) m[g + ⌘p v(µ! 1)].
As in the moneyless case, dh can be treated an auxiliary variable following:
⇥ ⇤
d˙h = ch (! + rdf + ⇡e ) ! rdf ⇡e dh g + ⌘p v(µ! 1) .

3.4.2 Endogeneous money creation

Let us track on Table 3.3 the path followed by money creation, that is, by an additional increment
of credit, `˙ = `˙h + `˙f , distributed at time t by the banking sector to, respectively, households
and firms. Suppose that households use this money injection in order to increase their current
consumption while firms use this additional funding to finance some additional investment. Suppose
also that, the surge in demand is fulfilled by firms by drawing on the inventories. The accounting
value of the stock of unsold production will be reduced by pc `˙h = ! `˙h . As a consequence, the
aggregate profit of the corporate sector will increase by ! `˙h . Suppose that a fraction ✓ 2 (0, 1)
of this additional profit is redistributed to the firms’ shareholders of firms. Therefore, households’
money deposits will increase by ✓`˙h . On the side of corporates, they receive the initial loan arising
from the banking sector, `˙f , which, being spent for investment, remains on the balance sheet of the
productive sector. Firms’ deposits, however, are also raised by the money left after dividends have
been paid, (1 ✓)`˙h .

97
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

Households Firms Banks Row Sum

Balance sheet
Capital stock +pK +pK
Inventory +cV +cV
Deposits +Mh +Mf M 0
Loans Lh Lf L 0
Equity E Ef Eb 0
Column sum (Net worth) Xh Xf = 0 Xb = 0 X
Transactions Current Capital
Consumption pC `˙h +pC+`˙h 0
Investment +pI+`˙f pI `˙f 0
Changes in Inventory +cV̇ ! `˙h cV̇ +! `˙h 0
Accounting memo [GDP] [Yn ]
Depreciation p K +p K 0
Wages +W W 0
Interest on loans rLh rLf +rL 0
Interest on deposits +rMh +rMf rM 0
Pre-dividend net profits ⇧f `˙h +⇧f +`˙h + ⇧b 0
Dividends + f +✓`˙h + b ˙
f ✓ `h b 0
Column sum (balances) Sh (1 ✓)`˙h 0 Sf +(1 ✓)`˙h pK̇ cV̇ Sb = 0 0
Flows of funds
Change in capital stock +pK̇ +pK̇
Change in inventory cV̇ cV̇
Change in deposits +Ṁh +✓`˙h +Ṁf +(1 ✓)`˙h + `˙f Ṁ `˙ 0
Change in loans L̇h `˙h L̇f `˙f +L+`˙ 0
Column sum (savings) Sh (1 ✓)`˙h Sf +(1 ✓)`˙h Sb pK̇ + cV̇
Change in firm equity +Ėf (Sf + ṗK + ċV ) 0
Change in bank equity +Ėb Sb 0
Change in net worth Ė + Sh ˙ ˙ )
(pK) + (cV

Table 3.3: Endogenous money creation

98
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

3.5 Long-run non-neutrality and monetary policy

We now turn to the long-run properties of our monetary economy.

3.5.1 Monetary steady states

First, observe that the equilibrium condition for the money ratio, ṁ = 0, implies

⌫(! ⇤ )
! ⇤ (1 yd⇤ )
g ⇤ := g(! ⇤ , ⇤
, d⇤f , m⇤ ) = i(! ⇤ ). (37)
m⇤
In words, growth is a non-trivial function of m⇤ , even at equilibrium. Therefore, unless the right-
hand side of (37) turns out to be independent from the path followed by m, money is non-neutral
even in the long-run.
In order to check whether (37) is path-dependent with respect to m, let us study the long-run
steady states with money. Under the same technical conditions as in section 2, (36) exhibits 7
meaningful types of equilibrium points which contain the monetary analogs of the three steady
states that had already emerged in the long-run pattern of the real economy.39 Only two of them,
however, turn out to be locally stable under reasonable conditions: the Solovian equilibrium with
money (!1⇤ , ⇤1 , d⇤f1 , m⇤1 ) and the debt-deflationary equilibrium with skyrocketing levels of debt and
money (0, 0, +1, +1) but with zero wage and employment. See Figure 3.7a for an illustration of
a trajectory towards the Solovian equilibrium, and 3.7b for the associated evolution of the money-
to-output ratio. Fig. 3.8a and 3.8b illustrate convergence towards a debt-deflationary crisis.
Therefore, once again, the third equilibrium (debt-deflation but with a finite, asymptotic level of
debt ratio) remains locally unstable for a large family of parameters. Consequently, in the sequel,
we focus on the first two zeroes of the vector field with money, which survive as meaningful and
locally stable long-run steady states.

As detailed in Appendix B.2.2, in a neighborhood of the monetary Solovian equilibrium

(!1⇤ , ⇤1 , d⇤f1 , m⇤1 ),
⇤ =1
- The expected profit rate converges to ⇡e1 !1⇤ rd⇤f1 + ⌫(!1⇤ );
- The households’ debt ratio converges to d⇤h1 ;40
- The demand-to-output ratio, yd , converges to yd⇤1 = ⌫(!1⇤ )(↵ + n + ) + ch (!1⇤ +
rd⇤f1 + ⇡e1
⇤ ), which can be either higher or lower than 1, depending on the aggregate

consumption function ch (·);41

39
Other steady states are structurally unstable and will not be studied further: a trivial Equilibrium (0, 0, 0, m⇤4 ),
a slavery equilibrium (0, ⇤5 , d⇤f5 , m⇤5 ) or any steady state with m = 0, such as (!6 , 6 , df6 , 0). See Appendix B.2.2 for
further details.
40
Indeed d⇤h1 can be deduced from others variables which converge. See Appendix C.1.4 for the demonstration,
which is still true after introducing money velocity in Appendix B.2.2.
41
With our leading parameterization, yd1 = 0.98, so the Solovian equilibrium is overproducing and consequently
the stock of inventories grows to infinity. Notice that inventories, V , grow to infinity along this long-run steady state
with overproduction (yd⇤1 < 1) but plummet if yd⇤1 > 1. With the consumption function of our numerical example,
the latter situation is quite unlikely to happen: yd⇤1 > 1 is equivalent to ch (!1⇤ + rd⇤f1 + ⇡e1 ⇤ ⇤
) + (⇡e1 ) > 1. In
⇤ 2
the neighborhood of the Solovian equilibrium, (⇡e1 ) = ⌫(↵ + n + ) ⇠ 10 under reasonable specifications. The
consumption function, ch (·), is a continuously strictly increasing function which tends to c1 !1 +c2 rdf1 , with c1 , c2 > 0.
Consequently yd⇤1 > 1 means c1 (1 ⇡e1 ⇤
rdf1 + ⌫(!1⇤ )) + c2 rdf1 + ⌫(↵ + n + ) > 1, where 1 c1 ⇠ 10 1 , c1 ⇡e1 ⇤
>0
and c1 ⌫(!1⇤ ) ⇠ 10 2 . Thus, in order to satisfy yd⇤1 > 1, we must have at least c2 c1 > 0 which implies that the
volume of wealth dedicated to consumption coming from the redistributed banks’ profit be greater than the volume
of income dedicated to consumption stemming from wages — which is unlikely in practice.

99
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS
yd⇤
- The income velocity of money converges to v1⇤ = m⇤1
1
> 0;
- Inflation converges to i⇤1 and the growth rate of the quantity of circulating money
M̂ = m̂ + p̂ + Ŷ converges to ↵ + n + i⇤1 . The standard monetarist equation with rational
expectations therefore holds at the monetary Solovian equilibrium: M̂ = i + g.
- In our numerical example (introduced in section 4.4.2), income velocity converges
towards ⇠ 100 around the monetary Solovian steady-state. Once v is introduced in
the price dynamics (35), inflation is therefore asymptotically 100 times larger than with
(21), while the asymptotic private debt ratio is 10 times smaller. In both cases, the
asymptotic employment rate stands above 0.96, there is overproduction and households
accumulate positive net savings.

(a) Trajectories from initial values: !0 = 0.8, 0 = 0.9 and (b) Evolution of the money-to-output ratio along
d0 = 0.1, m0 = 0.1 to final value: !1⇤ = 0.79113030 , ⇤1 = the path to the Solovian Equilibrium towards m⇤1 =
0.98529608, and d⇤f1 = 0.05865288 0.01076340.

Figure 3.7: Trajectories leading to the Solovian equilibrium of (27).

Along a path to the locally asymptotically stable monetary debt-deflationary steady state
(0, 0, +1, +1):42
- The expected profit rate, ⇡e , tends to 1;
- The investment share, (⇡e ), converges to its lower bound value, k0 ;
- The consumption share reaches its upper-bound, c+ ;43
- This drives the demand-to-output ratio to y d = k0 + c+ , a positive value which can be higher
or lower than 1. Because production is plummeting, the volume of demand Yd = yd Y tends to 0
(despite the fact that the consumption share stays positive).
- Since v = ymd , income velocity vanishes in a neighborhood of the asymptotic collapse;
- In our numerical example, asymptotic inflation is either nihil or negative (i = ⌘p = 0.03)
depending upon whether one adopts (21) or (35) for the price dynamics.
Equation (37) then implies that g = 0 at this debt-deflationary equilibrium. As a consequence,
money is non neutral in the long-run in as much various trajectories of m lead to di↵erent steady
42
See Appendix B.2.2 for details.
43
The consumption function here is driven by firms’ debt which is held as a liability by the banks’ shareholders.
Consequently, ch (! + rdf + ⇡e ) tends to its upper bound c+ , because df tends to +1 and < 1.

100
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

states, with distinct real growth rates. Notice, however, that the Monetarist master equation,
M̂ = g + i, being true at the monetary Solovian equlibrium, money looks neutral at this equilibrium.
This apparent long-run neutrality, however, is an artifact from static equilibrium models which
neglect the dynamical path leading to this particular equilibrium.

(a) Trajectories from initial values: !0 = 0.3, 0 = 0.5 and (b) Income velocity of money on a path to the debt-
d0 = 1 to the debt-deflationary equilibrium: (!, , df , m) = deflationary equilibrium.
(3.252 e 09, 1.916 e 05, 4.222 e+06, 9.588 e+03).

Figure 3.8: Trajectories leading to solutions of (27) to (0, 0, +1, +1).

3.5.2 Monetary persistence

This section shows how a well-tuned monetary policy and banking regulation can guarantee that
the economy does not reach a state of permanently zero employment. For that purpose, let us
extract from Appendix A the following technical assumptions:

Assumption A a) (ii) The short-run Phillips curve verifies (0) < ↵ (1 )⌘p .

Assumption B The investment function satisfies the following boundary condition:

b1/⌘
0 (0) > (↵ + n + )⌫(0) = (↵ + n + ) .
A
In the sequel, following [Smith and Thieme, 2011], a dynamical system is said to be X-UWP
(Uniformly Weakly Persistent with respect to the functional t 7! X(t)) if there exists " > 0 such
that lim X(t) ". The main result of this section is the following:
t!+1

Proposition 1. Whenever inflation is given by (16) and under Assumptions A 1)

(ii) and B, the monetary economy defined by (36) is -UWP, provided the central bank
implements an interest rate policy, t 7! r(t), such that

"
r(t) < , (38)
df (t)

as soon as df (t) > 0 and (t) < ", for some low " > 0.

101
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

Before entering into the proof, let us briefly comment this result. The main takeaway is that a
specific monetary policy is sufficient to drive a monetary economy out of the dangerous waters
of the liquidity trap each time the employment rate becomes too low. This policy consists in
lowering the short-run nominal interest rate so that (38) holds provided aggregate corporate debts
are positive. This is but a stylized version of the policy put into practice in a number of countries
after the Great recession (Japan, UK, US, euro zone...), except that, here, it is not framed as
a function of some inflation target — as in the conventional Taylor rule ( [Taylor, 1993])— but
according to the debt ratio of the private sector. To the best of our knowledge, this is the first
time such a criterion emerges as a condition for a successful monetary policy when an economy is
threatened by a liquidity trap.
That inflation be captured by (16) means that it does not depend on income velocity but does
not prevent the reverse causality to be true. Assumption A 1) (ii) is weak and standard in the
literature devoted to this class of macro-economic dynamics. Assumption B is somewhat stronger.
It complements A 2) (ii), which we use to prove the existence of both the monetary Solovian
equilibrium and some economically irrelevant (and structurally non-existent) steady states (see
Appendix B.2.2). A 2) (ii) basically says that, provided rdf remains bounded away from zero, the
investment-to-output ratio should be bounded from above by a cap given by

b1/⌘
(↵ + n + ) . (39)
A
On the contrary, B requires (normalized) aggregate investment to exceed this level whenever
rdf = 0. In words, if the central bank implements zirp, it should successfully encourage low-cost
borrowing up to the point that investment crosses what is, otherwise, an upper-bound of theirs.
Notice, however, that our main result does not require zirp but only a very low nominal interest
strategy, provided the private sector’s investment appetite is strong enough. According to standard
calibrations (including our numerical example), the threshold (39) is close to 10-15%. Evidently,
in some countries, the private sector is too weak to reach like 15% of its output — however un-
conventional the monetary policy might be. This suggests that, for such countries, either a public
intervention is needed along the lines advocated, e.g., by [Mazzucato, 2011] or negative interest
nominal short-term interest rates should be envisaged in order to escape from the liquidity trap.
Proof As in [Costa Lima and Grasselli, 2014], we proceed by contradiction. Suppose that, for
every " > 0, there exists a time t0 > 0 such that 8t t0 , (t) < ". Then, given that, obviously,
i(!) ⌘p , the di↵erential
⇥ equation⇤ followed by ! in (36) implies that
⌘
(t t ) (") ↵+(1 p )⌘
!(t) < !(t0 )e 0 ⌘+1 .
It follows from Assumption B a)(ii) that, whenever " is chosen small enough so that
(") ↵ + (1 )⌘p < 0,
there exists t1 > t0 such that, 8t t1 ,!(t) < ". Consequently, (20) and ⌘ > 0 imply that the
capital-to-output ratio verifies
⇣ ⌘1/⌘
⌫(!(t)) < A1 1 b " .
Now, (13) and (38) imply that the normalized expected profit satisfies
⇣ b ⌘1/⌘
⇡e (t) > 1 2" > 0. (40)
A 1 "
The last inequality arises from Assumption B.
On the other hand, the motion of the wage share, !, is such that

102
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

!˙ 1 h i
= ( ) ↵ (1 )i .
⌘!(1 !) (⌘ + 1)(1 !)
Consequently, for t > t1 ,

!(t)
˙ 1 h i
 (") ↵ + (1 )⌘p =: ".
⌘!(t)(1 !(t)) (⌘ + 1)(1 ")
It follows, again, from A a)(ii) that the right-hand side, ", is negative. Consequently, the motion
of the employment rate, , in (36) leads to
h i h i
(⇡e ) (⇡e )
(t t1 ) ⌫(") " (↵+n+ ) (t t1 ) ⌫(")
(↵+n+ )
(t) (t1 )e > (t1 )e .
Therefore, (40) guarantees that there exists t2 > t1 such that, for every t t2 , (t) > ". A contra-
diction.
⇤

3.6 Conclusion
In this paper, a first attempt at modelling a monetary economy according to the lines advocated
by [Werner, 2016] has been provided. We showed that it is possible to make explicit endogenous
money creation through banking credit in a macroeconomic dynamics so as to circumvent the
famous conundrums associated with inflation instability. No rational expectation is needed for that
purpose.

In the absence of monetary policy, the dynamics is primarily driven by the investment decisions of
firms based on profit levels and by the consumption decisions of households based on income: high
profits lead to high investment and capital expansion financed by increasing private debt levels; high
incomes lead to high e↵ective demand financed by increasing household debts. We have shown that
a key for letting money enter the picture is to drop Say’s law, so that the imbalance between debts
and savings must be filled with new credit banking. The dynamic analysis showed that the interplay
between e↵ective demand, production and money creation can lead either to an equilibrium with
a finite private debt-to-output ratio, positive but stable inflation and positive money velocity —
the monetary Solovian equilibrium— or to another steady state where debts become infinite while
the wage share, the employment rate, and money velocity both collapse to zero. Inflation either
collapses as well or can even become negative — a hallmark for the deflationary trap.

Moreover, money has been shown to be non-neutral both in the short- and the long-run, even
though the monetarist interpretation of the exchange equation happens to be true at the monetary
Solovian equilibrium — and only there. The fact that banks create credit and money out of nothing,
if used appropriately, results in non-hyper-inflationary growth, provided the economy crosses the
basin of attraction of the monetary Solovian equilibrium.

Eventually, we have shown that central bank intervention may have a clear positive e↵ect in
preventing a crisis characterized by collapsing employment rates. Monetary policy prevents this
outcome by conditioning the short-run nominal interest rate upon the private debt-to-outcome ratio:
The larger the latter, the smaller should the cost of money be. The persistence result of section 4
are much stronger than, say, the possible local, asymptotic instability of undesirable steady-states

103
CHAPTER 3. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE, STOCK-FLOW
CONSISTENT DYNAMICS

under typical parameter values of the model: central bank intervention, in the form of a responsive
enough monetary policy, prevents the economy from remaining permanently at arbitrarily low levels
of employment regardless of the initial conditions of the system. To echo [Costa Lima and Grasselli,
2014], it may be that stabilizing an unstable monetary economy is too tall an order for central bank,
but destabilizing a stable debt-deflationary crisis is possible.

There are, of course, a number of institutional aspects that we have neglected in this first tentative
pass: reserve requirements, capital adequacy requirements (such as the Basel iii/crr framework),
prudential regulation... [Bovari et al., 2020] have introduced some refinements in the modelling of the
banking system. within a “real” economy. Added to the present framework, these refinements could
enable us to tackle important issues around a counteryclical leverage-based regulation of banks.

Perhaps as important is to consider open economies and the interplay between non-neutral credit
creation out of thin air, exchange rates, and international borrowing. The main thesis defended
by [Werner, 2014b] should then be examined in order to check under which conditions developing
countries should have relied on domestic money creation rather than borrowing money from abroad,
at the risk of getting “ensnared in spiralling foreign currency debt, when actually the foreign banks
just created the money out of nothing, something the developing country could have ar- ranged
for through its own domestic banks” ( [Werner, 2016]). As made clear by this author, considering
endogenous money creation “also has implications for the question of who should pay for bank
bailouts, shifting the pendulum from burdening tax-payers towards central bankers”. We leave for
further research the task of investigating these fascinating questions.

104
Chapter 4

Directed Technical Change in an

inventory Stock-flow Consistent
Dynamics with Government: from
dirty to clean Capital

Abstract
The aim of this paper is to propose a simple model of energy transition from fossil fuels
to renewable energies.
We consider here the extreme case of two natures of capital, one brown (that needs
fossil energy to operate), the other, green (free of fossil energy), which, however, is less
productive than the dirty one. This dual capital structure is embedded into a stock-flow
consistent macrodynamics where investment can be financed by debt and the productive
sector adjusts shifts in demand thanks to inventories. The transition is said to take place
if the economy does asymptotically operate with green capital only.
Absent any public intervention, the economy converges to a purely brown production
sector: unsurprisingly, renewable energies are kicked out as long as green capital remains
less productive than the brown one. The inertia a↵ecting the transition of aggregate
investment from one type of capital to the other, however, turns out to destabilize the
debt-deflationary equilibria of the dynamics. Finally, we pinpoint two main levers for a
government intervention: increasing the energy price via a carbon tax or the output-to-
capital ratio of green capital via publicly subsidized directed technical innovation. The
first option runs the risk of reducing private profits, hence fostering private debts. The
second runs the dual risk of boosting public debt. We identify conditions under which
the transition to green capital takes place without leading to an overhang of debt, be it
private or public. Numerical simulations illustrate the properties of each equilibrium.

4.1 Introduction
Directing technical change towards alternative energy sources to fossil fuels is of crucial impor-
tance for the ecological transition. This paper aims at developing a simple framework for analysing
the driving forces shaping the biases of capital increase that may, or may not, favour the shift
towards renewable energies. We consider two types of capital, a “brown” (or dirty) capital which

105
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

needs fossil energy to operate, and a “green” (or clean) capital, which is less productive but does
not rely on fossil energy.1 Private investment is directed towards the most profitable capital type,
depending on the (exogenous) price of fossil energy. Both types of capital are putty-clay ( [Johansen,
1959]) with respect to energy: once a capital good is created, its capital to energy consumption ratio
is fixed once for all, and both types of capital are not substitutable. This production structure is
embedded into a stock-flow consistent prey-predatory macrodynamics [Goodwin, 1967] where pri-
vate investment can be financed by debt [Grasselli and Costa Lima, 2012]. The transition towards
renewable energies takes place if the economy does asymptotically operate with green capital only.
Two main drivers for this to happen are highlighted in this paper: the output-to-capital ratio of
brown versus green capital and the relative price of fossil energy compared with renewable energies.

Investment is directed towards the most profitable capital type, depending on its productivity and
the relative price of fossil energy, but does not adjust instantly (although the higher the gap between
the rates of profit, the quicker the adjustment). This inertia of investment can be thought of as
capturing complementarity e↵ects between the two types of capital (which are otherwise perfect
substitutes as production factors), specialization of workforce or infrastructure externalities. It
departs from the instantaneous shift from brown to green capital often postulated in the literature.2

Following [Grasselli and Nguyen-Huu, 2018], Say’s law is dropped here and public policy is in-
troduced through public spending and taxes in a way similar to [Costa Lima and Grasselli, 2014]).
The interplay between inventory dynamics and public policy leads to a rich dynamics similar to
a two-dimensional Aubry-Mather-like dynamics on the torus with a small cycle (linked to the un-
derlying prey-predatory dynamics between employment and wages) intertwined with a larger one
induced by the county-cyclical public response. Together with the dynamics of private debt, wage
share and underemployment, it leads to a 10-dimensional non-linear dynamical system. We show
that this system can be reduced to a 5-dimensional one. That system, however, is not autonomous
because it depends on the exogenously given price of fossil energy, which determines the allocation
of investment.

As in the literature devoted to this type of macrodynamics (see [Grasselli and Costa Lima,
2012] and [Giraud and Grasselli, 2019]), we show that, independently of the nature of capital
prevailing in the long-run, there are two types of steady states: a Solovian-like one, which is but
the extension to our setting of the (unique) equilibrium of the standard Solow-Swan dynamics; and
a debt-deflationary one, where an overhang of private debts leads the economy to a liquidity trap
where trades, production and employment collapse.

Under reasonable specifications of the short-run Phillips curve, as well as investment and con-
sumption functions, we show that debt-deflationary equilibria, whether dirty or clean, are locally
unstable. This stands in sharp contrast with the literature already mentioned. The force that desta-
bilizes deflationary equilibria turns out to be the inertia with which aggregate investment shifts from
one type of capital to another. [Costa Lima and Grasselli, 2014] have provided a completely di↵erent
mechanism that destabilizes the debt-deflationary crisis, based on public intervention. Our result
provides insight on this problem by suggesting that we might need to dig further the real causes of
the debt-deflation disease, hence the remedies that could help cure it.
1
To save notations, we do not make the dissipation of renewable energies explicit. Nevertheless the price of fossil
energy is to be understood as the relative price of fossil to renewable energies.
2
See, e.g. in Nordhaus’ DICE [Nordhaus, 2014] model; for a climate change model sharing our theoretical frame-
work, see [Bovari et al., 2018].

106
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

The main driver of the sustainability transition of the economy turns out to be the di↵erence
between the relative price of fossil energy and the productivity gap between the two types of capital.
If the fossil energy price remains too low with respect to the di↵erential in productivity between the
two types of capital, the economy does asymptotically operate exclusively with brown capital. The
cost of fossil energy is compensated in the long-run by a decrease in the share of wages courtesy
of the higher productivity of capital. Conversely, whenever the fossil energy price is sufficiently
deterrent, the investment ultimately concentrate on the clean capital. Numerical simulations show
the e↵ects of changes in the parameters and in the relative fossil energy price.

Consequently, the main levers for a government intervention consist in increasing either the
relative fossil energy price or the productivity of clean capital. Both tend to discourage investment
directed at brown factors, leading to directed technological innovation towards green infrastructures.
The transition is said to take place if the economy does asymptotically operate with green capital
only. We show that this happens if the carbon tax sufficiently increases the relative fossil energy
price while directed innovation boosts the productivity of clean capital. Moreover, when a carbon
tax is adequatly combined with public subsidies

The rest of the paper is organised as follows. The following section sets up the two types of
capital within an inventory dynamics with government intervention. It leads to the identification
of a quantitative criterion for the green transition to take place, namely that the relative fossil
energy price be higher than the productivity gap between the two types of capital. In section 3, we
construct the 10-dimensional dynamical system followed by this dual economy, and study the local
stability of its long-run steady states. In the next section, we refine the public policy of directed
technical change so as to identify conditions that ensure that the economy will ultimately shift
towards clean capital. The last section concludes.

4.2 Inventory dynamics with government intervention

We consider a four-sector closed economy, E, comprising households, a banking sector, a public
sector and private firms producing a single consumption good with brown or green capital. For the
sake of simplicity, we restrict ourselves to the extreme case of two types of capital: brown capital
on the one hand, which needs a constant flow of fossil energy to produce output ; and green capital
on the other hand, which operates without fossil energy but is less productive than dirty capital.

4.2.1 Dirty vs clean capital

Output, Y, can be produced in two ways: either courtesy of a stock, K1 , of brown capital with a
flow, E, of fossil energy and labour, L1 , or operating a stock of clean capital, K2 , with a flow, L2 ,
of labour alone.3 For each type k = 1, 2, the capital-to-output ratio ⌫k is constant. Installed capital
goods of any type cannot be converted into one another: If investors want to switch from one type
to another, they have to invest into new capital goods. As regards output production, both types
of outputs, Y1 and Y2 , are perfect substitutes:
K1 E K2
Y = Y1 + Y2 with Y1 = = = a1 L1 and Y2 = = a2 L2 , (1)
⌫1 ⌫1 ⌫2
3
To simplify notations, we don’t introduce explicitly renewable energies, which might be used by both types of
capital. such an addition could easily be done without significantly improving our results, nor their clarity.

107
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

where ⌫1 , ⌫2 , a1 , a2 > 0. Full employment of capital is assumed throughout.4 All quantities are
expressed in real terms, and are net quantities, meaning that intermediate revenues and expenditures
are deducted from the final yearly output.

As usually, capital accumulation depends upon investment, I, and a constant depreciation rate,
> 0, for both types of capital goods:

8 k = 1, 2 K̇k = Ik Kk . (2)

For the sake of our argument, the two types of capital di↵er only by their productivity: ⌫1 > ⌫2 .
Therefore, we impose that they have the same “capital-labor ratio”:5
K1 K2
= a1 ⌫ 1 = a2 ⌫ 2 = (3)
⌫1 ⌫2
Labor productivities, a1 and a2 , are assumed to grow exponentially:
a˙1 a˙2
= = ↵, (4)
a1 a2
with the same (constant) growth rate, ↵ > 0, of Harrod-neutral technical progress. Total labor
forces, N , grow at some deterministic time-dependent rate N̂ := Ṅ /N = n(t) 0. The employment
L(t)
rate is defined as := N (t) .

4.2.2 Expected profits, taxes and subsidies

Let us now introduce inventories along the lines of [Grasselli and Nguyen-Huu, 2018], as well
as government intervention (as in [Costa Lima and Grasselli, 2014]). For simplicity, let us restrict
government intervention to simple “base-level” taxes and subsidies targeting the net profits of private
non-financial firms. Getting rid of Say’s Law is a turnkey in order to tackle the issue of “degrowth”
and consumption sobriety. The household’s sector is introduced in the same way as in [Giraud and
Grasselli, 2019].
Let us denote total real demand by Yd :

Yd := C + I + GE , (5)

where I = I1 +I2 is the total investment volume for both types of capital goods, C denotes aggregate
consumption, and GE , government expenditures (also in real terms).

The di↵erence between output and demand determines changes in the level of inventories, V ,
held by firms:

V̇ := Y Yd . (6)
4
For more thought on this assumption, see the conclusion.
5
Another possible assumption would be that the productivity of labor is the same across the technologies, i.e.
a1 = a2 .

108
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Let p 0 denote the unit price level for both investment, government expenditure and consump-
tion good. Nominal sales are given by pYd = pC + pI + pGE . Let w designate the unitary money
wage, W := wL the wage bill, ! := wL/pY the wage share, and c := wL/Y = p!, the unit cost of
production. Nominal output is

Yn := pYd + cV̇ . (7)

In other words, we follow [Godley and Lavoie, 2006] (eq. (8.24) and (8.25)), [Grasselli and Nguyen-
Huu, 2018] by valuing inventory changes at cost c, consistently with current accounting practices:
cV is what is incurred by the acquisition department of a firm6 in order to purchase unsold goods
from the production department.7

Let r 0 be a constant, nominal short-run interest rate charging private debts while Df denotes
the nominal stock of firms’ debts. The pre-dividend net profit of non-financial firms, ⇧f (after
paying wages, interest on debt, energy purchase, taxes, T , and receiving subventions, G) is given
by:
⇧f := Yn W rDf pE E + G T. (8)
We must distinguish here this profit from their expected profit, ⇧e , defined as:

⇧e := pY W rDf pE E + G T. (9)

On the household side, as in [Giraud and Grasselli, 2019], the budget constraint implies that,
whenever nominal household consumption exceeds their disposable income, the di↵erence needs to
be financed by an increase, Ḋh , in debt. Conversely, if households’ disposable income exceeds nom-
inal household consumption, households’ saving is used to reimburse (part of) their net aggregate
debt. As banks are privately owned by households,8 let us denote by b the dividends received by
households from banks. Let also assume that public debt is directly hold by households and consti-
tutes another source of income. The households’ disposable income is W + b + rg Dg rDh . The
banks’ current profit is ⇧b := r(Df + Dh ).9 For simplicity again, we assume that, facing negligible
operating costs, banks redistribute all their profits, b := r(Df + Dh ).

Current cash-flows validate past liabilities and form the basis for investment decisions. At variance
with these authors, however, we introduce here a small twist by letting investment depend upon
expected sales, pY , and not Yn .10 This enables to circumvent any circularity in the definition of
profits (since Yn itself depends upon current investments) and, quite realistically, make investment
depend upon (a reduced-form of) entrepreneurs’ expectations. Once more, for simplicity, we assume
that non-financial firms inject all their profits, ⇧f , for investment purposes (no redistribution of
their expected profit to their shareholders).11 Whence, changes in corporate debts are given by

Ḋf := pI ⇧f . (10)
6
Represented in the capital account of Table [4.1], see below.
7
Therefore, Yn = pY if, and only if, either p = c or V̇ = 0.
8
For simplicity, at variance with [Giraud and Grasselli, 2019], we don’t draw any distinction between wage-earners
and “rentiers” earning dividends.
9
For the sake of simplicity, we assume the same interest rates for both households and firms debt.
10
As a consequence, whenever Yn < pY , firms’ expected profit turns out to have been over-optimistic: ⇧f < ⇧e .
11
See [Dossetto and Giraud, 2020] for an explicit treatment of dividends. By contrast, banks’ dividends must be
assumed to be redistributed in order to keep the stock-flow consistency of the dynamics.

109
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

The dynamics of households’ debt can now be written as the di↵erence between consumption
spending and disposable income:

Ḋh := pC (W + b + rg D g rDh ) = pC W rDf rg D g

= Yn cV̇ pI pGE W rDf rg D g (11)
= D˙f Ḋg cV̇ .

Changes in public debt, Dg , depend on the volume of public expenditures, GE , and the State net
revenues, T ⇤ , which take into account revenues from taxes, the public debt servicing and the cost
of subsidies to firms:
Ḋg = pGE T ⇤
(12)
with T ⇤ := T rg Dg G,
where rg is a constant, nominal short-run interest rate charged on sovereign debt, while G and T
are base-level subsidies and taxes. The dynamics of the latter primarily depends on the overall state
of the economy as measured by the level of output, the employment rate and non-financial firms’
profits:
Ġ(t) = ( )Y with ˙ ( ) < 0,
(13)
Ṫ (t) = ⇥(⇡f )Y ˙ f ) > 0.
with ⇥(⇡
We postulate the dynamics for expenditures to take the form: G˙E = E (!, , ⇡f , g, ⌧, GE , Y ).

4.2.3 Some accounting

The balance sheet, transactions, and flow of funds for this economy are described in Table [4.1].
Unless otherwise stated, all quantities are functions of time. Government debt, Dg , is the only
liability of the public sector.12 Following the simplifying assumptions of [Costa Lima and Grasselli,
2014], public sector does not keep bank deposits or make banks loans. Households’ assets consist
thus of government debt, firm equity, Ef , bank equity, Eb and current money deposits, Mh . Their
liabilities consist of loans, Lh , resulting in net worth (i.e., wealth) Xh := Eh + Mh Lh + Dg , with
Eh := Ef + Eb . Firms’ assets are the capital stock, pK, plus inventories, cV , and current money
deposits, Mf , whereas their liabilities are their loans, Lf , and equity is Ef := pK + cV + Mf Lf .13
Similarly, loans, L := Lh +Lf , are the assets of the banking sector, and its liabilities are the deposits,
M := Mh + Mf , whereas bank equity is Eb = L M . Hence, the net worth of firms and banks
remains Xf = Xb = 0. Again for simplicity, we assume that firms do not keep positive deposits,
preferring to use any balances to repay their loans instead. As already said, all their net profits are
reinvested: ⇧f = Sf . Savings of the banking sector are given by 14

Sb = rL rM b = r(Df + Dh ) b = 0. (14)

Finally, the financial balance row on Table [4.1] obeys the following ex post accounting identity
between total nominal savings and investment in the economy E:

S := Sh + Sf + Sb + Sg = pI + cV̇ pE E. (15)

12
Government debt is not hold by firms or banks.
13
Here, equity is treated as a balancing variable so that the net worth of non-financial firms is always Xf = 0.
14
To save notations, we assume that banks have no consumption. Several alternative specifications of banks’
behavior would be conceivable. Details can be found in [Giraud and Grasselli, 2019] .

110
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

We eventually get an expression for total wealth, X, in E:

X = X h + X g = Ef + Eb + Mh Lh + D g Dg
= pK + cV.
In words, total wealth equals the wealth of the households and the government, which in turn is
equivalent to the productive assets of the corporate sector.

Aggregate behavioural rules will now complete the structure provided by Table [4.1].

4.2.4 Aggregate behavioural rules

Business decisions will depend upon the rates of profit of each type of capital:
✓ ◆ ✓ ◆
pY1 W1 pE E 1 w pE 1 w
⇡K1 := = 1 and ⇡K2 := 1 .
pK1 ⌫1 a1 p p ⌫2 a2 p
The flow of investment, I, is allocated between the two types of capital goods according to the
following rule:

I = I1 + I2 where I1 = (1 ✓)I and I2 = ✓I, (16)

where the variable, ✓ 2 [0, 1], represents the proportion of investment directed towards green capital.
Its dynamics is determined by the gap between the two rates of profit ⇡K1 and ⇡K2 :
✓˙ = ✓(1 ✓)(⇡K2 ⇡K1 ), (17)
where ✓(1 ✓) characterizes the inertia of investment.15 When ! +1, investment adjusts
instantly as would be assumed in a neo-classical equilibrium model. Then, one would have ✓ = 1
as soon as ⇡K2 > ⇡K1 and ✓ = 0 otherwise.16
The aggregate investment function of firms is given by
I := (⇡e )Y, (18)
⇧e
where (·) is a continuous, increasing function of the normalized expected profit, ⇡e := pY .

The rate of change in money wages is a function of the current employment rate and inflation:
ẇ ṗ
:= ( ) + , (19)
w p
where (·) is an increasing short-run Phillips curve and 2 [0, 1] is a measure of money illusion,
where = 1 corresponds to no monetary illusion (see, e.g., [Gordon, 2011] for a historical survey,
[Gordon, 2013], [Mankiw, 2001] and [Mankiw, 2016], as well as [Grasselli and Nguyen Huu, 2016]).
15
This inertia of investment can be thought of as capturing complementarity e↵ects between the two types of
capital that are not formalized by the production function, specialization of workforce or infrastructure externalities.
Of course, a single parameter, , is not enough to translate the complexity of these e↵ects. In particular, infrastructure
externalities and learning-by-doing suggest that should depend on ". But this dependency need not be symmetrical
and we would thus have to distinguish between an adjustment speed, + , when the transition is going forward and
another one, , when the economy actually increases its investment in dirty capital. These improvements are outside
the scope of this paper. For now, let us just keep in mind that a small means that investment reacts very slowly to
a gap between the rates of profits.
16
Obviously, the speed of investment adjustment depends upon the liquidity of markets and money velocity, see
[Dossetto and Giraud, 2020]. These aspects would go beyond the scope of this paper, and will be treated as a
constant parameter.

111
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Households Firms Banks Gov. Row Sum

Balance sheet
Capital stock +pK +pK
Inventory +cV +cV
Public Debt Dg Dg 0
Deposits +Mh +Mf M 0
Loans Lh Lf L 0
Equity E Ef Eb 0
Column sum (Net worth) Xh Xf = 0 Xb = 0 Xg X
Transactions Current Capital
Consumption pC +pC 0
Investment +pI pI 0
Government expenditure +pGE pGE 0
Changes in inventory +cV̇ cV̇ 0
Accounting memo [GDP] [Yn ]
Wages +W W 0
Cost of Energy pE E +pE E 0
Interest on loans rLh rLf +rL 0
Interest on deposits +rMh +rMf rM 0
Interest on public debt +rg Dg rg D g 0
Subsidies +G G 0
Taxation T +T 0
Dividends b b 0
Column sum (balances) Sh ⇧f pI cV̇ + pE E Sb = 0 Sg 0
Flows of funds
Change in capital stock +pK̇ +pK̇
Change in inventory cV̇ cV̇
Change in deposits +Ṁh +Ṁf Ṁ 0
Change in loans L̇h L̇f +L̇ 0
Change in public debt +Ḋg Ḋg 0
Column sum (savings) Sh Sf Sb Sg pK̇ + cV̇
Change in firm equity +Ėf (Sf + ṗK + ċV ) 0
Change in bank equity +Ėb Sb 0
Change in net worth Ė + Sh 0 0 Sg ˙ + (cV
(pK) ˙ )

Table 4.1: Balance sheet, transactions and flow of funds for the four-sector, closed
economy with inventories and prices

112
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

As a proxy for aggregate consumption, we adopt a function, c(·), of disposable income (11):

C := c(! + rdf + rg dg )Y, (20)

Df Dg
where df := pY and dg := pY are the normalized levels of public and private debts held by
households.

A short-run price-wage dynamics is given by:17

ṗ
i(!) := = ⌘p (m! 1). (21)
p

Prices relax at speed 1/⌘p > 0 around the unit cost of production, c = p!, inflated by some
markup, m 1, whose magnitude depends upon the imperfection of competition on the commodity
market. The markup epitomizes the gap between the average cost of production, wL/Y = c, and
the (long-run) equilibrium price, p.

4.3 Long-run stability study

The real growth rate of the economy, g := Ŷ := Ẏ /Y , can be expressed as:
✓ ◆ 
1 d K1 K2 1 I1 K1 I 2 K2
g= + = +
Y dt ⌫1 ⌫2 Y ⌫1 ⌫2
(22)
(⇡e )
= ,
⌫✓
1 1 ✓ ✓
where ⌫✓ = ⌫1 + ⌫2 is the mean productivity of current investment.

The non-financial firms’ debt ratio dynamics, df = Df /pY , is defined by:

✓ ◆
d˙f D˙f Ẏ ṗ (⇡e ) ⇡f (⇡e )
= = +i , (23)
df Df Y p df ⌫✓
⇧f
where ⇡f := pY is the normalized corporate net profit.

The households’ debt ratio dynamics, dh = Dh /pY , follows:

d˙h Ḋh Ẏ ṗ
=
dh Dh Y p
✓ ◆ (24)
c(! + rdf + rg dg ) ! rdf (⇡e )
= +i .
dh ⌫✓
17
See for example [Godard, 2007], Section 8.3.

113
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

GE T
The kinetics of the ratio of government expenditure, ge := Y , tax ratio, ⌧ := pY , and subsidies
G
ratio, g := pY can be respectively written:
✓ ◆
E (!, , ⇡f , g, ⌧, GE , Y ) (⇡e )
g˙e = ge ,
Y ⌫✓
✓ ◆
(⇡e )
⌧˙ = ⇥(⇡f ) ⌧ +i , (25)
⌫✓
✓ ◆
(⇡e )
ġ = ( ) g +i .
⌫✓

Dg
Thus the public debt ratio, dg = pY , follows the ordinary di↵erential equation:

d˙g Ḋg Ẏ ṗ
=
dg Dg Y p
✓ ◆ (26)
(⇡e )
) d˙g = ge + g ⌧ + d g rg + i .
⌫✓

Notice that, courtesy of (1), " = YY1 = ⌫1EY , is directly related to the brown energy intensity of
production. The variable " can therefore serve as an index of the depth of the energy shift towards
low-carbon energies: " ! 1 means that the economy remains entirely fossil-based in the long-run.
On the contrary, " ! 0 means that, asymptotically, no dirty capital operates any longer. Its
dynamics is given by:
✓ ◆ ✓ ◆
"˙ Y˙1 Ẏ K̇1 (⇡e ) (1 ✓)I K1 (⇡e ) (1 ✓)I Y (⇡e )
= = = =
" Y1 Y ⌫ 1 Y1 ⌫✓ ⌫ 1 Y1 ⌫✓ ⌫ 1 Y Y1 ⌫✓
✓ ◆
1 ✓ "
, "˙ = (⇡e )
⌫1 ⌫✓
(27)
1 " " 1 1
Hence, denoting µ" := ⌫1 + ⌫2 and and ⌫ := ⌫1 ⌫2 ,
✓ ◆
(⇡e ) µ" 1
"˙ = .
⌫ ⌫✓ ⌫1 ⌫2
The new auxiliary variable, µ" , captures the mismatch between the share of green investment in
the economy and the productivity gap between the two types of capital.

W
The wage share, ! = pY , becomes:
✓ ◆ ✓ ◆ ✓ ◆
L1 + L2 L1 Y1 L2 Y2 " 1 "
!=w =w + =w +
pY Y1 pY Y2 pY a1 p a2 p
✓ ◆ (28)
w ⌫2 ⌫ 2 µ"
= " + (1 ") =w ,
a1 p ⌫1 a1 p

and consequently, its motion can be expressed as:

!˙ ẇ a˙1 ṗ µ˙" "˙ ⌫
= + = ( ) (1 )i ↵ . (29)
! w a1 p µ" µ"

114
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

As for the employment rate dynamics, it reads:

L LY ⌫ 2 µ" Y
== =
N Y N a1 N
˙ (30)
µ˙" a˙1 Ẏ Ṅ (⇡e ) "˙ ⌫
) = + = ↵ n .
µ" a1 Y N ⌫✓ µ"

The new auxiliary variable, ⌫, measures the productivity gap between the two types of capital.
It allows us to express the di↵erence between rates of profit as
pE
⇡K2 ⇡K1 = ⌫= E ⌫, (31)
p
pE
where E := p is the relative price of energy.

Thus, changes in the rate, ✓, which measures the proportion of investment directed towards K2 ,
can be written as:18

✓˙ = ✓(1 ✓)(⇡K2 ⇡K1 ) = ✓(1 ✓) ( E ⌫) . (32)

It now becomes clear that ⌫ plays the role of a threshold for the ‘oil price’:19 if E > ⌫, oil is
expensive enough to make the green technology attractive albeit being less productive. The share
of green investment rises and the transition towards a zero-carbon economy takes place. But if
E < ⌫, oil remains cheap enough to make the green investment unattractive and no exit from
the fossil fuel trap can spontaneously take place.
18
Its dynamics is determined by the gap between the two rates of profit weighted by the inertia of investment
19
At first glance, ⌫ and E should not be of the same dimension. But we normalized capital stock K in (2) so
that one unit of output invested provides one unit of capital, and energy consumption E in (1) so that one unit of
brown capital uses one unit of energy per unit of time. As a consequence, the productivity gap ⌫ and the relative
price of oil E both have the dimension, t 1 , of the inverse of time.

115
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

The overall dynamics therefore leads to a 10-dimensional non-linear system:20

8
> !˙ "˙ ⌫
>
> = ( ) (1 )i ↵
>
> ! µ"
>
>
>
> ˙ (⇡e ) "˙ ⌫
>
> = ↵ n
>
>
>
> ⌫✓ µ"
>
> ✓ ◆
>
> ˙ (⇡e )
>
> df = (⇡e ) (1 !)yd + E ⌫1 " g + ⌧ + df r + i
>
> ⌫✓
>
>
>
>
>
>
>
> ✓ ◆
>
>d˙h = c(! + rdf + rg dg ) ! rdf dh (⇡e )
>
> + i
>
> ⌫✓
>
>
>
>
>
>
>
> ✓ ◆
<
E (!, , ⇡, g, ⌧, GE , Y ) (⇡e )
g˙e = ge (33)
>
> Y ⌫✓
>
> ✓ ◆
>
> (⇡e )
>
> ˙
d g = g e + g ⌧ + d g rg + i
>
> ⌫✓
>
> ✓ ◆
>
>
>
> (⇡e )
>
> ġ = ( ) g + i
>
> ⌫✓
>
> ✓ ◆
>
> (⇡e )
>
> ⌧˙ = ⇥(⇡f ) ⌧ +i
>
> ⌫✓
>
>
>
>
>
>
>
> ✓ ◆
>
> (⇡e ) µ" 1
>
>
>
> "˙ =
>
> ⌫ ⌫✓ ⌫1 ⌫2
>
: ˙
✓ = ✓(1 ✓) ( E ⌫) ,

with the following auxiliary variables:

8
>
> ⇡e = 1 ! rdf E ⌫1 " + g ⌧
>
>
>
> 1 1 ✓ ✓
>
> = +
>
> ⌫ ⌫ 1 ⌫ 2
>
>
✓
>
> 1 " "
>
< µ" = +
⌫1 ⌫2 (34)
>
> 1 1
>
> ⌫=
>
> ⌫ 1 ⌫ 2
>
>
>
> Y d
>
> yd := = c(! + rdf + rg dg ) + (⇡e ) + ge
>
> Y
>
:
i = ⌘p (m! 1).

One readily sees that the last two di↵erential equations of (33) admit only two steady states
(", ✓) = (0, 1) or (1, 0). This corresponds respectively to an equilibrium operating and investing
only in green capital (✓ = 1 and " = 0) and another one with brown capital only (✓ = 0 and " = 1).
20
As shown in Appendix C.1.2, dh can be considered as an auxiliary variable of the system. For the same reason,
and following [Costa Lima and Grasselli, 2014], we solve the ge -equation separately, especially because ge does not
depend here on the public debt ratio dg . Note that, if this was the case, one could still solve separately the couple
(ge , dg ) as auxiliary variables depending on the rest of the system.

116
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

It can be shown that, under quite reasonable conditions, detailed in Appendix C.1.2,21 in each case,
eight equilibrium points emerge. Hence, the vector field (33) exhibits 16 long-run steady states.

The most desirable one is a “good” steady state corresponding to a positive wage share and em-
ployment rate, and a finite level of debt ratios, dg , dh and df , subsidies, g, and taxes, ⌧ . Following
a remark from [Giraud and Grasselli, 2019], it is the analog of the Solovian-like because the real
growth rate, at this steady state, equals labor productivity growth ↵ (technical progress) plus pop-
ulation growth, n. Moreover it verifies the Golden rule. Were we to use a ces production function,
Solow’s workhorse model would become g a particular case of the present one,22 this equilibrium can
be interpreted as the analog of the embedding of the unique equilibrium exhibited by the standard
Solow model within our broader setting. This steady state, however, is no longer globally stable as
in the Solow model, but turns out to be locally stable under standard assumptions.23

The second long-run equilibrium of (33) is a debt-deflationary steady state associated with a
skyrocketing level of debt ratios while wages and employment shrink to zero. This long-term steady
state is locally asymptotically unstable — which sharply contrasts with the takeaway of the previous
literature where, usually, both the Solovian-like and the debt-deflationary equilibria are stable under
reasonable conditions. As shown in Appendix C.1.4, the main reason for the instability of debt-
deflationary equilibria is the introduction of energy in the production process.

Two other equilibria correspond to deflationary states with a finite debt level – one with zero
wage, the other one with positive wages. Their local asymptotic stability cannot be judged a priori
and must be checked on a case-by-case basis. The same holds the so-called slavery equilibrium
(with no wages but employment). Eventually, there are three other infinite-valued equilibria. All
of them, however, are structurally unstable and therefore irrelevant.

Leading numerical Example

The baseline parameters for our numerical example are provided in Table [4.2].24 We use a
short-run Phillips curve of the form:

1
( )= 0, (35)
(1 )2

with parameters specified below. The investment function is taken from [Grasselli and Costa Lima,
2012]: (⇡e ) = k0 + k1 ek2 ⇡e . The tax and subsidies function are taken from [Costa Lima and
Grasselli, 2014] = ⇥(⇡f ) = ✓0 + ✓1 e✓2 ⇡f and ( ) = 0 (1 ). Following [Grasselli and Nguyen-
Huu, 2018], the consumption ratio is given by pC = c1 W + c2 Df . These functions satisfy the
assumptions in Appendix C.1.2.
21
See Appendix C.1.2 for a complete analysis of the existence and stability of equilibria.
22
It could then be recovered by putting r = = 0, m = 1, ch (x) ⌘ x, (⇡e ) ⌘ ⇡e , and ' ⌘ 0.
23
See Figure 4.1a for the trajectories of the brown Solovian-like equilibrium and Figure 4.1b for the green one.
24
See [Grasselli and Costa Lima, 2012], [McIsaac, 2016], [Grasselli and Nguyen Huu, 2016] and [Giraud and Grasselli,
2019] for exhaustive justifications of these values.

117
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Symbol & Definition Value Symbol & Definition Value

0.04
↵ Labor productivity growth rate 0.025 0 Phillips curve ( ) parameter 1 0.042
0.043
n Population growth rate 0.02 1 Phillips curve ( ) parameter 1 0.04
Depreciation rate 0.01 k0 Investment function (⇡e ) constant -0.0065
⌫1 brown capital-to-output ratio 2 k1 Investment function (⇡e ) parameter e 5
⌫2 green capital-to-output ratio 3 k2 (⇡e ) exponential parameter 20
r Nominal interest rate 3% rg Public interest rate 0%
Degree of monetary illusion 0.08 m Prices mark-up 1.8
c1 Fraction of income consumed 0.95 c2 Fraction of wealth consumed 0.1
⌘p Speed of lagged adjustment 0.03 0 Subsidies function ( ) constant 0.0045
✓0 Tax function ⇥(⇡f ) constant - 0.2 ✓1 Tax function ⇥(⇡f ) parameter 0.2
✓2 ⇥(⇡f ) exponential parameter 0.01 ✓1 Share of profits saved by firms 0.2
pE Cost of brown energy 0.18 Inertia of green investments 1

Table 4.2: Parameters

As illustrated in Figure 4.1b, an initial positive value of g ⌧ ensures the convergence of these
trajectories to a green equilibrium. These values will decrease as the transition is e↵ective (" ! 0).
The private debt ratio is also declining in the long-run because profits are not impinged by fossil
energy prices.

(a) Trajectory for pE = 0.13 from initial values: (b) Trajectory for pE = 0.18 from initial values: ! =
! = 0.6, = 0.9, d = 1, " = 0.9, ✓ = 0.3, 0.6, = 0.9, d = 1, " = 0.9, ✓ = 0.3, g = 0.09, ⌧ = 0.08
g = 0.003, ⌧ = 0.08 to Brown Solovian-like equi- to Green Solovian-like equilibrium with final values:
librium with final values: !1 = 0.55 , 1 = 0.97 , !1 = 0.82 , 1 = 0.97 , d1 = 0.16, "¯ = 0.02, ✓¯ = 0.99,
d1 = 1.48, "¯ = 0.99, ✓¯ = 1.8 E 07, g1 = 0.003, g1 = 0.002, ⌧1 = 0.005, ⇡1 = 0.16
⌧1 = 0.006, ⇡1 = 0.14

Figure 4.1: Trajectories to the locally stable equilibria of (33)

As detailed in Appendix C.1.2, this Leading Example delivers some quantitative results in a
¯ be it brown or
neighborhood of the analog of the Solovian-like equilibrium (!1 , 1 , ⇡e1 , "¯, ✓),
green:
– The expected profit rate converges to ⇡e1 that is higher in the Green Solovian-like equilibrium

118
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

than in the Brown Solovian-like equilibrium where it is impinged by the price of fossil energy.

– The demand-to-output ratio, yd , converges to yd1 = ⌫✓ (↵ + n + ) + c(!1 + rdf1 ) + ḡe , which

can be either higher or lower than 1, depending on the aggregate consumption function. With
our Leading example, yd1 = 0.83 in the case of brown equilibrium whereas yd1 = 0.99 in
the case of green equilibrium. Consequently, the Brown Solovian-like equilibrium is more
overproducing than the Green Solovian-like equilibrium.25 Notice that inventories, V , grow
to infinity along this long-run steady state with overproduction (yd1 < 1) but plummet if
yd1 > 1. With the consumption function of our Leading Example, the latter situation is quite
unlikely to happen.26

– The households’ debt ratio converges to dh1 = 1.65 in the neighborhood of the Brown Solovian-
like equilibrium and dh1 = 0.18 in the neighborhood of the green one.27

– Public debt ratio converges to dg1 = 0.24 in the neighborhood of the Brown Solovian-like
equilibrium and dg1 = 0.13 in the neighborhood of the green one.

– The pre-dividend profit rate of firms, ⇡f , converge to ⇡f1 = ⇡e1 (1 !1 )(1 yd1 ) = 0.05 in
the neighborhood of the Brown Solovian-like equilibrium and ⇡f1 = 0.14 in the neighborhood
of the green one.

– The inflation rate converges to i1 = 1.5% in the neighborhood of the Brown Solovian-like
equilibrium and i1 = 1.4% in the neighborhood of the green one.

4.4 Introducing government levers

We are now interested in introducing governmental levers acting on the master transition criterion:

E ⌫ > 0.
This can be achieved either by influencing the relative energy price, E , through a carbon tax
(according to, e.g., the corridor of values established by [Stiglitz et al., 2017]) or by reducing the
di↵erence, ⌫, in capital-to-output ratios courtesy of publicly subsidized directed technical change
targeting the output-to-capital ratio of green capital. Although these two policy tools have been
widely advocated for in the literature, their long-run efficiency has not been clearly demonstrated
so far. When high enough, a carbon tax provides a strong incentive to shift from fossil fuel to
renewable energies but at the cost of penalizing firms, reducing their profits, hence potentially
fostering corporate debts. Conversely, public subsidies may boost sovereign indebtedness up to a
non-sustainable level. Under which conditions will either of these forces prevail?
25
This implies that the stock of inventories grows to infinity.
26
Indeed, yd1 > 1 is equivalent to c(!1 +rdf1 )+(⇡e1 )+ḡe > 1. In the neighborhood of the Solovian-like equilibrium,
(⇡e1 ) = ⌫✓ (↵ + n + ) ⇠ 10 2 under reasonable specifications. The consumption function, c(·), is a continuously
strictly increasing function which tends to c1 !1 + c2 rdf1 , with c1 , c2 > 0, and we can reasonably assume ḡe ⇠ 10 2
(government expenditures other than remuneration). Consequently yd1 > 1 means c1 (1 ⇡e1 rdf1 pE ⌫1 "¯ ⌧1 +
g1 ) + c2 rdf1 + ⌫✓ (↵ + n + ) + ḡe > 1 , (c2 c1 )rdf1 + ḡe > 1 c1 + c1 (⇡e1 + pE ⌫1 "¯ g1 + ⌧1 ) ⌫✓ (↵ + n + ), where
1 c1 ⇠ 10 1 , c1 (⇡e1 + pE ⌫1 "¯ g1 + ⌧1 ) > 0 and ⌫✓ (↵ + n + ) ⇠ 10 1 . Thus, in order to guarantee yd1 > 1 for a broad
range of parameter, we need c2 c1 > 0 which implies that the volume of wealth dedicated to consumption coming
from the redistributed banks’ profit be greater than the volume of income dedicated to consumption stemming from
wages – which is very unlikely in practice.
27
Indeed d⇤h1 can be deduced from others variables which converge.

119
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

4.4.1 Introduction of a carbon tax

We first use the following simple form of carbon tax:

Tc = k c E (36)
where kc > 0 is a constant, so that Tc is directly proportional to the flow of fossil energy entering
the economy. The stock-flow matrix with this carbon tax is displayed in Table 4.3 (see in red the
changes made).

This carbon tax impacts the net profit of firms, ⇡f , as well as their expected profit, ⇡e :

⇡f = (1 !)yd rdf E ⌫1 " +g ⌧ ⌧c ,

(37)
⇡e = 1 ! rdf E ⌫1 " +g ⌧ ⌧c ,
kc kc
where ⌧c = Tc /pY = p ⌫1 " = c ⌫1 " with c := p .

Consequently, the di↵erence between the two rates of profit becomes

w w
1 pE + k c
a2 1 a1
⇡K2 ⇡K1 = + = E + c ⌫
⌫2 ⌫1 p
⇤
and, thus, the new dynamics of green investment, ✓ , is:

✓˙⇤ = ✓⇤ (1 ✓⇤ )( E + c ⌫). (38)

Unsurprisingly, the carbon tax also reduces the public debt ratio:28
✓ ◆
˙ (⇡e )
d g = g e + g ⌧ ⌧c + d g rg + i .
⌫✓
Eventually, expected profit in (4) is also modified by c:
✓ ◆
(⇡e )
⇡˙e = !( ( ) (1 )i ↵) + ( ) ⇥(⇡f ) + [1 ! ⇡e ( E + c )⌫1 "] +i
⌫✓
✓ ◆
⌫
r [(⇡e ) ⇡e + (1 !)(1 yd )] + "˙ ! pE ⌫ 1 .
µ"
(39)

Apart from these mild changes, the di↵erential system is nearly identical to (33).

The Jacobian matrix of this new system is similar to (40) (see Appendix C.1 for the whole study)
except:

@f5
= ✓⇤ (1 ✓⇤ ) c ⌫1 (40)
@"
For each equilibrium of the system (33), ✓⇤ ! 0 or 1, making the above term null. Consequently
the Jacobian matrix at each equilibrium is the same as before, apart of the last eigenvalue. This
last eigenvalue is now (1 2✓)(¯ E+ c ⌫) at Solovian-like and debt-deflationary equilibria,
whether brown or green.
28
Despite this modification, dg remains an auxiliary variable of the system and can therefore be removed from the
stability study.

120
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Households Firms Banks Gov. Row Sum

Balance sheet
Capital stock +pK +pK
Inventory +cV +cV
Public Debt Dg Dg 0
Deposits +Mh +Mf M 0
Loans Lh Lf L 0
Equity E Ef Eb 0
Column sum (Net worth) Xh Xf = 0 Xb = 0 Xg X
Transactions Current Capital
Consumption pC +pC 0
Investment +pI pI 0
Government expenditure +pGE pGE 0
Changes in inventory +cV̇ cV̇ 0
Accounting memo [GDP] [Yn ]
Wages +W W 0
Cost of Energy pE E +pE E 0
Interest on loans rLh rLf +rL 0
Interest on deposits +rMh +rMf rM 0
Interest on public debt +rg Dg rg D g 0
Subsidies +G G 0
Taxation T Tc +T +Tc 0
Dividends b b 0
Column sum (balances) Sh ⇧f pI cV̇ + pE E Sb = 0 Sg 0
Flows of funds
Change in capital stock +pK̇ +pK̇
Change in inventory cV̇ cV̇
Change in deposits +Ṁh +Ṁf Ṁ 0
Change in loans L̇h L̇f +L̇ 0
Change in public debt +Ḋg Ḋg 0
Column sum (savings) Sh Sf Sb Sg pK̇ + cV̇
Change in firm equity +Ėf (Sf + ṗK + ċV ) 0
Change in bank equity +Ėb Sb 0
Change in net worth Ė + Sh 0 0 Sg ˙ + (cV
(pK) ˙ )

Table 4.3: Balance sheet, transactions and flow of funds for the four-sector, closed
economy with inventories, prices and carbon tax

121
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

As a conclusion, the transition and the local stability criteria are identical: E + c ⌫ > 0. For
a sufficiently high c := kc /p, this criterion is satisfied, the green Solovian-like steady state is locally
asymptotically stable, and if the economy converges, it must become asymptotically green (which
means that it started in the basin of attraction of this equilibrium) because the dirty equilibria
are unstable. Conversely, if the carbon tax remains too low, the economy cannot help but, if it
converges, do so towards one of the several fossil-fuel intensive equilibria because green steady-states
are unstable.

With our Leading Example, we can display the following trajectories with di↵erent values of c :
c = 0.03, no transition occurs or c = 0.06, the transition criterion is satisfied (See Figure 4.2).

(a) Trajectories with c = 0.03: no transition (b) Trajectories with c = 0.06: the transition
occurs, the system asymptotically converges to criterion is satisfied, E + c ⌫ > 0, and
the Brown Solovian-like equilibrium as E + c the system asymptotically converge to the Green
⌫<0 Solovian-like equilibrium

Figure 4.2: Trajectories along a path to the Brown Solovian-like equilibrium of (55),
E = 0.13, with carbon tax only.

4.4.2 Innovation and directed technical change

In order to satisfy the transition criterion E ⌫ > 0, the previous section introduced a carbon
tax so as to rise the fossil energy price. Another way to guarantee that the economy goes green
consists in reducing the capital-to-output ratio of green capital. As ⌫ = ⌫11 1
⌫2 will shrink, the
transition criterion will be more easily satisfied at the cost of more public debt.

Inspired by the classical capital accumulation equation, we propose the following dynamics for a
stock of knowledge which impacts the labor productivity and so the capital-to-output ratio, ⌫2 , of
the green capital. The dynamics of knowledge is assumed to follow:

Ṙ = IRD RD R, (41)

where Ṙ measures the speed at which innovations emerge and IRD denotes investment in R&D:

IRD = ✓RD I2 + RD ⌧c Y.

122
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

In words, a share, ✓RD 2 [0, 1], of investment in green capital, I2 , is dedicated to R&D while
a fraction, RD 2 [0, 1], of the carbon tax revenues is used to fund further investments. The
parameter, RD > 0, captures the pace at which knowledge becomes obsolete.29

Relaxing the assumption that aa˙11 = aa˙22 , and keeping a1 still equal to a1 (t) = A1 e↵t , we get the
transmission of this knowledge to the labor productivity:
a2 (t) = ⇢ ⌫ A2 e↵t (42)
where ⇢ := R/K2 is the ratio of knowledge compared to green capital, and ⌫ is a coefficient
representing the incomplete transmission of knowledge.

a1 ⌫1
It follows from (3), that the dynamics of the green capital, ⌫2 = a2 , can be written as a function
of the evolution of knowledge:
⌫˙2 a˙1 a˙2
= =↵ (↵ + ⌫⇢
ˆ) = ⌫⇢
ˆ (43)
⌫2 a1 a2
Note that the evolution of the knowledge-to-capital ratio is given by:
✓RD ✓(⇡e ) + RD ⌧c ⇢✓(⇡e )
⇢ˆ = R̂ K̂2 ) ⇢˙ = . (44)
⌫2 (1 ")
The stock-flow matrix with carbon tax and directed technical change capturing green R&D is
displayed in Table 4.4 (see in red the changes made). Firms’ assets are still the capital stock,
pK, plus inventories, cV , and current money deposits, Mf , whereas their liabilities are their loans,
Lf , and equity is Ef := pK + cV + Mf Lf , but, now, the investment of firms is either capital
investment, I = I1 + I2 (1 ✓RD ), or R&D investment, IRD , as defined in (41).

Dynamics with innovation and public subsidies

The real growth rate of the economy, g, can now be expressed as:
✓ ◆ 
1 d K1 K2 1 I1 K1 I 2 K2 ⌫˙2 K2
g= + = +
Y dt ⌫1 ⌫2 Y ⌫1 ⌫2 ⌫22
(45)
(⇡e )
= + ⌫ ⇢ˆ(1 "),
⌫✓
where ⇡e , as before, depends upon the carbon tax: ⇡e = 1 ! rdf E ⌫1 " +g ⌧ ⌧c .

Consequently, the non-financial firms’ debt ratio dynamics, df = Df /pY , is defined by:
✓ ◆
d˙f D˙f ṗ (⇡e ) ⇡f (⇡e )
= g = + ⌫ ⇢ˆ(1 ") + i . (46)
df Df p df ⌫✓

In the same way, the households’ debt ratio dynamics, dh = Dh /pY , follows:
d˙h Ḋh ṗ
= g
dh Dh p
✓ ◆ (47)
c(! + rdf + rg dg ) ! rdf (⇡e )
= + ⌫⇢
ˆ(1 ") + i .
dh ⌫✓
29
A more realistic model would assume that the obsolescence rate of knowledge it itself a function of the a speed of
innovations. This type of refinement is left for further research.

123
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Households Firms Banks Gov. Row Sum

Balance sheet
Capital stock +pK +pK
Inventory +cV +cV
Public Debt Dg Dg 0
Deposits +Mh +Mf M 0
Loans Lh Lf L 0
Equity E Ef Eb 0
Column sum (Net worth) Xh Xf = 0 Xb = 0 Xg X
Transactions Current Capital
Consumption pC +pC 0
Investment +pI + pIRD pI p✓RD I2 p RD Tc 0
Government expenditure +pGE pGE 0
Changes in inventory +cV̇ cV̇ 0
Accounting memo [GDP] [Yn ]
Wages +W W 0
Cost of Energy pE E 0
+pE E
Interest on loans rLh rLf +rL 0
Interest on deposits +rMh +rMf rM 0
Interest on public debt +rg Dg rg D g 0
Subsidies +G G 0
Taxation T Tc +T +Tc 0
Dividends b b 0
Column sum (balances) Sh ⇧f pI p✓RD I2 Sb = 0 Sg 0
cV̇ + pE E
Flows of funds
Change in capital stock +pK̇ +pK̇
Change in inventory +cV̇ +cV̇
Change in deposits +Ṁh +Ṁf Ṁ 0
Change in loans L̇h L̇f +L̇ 0
Change in public debt +Ḋg Ḋg 0
Column sum (savings) Sh Sf Sb Sg pK̇ + cV̇
Change in firm equity +Ėf Sf ċV ṗK 0
Change in bank equity +Ėb Sb 0
Change in net worth Ė + Sh Sg ˙
(pK) +
˙
(cV )

Table 4.4: Balance sheet, transactions and flow of funds for the four-sector, closed
economy with inventories, prices and carbon tax

124
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

GE T
The kinetics of the government expenditure ratio, ge := Y , tax ratio, ⌧ := pY , and subsidy
G
ratio, g := pY , can be respectively written:
✓ ◆
E (!, , ⇡f , g, ⌧, GE , Y ) (⇡e )
g˙e = ge + ⌫⇢
ˆ(1 ") ,
Y ⌫✓
✓ ◆
(⇡e )
⌧˙ = ⇥(⇡f ) ⌧ + ⌫ ⇢ˆ(1 ") + i ,
⌫✓ (48)
✓ ◆
(⇡e )
ġ = ( ) g + ⌫ ⇢ˆ(1 ") + i .
⌫✓

kc kc
Since the carbon tax is unchanged: ⌧c = Tc /pY = p ⌫1 " = c ⌫1 " with c := p , the public debt
Dg
ratio, dg = pY , evolves according to

d˙g Ḋg ṗ
= g
dg Dg p
✓ ◆ (49)
(⇡e )
) d˙g = ge + g ⌧ ⌧c + d g r g + ⌫⇢
ˆ(1 ") i .
⌫✓

The dynamics of the share of output produced out of brown capital, " := Y1 /Y , is given by:
✓ ◆
"˙ Y˙1 Ẏ K̇1 (⇡e )
= = + ⌫ ⇢ˆ(1 ")
" Y1 Y ⌫ 1 Y1 ⌫✓
✓ ◆
(1 ✓)I K1 (⇡e ) (1 ✓)I Y (⇡e )
= + ⌫ ⇢ˆ(1 ") = ⌫⇢
ˆ(1 ") (50)
⌫ 1 Y1 ⌫✓ ⌫ 1 Y Y1 ⌫✓
✓ ◆ ✓ ◆
1 ✓ " (⇡e ) µ" 1
) "˙ = (⇡e ) ⌫⇢ ˆ(1 ")" = ⌫⇢
ˆ(1 ")".
⌫1 ⌫✓ ⌫ ⌫✓ ⌫1 ⌫2

W
The wage share, ! = pY , still equals:
✓ ◆ ✓ ◆ ✓ ◆ ✓ ◆
L1 + L2 L1 Y1 L2 Y2 " 1 " w ⌫ 2 µ" ⌫2
!=w =w + =w + = " + (1 ")
, =w
pY Y1 pY Y2 pY a1 p a2 p a1 p a1 p ⌫1
(51)
where µ" = 1⌫1" + ⌫"2 and consequently µ˙" = ⌫ "˙ + ⌫⌫2⇢ˆ" . Consequently, the wage share dynamics
can now be expressed as:
!˙ ẇ ⌫˙2 a˙1 ṗ µ˙" "˙ ⌫ ⌫⇢ˆ"
= + + = ( ) (1 )i ↵ ⌫⇢
ˆ + . (52)
! w ⌫2 a1 p µ" µ" ⌫ 2 µ"

The employment rate dynamics becomes:

L LY ⌫ 2 µ" Y
== =
N Y N a1 N
˙ (53)
µ˙" ⌫˙2 a˙1 Ṅ (⇡e ) "˙ ⌫ ⌫⇢ˆ"
) = + +g = + ⌫⇢
ˆ(1 ") ↵ ⌫⇢
ˆ n + .
µ" ⌫ 2 a1 N ⌫✓ µ" ⌫ 2 µ"

125
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL
w w
1 a2
1 a1 pE +kc
The gap between rates of profit remains unchanged: ⇡K2 ⇡K1 = ⌫2 ⌫1 + p =
E + c ⌫ and so is the motion of the share ✓⇤ , of green investment:

✓˙⇤ = ✓⇤ (1 ✓⇤ )( E + c ⌫). (54)

The di↵erential system followed by our economy is now 11-dimensional:

8 ✓ ◆
>
> !˙ "˙ ⌫ "
>
> = ( ) (1 )i ↵ ⌫⇢ˆ 1
>
> ! µ" ⌫ 2 µ"
>
> ✓ ◆
>
> ˙ (⇡e ) "˙ ⌫ 1
>
> = ↵ n
>
> ⌫⇢ˆ" 1
>
> ⌫✓ µ" ⌫ 2 µ"
>
> ✓ ◆
>
> (⇡e )
>
> ˙
df = (⇡e ) (1 !)yd + E ⌫1 " g + ⌧ + ⌧c + df r + ⌫⇢
ˆ(1 ") i
>
> ⌫✓
>
>
>
>
>
>
>
> ✓ ◆
>
> (⇡e )
>d˙ = c(! + rd + r d ) ! rd
> dh + ⌫ ⇢ˆ(1 ") + i
>
> h f g g f
⌫✓
>
>
>
>
>
>
>
> ✓ ◆
>
>
>
> E (!, , ⇡, g, ⌧, ⌧c , GE , Y ) (⇡e )
>
> g˙e = ge + ⌫ ⇢ˆ(1 ")
>
< Y ⌫✓
✓ ◆
˙ (⇡e ) (55)
>
> d = g e + g ⌧ ⌧c + d g r g + ⌫⇢ ˆ(1 ") i
> g
> ⌫✓
>
> ✓ ◆
>
> (⇡e )
>
> ġ = ( ) g + ⌫ ⇢ˆ(1 ") + i
>
> ⌫✓
>
>
>
> ✓ ◆
>
> (⇡e )
>
> ⌧˙ = ⇥(⇡f ) ⌧ + ⌫ ⇢ˆ(1 ") + i
>
> ⌫✓
>
>
>
>
>
>
>
> ✓ ◆
>
> (⇡e ) µ" 1
>
> "˙ = ⌫⇢
ˆ(1 ")"
>
> ⌫ ⌫✓ ⌫1 ⌫2
>
>
>
>
>
>
> ✓˙⇤ = ✓⇤ (1 ✓⇤ )( E + c ⌫)
>
>
>
>
>
>
>
> (✓RD ⇢)✓(⇡e ) + RD ⌧c
>
>
: ⇢˙ =
⌫2 (1 ")

126
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Auxiliary variables
8
>
> ⌫ˆ2 = ⌫⇢
ˆ
>
>
>
> ⌧c = c ⌫1 "
>
>
>
>
>
> ⇡e = 1 ! rdf E ⌫1 " +g ⌧ ⌧c
>
>
>
> i = ⌘p (m! 1).
>
>
>
>
>
> 1 1 ✓ ✓
>
< = +
⌫✓ ⌫1 ⌫2
1 " " (56)
>
> µ" = +
>
>
>
> ⌫1 ⌫2
>
> 1 1
>
>
>
> ⌫=
>
> ⌫1 ⌫2
>
>
>
> Yd
>
> yd := = c(! + rdf + rg dg ) + (⇡e ) + ge
>
> Y
>
:
⇡e ⇡f = (1 !)(1 yd )

Under the same reasonable conditions as before (detailed in Appendix C.1.2),30 the vector field
(55) exhibits the same types of steady states —be they brown or green depending upon to the
solution of the (", ✓)-equations. The main departure with the previous section, however, arises from
the introduction of the variable ⌫2 : new equilibria appear in which " and ✓ need not equal 0 nor
1. These new steady states are less manichean than before, since they are compatible with the
coexistence of green and brown capital in the long run. But they raise, of course, new questions:
under which conditions will the economy converge to a zero-carbon sustainable state, as opposed
to a low-carbon situation where, even in the long-term, dirty capital would still operate, albeit in a
moderate way?
For each of the clean variants of the three nature of equilibria, we still have locally stable equilibria
under standard assumptions : the desirable situation with positive wages and employment, and a
finite level of debt ratios, dg , dh and df , subsidies g and tax ⌧ . As before debt-deflationary steady
state associated with a skyrocketing level of debt ratios while wages and employment shrink to zero
are locally unstable. What is new, however, is that, under the following technical assumption, the
brown Solovian-like steady state turns out to be locally unstable as well:

Assumption C The ratio, ⇢ := R/K2 , is such that its time-derivative, ⇢,

˙ remains
+
bounded from above whenever K2 ! 0 .

This restriction amounts to saying that, whenever clean capital shrinks to zero (because of the
lack of green investment and exponential depreciation), the ecological know-how encapsulated in R
shrinks even more rapidly. This can be achieved, e.g., by imposing that the decay rate of knowledge,
31
RD , in (41) be such that RD > .

Leading numerical Example

30
See Appendix C.2 for a complete analysis of the existence and stability of equilibria.
31
The other equilibria also appear to be locally unstable, and will not be further considered. We let the reader
verify they are still locally or structurally unstable with the new system (55).

127
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Symbol & Definition Value Symbol & Definition Value

c Carbon tax ratio 0.005 ✓RD Share of investment in R&D 0.25
⌫ Knolwedge transmission factor 0.05 RD Carbon tax share to public R&D 0.25
lag Delay before intervention 0

Table 4.5: New Parameters

The baseline parameters for our simulations are the same as in Table [4.2] 32 to which we have
added the parameterization of Table [4.5]. Depending on these values, transition from a initial state
leading to brown equilibria ( e = 0.13, ⌫1 = 2, ⌫2 = 3) occurs or not.

Assuming first, that public and private fundings for green R&D are equal, we find that transition
occurs whatever is the delay of intervention (lag = 0 in Figure 4.3a or lag = 200 - which corresponds
to almost a century - Figure 4.3b) and whether there is a carbon tax or not. The more the
intervention is delayed, the longer the transition is (around half a century with no lag while it last
almost two century with lag = 200) and consequently the volume of private and public debt is even
more important.

(a) Trajectories with no carbon tax c = 0 and (b) Trajectories with no carbon tax c = 0 and
no delay in intervention lag = 0: transition an important delay lag = 200: transition occurs
occurs, the system asymptotically converges to because the criterion will be satisfied, E + c
the Green Solovian-like equilibrium because ⌫2 ⌫ > 0, and eventually the system asymptoti-
is such that E ⌫ becomes positive cally converge to the Green Solovian-like equilib-
rium

Figure 4.3: Trajectories of (55) from a path leading to the Brown Solovian-like equilib-
rium, E = 0.13 to the Green Solovian-like equilibrium, with no carbon tax

The introduction of carbon tax reduces the public debt during the transition phase and accelerates
the transition. See Figure 4.4a with a small carbon tax, c = 0.005, where the transition phase
lasts 150 time steps (around 75 years) compared with Figure 4.4b with an important carbon tax,
c = 0.03, where the transition phase lasts ’only’ 50 time steps (around 25 years).
32
See [Grasselli and Costa Lima, 2012], [McIsaac, 2016], [Grasselli and Nguyen Huu, 2016] and [Giraud and Grasselli,
2019] for exhaustive justifications of these values.

128
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

(a) Trajectories with small carbon tax c = (b) Trajectories with important carbon tax c =
0.005 and no delay in intervention lag = 0: tran- 0.03 and no delay lag = 0: transition quickly oc-
sition slowly occurs curs

Figure 4.4: Trajectories of (55) from a path leading to the Brown Solovian-like equilib-
rium, E = 0.13 to the Green Solovian-like equilibrium, with di↵erent carbon taxes

Hybrid equilibria
Apart from the simplistic equilibria, all-green or all-brown, of the previous section, the dynamical
system with government intervention (55) displays new long-run steady states where both brown
and green capital coexist (" 6= 0 or 1 and ✓ 6= 0 or 1).

These hybrid equilibria are not locally stable in our numerical Example but can be turned into
stable ones with alternate parameterizations. This can be achieved by assuming, for instance, that
public investment in green R&D is low RD = 0.05 (instead of RD = 0.25) and that private
investment is even lower ✓RD = 0.01 (instead of ✓RD = 0.25, one can display Figure 4.5a with
c = 0.02 or Figure 4.5b with c = 0.025. Small variations of carbon tax here a↵ect the equilibrium
values of the share of output produced from brown capital, " and the share of green investments, ✓.

129
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

(a) Trajectories with c = 0.02 lead towards (b) Trajectories with c = 0.025 lead towards
equilibrium values: " = 0.56 and ✓ = 0.54 equilibrium values: " = 0.4 and ✓ = 0.7

Figure 4.5: Trajectories of (55) with low private and public investments, ✓RD = 0.01 and
RD = 0.05 leading from the Brown Solovian-like equilibrium ( E = 0.13) to intermediate
Solovian-like equilibria.

4.5 Concluding remarks

In this paper, we first built a simple transition model from fossil fuel-intensive to fossil-free capital
in a dynamics with public and private debts, inventories and government intervention. Multiple
long-run equilibria are the rule so that the main question of economic policy becomes: how to drive
an economy towards the basin of attraction of some desirable long-run steady state?

We identified conditions under which the transition to green capital can be implemented without
leading to an overhang of debt, be it private or public. The key parameter driving whether transition
towards renewable energies takes place is the di↵erence between the price of fossil energy and the
productivity gap between the two types of capital: E ⌫. Surprisingly, because of inertia in the
allocation of investment, the debt-deflationary equilibria, whether brown or green, turn out not to
be locally stable. Each of them is crossed by a one-dimensional repulsive submanifold. This raises
an interesting question for future research: what are the forces that drive an economy towards
debt-deflation in a world when multiple types of capital coexist and investment exhibits some kind
of inertia?

Introducing governmental levers is done in this paper both by influencing the energy price through
a carbon tax (according to the corridor of values established by [Stiglitz et al., 2017]) and by reducing
the capital-to-output gap between the two types of capital.

Whenever only the carbon tax tool is used, we have shown that the transition and the local
stability criteria happen to be identical: E + c ⌫ > 0. For a sufficiently high carbon tax, this
criterion is satisfied, the green Solovian-like steady state is locally asymptotically stable, and if the
economy converges, it must become asymptotically green because the dirty equilibria turn out to be
asymptotically unstable. Conversely, if the carbon tax remains too low, the economy cannot help
but, if it converges, do so towards one of the several fossil-fuel intensive equilibria because green
steady-states become unstable.

130
CHAPTER 4. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

In order to capture directed technical change, we introduced public subsidies of green innovation,
with the hope to be able to ensure that the transition will take place, whatever being the starting
point of the economy.

We exhibited a particular parameterization leading to a transition with a low price of fossil

energy. In this example, public and private funding for green R&D ensure that the transition
is achieved, even if the economy was originally following a dirty path. Introducing a carbon tax
accelerates the transition but does not guarantee it. As in [Acemoglu et al., 2012], any delay in the
implementation of public transition policies can be very costly in time and money. Contrary to the
result in [Acemoglu et al., 2012], however, here, no transition can occur under laissez-faire.

Moreover, we have shown that, in general, new types of long-run steady states emerge, where
brown and green capital coexist. In our specific example, this is due to the low level of public
and private investments in green R& D. While less simplistic than our previous 0-1 equilibria,
these hybrid steady states suggest that, not only is laissez-faire insufficient but a too timid public
intervention may fail to drive the economy towards a zero-carbon long-run situation. A surprising
result, eventually, is that, under a realistic assumption (Assumption C), the dirty Solovian-like
equilibrium becomes asymptotically unstable.

Several improvements of this paper should be considered. The price dynamics should obviously
incorporate the cost of energy and be at least partially demand-driven. This would allow consump-
tion prices to reflect the increasing cost of fossil fuel whenever a carbon tax is put into practice. The
impact on e↵ective demand would presumably be non-negligible. The main question would then
become whether a depressed demand would be able to prevent the private sector from investing
sufficiently in green infrastructures so as to reach the clean Solovian-like equilibrium. The same
question would be probably reinforced if households were to consume energy as well (and not just
the production sector) and would therefore directly bear part of the additional cost induced by a
carbon tax.

On the other hand, we have focused exclusively, here, on the long-run analysis. A short-run
analysis would also help understand the pros and cons of the political levers considered in this
paper and which have been much debated in the U.S. in the first part of the Biden mandate.
In particular, the asymptotic instability of the dirty Solovian-like steady state should be put in
perspective with the time urgency induced by global warming.

Another improvement of this paper is left for further work, which consists in endogenizing the
price, pE , of fossil energy. Eventually, building on the present work, one might introduce a variety of
di↵erent types of capital in a full-blown Putty-Clay production sector similar to the one introduced
by [Akerlof and Stiglitz, 1969] in order to get a more realistic setting than the extreme case envisaged
here.

131
Chapter 5

Conclusion

The aim of our work was to develop a robust transition model that should be able to provide
information on public policies to avoid either an attractive equilibrium with fossil fuel-intensive
capital only or a collapse due to a climate-induced Minsky moment. The construction of such a
model was based on [Grasselli and Costa Lima, 2012], which belongs to the family of disequilibrium
models. This model was selected because of its consistent stock-flow structure with nonlinear
dynamics leading to multiple equilibria, which can incorporate this Minsky moment.

To cope with the criticisms formulated against this family of model, especially concerning their
strong dependence on parameters, we developed a sensitivity analysis of the parameters of [Grasselli
and Costa Lima, 2012]. We study the relative influence of uncertainties on all inputs’ parameters
with a variance-based Global Sensitivity Analysis (GSA), because this model contains a large num-
ber of nonlinearities and apriori independent inputs. Due to its small number of inputs (and
consequently its low computational cost), we chose a GSA with Sobol’ indices. The calculation of
the Sobol’ indices was based on combinations of inputs values obtained from a Design of Exper-
iment (DoE): an Optimal Latin Hypercube Sampling, according to the [Saltelli, 2002] method of
estimation.

We noted that the most influential parameters on the model outputs are the labor productivity
growth rate, the population growth rate, the depreciation rate and the capital-to-output ratio.
The last one has the maximal first-order Sobol’ indice (less than 0.2). These values are not very
dependent on the number of simulations as we show it by successively increasing their number until
the satisfaction of a certain criterion about the standard deviation of our results (less than 0.05).
Two of these most influential parameters can be assessed with a certain confidence as detailed
in [McIsaac, 2016] (from demographic data from United Nations time series for the population
growth rate and by averaging the time series of the labour productivity growth rate). On the
contrary, the capital-to-output ratio and the depreciation rate are hard to assess that can impede
the robustness of the model.

Although we cannot directly compare these results to other similar GSA on DSGE, RBC and
IAM models, because the outputs and inputs of each study are di↵erent, we can note that the model
observed in [Ratto, 2008], [Harenberg et al., 2017] or [Nordhaus, 2008] are highly sensitive to a single
parameter that is generally hard to assess while the sensitivity of the [Grasselli and Costa Lima, 2012]
model with a Leontief production function is more balanced between four parameters. Moreover
when studying the original [Goodwin, 1967] model, we noticed that its sensitivity was dependent

132
CHAPTER 5. CONCLUSION

on the same parameters as its extension, that is quite reassuring that refining the model does not
seem to a↵ect the deep nature of its sensitivity,

As the extensions of this model do not seem to fundamentally alter its robustness, improvements
have been made in particular concerning the relaxation of Say’s “law” through the introduction of
inventories, which is essential for studying money creation. The dynamic analysis of this “extended”
model showed that the interplay between e↵ective demand, production and money creation can lead
either to an equilibrium with a finite private debt-to-output ratio, positive but stable inflation and
positive money velocity — the monetary Solovian equilibrium— or to another steady state where
debts skyrocket while the wage share, the employment rate, and money velocity both collapse.
Inflation either collapses as well or can even become negative — a hallmark for the deflationary
trap.

Moreover, money has been shown to be non-neutral both in the short- and the long-run, even
though the monetarist interpretation of the exchange equation happens to be true at the monetary
Solovian equilibrium — and only there. The fact that banks create credit and money out of nothing,
if used appropriately, results in non-hyper-inflationary growth, provided the economy crosses the
basin of attraction of the monetary Solovian equilibrium. We also showed that central bank inter-
vention may have a clear positive e↵ect in preventing a crisis characterized by collapsing employment
rates. According to the persistence result, central bank intervention, in the form of a responsive
enough monetary policy, prevents the economy from remaining permanently at arbitrarily low levels
of employment regardless of the initial conditions of the system.

Eventually a simple transition model is constructed with two natures of capital (can be generalized
in a ”Putty-Clay” framework in the sense of [Akerlof and Stiglitz, 1969], [Cass and Stiglitz, 1969]),
from fossil fuel-intensive to fossil-free capital in a dynamics with public and private debts, inventories
and government intervention in order to answer the question: how to drive an economy towards the
basin of attraction of some desirable long-run steady state? We identified conditions under which the
transition to green capital can be implemented without leading to an overhang of debt, be it private
or public. The key parameter driving whether transition towards renewable energies takes place is
the di↵erence between the price of fossil energy and the productivity gap between the two types
of capital. Surprisingly, because of inertia in the allocation of investment, the debt-deflationary
equilibria, whether brown or green, turn out not to be locally stable.

Introducing governmental levers is done in this paper both by influencing the energy price through
a carbon tax (according to the corridor of values established by [Stiglitz et al., 2017]) and by
reducing the capital-to-output gap between the two types of capital. Whenever only the carbon
tax tool is used, we have shown that the transition and the local stability criteria happen to be
identical: E + c ⌫ > 0. For a sufficiently high carbon tax, this criterion is satisfied, and if the
economy converges, it must become asymptotically green because the dirty equilibria turn out to be
asymptotically unstable. Conversely, if the carbon tax remains too low, the economy cannot help
but, if it converges, do so towards one of the several fossil-fuel intensive equilibria because green
steady-states become unstable.

In order to capture directed technical change, we introduced public subsidies of green innovation,
with the hope to be able to ensure that the transition will take place, whatever being the starting
point of the economy. We exhibited a particular parameterization leading to a transition with a

133
CHAPTER 5. CONCLUSION

low price of fossil energy. In this example, public and private funding for green R&D ensure that
the transition is achieved, even if the economy was originally following a dirty path. Introducing
a carbon tax accelerates the transition but does not guarantee it. As in [Acemoglu et al., 2012],
any delay in the implementation of public transition policies can be very costly in time and money.
Contrary to the result in [Acemoglu et al., 2012], however, here, no transition can occur under
laissez-faire.

Moreover, we have shown that, in general, new types of long-run steady states emerge, where
brown and green capital coexist. In our specific example, this is due to the low level of public
and private investments in green R&D. While less simplistic than our previous 0-1 equilibria, these
hybrid steady states suggest that, not only is laissez-faire insufficient but a too timid public in-
tervention may fail to drive the economy towards a zero-carbon long-run situation. A surprising
result, eventually, is that, under a realistic assumption, the dirty Solovian-like equilibrium becomes
asymptotically unstable.

Several improvements of this work should be considered. About the introduction of money,
there are, of course, a number of institutional aspects that we have neglected in this first tentative
pass: reserve requirements, capital adequacy requirements (such as the Basel iii/crr framework),
prudential regulation... Added to the present framework, these refinements could enable us to tackle
important issues around a counteryclical leverage-based regulation of banks. About transition, we
focused exclusively, here, on the long-run analysis. A short-run analysis would also help understand
the pros and cons of the political levers considered that should be put in perspective with the time
urgency induced by global warming.

134
List of Figures

1.1 Modeling after [Rosen, 1991] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

2.1 Trajectories leading to the locally stable equilibria of system of the [Grasselli and
Costa Lima, 2012] model with the set of parameters defined in Table 2.1. . . . . . . 55
(a) Trajectory from initial values: !0 = 0.8, 0 = 0.9 and d0 = 0.1 to the Good
Equilibrium with final values: (!1 , 1 , d1 ) = (0.8361, 0.9686, 0.0702) . . . . . . . 55
(b) Trajectory from initial values: !0 = 0.7, 0 = 0.7 and d0 = 0.1 to the Explosive-
debt Equilibrium with final values (0, 0, 9 · 10101 ) . . . . . . . . . . . . . . . . . 55
2.2 Basin of attraction of the Good Equilibrium for the set of parameters defined in Table
2.1 ; each green point is a set of initial values for the inputs (!; ; d) leading to the
Good Equilibrium ; the rest of the space leads to the Explosive Debt Equilbrium . . 56
2.3 Decision diagram for the choixe of a SA method (from [De Rocquigny, 2008] . . . . . 59
2.4 Comparaison of samplings with only three possible levels of two inputs . . . . . . . . 65
2.5 Convergence of the Sobol’ indices of the main influent inputs (↵ ; ; ; ⌫) on (!1 ; ; d1 ) 69
(a) First-order indices of (↵ ; ; ; ⌫) on (!1 ; 1 ; d1 ) for various number of simulations 69
(b) First-order indices of (↵ ; ; ; ⌫) on !1 for various number of simulations . . 69
(c) First-order indices of (↵ ; ; ; ⌫) on 1 for various number of simulations . . 69
(d) First-order indices of (↵ ; ; ; ⌫) on d1 for various number of simulations . . 69
2.6 Sobol’ indices of [Grasselli and Costa Lima, 2012] for the most influential parameters
and their standard deviation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
(a) First-order Sobol’ indices and standard deviation of (↵ ; ; ; ⌫) on (!1 ; 1 ; d1 ) 70
(b) Total-e↵ect Sobol’ indices and standard deviation of (↵ ; ; ; ⌫) on !1 . . . 70

3.1 Income velocity in US Dollars, 2008-2019. . . . . . . . . . . . . . . . . . . . . 80

3.2 Income velocity in Czech Koruna, 2008-2018. . . . . . . . . . . . . . . . . . . 80
3.3 Private debt/GDP Canada, 2008-2014. . . . . . . . . . . . . . . . . . . . . . . 81
3.4 Trajectories leading to the locally stable steady states of (26). . . . . . . . . 94
(a) Trajectories from initial values: !0 = 0.8, 0 = 0.9 and d0 = 0.1, to the Solovian
equilibrium with final value: !1 = 0.79 , 1 = 0.97 , df1 = 0.602 . . . . . . . . 94
(b) Trajectories from initial values: !0 = 0.3, 0 = 0.5 and d0 = 1 to the debt-
deflationary equilibrium: !p2 = 0, 2 = 0 , df2 = 9.10101 . . . . . . . . . . . . . 94
3.5 Inflation and demand-to-output ratio . . . . . . . . . . . . . . . . . . . . . . . . 95
(a) Convergence of inflation, i, to its Solovian asymptotic value: i1 = 1.2% . . . . 95
(b) Convergence of inflation, i, to its debt-deflationary asymptotic value i = 3% . 95
(c) Convergence of the demand-to-output ratio, yd , along a path to the Solovian
equilibrium, with final value: yd1 = 0.83 . . . . . . . . . . . . . . . . . . . . . . 95
(d) Convergence of the demand-to-output ratio, yd , to its upper-bound, c+ + k0 ,
along a path to the debt-deflationary equilibrium . . . . . . . . . . . . . . . . . 95

135
LIST OF FIGURES

3.6 Influence of imperfect competition on the basins of attraction. . . . . . . . 95

(a) Basin of attraction of the Solovian equilibrium with perfect competition: =
µ = 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
(b) Basin of attraction of the Solovian equilibrium with imperfect competition:
> µ > 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.7 Trajectories leading to the Solovian equilibrium of (27). . . . . . . . . . . . . 100
(a) Trajectories from initial values: !0 = 0.8, 0 = 0.9 and d0 = 0.1, m0 = 0.1 to
final value: !1⇤ = 0.79113030 , ⇤1 = 0.98529608, and d⇤f1 = 0.05865288 . . . . . 100
(b) Evolution of the money-to-output ratio along the path to the Solovian Equilib-
rium towards m⇤1 = 0.01076340. . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
3.8 Trajectories leading to solutions of (27) to (0, 0, +1, +1). . . . . . . . . . . . 101
(a) Trajectories from initial values: !0 = 0.3, 0 = 0.5 and d0 = 1 to the debt-
deflationary equilibrium: (!, , df , m) = (3.252 e 09, 1.916 e 05, 4.222 e+06, 9.588 e+03).101
(b) Income velocity of money on a path to the debt-deflationary equilibrium. . . . 101

4.1 Trajectories to the locally stable equilibria of (33) . . . . . . . . . . . . . . . 118

(a) Trajectory for pE = 0.13 from initial values: ! = 0.6, = 0.9, d = 1, " = 0.9,
✓ = 0.3, g = 0.003, ⌧ = 0.08 to Brown Solovian-like equilibrium with final
values: !1 = 0.55 , 1 = 0.97 , d1 = 1.48, "¯ = 0.99, ✓¯ = 1.8 E 07, g1 = 0.003,
⌧1 = 0.006, ⇡1 = 0.14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
(b) Trajectory for pE = 0.18 from initial values: ! = 0.6, = 0.9, d = 1, " = 0.9,
✓ = 0.3, g = 0.09, ⌧ = 0.08 to Green Solovian-like equilibrium with final values:
!1 = 0.82 , 1 = 0.97 , d1 = 0.16, "¯ = 0.02, ✓¯ = 0.99, g1 = 0.002, ⌧1 = 0.005,
⇡1 = 0.16 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
4.2 Trajectories along a path to the Brown Solovian-like equilibrium of (55),
E = 0.13, with carbon tax only. . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
(a) Trajectories with c = 0.03: no transition occurs, the system asymptotically
converges to the Brown Solovian-like equilibrium as E + c ⌫<0. . . . . 122
(b) Trajectories with c = 0.06: the transition criterion is satisfied, E + c
⌫ > 0, and the system asymptotically converge to the Green Solovian-like
equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
4.3 Trajectories of (55) from a path leading to the Brown Solovian-like equi-
librium, E = 0.13 to the Green Solovian-like equilibrium, with no carbon
tax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
(a) Trajectories with no carbon tax c = 0 and no delay in intervention lag = 0:
transition occurs, the system asymptotically converges to the Green Solovian-
like equilibrium because ⌫2 is such that E ⌫ becomes positive . . . . . . . 128
(b) Trajectories with no carbon tax c = 0 and an important delay lag = 200:
transition occurs because the criterion will be satisfied, E + c ⌫ > 0,
and eventually the system asymptotically converge to the Green Solovian-like
equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
4.4 Trajectories of (55) from a path leading to the Brown Solovian-like equi-
librium, E = 0.13 to the Green Solovian-like equilibrium, with di↵erent
carbon taxes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
(a) Trajectories with small carbon tax c = 0.005 and no delay in intervention
lag = 0: transition slowly occurs . . . . . . . . . . . . . . . . . . . . . . . . . . 129
(b) Trajectories with important carbon tax c = 0.03 and no delay lag = 0: tran-
sition quickly occurs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

136
LIST OF FIGURES

4.5 Trajectories of (55) with low private and public investments, ✓RD = 0.01 and
RD = 0.05 leading from the Brown Solovian-like equilibrium ( E = 0.13)
to intermediate Solovian-like equilibria. . . . . . . . . . . . . . . . . . . . . . . 130
(a) Trajectories with c = 0.02 lead towards equilibrium values: " = 0.56 and
✓ = 0.54 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
(b) Trajectories with c = 0.025 lead towards equilibrium values: " = 0.4 and ✓ = 0.7130

A.1 Evolution of a Predator-Prey population (interpreted from [Lotka, 1920] and [Volterra,
1928]) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
A.2 Trajectories of the locally stable equilibria of the 2D-model with the set of parameters
defined in [Grasselli and Costa Lima, 2012]. . . . . . . . . . . . . . . . . . . . . . . 147
(a) Evolution to the trajectories in the 2D model for initial values: !0 = 0.75,
0 = 0.75. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
(b) Evolution to the trajectories in the 2D model for initial values: !0 = 0.8, 0 = 0.9.147
A.3 Trajectories of the two locally stable equilibria of the [Grasselli and Costa Lima,
2012] model with the set of parameters defined in Table 2.1 . . . . . . . . . . . . . . 149
(a) Evolution to the Good Equilibrium for initial values: !0 = 0.8, 0 = 0.9 and
d0 = 0.1 and final values: (!1 , 1 , d1 ) = (0.8361, 0.9686, 0.0702) . . . . . . . . . 149
(b) Evolution to the Bad Equilibrium for initial values: !0 = 0.7, 0 = 0.7 and
d0 = 0.1 and final values (0, 0, +1) . . . . . . . . . . . . . . . . . . . . . . . . 149
A.4 Performances of MC simulations [Lomas and Eppel, 1992], simple random sampling
and LHS [Iman, 1999] according to the number of simulated samples. . . . . . . . . . 155
(a) Relationship between normalized confidence interval and number of Monte-
Carlo simulations [Lomas and Eppel, 1992] . . . . . . . . . . . . . . . . . . . . 155
(b) Sampling coverage for simple random sampling and LHS [Iman, 1999], showing
the number of samples on the x-axis, and the fraction of the sampling coverage
on the y-axis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
A.5 SA methods graphical synthesis [Iooss and Lemaı̂tre, 2015]. . . . . . . . . . . . . . . 171

B.1 Goodwin with Normalized Philips Curve !0 = 0.8 and 0 = 0.9 . . . . . . . . . . . . 197
B.2 Evolution of the function f for = 0.98 . . . . . . . . . . . . . . . . . . . . . . . . . 199
B.3 Evolution of the function f for = 0.95 . . . . . . . . . . . . . . . . . . . . . . . . . 199
B.4 Evolution of the function f for = 0.95 . . . . . . . . . . . . . . . . . . . . . . . . . 199

C.1 Trajectories to the brown locally stable equilibria of (33) . . . . . . . . . . . 209

(a) Trajectory from initial values: ! = 0.6, = 0.9, d = 1, " = 0.9, ✓ = 0.3, g =
0.003, ⌧ = 0.08 to Solovian-like equilibrium with final values: !1 = 0.5546858
, 1 = 0.9733608 , d1 = 1.478804, "¯ = 0.9999996, ✓¯ = 1.829661e 07, g1 =
0.003236408, ⌧1 = 0.006333863, ⇡1 = 0.1378527 . . . . . . . . . . . . . . . . . . 209
(b) Trajectory (with k0 = 0.05) from initial values: ! = 0.53, = 0.9, d = 1,
" = 0.9, ✓ = 0.3, g = 0.09, ⌧ = 0.08 to debt-deflationary equilibrium with final
values: !2 = 0 , 2 = 0 , d2 = 351.99, "¯ = 1, ✓¯ = 0, g2 = 26, 99, ⌧2 = 0.67 . . 209
C.2 Trajectories to the green locally stable equilibria of (33) . . . . . . . . . . . 209
(a) Trajectory from initial values: ! = 0.6, = 0.9, d = 1, " = 0.9, ✓ = 0.3, g =
0.09, ⌧ = 0.08 to Solovian-like equilibrium with final values: !1 = 0.8226136
, 1 = 0.9713984 , d1 = 0.1589752, "¯ = 0.0217836, ✓¯ = 0.9888603, g1 =
0.0021762718, ⌧1 = 0.005458815, ⇡1 = 0.1614925 . . . . . . . . . . . . . . . . . 209

137
LIST OF FIGURES

(b) Trajectory (with k0 = 0.05) from initial values: ! = 0.5, = 0.9, d = 1,

" = 0.9, ✓ = 0.3, g = 0.09, ⌧ = 0.08 to debt-deflationary equilibrium with final
values: !2 = 0 , 2 = 0 , d2 = 1.1 E12, "¯ = 0, ✓¯ = 1, g2 = 0.12, ⌧2 = 6.8 E10 . . 209

138
List of Tables

2.1 Canonical parameterization of [Grasselli and Costa Lima, 2012]. . . . . . . 55

2.2 Definition and Range of the inputs of our study . . . . . . . . . . . . . . . . . . . . . 64
2.3 Sobol’ indices and standard deviations (with 10 bootstrap replicates) of all inputs
(ranked from #1 the main influent parameter to #10 the less influent one) of the
[Grasselli and Costa Lima, 2012] model on the outputs: the Good Equilibrium values
of wage share !1 , employment rate 1 and debt ratio d1 . . . . . . . . . . . . . . . . 67
2.4 Sobol’ indices and standard deviations (with 10 bootstrap replicates and 500 simu-
lations) of the main influent inputs of the [Grasselli and Costa Lima, 2012] model on
the outputs: the Good Equilibrium values of wage share !1 , employment rate 1 and
debt ratio d1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2.5 Parameters selected for the analysis and their distributions (from [Nordhaus, 2008]). 75

3.1 Balance sheet, transactions and flow of funds for the three-sector, closed
economy with inventories and prices . . . . . . . . . . . . . . . . . . . . . . . . 88
3.2 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.3 Endogenous money creation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

4.1 Balance sheet, transactions and flow of funds for the four-sector, closed
economy with inventories and prices . . . . . . . . . . . . . . . . . . . . . . . . 112
4.2 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
4.3 Balance sheet, transactions and flow of funds for the four-sector, closed
economy with inventories, prices and carbon tax . . . . . . . . . . . . . . . . 121
4.4 Balance sheet, transactions and flow of funds for the four-sector, closed
economy with inventories, prices and carbon tax . . . . . . . . . . . . . . . . 124
4.5 New Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

A.1 Comparison of local and global SA methods [Hoes and De Vann, 2005] and [Hopfe
et al., 2007]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
A.2 Comparison of four di↵erent methods to conduct global sensitivity analysis (European
commission, 2005) [Hopfe et al., 2007]. . . . . . . . . . . . . . . . . . . . . . . . . . . 158
A.3 Part 1 / Second-order Sobol’ indices and standard deviations (with 10 bootstrap
replicates and 100 simulations) of all inputs in the [Grasselli and Costa Lima, 2012]
model on the outputs: the Good Equilibrium values of wages share !1 and employ-
ment rate 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
A.4 Part 2 / Second-order Sobol’ indices and standard deviations (with 10 bootstrap
replicates and 100 simulations) of all inputs in the [Grasselli and Costa Lima, 2012]
model on the outputs: the Good Equilibrium values of wages share !1 and employ-
ment rate 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173

139
LIST OF TABLES

A.5 First-order Sobol’ indices and standard deviations (10 bootstrap replicates for each
200 & 300 simulations) of the main influent inputs of the [Grasselli and Costa Lima,
2012] model on the outputs: the Good Equilibrium values of wages share !1 , employ-
ment rate 1 and debt ratio d1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
A.6 First-order Sobol’ indices and standard deviations (10 bootstrap replicates of each 400
& 500 simulations) of the main influent inputs of the [Grasselli and Costa Lima, 2012]
model on the outputs: the Good Equilibrium values of wages share !1 , employment
rate 1 and debt ratio d1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
A.7 Second-order Sobol’ indices and standard deviations (10 bootstrap replicates and 500
simulations) of the main influent inputs of the [Grasselli and Costa Lima, 2012] model
on the outputs: the Good Equilibrium values of wages share !1 , employment rate 1
and debt ratio d1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
A.8 Definition and Range of the inputs of the GSA for the [Goodwin, 1967] model . . . . 175
A.9 Sobol’ indices and standard deviations of all inputs of the [Goodwin, 1967] model on
the outputs: the Good Equilibrium values of wages share, !1 , and employment rate,
1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
A.10 Second-order Sobol’ indices and standard deviations of all inputs of the [Goodwin,
1967] model on the outputs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
A.11 Definition and Range of the inputs of the GSA for the [Grasselli and Costa Lima,
2012] model with CES production function . . . . . . . . . . . . . . . . . . . . . . . 177
A.12 Sobol’ indices and standard deviations (with 10 bootstrap replicates) of all inputs of
the [Grasselli and Costa Lima, 2012] model with a CES production function on the
outputs: the Good Equilibrium values of wages share !1 , employment rate 1 and
debt ratio d1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
A.13 Part 1 / Second-order Sobol’ indices and standard deviations (with 10 bootstrap
replicates and 100 simulations) of all inputs in the [Grasselli and Costa Lima, 2012]
model with CES production function on the outputs: the Good Equilibrium values
of wages share !1 and employment rate 1 . . . . . . . . . . . . . . . . . . . . . . . . 179
A.14 Part 2 / Second-order Sobol’ indices and standard deviations (10 bootstrap replicates,
100 simulations) of all inputs in the CES [Grasselli and Costa Lima, 2012] model on
the outputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180

140
APPENDIX

141
Appendix A

Building a global sensitivity analysis

to quantify the robustness of
macro-economic models

A.1 Overview of macro-economic modeling: Features and criti-

cisms of the [Grasselli and Costa Lima, 2012] model
In regard to various possible classifications, [Grandjean and Giraud, 2017] displayed a comparison
between climate and macro-economic models to explain the problems that modelers in Climate
Change Economics are facing. As written in introduction,1 they classified models between:

• New-keynesian based models (e.g.,Mésange of the French Ministry of Economy) in which

temporary disequilibrium is possible due to the inertia of quantities adjustments;

• Computable General Equilibrium models (such as Integrated Assessment Models - IAM, usu-
ally). Generally, institutions use both an estimated Dynamic Stochastic General Equilibrium
or DSGE (such as Global Integrated Monetary and Fiscal model - GIMF, New Area-Wide
Model of The Euro area - NAWM, ...) and a neo-keynesian model for their forecast;

• Disequilibrium models, and more specifically those able to represent the Minsky’s Finance
Instability Hypothesis (FIH) and predict crashes such as GEMMES (General Monetary and
Multisectorial Macrodynamics for the Ecological Shift). Their stochastic dynamic allows the
emergence in aggregated indicators, but they are very criticized due to the high-dependency on
their parameters, their myopic nature (without expectations) and exogenous growth. However
they seem the most able to represent the trajectories we want to explore, if we can address
these criticisms. An estimated version of the model with a Climate Feedback Loop exists as
a result of the work of [Bovari et al., 2018] and a first approach dedicated to the brazilian
economy was implemented by [Bastidas et al., 2017].

In their analysis, [Grandjean and Giraud, 2017] displayed di↵erent features of these models which
lead to uncertainties in regard to their ability to predict reliable trajectories. These uncertainties
1
See section 1.3.1.

142
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

are linked to their inputs definition2 , the ambiguity of their output indicators (GDP3 ; Cost of
public policy; Unemployment rate4 .), their lack of back-testing, their potentially large sensitivity
to calibrated parameters and their inability to represent out-of-equilibrium trajectories to multiple
equilibria, especially to a global collapse.

The more interesting typology of models for our study seems to be the disequilibrium models
(such as Three-Me and its extension IMACLIM-3ME [Ghersi, 2020], or [Akerlof and Stiglitz, 1969],
or [Grasselli and Costa Lima, 2012] based on a Lotka-Volterra approach between wages and employ-
ment rate5 ). These models assess by construction the trajectories to multiple equilibria and are able
to describe global out-of-equilibrium dynamics. The GEMMES model6 is an application of this type
of model about various geography [Bastidas et al., 2017] and production functions [McIsaac, 2016],
with climate feedback loop [Bovari et al., 2018], and features (Government intervention [Costa Lima
and Grasselli, 2014], Inventories [Grasselli and Nguyen Huu, 2016], Inequalities [Giraud and Gras-
selli, 2017], etc.).

A.1.1 Features and criticisms of the [Grasselli and Costa Lima, 2012] model
To address the need for macro-economic models to assess possible trajectories to a climate driven
Minsky moment, I will focus my research on a particular family of disequilibrium models. This
family, and especially the [Grasselli and Costa Lima, 2012] model, can indeed answer to some of
the criticisms of a large majority of usual models such as:

• Not taking into account energy or matter in their production function;

• Neutral or exogenous money;

• Generally no global out-of-equilibrium dynamics;

• Mainly linear in the presumed dynamics;

• Without private debt or massive unemployment;

• Without banking sector;

• Using rational expectations that constrain trajectories.

The selected model is based on the work of [Goodwin, 1967], with debt introduction [Keen, 1995],
of which the nature and stability of the equilibria were analyzed in [Grasselli and Costa Lima,
2012] and [McIsaac, 2016] for the Van Der Ploeg’s Extensions7 . Based on the Lotka-Volterra
Predator-Prey Logic, with a classical capital’s accumulation equation and two behavioral functions
(about, on the one hand, investment depending on profits, and, on the other hand, the link between
employment rate and wages share, through a short-term Phillips curve), the model takes into
2
E.g.,the Cambridge capital controversy in the late 1950s expresses the gap between definition of the nature and
role of capital goods [Stiglitz, 1974b].
3
See a review of the criticisms in [Jany-Catrice and Méda, 2015] or the proposition of [Piketty, 2009].
4
The ECB experts estimate the labour market slack around 18% in the Euro Zone, almost twice greater than
the official rate from Eurostat around 9.5%. This shows the current paradox about the calculation of this indicator:
http://bit.ly/2q2bVBz .
5
Inspired from [Goodwin, 1967] and developed by [Keen, 1995].
6
Developed at the French Development Agency under the direction of Gaël Giraud.
7
Introducing a CES production function instead of the Leontief production function [Van der Ploeg, 1985].

143
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

account debt variation (as the di↵erence between investment and profits), but with an exogenous
population and labor productivity growth rate. The major criticisms of this type of model lie in
its exogenous growth model, the use of the Say’s law (which can be relaxed by the introduction
of inventories [Grasselli and Nguyen Huu, 2016]) its myopic nature (without expectations) and its
high parameter-dependency8 .

In conjunction with the current models (used by institutions such as the French Ministry of
Economy or Environment), this work aims at providing a model which gives a broader view of the
transition challenges by o↵ering:

• A more robust, non-linear, global out-of-equlibrium dynamic;

• A Stock-Flow Consistency, with two natures of capital;

• An analysis of multiple equilibria and their stability criterion;

• Various possible extensions: climate feedback loop [Bovari et al., 2018], government interven-
tion [Costa Lima and Grasselli, 2014], inventories [Grasselli and Nguyen Huu, 2016], inequal-
ities [Giraud and Grasselli, 2017], etc., and applications (Brazil: [Bastidas et al., 2017]).

Features and underlying assumptions of the [Grasselli and Costa Lima, 2012] model
Based on the Lotka-Volterra Predator-Prey Logic9 , the original model follows the 2D (two di-
mensions) dynamical system of the form:

ẋt = xt ( (yt ) ↵) , (1)

y˙t = yt ( (xt )) , (2)

in which the predator population dynamics, xt , is increased by a function depending on the number
of preys, and decreased by a loss rate of the predators due to either natural death or emigration,
whereas preys are decreased in function of the number of predators and increased through a natural
(exponential) growth rate . These equations have a periodic solution (similar to a simple harmonic
motion if linearized).

Figure A.1: Evolution of a Predator-Prey population (interpreted from [Lotka, 1920] and [Volterra,
1928])

8
See also the study of [Pottier and Nguyen-Huu, 2017] about the dependency of the model dynamic to the selected
investment function.
9
An example of a Kolmogorov model [Hoppensteadt, 2006] which is a more general framework that can model the
dynamics of ecological systems with Predator-Prey interactions, competition, disease, and mutualism.

144
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

The general frame, built on this logic, contains a production function (originally: a Leontief
production function but [Van der Ploeg, 1985] o↵ers an extension with a Constant Elasticity of
Substitution production function) such as:
⇢
K
Y = min , aL . (3)
⌫

As in [Grasselli and Costa Lima, 2012], Y is the total yearly output, K is the stock of capital, ⌫
is a constant capital-to-output ratio, L is the number of employed workers, and a is the labour
productivity, that is to say, the number of units of output per worker per year. All quantities
are assumed to be quoted in real rather nominal terms, thereby already incorporating the e↵ects of
inflation, and are net quantities, meaning that intermediate revenues and expenditures are deducted
from the final yearly output. For the sake of simplicity in the calculation of growth rate, demography
and labor productivity are canonically given by the following exponential shaped functions10 :
t
N (t) = N0 e , (4)
↵t
a(t) = a0 e , (5)

where N0 and a0 are initial values (whatever is the initial time) that would not impact the calculation
of growth rate ȧa = ↵ or Ṅ
N = . These constant growth rates are tested in our sensitivity analysis.
For macroeconomists, their sum recalls the Good balanced growth path (which equals the sum of
the technological progress and population growth rate).

L(t)
Defining the employment rate = N (t) and the full capital utilization assumption allows us to
write:
⇢
K K
Y = min , aL = = aL = a N . (6)
⌫ ⌫

Two key assumptions are made in the model:

• On the one hand, about the link between wages w (or inflation rate) and employment rate .
This first behavioral function assumes that the rate of change in real wages w is an increasing
function of the employment rate . This relation is known as the Phillips curve :

ẇ = ( )w (7)
1 0.04 0.043
where : ( ) = 0 with 0 = , 1 = . (8)
(1 )2 1 0.042 1 0.042

This form and parameterization of the Phillips Curve guarantee that 8 t, 0  (t)  1
because (as shown in [Desai et al., 2006]) is a continuously di↵erentiable function on [0, 1]
and
˙ ( ) > 0 on [0, 1] ,
(0) < ↵ , (9)
lim ( ) = +1 .
!1
10
[McIsaac, 2016] o↵ered an extension with a population dynamic based on the recommended United Nations
sigmoid.

145
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

• On the other hand, assuming the Say’s Law, which can be summarized by“All wages are
consumed and all profits are reinvested”, total sales demand Yd is equal to the output Y (so
consumption C acts only as an adjusting variable):

C + I = Yd = Y = W + ⇧ , (10)

where ⇧ is the net profit for firms after paying wages W .

Assuming also a classical capital accumulation’s equation:

K̇ = (Y wL) K = (1 !)Y K , (11)

where is the depreciation rate of capital (assumed to be constant), and the wage share, ! :=
W/Y = wL/Y , one can deduce the following wage share (predator) and employment rate (prey)
evolution:
!˙ ẇ ȧ
= = ( ) ↵,
! w a
˙ (12)
Ẏ ȧ Ṅ 1 !
= = ↵ ,
Y a N ⌫
8
>
<!˙ = ! ( ( ) ↵)
✓ ◆
, 1 ! (13)
>
: =˙ ↵ .
⌫

This di↵erential system (13) leads to two di↵erent equilibria:

• Trivial equilibrium: (!0 , 0) = (0, 0);

• Non trivial equilibrium: (!1 , 1) = (1 ⌫(↵ + + ), 1 (↵)).

The stability of the non trivial equilibrium is studied in [Grasselli and Costa Lima, 2012] through
the Lyapunov function, H(·), associated to the system (using separating variables and integrating
the equation for d /d!):
Z
(s) ! 1
H(!, ) = ds ↵ ln( ) + ( ↵ ) ln(!) , (14)
0
s ⌫ ⌫
where Ḣ(!, ) = rH · (!, ˙
˙ )=0, (15)

which corroborates the conservative behavior of the system, observed on Figure A.2b.

With the parameterization of [Grasselli and Costa Lima, 2012], one can observe the limit circle
trajectory on Figure A.2. One obvious drawback of the model is that it does not constrain the
variables ! and to remain in the unit square, as should be the case given their economic interpre-
tation. For example, for initial values ! = 0.75 and = 0.75 , we obtain Figure A.2a, whereas for
initial values ! = 0.8 and = 0.9, we obtain the constrained Figure A.2b. In order to avoid ! > 1,
we can choose a normalized Phillips curve to constraint ! 2 [0, 1] (see appendix B.3 for the study
of the system with the normalized Phillips curve).

146
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

(a) Evolution to the trajectories in the 2D model for (b) Evolution to the trajectories in the 2D model for
initial values: !0 = 0.75, 0 = 0.75. initial values: !0 = 0.8, 0 = 0.9.

Figure A.2: Trajectories of the locally stable equilibria of the 2D-model with the set of parameters
defined in [Grasselli and Costa Lima, 2012].

Introducing a banking sector to finance new investments

Inspired from [Minsky, 1982], and formalized by [Keen, 1995], [Grasselli and Costa Lima, 2012]
modeled investment as a function of the net profit after paying wages and interests on debt ⇧ =
Y W rD (where D is the amount of debt in real term and r a constant interest rate):

I = (⇡)Y , (16)
⇧
with ⇡, the net profit share: ⇡ = Y.

Investments can now exceed profits, thus generating debt. With this debt (D amount of debt in
real terms), the net profit after paying wages and interest on debt is:

⇧ = (1 ! rd)Y = ⇡Y , (17)

where ⇡ is the net profit share, r is a constant short-term real interest rate and d = D/Y is the debt
ratio in the economy. The change in capital stock depends on the investment I and the depreciation
of this capital:

K̇ = I K = (1 ! rd)Y K . (18)

To guarantee the existence of the relevant equilibria we made the following additional assumptions
on (·), a continuously increasing di↵erentiable function such that:

̇(⇡) > 0 on [0, 1] ,

lim (⇡) = ( 1)  ⌫(max(r, ↵ + ) + )  lim (⇡)  1 , (19)
⇡! 1 ⇡!+1
2
lim ⇡ ̇(⇡) < +1 .
⇡! 1

147
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

In [Grasselli and Costa Lima, 2012], the authors take for instance11 the following form of ,
complying with the previous assumptions:

(⇡) = k0 + k1 ek2 ⇡ . (20)

The variation of D is consequently built on the di↵erence between the demand in investment and
profits:

Ḋ = I ⇧. (21)

With this definition of capital accumulation, the growth rate of the economy and the employment
rate dynamics become:

Ẏ (⇡)
= , (22)
Y ⌫
˙ (⇡)
= ↵ . (23)
⌫

Defining d = D/Y , the debt ratio of the economy, one can express its dynamics through:
✓ ◆
d˙ Ḋ Ẏ ((⇡) ⇡)Y Ẏ (⇡) ⇡ (⇡)
= = = . (24)
d D Y D Y d ⌫

The [Grasselli and Costa Lima, 2012] model now owns three state variables: the wage share !,
the employment rate and the debt ratio of the economy d. The 3D di↵erential system 12 becomes:
8
>
> !˙ = ! [ ( ) ↵] ,
>
> 
>
<˙ (⇡)
= ↵ ,
⌫ (25)
>
> 
>
> (⇡)
>
: d˙ = (⇡) ⇡ d .
⌫
The basic intuitions from this di↵erential system are the following truisms:

• Wages share ! % if wage rise exceeds growth in labor productivity;

• Employment rate % if economic growth > the sum of pop. + labor productivity growth;

• Debt-to-output ratio d % if rate of debt growth exceeds rate of growth of GDP.

11
A more exhaustive study of the di↵erent form of  is done in [Pottier and Nguyen-Huu, 2017].
12
With the following notation:

! wage share w(t)/a(t) ↵ labor productivity rate

employment rate L(t)/N (t) population growth rate
d debt ratio D/Y depreciation rate
⇡ net profit share ⇧/Y ( ) Short-term Phillips curve
(⇡) Investment function ⌫ capital-to-output ratio

148
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

The analysis of the system leads to the identification of two economically meaningful long-run
equilibria:

• A Good equilibrium, locally stable: in which final values of employment rate and wage share
are positive with finite level of debt, following a Solovian balanced growth13 . One can calculate
the equilibrium for each State Variable (!1 , 1 , d1 ):
1
1 = (↵) ,
(⇡1 ) ⇡1
d1 = ,
↵+ (26)
!1 = 1 ⇡1 rd1 ,
1
with ⇡1 =  (⌫(↵ + + )) .

• A Slavery equilibrium (with no wages but employment), structurally unstable and economi-
cally meaningless;

• A Explosive-Debt equilibrium, locally stable: in which employment rate and wage share are
collapsing with an infinite level of debt. It is the situation of economic crisis.

Graphical representations and mathematical proofs of these results are presented in [Grasselli
and Costa Lima, 2012]. Figures A.3a and A.3b show the trajectories of these two locally stable
equilibria.

(a) Evolution to the Good Equilibrium for initial val- (b) Evolution to the Bad Equilibrium for initial val-
ues: !0 = 0.8, 0 = 0.9 and d0 = 0.1 and final values: ues: !0 = 0.7, 0 = 0.7 and d0 = 0.1 and final values
(!1 , 1 , d1 ) = (0.8361, 0.9686, 0.0702) (0, 0, +1)

Figure A.3: Trajectories of the two locally stable equilibria of the [Grasselli and Costa Lima, 2012]
model with the set of parameters defined in Table 2.1

The methodology used to study the stability of each equilibria is the following:
13
The growth rate of the economy equals in the long-run to the sum of the technical progress growth and the
population growth.

149
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

• Building the Jacobian Matrix of the system:

0 1
( ) ↵ ! ˙( ) 0
B C
B C
B ̇(⇡) (⇡) ⌫(↵+ + ) r ̇(⇡) C
J =B ⌫ ⌫ ⌫ C ; (27)
B C
@ A
(d ⌫)̇(⇡)+⌫ r(d ⌫)̇(⇡) (⇡)
⌫ 0 r+ + ⌫

• For each equilibrium, computing the Jacobian Matrix (with the change of variable u = 1/d in
Explosive Debt case. For instance, in the Good Equilbrium, the Jacobian Matrix becomes:
0 1
0 K0 0
B C
B C
B
J = B K1 0 rK1 C , (28)
C
@ A
K2 0 rK2 (↵ + )

where:
K0 = ! ˙ ( ) ,
̇(⇡)
K1 = , (29)
⌫
(d ⌫)̇(⇡) + ⌫
K2 = ;
⌫
• Extracting the characteristic polynomial P (X) = det(J XId), here:

P (X) = X 3 + (↵ + rK2 )X 2 + K0 K1 X + K0 K1 (↵ + ) ; (30)

• Computing a stability criterion (in our case: Routh-Hurwitz [Routh, 1877] and [Hurwitz,
1963]):“Necessary and sufficient condition for all roots of the characteristic polynomial to
have negative real parts is”:

̇(⇡1 )
(⇡1 + ⌫(↵ + ) (⇡1 )) (↵ + ) > 0 ; (31)
⌫

• Verifying the criterion is satisfied by the parameterization14 or not.

Basins of Attraction
[McIsaac, 2016] studied the basins of attraction of these equilibria. Their dimensions depend on the
set of parameters, as they are modifying the condition of stability (Routh-Hurwitz criterion [Routh,
1877] and [Hurwitz, 1963]).15 .
14
See section 4.4.2 for the definition of our set of parameters in this work and [Grasselli and Costa Lima, 2012],
[McIsaac, 2016], [Grasselli and Nguyen Huu, 2016] and [Giraud and Grasselli, 2019] for exhaustive justifications of
these values.
15
Figure 2.2 represents the basin of attraction — i.e. all initial values of input’s variables leading to this equilibrium
according to our set of parameters — of the Good Equilibrium. The figures is displayed with the .R code of [Augier,
2018].

150
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

Many criticisms to this 3D-model can be highlighted, such as the absence of government, the
assumption of the Say’s law, which are relaxed with the method of [Costa Lima and Grasselli,
2014] and [Grasselli and Nguyen Huu, 2016], and the impossibility to study a transition between
di↵erent natures of capital (due to the sole nature of capital here). The aim of our last chapter is
to address this last point. Another usual criticism of this 3D-model is the absence of substitution
between factors. This issue was addressed through the extension [Van der Ploeg, 1985], developed
in [McIsaac, 2016].

Introducing a new production function (with a Constant Elasticity of Substitution)

Implemented by [McIsaac, 2016], following the work of [Van der Ploeg, 1985], this extension incor-
porates substitution between labor and capital into the original model by using the CES production
function:
1
⌘ ⌘
Y (t) = C(bK + (1 b)(a(t)L) ) ⌘ , (32)

where C > 0 and 0 < b < 1 are constants and the elasticity of substitution is defined by: =
1 @Y
1+⌘ . Using the hypothesis @L = w, the author finds that the optimal capital-to-output ratio and
productivity are given by:
✓ ◆ 1
K(t) 1 !(t) 1⌘
⌫(t) = = , (33)
Y (t) C b
✓ ◆1
Y (t) ↵t !(t) ⌘
a(t) = = a0 e . (34)
L(t) 1 b

As previously we obtain a slighlty modified wage share and employment rate dynamics:

ẇ = ( )w ,
w(t)
!= ,
a(t)
!˙ ẇ ȧ ⌘
= = ( ( ) ↵) ,
! w a ⌘+1
(35)

˙ L̇ Ṅ K̇ ⌫˙ ȧ Ṅ
= =( ) ,
L N K ⌫ a N
˙ 1
1+ ⌘1 1 !˙
= Cb ⌘ (1 !) ↵ .
⌘ !(1 !)

Consequently the 2D-di↵erential system becomes:

!˙ ⌘
= ( ( ) ↵) , (36)
! ⌘+1
˙ 1
1+ ⌘1 1 !˙
= Cb ⌘ (1 !) ↵ . (37)
⌘ !(1 !)

151
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

To extend the basic two dimensional model, we introduce debt (D amount of debt in real terms).
The net profit after paying wages and interest on debt, change in capital stock, is as before:

⇡ = (1 ! rd)Y ,
K̇ = (⇡)Y K , where (⇡) = k0 + k1 ek2 ⇡ (38)
K̇ (⇡)
) = .
K ⌫
Consequently, growth rates and employment rate dynamics become:

˙ ✓ ◆1
1 ! ⌘ !˙
= (⇡)C ↵ , (39)
b ⌘ !(1 !)
✓ ◆1
Ẏ 1 ! ⌘ !˙
g(!, d) = = (⇡)C . (40)
Y b ⌘(1 !)

It follows the debt ratio dynamics:

✓ ◆1
d˙ Ḋ Ẏ (⇡) ⇡ 1 ! ⌘ !˙
= = (⇡)C + + . (41)
d D Y d b ⌘(1 !)

Thus the 3D-di↵erential system becomes:

8 ⌘
>
> !˙ = ! ( ( ) ↵)
>
> ⌘+1
>
> ✓ ◆1 !
>
>
<˙ 1 ! ⌘ !˙
= (⇡)C ↵
b ⌘ !(1 !) . (42)
>
>
>
> ✓ ◆1 !
>
> 1 ! ⌘ !
˙
>
: d˙ = d r (⇡)C
> + + + (⇡) (1 !)
b ⌘(1 !)

The nature of the equilibria is the same as previously: Good, Slavery and Explosive Equilibrium.
Denoting ⇡ = 1 ! r d , the Jacobian for the di↵erential system with a price wage dynamics
becomes:
0 @f1 ⌘ ˙
1
@! ! ⌘+1 ( ) 0
B C
B C
B @f2 @f2 1 ! 1/⌘ C
J(!, , d) = B r ̇(⇡)C C , (43)
B @! @ b C
@ A
@f3 @f3
@! 0 @d

152
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

where
@f1 ⌘
= ( ( ) ↵) ,
@! ⌘+1
✓ ◆ ✓ ◆ !
@f2 (⇡)C 1 ! 1/⌘ 1 ! 1/⌘ !⌘(1
˙ 2!)
= + ̇(⇡)C + ,
@! ⌘(1 !) b b (!⌘(1 !))2
✓ ◆
@f2 1 ! 1/⌘ !˙
= (⇡)C (↵ + + ) + , (44)
@ b !⌘(1 !)
✓ ◆ ✓ ◆ !
@f3 (⇡)C 1 ! 1/⌘ 1 ! 1/⌘ !⌘(1
˙ 2!)
=d + ̇(⇡)C ̇(⇡) + 1 ,
@! ⌘(1 !) b b (!⌘(1 !))2
✓ ◆ ✓ ◆
@f3 1 ! 1/⌘ !˙ 1 ! 1/⌘
=r+ (⇡)C + + rd̇(⇡)C r̇(⇡) .
@d b ⌘ !(1 !) b

With the same methodology, we identifiy the stability criterion of each equilibria. The characteris-
tic polynomial of this Good Equilibrium fulfills already some of the conditions of the Routh-Hurwitz
criterion with the same parameterization as before. The new necessary and sufficient condition be-
comes (also satisfied with our parameters):
✓ ◆1/⌘
(⇡1 )C 1 !p1
K0 K2e ( rK3e ) + rK0 p1 (̇(⇡1 ) 1) > 0 . (45)
⌘(1 !p1 ) b

The first two eigenvalues of the Jacobian Matrix of the Explosive Debt Equilibrium are negative.
1/⌘
So this equilibrium is stable if and only if: k0 C 1b (r + ) < 0 which is satisfied with our model
parameters.

Conclusion The 3D base-model of [Grasselli and Costa Lima, 2012] is the core of our study
because of its ability to take into account private debt and represent trajectories towards multiple
equilibria with a non-linear dynamic. Two main locally stable equilibria exist: one, looking like a
Solovian balanced growth path in the long-run, with a growth rate equals to the sum of the technical
progress and population growth rates, and the other one, with an explosive level of debt, no wages
share and no employment rate.

The introduction of di↵erent features in the model is possible. They sometimes increase the
number of equilibria and impact their stability criterion. For this reason, [Costa Lima and Grasselli,
2014] tried to destabilize the Explosive Debt Equilibrium with a government intervention. Other
possible extensions were implemented such as climate feedback loop [Bovari et al., 2018], Inventories
[Grasselli and Nguyen Huu, 2016], or Inequalities [Giraud and Grasselli, 2017].

Moreover, this type of model provides information about the initial values leading to each equil-
brium, once the stability criterion are satisfied. These basins of attraction are studied by [McIsaac,
2016] and show how little is the volume leading to the Good Equilibrium. These volumes of course
depend on the set of parameters.

153
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

A.2 Some reminder on sensitivity analysis (SA)

Sensitivity analysis (SA) is designed to assess (qualitatively or quantitatively) the outputs vari-
ation due to inputs variation [Lam and Hui, 1995] , [Macdonald et al., 1999], [Helton et al.,
2006], [Saltelli et al., 2008a]. One can thus considered the SA as an ”input-output” analysis of
the simulation model [Corson, 1992], leading to a classification of the inputs as ”sensitive” or ”ro-
bust” (or ”insensitive”) to the outputs. Sensitive inputs means that changes in their values have an
important impact on the outputs values, while robust inputs have negligible impacts [Macdonald
et al., 1999]. The influence of sensitive inputs depends on the choice of value ranges: if a sensitive
input is known to within a close tolerance, the predicted outputs will not be severly disturbed
by its uncertainty (the standardized regression coefficient calculated to measure its sensitivity will
corresponding become smaller) [Hamby, 1994]

SA are most often run together with an uncertainty analysis (UA), where the SA is used to
rank the uncertainty sources (identified by UA), according to their influences on outputs [Saltelli
et al., 2008a].The di↵erence between the definition of uncertainty analysis (UA) and sensitivity
analysis (SA) is stated by [Helton et al., 2006] as that: ”Specially, uncertainty analysis refers to the
determination of the uncertainty in analysis results that derives from uncertainty in analysis inputs,
and sensitivity analysis refers to the determination of the contributions of individual uncertain inputs
to the uncertainty in analysis results.”

The following subsections aim at giving definitions, categories, techniques and applications of
uncertainty and sensitivity analysis in order to motivate the selected method in our work.

A.2.1 Uncertainties Analysis (UA): exploring the space of the input factors
The analysis of the uncertainties contains two steps: first, we estimate the probability density of
the input with statistical estimation techniques or through experts, and then quantify its propaga-
tion in the model through a wide range of possible methods.16 Each of them belongs to the following
classification, regarding the output Y of a mathematical model f (X) where X is the p-inputs vector
X = {X1 , X2 , ..., Xp } defined on the identity hypercube ⌦p = [0, 1]p .

Monte-Carlo (MC) Analysis

In litterature, one can find deterministic and probabilistic form of the Monte-Carlo analysis, which
is the most widely used method for sampling, due to its fast computation and easy understanding
[Tian, 2013]. The concept is to analyse the approximate distribution of possible outputs with
probabilistic variables [Lomas and Eppel, 1992] [Hopfe et al., 2007]. The total output uncertainty
is assessed by the total inputs uncertainties regardless of the amount of the variables and their
possible interactions. The principle is the following:

1. Generate a sampling of N-values of the p-inputs vector X = (X1 , ..., Xp ) on the basis of the
density of their joint probability;

2. Calculate the deterministic output for each set of parameter;

3. Assess the statistical properties of the output: mean, variance, density of probability.
16
See appendix A.2.1 for a more precise description of these methods.

154
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

Thus we assess the expectation and variance of the output such as:
N
1 X
E[Y ] ⇡ f (Xi )
N
i=1
(46)
N
X
1
V [Y ] ⇡ (f (Xi ) E(X))2
N 1
i=1

[Lomas and Eppel, 1992] stated that with a sufficiently high number of repetitions, the Gaussian
distribution of outputs will be exhibited, regardless of the types of variables probability distributions:
the performance of MCA can be marginally improved after 60 to 80 simulations (see Figure A.4a)

(a) Relationship between normalized confidence in- (b) Sampling coverage for simple random sampling and LHS
terval and number of Monte-Carlo simulations [Lo- [Iman, 1999], showing the number of samples on the x-axis,
mas and Eppel, 1992] and the fraction of the sampling coverage on the y-axis

Figure A.4: Performances of MC simulations [Lomas and Eppel, 1992], simple random sampling
and LHS [Iman, 1999] according to the number of simulated samples.

To diminish the number of repetitions and improve sample coverage, three sampling techniques are
widely used: simple random sampling, stratified sampling and Latin Hypercube sampling (LHS)
[Wang, 2014]:
• Simple random sampling: a basic sampling technique, working through randomly generating
samples and adjusting them to the target output, by means of the probability distributions
of samples;

• Stratified sampling: an improvement sampling method over the simple random sampling, forc-
ing samples to conform to the whole distribution being analysed. To achieve this, each sample
value is randomly selected within each stratum, where several strata have equal probability,
divided according to the probability distribution of the target output.

• Latin hypercube sampling (LHS): an evolution of the stratified sampling, dividing the input
variables into strata, and then generating samples that have variable values from di↵erent
stratum.

Stratified sampling are generally used to improve the inputs space coverage, especially when pro-
cessing is expensive due to a high number of parameters or wide ranges of uncertainty. However,

155
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

according [Macdonald et al., 1999] and [Iman, 1999], results from di↵erent sampling methods are
quite similar (at a 5% level see Figure A.4b) in increasing the robustness of MCA, after approxi-
mately 60 simulations. [Saltelli et al., 2008a] state that one can not make firm conclusions about
the superiority of LHS over random sampling method. Thus, stratified sampling are used in the
case of fewer simulations required.

Latin Hypercube Sampling (LHS)

Samples generated from MC are likely grouped in an inputs region of highest probability, leading to
a lack of information in others regions. LHS, as a stratified sampling, give a more uniform coverage
of the inputs space. This sampling technique described by [McKay et al., 2000] is one of most widely
used in litterature [Dubus et al., 2003] [McMahon et al., 2001]. The principle of the method is the
following:

1. Stratify the random variables repartition in N equiprobable intervals (with N, an apriori

setting).

2. For each input, random sampling is done on every interval, so all the possible levels of the
input are explored. N values are so obtained for each input, indiced by the corresponding
interval. Random permutations of these intervals are done to create a new sample

3. Obtained values for each input are then associated to give N n-uplets.

Thus, no samples can contain the same value for a given parameter. The sample is evenly spread
on the range of each inputs variation and so the distribution of the output will be more precise, as
shown in [McKay et al., 2000].

Optimization process can be add to the LHS to deliver even better coverage, they are called
Optimal Latin Hypercube Sampling (OLHS). The construction of an LHS is based on a first step
of randomization followed by a stochastic optimization process which aims at spreading points as
evenly as possible within the design space (according to a optimality criteria such as the maximin
distance criteria). One can notice that this type of matrix is not reproducible (unless using the
same random seed). [McKay et al., 2000] compared LHS with MC sampling and conclude
that LHS gives better statistical results on the outputs.

Bayesian approach
Bayesian methods are now widely used for the estimation and evaluation of dynamic stochastic
general equilibrium (DSGE) models [Del Negro and Schofheide, 2008]. This approach has been
discussed by many authors in the literature in the last few years (e.g. [Schorfheide, 2000], [Lubik
and Schorfheide, 2007]). A comprehensive survey is provided in [An and Schorfheide, 2007]. Despite
their importance, the literature has paid little attention to the systematic elicitation of priors. Prior
distributions either reflect subjective opinions or summarize information derived from data not
included in the estimation sample (recent exceptions in medical studies [McCandless and Gustafson,
2017], computing science [Le Gratiet et al., 2018] and macroeconomics [Jacobi et al., 2019]).

Due to the formulation of priors and the need to linearize first the model, the bayesian approaches
can not be good candidates for our study, as we desire especially to study the robustness of non-linear
models without linearly approximating it. Nevertheless, Bayesian approaches have multiple success
and applications such as in the DYNARE Software of the European Commission (a DSGE and

156
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

overlapping generations model controlled by a sensitivity analysis [Adjemian et al., 2011]) or [Smets
and Wouters, 2003], [Oakley and O’Hagan, 2004]

Conclusion Conducting our uncertainty analysis on a very non-linear model, without linear ap-
proximations, leads to consider a sampling method with stratification, such as Latin Hypercube
Sampling, because of its features [McKay et al., 2000] and its ability to be optimized.

A.2.2 Sensitivity Analysis (SA): from uncertainties in the input to uncertainty

in the output
A possible definition of sensitivity analysis (usually conducted in parallel of an UA) is the fol-
lowing: The study of how uncertainty in the output of a model (numerical or otherwise) can be
apportioned to di↵erent sources of uncertainty in the model input [Saltelli et al., 2008b]. A wide
range of techniques are available to conduct sensitivity analysis. This part will present a brief review
of them. We develop some of these techniques in next parts.

According to [Saltelli et al., 2008a] SA can be classified into qualitative SA (such as Screening
Methods) or Quantitative SA which can be carried for a single input parameter or multiple param-
eters. Quantitative SA are commonly grouped into local and global methods [Lomas and Eppel,
1992], which distinctions are listed by [Hoes and De Vann, 2005] in Table A.1.

The principle of SA is to regard the mathematics of simulation as a ”black box”. We only

alter the model descriptions such as the initial conditions, the boundary constraints and solution
methods and run repeatedly the simulations of each set of inputs as a separate model [Wang, 2014].
Statistical techniques are adopted to analyse the relationship between the perturbations (di↵erent
sets of input variables) and outputs and so quantify the uncertainty [MacDonald, 2002].

Local SA assesses output uncertainty by increasing a small amount of each input value. For
global SA, the input space is sampled and the impacts of all inputs on the output uncertainty are
assessed. Most common methods in SA are Screening methods and Finite-di↵erence approximation,
Monte-Carlo Analysis, ANalysis Of VAriance (ANOVA).

Even if the methods of local SA are considered to be computational faster and easier. Less
accurate is their inherent problem, compared to the sophisticated global SA methods. According to
literatures, it also has been stated that the global SA should be used, when input variables are in
a non-linear model and from di↵erent magnitudes of uncertainty sources [Cukier et al., 1973]. The
main methods of SA are further compared in Table A.2.

A.2.3 Screening Methods

Screening methods are presented, inter alia, by [Campolongo, 2005], they are meant to classify
inputs of the model by their impact on outputs. They explores the behavior of the system in presence
of a high number of inputs by using discrete sampling techniques on each input. The discretization
of inputs in di↵erent values (called levels) make them so-called deterministic methods, because they
do not use probability on inputs. This is the reason why they are usually easy to implement and
faster in terms of computation time than other SA methods. These methods are usually classified
as qualitative and would not be used in our quantitative study. However the principle of screening

157
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

Items Local SA Global SA

It is meant for the determination of It is meant for the determination
Aim partial derivation of the output in of output uncertainty in relation
relation to a specific input variable. to the overall input variables.
Input variables are sampled
Input variables are sampled
Input variables one-at-once (remaining variables stay
simultaneously.
at their ’base-case’ values).
Relationship
A linear relationship is not
between input A linear relationship is assumed
necessarily assumed between
variables and between input variables and outputs.
input variables and outputs.
outputs
Each input variable can be
There is only one assigned distribution
Distribution of assigned to di↵erent
to input variables (typically a normal
input variables variation/distribution (typically a
distribution).
normal distribution).
On average, a large number of
Numbers of The number of simulations in
simulations are required (but the exact
simulations local SA is less than in global SA.
numbers depend on the LSA method).

Table A.1: Comparison of local and global SA methods [Hoes and De Vann, 2005] and [Hopfe et al.,
2007].

Method Name I II III IV

Monte-Carlo Analysis (MCA) Yes Yes Yes No
Variant based methods (e.g.
Yes Yes Yes Yes
ANOVA)
Screening methods No/Yes Yes Yes Yes
Local methods (e.g.
No No No Yes
Finite-di↵erence approximation)
The influences of input variables (on outputs) will
incorporate the e↵ects of the range of input variation and
I: Input variables
the probability density function (particularly, the normal
or uniform distribution).
Di↵erent to partial derivatives, the variation of one input
II: Variation of input variables
variable will cause the changes of others.
The performance of sensitivity techniques will not be
III: Model independence a↵ected by model nonlinearity, if input variables are
independent.
The influence of individual variable should be evaluated,
IV: Treat group variables as
even variables are varied as a group. It improves the
they were individuals
agility of results interpretation.

Table A.2: Comparison of four di↵erent methods to conduct global sensitivity analysis (European
commission, 2005) [Hopfe et al., 2007].

158
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

through a Design of Experiment (DoE) with a Optimal Latin Hypercube Sampling inspired the first
step of our analysis.

Simple or High dimensional Screenings’ methods

This type of techniques comes from the field of Design of Experiment (DoE) and are easy to
implement as they do not need binding assumptions about the nature of the model. Among these
assumptions, independence of the inputs, monotony of the output variation in relation with each
input and small number of inputs impacting the system (compared to the total number of inputs)
are the most famous in litterature [Iooss, 2011] leading to the following methods:

• Fractional Factorial Screening: in the field of DoE, we commonly have a greater number of
observations than the number of inputs to study. This leads to the probabilistic estimation of
the coefficients representing the influence of inputs on output. These influence are categorized
in: main e↵ects (or first-order e↵ects), joint e↵ect e↵ects (second-order e↵ects), etc., which
can be analyzed with a Full Factorial DoE (where every possible combination of levels of each
inputs is tested). However, in a lot of applications, a sufficient number of observations can be
hard to provide or might be biased, due to a lack of information. In this context, Fractional
Factorial screening (as a part of the so-called oversatured designs) allows to determine the
main e↵ects with a smaller number of simulations. The technique is detailed by [Burman and
Plackett, 1946], improved by [Box and Behnken, 1960] and Taguchi methods [Karna et al.,
2012].

• Group Screening: More adapted to numerical experiments, it consists in the division of

inputs into group to identify the more influent ones. This is meant to multiply the group
composition by discarding the less influent until the most influent one in obtained. The
principle is the following:

1. Independent inputs are grouped on the apriori intuition of the researcher. Each group-
input is then studied as a input with two levels:
– Group-input is at the + level when every input in the group are at their + level;
– Group-input is at the - level when every input in the group are at their - level;
2. Once the most influent group are identified, non influent are discarded of the second step
where each group is either divided into smaller groups with less inputs, or each input in
the most influent group is one-at-a-time tested.

This method need however a prior knowledge of the change direction of the output according
to the change direction of each input, which is not the case in very non-linear macro-economic
models.

• Sequential Bifurcation for Factor-Screening: A two-groups screening without DoE,

which uses regression analysis or variance-based analysis. This approach is close to dichotomic
search (or binary search), where experiments are sequentially chosen by using previous results
to guide the research of most influent inputs. As for group screening, the computation cost
of this function depends on the number of influent inputs, the classification strategy ad the
ability to gather most influent inputs in the same group.

159
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

Morris’ Method
Campolongo’s Screening methods have been developped and improved by Morris, relaying on the
idea that a model with a high number of inputs is hard to explore, but a small number of inputs
are often very influent. A qualitative approach of the Morris’ method consist in identifying them
first, according to their relative e↵ets:

• negligible e↵ects;

• linear e↵ects without joint e↵ect;

• non-linear e↵ects with joint e↵ect.

With a finite-element approximation, Morris improved the common derivative methods, and can
explore bigger variation of the inputs. This lead to what he called ”elementary e↵ect”, which the
total impact on the output due to the variation of the inputs. By repeating the study, he computes
the statistical properties of this elementary e↵ect.

The calculation of Morris is thus a mean and variance of derivative calculations for di↵erent
combinations of the inputs rather than a fine study of a particular ”nominal point” (combination of
the input) conducted by the common derivative methods. This technique is explored in the global
sensitivity analysis subsection, as a Quantitative SA.

A.2.4 Local Sensitivity Analysis (LSA)

LSA focuses on the local impact of input variables on model outputs. It is usually carried out
by assuming the linear or monotonous model, or computing partial derivatives of the outputs, with
respect to a small interval fractional variation of an input variable around its normal value [Saltelli
et al., 2008a]. The interval could be the same for all of input variables, but it is not related to the
knowledge degree of input variables. Since the model we want to study contains nonlinearities, we
should avoid Local Sensitivity Analysis, but we review some of these methods here as they pave the
way for global sensitivity analysis.

For now, assume that the parameters are all statistically independent and that any combination
of parameter values Xr with r = 1...p is equally likely. The p sampled parameter values define
a point in p-dimensional space, and sample points are evenly spread throughout a p-dimensional
hypercube of side 1.

Linear model
As a first approximation, the dependency can be viewed as linear in each parameter. So we can
write the model as:
Xp
Y = b0 + bi X i (47)
i=1

As the p-inputs are assumed to be independent, the variance is calculated by:

p
X
V (Y ) = b2i V (Xi ) (48)
i=1

160
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

where b2i V (Xi ) is the share of the variance due to the Xi input. So we can build an indice of the
variance of the output due to an input through the Standard Regression Coefficient (SRC) indice
which is always positive (SRCi 2 [0, 1]):

b2i V (Xi )
SRCi = (49)
V (Y )

We can notice that this indice equals the square of the linear joint e↵ect coefficient ⇢Xi ,Y . It is
indeed easy to verify that:

Cov(Xi , Y ) = bi V (Xi )
s
Cov(Xi , Y ) V (Xi ) (50)
) ⇢Xi ,Y = p = bi
V (Xi )V (Y ) V (Y )

However, it is sometimes hard to assess the SRCi coefficient when the repetition of the simulation
are not done on the same set of inputs values. joint e↵ect (second-order e↵ect) between inputs can
appears and a↵ect the share of the variance due to a particular input [Saporta, 1990]. This is the
reason why we use the Partial joint e↵ect Coefficient (PCC) indice:

Cov(Xi , Y |X⇠i )
P CCi = ⇢Xi ,Y |X⇠i = p (51)
V (Xi |X⇠i )V (Y |X⇠i )

Where X⇠i is the X vector without its i-th component. These indices are explored in [Saporta,
1990]. If their results are not equal, the classification they imply is the same (by taking P CC in
absolute terms as it can be negative).

One-at-a-time (OAT) sampling

When solving the matrix problem representing the linear and/or monotonous model Y = b0 +
p
P
bi Xi with N simulations
i=1
2 3 2 32 3
y1 1 x11 ... x1p b0
6y2 7 6 1 x21 ... x2p 7 6 b1 7
6 7=6 76 7 (52)
4 ... 5 4... ... ... ... 5 4...5
yp 1 xN1 ... xN
p bp

one can carefully select the data points in order to limit numerical problem. One way of simplifying
the matrix Xij is to use an One-at-a-time (OAT) design where only one input is changed at each
simulation. As an example with Xi 2 {0, 1}, assume the following first set of simulations:
2 3
2 3 1 0 0 ... 0 2 3
y1 6 7 b0
6y2 7 6 1 1 0 ... 0 7 6b1 7
6 7 = 6 1 1 1 ... 0 7 6 7 (53)
4 ... 5 6 74 5
4... ... ... ... ...5 ...
yp bp
1 1 1 ... 0

161
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

This system can thus easily simplified by operating the rows into:
2 3
2 3 1 0 0 ... 0 2 3
y1 60 1 0 b0
6 y2 y1 7 6 ... 0 77 6 b1 7
6 7 = 60 0 1 ... 0 7 6 7 (54)
4 ... 5 6 7 4...5
4... ... ... ... ...5
yp ykp1 bp
0 0 0 ... 1

[Saltelli et al., 2008a] explains by this equation that if there is any change in value between yi and
yi+1 , it can only be attributed to a change in parameter Xi (complicated by random e↵ects if the
model is stochastic). So the quantity yi = yi+1 yi is an estimate of the e↵ect on y by changing
Xi from 0 to 1. It is applicable everywhere if the linear model is appropriate, and for some region
around the current sample point otherwise.17

However OAT sampling is inefficient when the number of parameters p is large and only a few of
them are influential [Saltelli et al., 2008a] (about chemical kinetics, see the comparison of [Zádor
and Turány, 2006]). As the macro-economic models we are studying is very non-linear and as we
can not apriori define the number of influential input, we do not use this family of methods.

Extensions with di↵erent types of sampling, such as Monte-Carlo sampling, Fractional Factorial
DoE, Stratified or Latin Hypercube Sampling exists and have been described above, and are appli-
cable in the frame of linear and/or monotonous model. Generalization of the method exists known
as the Finite di↵erences analysis. However, to capture joint e↵ect with other inputs or nonlineari-
ties (which can hardly be identified with the OAT sensitivity analysis) a combination with scenario
analysis, discussed next, might be necessary.

Scenario analysis
In this very common LSA in economics, several inputs values are simultaneously changed to
figure changes in economy, allowing to capture some joint e↵ect between inputs. This intuitively
appealing method remains hard to apply in order to isolate the e↵ect of each joint e↵ect.18 As
explained in [Harenberg et al., 2019], one solution consists in combining scenario analysis with
17
This is the work developed with success by [Erb and Michaels, 1998] about biological models with closed time
interval and small distance to a ”nominal point”. In our example, we only take two possible extreme value of the
domain of each input. More complicated paths through the sample space would involve much smaller changes to
the parameter values, which would allow the estimation of elementary e↵ects, as previously explained in the Morris’
method.
18
About the canonical RBC model, evoked in section 2.4.2 [Harenberg et al., 2019], the authors explain that
scenarii can not capture a particularly plausible economic environment, but rather to exemplify the approach. One
drawback that becomes apparent is the high level of discretion typically involved in choosing scenarii and corresponding
parameter values, which is due to the local nature of this sensitivity measure. As an example, the scenarii can be:
• ”Baseline” Scenario: where all inputs are kept at their baseline values.
• ”High risk and risk aversion” Scenario: where , and ⇢, the standard deviation and autojoint e↵ect of the total
factor productivity (TFP),are at their upper bounds and ⌧ , the intertemporal elasticity of substitution (IES),
is at its lower bound, so that risk aversion 1 ⌧ is high. All other inputs are kept their baseline values.
• ”High capital utilization and frictions” Scenario: where ↵, the capital share, , the depreciation rate and ,
capital adjustment cost, are at their upper bounds, and all other inputs are kept at their baseline values.
Displaying the observed output from these di↵erent scenarii allows us to compare the scenarii and evaluate the impact
of joint parameter changes on the outputs. However, scenario analysis can not specify which parameter or which
interaction between parameters is important in each case.

162
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

OAT finite di↵erences to tease out individual parameter e↵ects and interactions, which is known
as scenario decomposition or generalized Tornado diagrams carried out with success in [Borgonovo
et al., 2011] and [Borgonovo and Plischke, 2016]. However, such Tornado diagrams are rarely
encountered in economic studies and are outside the scope of this study, since they also su↵er from
the fact that they are local and linear.19

Partial derivatives
Computing partial derivatives of the outputs, with respect to a small interval fractional variation
of an input variable around its normal value, can be used as a sensitivity measure. It delivers
information on the influence of an input in the local neighborhood of a ”nominal point”.

Among the advantages of the method over common screening methods are that it covers the
entire input space and that it is not restricted to linear influences. Partial di↵erences at di↵erent
points in the input space, so-called elementary e↵ects by [Morris, 1991], are computed:

|Y (x1 , ..., xi + p , ..., xk ) Y (X)|

di (X) = (55)
p

where {x1 , ..., xk } is sampled on a grid of p levels, xi is the value of the i-th component of X and p is
a predetermined multiple of p 1 1 . The distribution of the di provides information on the behavior of
Xi , and we can study its statistical properties: overall influence of the input is characterized by the
mean of di (X), so called Morris Importance Measure, standard deviation characterizes the linearity
of the Xi influence on the output Y . One can use various sampling techniques (Random Values,
Stratified Sampling, Monte-Carlo Sampling, LHS,...) to elaborate designs in order to calculate this
distribution of di .

Eventually, computing the Jacobian matrix of a system at various particular points should gives
a global sensitivity measure. This is the idea of Derivative-based Global Sentivity Measure (DSGM)
as developed by [Sobol’ and Kucherenko, 2009] and improved by [Lamboni et al., 2013].

A.2.5 Global Sensitivity Analysis (GSA)

To allievate the lacks of Local Sensitivity Analysis, when studying very non-linear models, or
exploring large inputs spaces, one can conduct a Global Sensitivity Analysis. A wide range of GSA
methods are existing. In this part, we motivate the selected method of our study, that is a GSA
with Sobol’ indices, by reviewing some of the most famous methods, such that:

• Derivative-based Global Sensitivity Measures (DGSM)20 [Lamboni et al., 2013]

19
However some of these techniques are still successfully used in linear macro-economic models or by approximating
nonlinearities with small perturbations spaces around the equilibrium, such as e.g. [Iskrev, 2010], [Gourieroux et al.,
2000] or the DSGE and overlapping generations models from the European Commission [Adjemian et al., 2011]. These
techniques often used partial derivatives to conduct their sensitivity analysis.
20
Defined in [Saltelli et al., 2008b] as:
“derivatives are only informative at the base point where they are computed and do not provide for an exploration of
the rest of the space of the input factors. This would matter relatively little for linear systems, in which the property
at a point away from the baseline can be computed quickly by linear extrapolation using first-order point derivatives,
but it would matter greatly for nonlinear ones”, as can be macro-economic models.

163
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

• Variance-based sensitivity indices (Sobol’ indices) for independent inputs and its extension for
dependent inputs (Shapley e↵ects)

• Frequency-based GSA: Fourier Amplitude Sensitivity Test (FAST) [Cukier et al., 1973]

• Other GSA in macroeconomics: Quantitative Meta-model (based on Chaos polynomial) [Haren-

berg et al., 2019] , [Fajraoui, 2014]

Derivative-based Global Sensitivity Measures The original [Morris, 1991] method aims at
measuring local sensitivity through ”elementary e↵ect” with a finite element approximation. A
generalization of this method to global measurement was given in [Sobol’ and Kucherenko, 2009]
in order to capture nonlinearities between output and inputs. This technique knew some successful
studies e.g. in finance about Option pricing [Campolongo et al., 2006] or Asset Allocation [Man-
ganelli, 2004].

Before [Sobol’ and Kucherenko, 2009], this Morris’ method with finite elements approximation
was assumed to be reliable to assess the total indices of the inputs influence, with comparable results
to the total Sobol’ indices STi (see next paragraph). But their numerical experiments shows that
SA based on vi can not allow to classify inputs by order of influence.21 This is why this method are
usually classified in screening methods. Even if they are less time-costly than total Sobol’ indices
calculation methods they can not be appropriate to our study.

Frequency-based GSA: Fourier Amplitude Sensitivity Test (FAST) The FAST method
was first developed by [Cukier et al., 1973], [Cukier et al., 1978] and [Schaibly and Shuler, 1973]. It
approximates the pluri-dimensionnal decomposition with mono-dimensionnal decompositions along
a curve on the domain [0, 1]p . This curve is defined by parametric equations:

xi (s) = gi (sin(!i s)) for i = 1, ..., p (56)

where gi are the function to determine in order to guarantee an uniform coverage of the identity hy-
percube (see example below), (!1 , ..., !p ) 2 Np are integer frequency linearly independent (no linear
combinations between them). Thus, when s is varying in R, the vector (x1 (s), ..., xp (s)) is varying
on the curve in [0, 1]p . [Saltelli and Bolado, 1998] showed, with a high number of experiments that
the results about FAST sensitivity indices are equivalent to those obtained with Sobol’ methods.22
21
The demonstration of this result is available in [Lamboni et al., 2013]. Indeed, the total-impact indices of this
R ⇣ @f (x) ⌘2
derivative-based method, vi , approximates vi = ⌦n @xi
dx and the authors showed that the total Sobol’ indices,
STi  4Ci vVi with Ci = 4⇡1 2 . Consequently, if Xi is not influent, its total-impact indices, vi ⇡ 0, showing that the
Morris’ method is useful when assessing what inputs are not influent.
22
R 1
RT
[Cukier et al., 1973] show that f0 = [0,1]p f (x)dx = lim 2⇡ T
f (x(s))ds. With integer frequency, the
T! 1
1
R⇡
curves are 2⇡-periodic: f0 = 2⇡ 2⇡ f (x(s))ds. These calculations are applied to compute the variance of the model
Y = f (X1 , ..., Xp ). [Saltelli et al., 1999] introduced an Extended FAST method, to compute total sensitivity indices
by assessing the share of the variance due to all the inputs but Xi as the sum of the square of all Fourier coefficients
linked to frequencies di↵erent from !i and its harmonic.

164
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

However the estimation of the indices involves to define gi , !i 23 and M , an upper bound to
assess the coefficient, which must also be chosen, as infinite sums are not computable.24 M is so a
compromise between the quality of the indices and their computation cost.25

One of the advantages of this method is that indices can be calculated independently from each
others, with the same sample, whereas Sobol’ method involves at least two samples. However,
the Sobol’ method is stochastic, it allows to display confident interval about the estimation of
the indices, which is impossible with FAST, as the estimation is deterministic once the set of
frequency is chosen. [Saltelli and Bolado, 1998] compared FAST and Sobol’ methods on a various
number of models and concluded that FAST was computationally less complex and less time-costly.
Nevertheless, FAST can be biased due to the choice of the frequency, whereas Sobol’ always converge
to the actual value of the sensitivity indices.

In our study, the complexity of the model due to the number of inputs and nonlinearities is not
too important to motivate a FAST. To avoid biases due to the choice of frequency we will so not
use a FAST to analyse the sensitivity of our model.

Variance-based GSA: ANalysis Of the VAriance (ANOVA) with Sobol’ indices As

recommended in [Saltelli et al., 2008b], the relative influence of the inputs variance on output
variance allows better results about the global sensitivity measure of a very non-linear system. This
is in order to exploit the statistical properties of variance [...] to investigate how factors contribute
to the variance. Interesting features of variance-based methods are:

• Model independence: the sensitivity measure is model-free;

• Capacity to capture the influence of the full range of variation of each input factor;

• Appreciation of interaction e↵ects among input factors;

23
The definition of the function gi have been proposed with di↵erent techniques. One of the most known remains
the one of [Saltelli et al., 1999], because it gives the best space coverage with a uniform distribution of the samples:
1 1
xi (s) = gi (sin(!i (s))) = + arcsin(sin(!i (s))) (57)
2 ⇡
For the frequency, the method consists in selecting a high frequency for the observed input Xi and smaller frequency
for the others inputs. But this kind of method tend to overestimate the influence of the inputs on the output
variance [Saltelli et al., 1999]. The algorithm proposed by the authors is:
1. Choosing !i frequency according to the impact on the minimum number of simulations N ;
!i
2. Determining the maximum value of the others frequency values: max(!⇠i ) = 2M
;
3. All the other frequencies are chosen between 1 and max(!⇠i ) with two conditions:
(a) Interval between each frequency should be maximal;
(b) Number of variables Xi with the same frequency should be minimal.
[Saltelli et al., 1999] determine thus the minimal sampling dimension pour the assessment of these indices:

N = 2M max(!⇠i ) + 1 (58)

.
24
M acts as a maximum harmonic in the calculation, and is chosen according to: The bigger, the more reliable will
be the indices and the smaller, the faster simulations will be.
25
[Cukier et al., 1975] estimates empirically that M should be in 4 or 6, whatever is the dimension of the model.

165
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

• Capacity to tackle groups of input factors: uncertain factors might pertain to di↵erent logical
types, and it might be desirable to decompose the uncertainty according to these types.

The main drawbacks of variance-based measures are their computational cost, which is not a big
issue for our model (less than ten inputs) and the high number of possible methods, leading to
di↵erent rankings of factors according to their objectives.

The application principle of the ANOVA with Sobol’ indices is described in section 2.2.3. The
estimation of these indices remains challenging (especially for second-order Sobol’ e↵ect, represent-
ing joint e↵ect between output26 ) and is developed in the next subsection A.2.6. Various successful
studies were conducted about DSGE models (e.g. [Ratto, 2008]).

Suppose that we can express a square integrable random variable Y as Y = f (X) where X =
(X1 , ..., Xp ) is a vector of p independent random variables uniformly distributed on [0, 1]. From the
so-called [M. Sobol, 1990] decomposition, it is possible to expand f 2 L2 ([0, 1]p , dx) into orthogonal
summands of increasing dimensions

p
X p X
X p
f (x1 , ..., xp ) = f0 + fi (xi ) + fi,j (xi , xj ) + ... + f1,...,p (x1 , ..., xp ) (59)
i=1 i=1 j>i

and to obtain the classical ANOVA decomposition

p
X p X
X p
V (Y ) = Vi + Vi,j + ... + V1,...,p
i=1 i=1 j>i

m (60)
p
X p
X X Vi,j p
Vi V1,...,p
1= + + ... +
V (Y ) V (Y ) V (Y )
i=1 i=1 j>i

where

Vi = V (E(Y |Xi ))
Vi,j = V (fi,j (Xi , Xj )) = V (E(Y |Xi , Xj )) V (E(Y |Xi )) V (E(Y |Xj )). (61)
..
.
26
The issue of GSA with dependent variables has been the object of intense recent research but is not considered in
our study as we make the classical assumptions of apriori independent inputs at a first step of the study. A prospect of
this work can be to check the independence of these inputs. We review here the di↵erent improvements made to take
into account dependent inputs. About Sobol’ indices, [Xu and Gertner, 2008] exhibits a Sobol’ indices decomposition
into dependent/correlated and independent/uncorrelated components for linear models.With this decomposition, [Li
et al., 2010] build the Sobol’ indices for a general model. An other technique is proposed by [Mara and Tarantola,
2011], using the Gram-Schmidt process to decorrelate the inputs variables. They also propose to define new indices
through the Sobol’ indices of the decorrelated problem. [Chastaing et al., 2013] provide a theoretical framework to
generalize the ANOVA decomposition to problems with dependent variables, based on the work of [Hooker, 2007]. In
contrast to the other works which focus on generalizing the ANOVA decomposition, [Kucherenko et al., 2012] develop
the Sobol’ indices via the law of total variance. Recent work of [Tarantola and Mara, 2017] considered estimating
Sobol’ indices with dependent variables using the Fourier Amplitude Sensitivity Test. In our study, inputs of the
model are the exogenous parameters of [Grasselli and Costa Lima, 2012], assumed to be independent. Thus we can
only consider the core method of the Sobol’ indices to assess its sensitivity to its inputs. But these extensions can be
useful when using e.g. a Taylor’s rule to endogeneize the interest rate r.

166
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

In this way, we can define the Sobol indices: 8s 2 {1, ..., p}, 8i1 < i2 < ... < is 2 {1, ..., p}s

Vi1 ,...,is
Si1 ,...,is = . (62)
V (Y )

The general principle of the estimation of these indices is usually based on a Monte-Carlo anal-
ysis27 with independent inputs (also called pick-freeze method) with the scheme of [Sobol, 1993]
and [Saltelli, 2002]28 or using replicated array-based Latin hypercube sample [McKay et al., 2000].29

Other GSA in macroeconomics: Beside variance-based analysis (or frequency-based analyis)

one can use emulators build on orthogonal polynomials (Hermite, Legendre, ...). Among them,
Chaos Polynomial Expansion (CPE) are the most famous in the litterature due to their efficience
in convergence (see [Harenberg et al., 2019] for a recent succes about canonical real business cycles
models). This development is based on the theory of [Wiener, 1938] and the works of [Cameron and
Martin, 1947], with a recent interest in numerical stochastic process.30 The principle is to build a
meta-model through CPE, which has the features to be reversible, and the disadvantages to need
truncatures (in a same way as FAST) involving an apriori arbitrary choice about the degree of the
polynomial.

The technique involves a lot of similarity with the ANOVA decomposition since the CPE are
orthogonal for a Gaussian process. The estimation of the coefficient is still challenging. There are
two group of method:

• Intrusive methods [Le Maı̂tre et al., 2002] [Matthies and Keese, 2003]: they consists in intro-
ducing calculation of the polynomial coefficient in the model. They allow thus to estimate
these coefficients with a single run of the simulation. But the method is vey complex to
implement depending on the degree of nonlinearities of the model [Berveiller, 2005];

• Non-intrusive methods [Li and Zhang, 2007], [Blatman and Sudret, 2010]: they consider the
model as a ”black box”, and estimate the coefficient through numerical simulation in dedi-
cating points of the inputs spaces (such as a Latin Hypercube Sampling), but their reliability
is highly dependent on the number of simulations [Fajraoui, 2014], and of the choice of the
DoE [Sudret, 2008].
27
Others methods consists in estimating Sobol’ first order indices with with B-spline Smoothing [Ratto and Pagano,
2010] or with Saltelli’s so-called ”extended-FAST” method [Saltelli et al., 1999].
28
The seminal method of [Sobol, 1993] to compute the indices given by the variance decomposition up to a specified
order, improved with the Saltelli’s scheme [Saltelli, 2002] to compute first order, second order and total indices has
many variations: inspired from the work of Jansen-Sobol’s [Jansen, 1999], Mauntz-Kucherenko’s scheme computes
first order and total indices using improved formulas for small indices [Tarantola et al., 2007] ; the Martinez’s scheme
uses joint e↵ect coefficient-based formulas associated with theoretical confidence [Martinez, 2011] ; Janon-Monod uses
optimal asymptotic variance [Monod et al., 2006] [Janon et al., 2014] ; or via unique matrix permutations [Alex Mara
and Rakoto Joseph, 2008] ; via 3 input independent matrices [Owen, 2013] for first-order and total-e↵ect indices using
Liu-Owen’s scheme [Liu and Owen, 2006] and pick-freeze scheme [Fruth et al., 2014].
29
Other sampling methods can be used for the estimation of the Sobol’ first order and closed second order in-
dices with, for example, replicated orthogonal array-based Latin hypercube sample [Tissot and Prieur, 2015], or
under inequality constraints by extension of the replication procedure or with kriging-based global sensitivity analy-
sis [Le Gratiet and Garnier, 2014] - available in the [Pujol et al., 2015] package.
30
See e.g. [Ghanem and Spanos, 1991]

167
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

Conclusion: Since the model we study contains nonlinearities and we want to assess the relative
e↵ects of its parameters on the output, we should avoid Local Sensitivity Analysis. We can conduct
a screening method at a first step but we overall must conduct a Global Sensitivity Analysis. The
first screening method can be based on Latin Hypercube Sampling, according to its features [McKay
et al., 2000].

The Global Sensitivity Analysis (GSA) can either be conducted with variance-based, frequency-
based or meta-models methods. Their results are quite equivalent [Saltelli and Bolado, 1998], but
the last two need the apriori choice of an upper bound (in the degree of the polynomial or the
harmonic), which can introduce biases in the study to assess actual values of the sensitivity indices.
Consequently we use a GSA with the Sobol’ indices methods.

As the inputs are assume to be independent, we do not need the introduction of the possible
extension, and we only have to study the di↵erent method to estimate the Sobol’ indices. These
methods are commonly based on Monte-Carlo or quasi-Monte-Carlo sampling. But [McKay et al.,
2000] introduces an improvement of the space coverage with the help of Latin Hypercube Sampling.
This is the object of the next subsection.

A.2.6 Estimation of the indices

Sobol’ indices can sometimes be formally calculated when the analytic form of the model is
known and simple. But, in the case of our study, the model is represented as a black box because
this function is either unknown or too complex. We have to approximate these indices using fast
numerical tools. Let us remark that the Sobol’ indices may be expressed using ratio of variances
of various conditional expectations. Thus, the use of basic Monte Carlo methods to approximate
them seems to be prohibitive in this case due to a quadratic time complexity.

In this subsection, we review an approach adapted to our choice of analysis: the Sobol’ estimation
method, improved by [Saltelli, 2002] and [McKay et al., 2000] leading to a linear time complexity.
Some success-studies with this estimation methods have been conducted by [Canova, 1995] about
DSGE, in finance with [Kucherenko, 2010] about GSA with Monte-Carlo pricing and [Gilquin et al.,
2016] about a Land use and Transport Integrated Model.

We generate31 two independent input sample matrix of size N for X = (X1 , ..., Xp ):

0 (1) (1) (1)

1 0 (2) (2) (2)
1
x11 x12 ... x1p x11 x12 ... x1p
B . . . . C B . . . . C
B C B C
B C B C
X(1) =B . . . . C X(2) =B . . . . C (63)
B C B C
@ . . . . A @ . . . . A
(1) (1) (1) (2) (2) (2)
xN 1 xN 2 ... xN p xN 1 xN 2 ... xN p
31
Even if classical random number generators can be used, Quasi Monte-Carlo sampling can be more efficient
according to [Sobol, 1998] , [Bratley et al., 1992] and [Joe and Kuo, 2008], because it o↵ers a better speed of
convergence. What is more, [McKay et al., 2000] develops an improved method based on Latin Hypercube Sampling
instead of Monte-Carlo sampling in order to generate sets of inputs to test with a better space coverage for a smaller
number of simulations needed. The latter is implemented in this work.

168
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

For j 2 {1, ..., p}, the first order Sobol’ indice Sj measuring the single impact of Xj on the total
V V (E(Y |Xj ))
variance is given by Sj = V (Yj ) = V (Y ) where

Vj = E( E(Y | Xj )2 ) E(Y )2 . (64)

| {z }
Uj

Classicaly, E[Y ] and V (Y ) are respectively approximated using

1 X⇣ ⌘
N
(1) (1) (1) (2) (2) (2)
Y = f (xi1 , ..., xij , ..., xip ) + f (xi1 , ..., xij , ..., xip )
2N
i=1

and
1 X ⇣ 2 (1) ⌘
N
(1) (1) (2) (2) (2) 2
V̂ = f (xi1 , ..., xij , ..., xip ) + f 2 (xi1 , ..., xij , ..., xip ) Y
2N
i=1

while the term Uj is estimated, using [Saltelli, 2002] fixing method (see also [Jacques, 2011]), by
N
1 X (1) (1) (1) (2) (2) (1) (2) (2)
Ûj = f (xi1 , ..., xij , ..., xip ) ⇥ f (xi1 , ..., xi(j 1) , xij , xi(j+1) , ..., xip ).
N
i=1

Thus, only using N (p + 2) runs of the model, the first order Sobol’ indices may be estimated using
2
Ûj Y
Ŝj = . (65)
V̂
In the same way, for j < k 2 {1, ..., p}2 , the second-order Sobol’ index Sj,k measuring the joint
impact of (Xj , Xk ) on the total variance (removing the first order impacts of Xj and Xk ) is given
by
Uj,k
z }| {
2
V (E(Y |Xj , Xk )) V (E(Y |Xj )) V (E(Y |Xk )) E( E(Y | Xj , Xk ) ) Uj Uk + E(Y )2
Sj,k = =
V (Y ) V (Y )

and may be approximated by

2
Ûj,k Ûj Ûk + Y
Ŝj,k = (66)
V̂
where

N
1 X (1) (1) (1) (1) (2) (2) (1) (2) (2) (1) (2) (2)
Ûj,k = f (xi1 , ..., xij , ..., xik , ..., xip ) ⇥ f (xi1 , ..., xi(j 1) , xij , xi(j11) , ..., xi(k 1) , xik , xi(k+1) , ..., xip ).(67)
N
i=1

Obviously, the approximation of these second-order indices relies on the estimation of the first-order
ones. More generally, the computation of higher-order indices becomes more and more cumbersome
and difficult to handle in practice when the number p of inputs is large. Nevertheless, in order

169
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

to capture the global impact (sum of single and related joint e↵ects) of Xj on the variance of Y ,
remind that the total-e↵ect Sobol’ index may be expressed as
Uj̃
z }| {
E(V (Y |Xj̃ )) V (E(Y |Xj̃ )) E(E(Y |Xj̃ )2 ) E(Y )2
ST j = =1 =1 (68)
V (Y ) V (Y ) V (Y )

where Xj̃ = (X1 , ..., Xj 1 , Xj+1 , ..., Xp ). Thus, at a reasonable computational cost (only using
N (p + 2) runs of the model), total-e↵ect Sobol’ indices may be estimated as follows

2
Ûj̃ Y
ŜTj = 1 (69)
V̂
where
N
1 X (1) (1) (1) (1) (1) (2) (1) (1)
Ûj̃ = f (xi1 , ..., xij , ..., xip ) ⇥ f (xi1 , ..., xi(j 1) , xij , xi(j+1) , ..., xip ).
N
i=1

A.2.7 Synthesis and selection of the GSA: Sobol’ indices estimated by Mc Kay’s
method
Even if this list of UA and SA techniques is not exhaustive, this review has illustrated the great
variety of available methods, positioning in terms of assumptions and type of results. A canonical
synthesis was provided in [Iooss and Lemaı̂tre, 2015] (see Figure A.5) which has several levels of
reading:

• Distinction between screening methods (identification of non-influential variables among a

large number) and more precise variance-based quantitative methods;

• Positioning methods based on their cost in terms of model calls number (which linearly de-
pends in the number of inputs for most of the methods);

• Positioning methods based on their assumptions about the model complexity and regularity;

• Distinction between the type of information provided by each method;

• Identification of methods which require some a priori knowledge about the model behavior.

Based on the characteristics of the di↵erent methods, some authors [De Rocquigny, 2008] have
proposed decision trees to help the practitioner to choose the most appropriate method for its
problem and its model. Figure 2.3 reproduces the flowchart of [De Rocquigny, 2008]. Although
useful to fix some ideas, such diagrams are rather simple and should be used with caution.

In many applications, outputs can be vector or function (temporal, spatial, spatio-temporal,...),32

The case of functional inputs also receives a growing interest33 but its treatment in a functional
statistical framework remains to be done [Iooss, 2011].
32
see [Campbell et al., 2006], [Auder et al., 2010], [Lamboni et al., 2013] who have produced first SA results on such
problems.
33
[Iooss and Ribatet, 2009], [Saint-Geours et al., 2011].

170
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

Figure A.5: SA methods graphical synthesis [Iooss and Lemaı̂tre, 2015].

To study the [Grasselli and Costa Lima, 2012] macroeconomic (apriori non monotonic) model,
with a non-linear dynamic, partial derivative equations, and around 10 inputs (parameters of the
model), we have thus:

1. (Qualitative SA and UA) To conduct first a screening method with a Latin Hypercube Sam-
pling (according to [Iman, 1999]), in order to identify the propagation of uncertainties and
the inputs with negligible e↵ects;

2. (Quantitative SA) Then, to refine our study, by assessing the First order, Second order and
Total e↵ect Sobol’ indices with the [Saltelli, 2002] estimation method.

171
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

A.3 Second-order Sobol’ indices of the [Grasselli and Costa Lima,

2012] model with 100 simulations
In this appendix, we display the second-order Sobol’ indices of the [Grasselli and Costa Lima,
2012] model, computed as explained in Appendix A.2.6, with 100 simulations for the parameters
described in the Table 2.3 of the Subsection 2.3.3. These results, as reported in Tables A.3 and
A.4, are coherent with the properties of the Sobol’ indices (see subsection 2.2.3 about the Variance-
based GSA) where the total-e↵ect Sobol’ indices of one input is always greater than the sum of the
contribution of the first-order Sobol’ indice of this input and its joint e↵ects (second-order Sobol’
indices). Note that The main influent joint e↵ect between inputs is due to the same four parameters
highlighted at first-order: ↵ ⌫ for the wage share, around 0.2346, ↵ for the employment rate,
around 0.1982, and ⌫ for the debt ratio, around 0.1897. The results are once again relatively
stable independently of the outputs.

N = 100
Second-order Sobol’ indices Sij of the [Grasselli and Costa Lima, 2012] model
simulations
Joint-e↵ect Joint-e↵ect
!1 1 d1 !1 1 d1
of inputs of inputs
0.0797 0.1769 0.1237 0.0014 0.0012 0.0017
↵- k0 - r
(0.0439) (0.0818) (0.0736) (0.0014) (0.0006) (0.0032)
0.1082 0.1353 0.13476 0.0005 0.0029 0.0017
- ↵- 0
(0.1044) (0.0827) (0.1235) (0.0002) (0.0042) (0.0033)
0.1913 0.1191 0.1314 0.0005 0.0058 0.0017
-⌫ - 1
(0.1142) (0.0793) (0.0852) (00002) (0.0121) (0.0034)
0.0015 0.0406 0.03499 0.0007 0.0018 0.0024
⌫- 0 - k0
(0.0016) (0.0705) (0.0274) (0.0006) (0.0012) (0.0034)
0.0004 0.0013 0.0016 0.0020 0.0976 0.1310
0 - 1 ⌫ - k1
(0.0002) (0.0011) (0.0032) (0.0006) (0.0926) (0.0976)
0.0007 0.0012 0.0017 0.0005 0.0044 0.0017
1 - k0 0 - k2
(0.0005) (0.0008) (0.0030) (0.0002) (0.0095) (0.0034)
0.0007 0.0012 0.0019 0.0004 0.0008 0.0017
k0 - k1 1 -r
(0.0006) (0.0008) (0.0031) (0.0002) (0.0006) (0.0032)
0.0432 0.0120 0.0455 0.0380 0.0483 0.0429
k1 - k2 ↵- 1
(0.0306) (0.0058) (0.0356) (0.0236) (0.0599) (0.0297)
0.0018 0.0858 0.0017 0.0911 0.0014 0.0021
k2 - r - k0
(0.0012) (0.0576) (0.0034) (0.0712) (0.0006) (0.0061)
0.0877 0.1982 0.1208 0.0406 0.0246 0.0341
↵- - k1
(0.0527) (0.1151) (0.0886) (0.0188) (0.0245) (0.0254)
0.2211 0.1403 0.1897 0.0017 0.0012 0.0276
-⌫ ⌫ - k2
(0.1112) (0.0535) (0.0765) (0.0011) (0.0007) (0.0311)
0.0675 0.0047 0.0642 0.0005 0.0012 0.0017
- 0 0 -r
(0.0506) (0.0092) (0.0537) (0.0002) (0.0007) (0.0032)

Table A.3: Part 1 / Second-order Sobol’ indices and standard deviations (with 10 bootstrap repli-
cates and 100 simulations) of all inputs in the [Grasselli and Costa Lima, 2012] model on the outputs:
the Good Equilibrium values of wages share !1 and employment rate 1

172
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

N = 100
Second-order Sobol’ indices Sij of the [Grasselli and Costa Lima, 2012] model
simulations
Joint-e↵ect Joint-e↵ect
!1 1 d1 !1 1 d1
of inputs of inputs
0.0004 0.0077 0.0017 0.0217 0.034 0.0387
⌫- 1 ↵ - k0
(0.0002) (0.0179) (0.0033) (0.0154) (0.0510) (0.0261)
0.0004 0.0017 0.0017 0.0306 0.0229 0.0228
0 - k0 - k1
(0.0003) (0.0017) (0.0032) (0.0256) (0.0258) (0.0249)
0.0004 0.0027 0.0017 0.0993 0.0387 0.0631
1 - k1 - k2
(0.0003) (0.0049) (0.0033) (0.0798) (0.0406) (0.0489)
0.0008 0.0014 0.0020 0.0028 0.0921 0.0006
k0 - k2 ⌫-r
(0.0006) (0.0004) (0.0032) (0.0011) (0.0747) (0.0004)
0.0012 0.001 0.0017 0.0464 0.0299 0.0178
k1 - r ↵ - k1
(0.0009) (0.006) (0.0032) (0.0324) (0.0338) (0.0110)
0.2346 0.1602 0.1842 0.0184 0.1341 0.1193
↵-⌫ - k2
(0.1655) (0.0873) (0.1272) (0.0213) (0.0321) (0.1152)
0.0004 0.0036 0.0017 0.0018 0.0375 0.0315
- 0 -r
(0.0002) (0.0702) (0.0034) (0.0009) (0.0619) (0.0401)
0.0013 0.0410 0.0352 0.0972 0.0229 0.0904
- 1 ↵ - k2
(0.0015) (0.0718) (0.0273) (0.0641) (0.0112) (0.0878)
0.0245 0.0387 0.0434 0.0018 0.0009 0.0020
⌫ - k0 -r
(0.0090) (0.0569) (0.0403) (0.0010) (0.0007) (0.0034)
0.0005 0.0022 0.0017 0.002 0.0012 0.0017
0 - k1 ↵-r
(0.0002) (0.0354) (0.0032) (0.0015) (0.0007) (0.0033)
0.0004 0.0094 0.00017
1 - k2
(0.0002) (0.0244) (0.0034)

Table A.4: Part 2 / Second-order Sobol’ indices and standard deviations (with 10 bootstrap repli-
cates and 100 simulations) of all inputs in the [Grasselli and Costa Lima, 2012] model on the outputs:
the Good Equilibrium values of wages share !1 and employment rate 1

A.4 Convergence of the Sobol’ indices of the [Grasselli and Costa Lima,
2012] model with a Leontief production function
In this appendix, we refine the meshing of our inputs (labor productivity growth rate ↵; population
growth rate ; depreciation rate and capital-to-output ration ⌫) to give a better assessment of the
Sobol’ indices of the [Grasselli and Costa Lima, 2012] model outputs: the Good Equilibrium values
of the wage share !1 , the employment rate 1 and the debt ratio of the economy d1 . Following
[Harenberg et al., 2017], we choose to stop this exploration when the standard deviation of every
indices is less than 0.05. These exploration with an increasing number of simulations (200, 300,
400, 500) leads to the results displayed in Tables A.5, A.6 and A.7). Once again, one can notice the
stability of the results whatever is the number of simulations.

173
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

First-order Si N= 200 simulations N= 300 simulations

Input Description !1 1 d1 !1 1 d1
labor productivity 0.083 0.127 0.093 0.080 0.128 0.091
↵
growth rate (0.0406) (0.0677) (0.0783) (0.0440) (0.0509) (0.0570)
0.069 0.076 0.083 0.078 0.063 0.082
population growth rate
(0.0500) (0.0358) (0.0839) (0.0499) (0.0306) (0.0618)
0.057 0.111 0.066 0.054 0.118 0.080
depreciation rate
(0.0376) (0.0746) (0.0384) (0.0347) (0.0685) (0.0282)
0.185 0.051 0.132 0.180 0.052 0.113
⌫ capital-to-output ratio
(0.0551) (0.0457) (0.0585) (0.0533) (0.0278) (0.0456)

Table A.5: First-order Sobol’ indices and standard deviations (10 bootstrap replicates for each 200
& 300 simulations) of the main influent inputs of the [Grasselli and Costa Lima, 2012] model on
the outputs: the Good Equilibrium values of wages share !1 , employment rate 1 and debt ratio d1

First-order Si N= 400 simulations N= 500 simulations

Input Description !1 1 d1 !1 1 d1
labor productivity 0.086 0.144 0.101 0.088 0.133 0.105
↵
growth rate (0.0351) (0.0597) (0.0511) (0.0470) (0.0482) (0.0344)
0.075 0.072 0.085 0.077 0.070 0.090
population growth rate
(0.0430) (0.0487) (0.0489) (0.0421) (0.0260) (0.0376)
0.053 0.101 0.081 0053 0.100 0.073
depreciation rate
(0.0497) (0.0543) (0.0511) (0.0227) (0.0248) (0.0447)
0.186 0.054 0.121 0.189 0.060 0.122
⌫ capital-to-output ratio
(0.0516) (0.0363) (0.0530) (0.0498) (0.0457) (0.0403)

Table A.6: First-order Sobol’ indices and standard deviations (10 bootstrap replicates of each 400
& 500 simulations) of the main influent inputs of the [Grasselli and Costa Lima, 2012] model on
the outputs: the Good Equilibrium values of wages share !1 , employment rate 1 and debt ratio d1

N = 500 simulations Second-order Sobol’ indices Sij

Joint-e↵ect of inputs !1 1 d1
↵- 0.053 (0.0331) 0.068 (0.0548) 0.056 (0.0307)
- 0.070 (0.0530) 0.091 (0.0532) 0.068 (0.0603)
-⌫ 0.064 (0.0463) 0.061 (0.0352) 0.066 (0.0504)
↵- 0.054 (0.0449) 0.098 (0.05997) 0.058 (0.0469)
-⌫ 0.076 (0.0505) 0.064 (0.0516) 0.079 (0.0543)
↵-⌫ 0.067 (0.0491) 0.057 (0.0341) 0.073 (0.0467)

Table A.7: Second-order Sobol’ indices and standard deviations (10 bootstrap replicates and 500
simulations) of the main influent inputs of the [Grasselli and Costa Lima, 2012] model on the
outputs: the Good Equilibrium values of wages share !1 , employment rate 1 and debt ratio d1

174
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

A.5 Sobol’ indices of the [Goodwin, 1967] model

Inspired from the description of the [Goodwin, 1967] model by [Grasselli and Costa Lima, 2012],
we undertake a GSA of the model with the same method: GSA with Sobol’ indices estimated
by [Saltelli, 2002] method. The outputs of our study are the Good Equilibrium values of the wage
share !1 and the employment rate 1 , because there is no debt in this model. The inputs of our study
are all the parameters of the model with the same ranges (see Table A.8) as in the GSA of section
2.3. The results are reported in Table A.9 and show that, at first-order and for only 100 simulations,
the [Goodwin, 1967] model is more robust to its inputs than the [Grasselli and Costa Lima, 2012]
(the capital-to-output ratio ⌫ is slightly less sensitive), at a first-order and total-e↵ect point of view.
This last result is coherent with the idea that more nonlinearities and interactions between inputs
are possible in the [Grasselli and Costa Lima, 2012] than in [Goodwin, 1967]. These results makes
us wonder about a possible compromise between the robustness and the realism of a model to find
an equilibrium between a very robust model unable to predict the reality (such as the [Goodwin,
1967] model), and a very fitted and very sensitive model (such as the DSGE of [Ratto, 2008]. From
this point of view, the [Grasselli and Costa Lima, 2012] model seems an interesting compromise,
with a finer description of the economy and a more robust structure.

Symbol Definition Min Value Max Value Range Nom. Value

↵ labor productivity growth rate 0 0.03 0.03 0.025
population growth rate 0.015 0.03 0.015 0.02
depreciation rate 0.005 0.03 0.025 0.01
⌫ capital-to-output ratio 2 4.5 2.5 3
0.04
0 Philips Curve ( ) parameter 0.038 0.042 0.004 1 0.042
0.043
1 Philips Curve ( ) parameter 6.333 E-5 7 E-5 6.67 E-6 1 0.04

Table A.8: Definition and Range of the inputs of the GSA for the [Goodwin, 1967] model

[Goodwin, 1967] - 100 simulations First-order Si Total e↵ect SiT

Input Description !1 1 !1 1
0.068 0.050 0.630 0.955
↵ labor productivity growth rate
(0.0388) (0.0448) (0.0484) (0.0317)
0.050 0.074 0.618 0.726
population growth rate
(0.0183) (0.0693) (0.2825) (0.3882)
0.084 0.060 0.677 0.849
depreciation rate
(0.0544) (0.0431) (0.0474) (0.0562)
0.240 0.056 0.857 0.926
⌫ capital-to-output ratio
(0.0581) (0.0486) (0.0428) (0.2833)
0.028 0.012 0.297 0.115
0 Philips Curve ( ) parameter
(0.0523) (0.0263) (0.0443) (0.2798)
0.028 0.012 0.297 0.115
1 Philips Curve ( ) parameter
(0.0523) (0.0263) (0.0443) (0.0444)

Table A.9: Sobol’ indices and standard deviations of all inputs of the [Goodwin, 1967] model on
the outputs: the Good Equilibrium values of wages share, !1 , and employment rate, 1 .

175
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

Second-order Sobol’ indices Sij of

N = 100 simulations
the [Goodwin, 1967] model
Joint-e↵ect of inputs !1 1
↵- 0.057 (0.0103) 0.076 (0.0204)
- 0.085 (0.0153) 0.102 (0.0190)
-⌫ 0.106 (0.0173) 0.066 (0.0161)
⌫- 0 0.032 (0.0161) 0.017 (0.008)
0 - 1 0.023 (0.0129) 0.016 (0.0259)
↵- 0.098 (0.0174) 0.098 (0.0240)
-⌫ 0.087 (0.0174) 0.137 (0.0089)
- 0 0.034 (0.0168) 0.017 (0.0085)
⌫- 1 0.029 (0.0147) 0.017 (0.0168)
↵-⌫ 0.085 (0.0213) 0.097 (0.0083)
- 0 0.032 (0.0161) 0.017 (0.0089)
- 1 0.035 (0.0180) 0.017 (0.0084)
↵- 0 0.020 (0.0130) 0.017 (0.0083)
- 1 0.033 (0.0166) 0.017 (0.0084)
↵- 1 0.035 (0.0179) 0.017 (0.0084)

Table A.10: Second-order Sobol’ indices and standard deviations of all inputs of the [Goodwin,
1967] model on the outputs.

What is remarkable here is that the most influential parameters are the same for both models, be
it [Goodwin, 1967] or [Grasselli and Costa Lima, 2012]: the labor productivity, ↵, the population
growth rate, , the depreciation rate, , and the capital-to-output ratio, ⌫. This result seems to
confirm the idea that refining this model(e.g. by introducing a debt ratio as in [Grasselli and
Costa Lima, 2012]) does not fundamentally a↵ect the nature of its sensitivity.

A.5.1 Sobol’ indices of the [Grasselli and Costa Lima, 2012] model with CES
production function
In order to improve the performance of the model in terms of robustness, one can decide to
switch the production function from a Leontief shape to a Constant Elasticity of Substitution (CES)
production function as in the work of [McIsaac, 2016] based on the Van der Ploeg’s extension. This
transformation endogeneizes the capital-to-output ratio ⌫ responsible for the main impact in terms
of Sobol’ indices on the output of our study.

[Van der Ploeg, 1985] indeed relaxed the assumption that capital and labor cannot be substituted
with each other by endowing the economy with a CES production function:
⌘ ⌘ 1/⌘
Y = C[bK + (1 b)(aL) ] (70)

where C > 0 is the factor productivity, b 2 [0, 1] is the share of capital and the short-run elasticity
1
of substitution between factors is = 1+⌘ . One can notice that when ⌘ ! +1, one can obtain a
Leontief production function without substituability (where C = ⌫1 ), and if ⌘ ! 0 this form leads
to a Cobb-Douglas production function. The first-order condition of the profit maximization leads

176
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

to the following shape of ⌫(t):

✓ ◆ 1/⌘
1 1 !(t)
⌫(t) = (71)
C b

Instead of ⌫, the study owns now three more parameters: the factor productivity, C, the share
of capital, b, and a parameter, ⌘, controlling the elasticity of substitution with the following ranges
(see Table A.11). When applying the same methodology to this model, we obtain the following

Symbol Definition Min Value Max Value Range Nom. Value

C Factor productivity 1/4.5 1/2 0.2778 1/3
b Share of capital 0.1 0.3 0.2 0.135
Controlling the elasticity of
⌘ 0 600 600 50
substitution

Table A.11: Definition and Range of the inputs of the GSA for the [Grasselli and Costa Lima, 2012]
model with CES production function

Sobol’ indices (Tables A.12, A.13 and A.14).

The results of this analysis shows that at first-order and for 100 simulations (10 bootstrap),
the [Grasselli and Costa Lima, 2012] model with CES production is slightly more robust to its inputs
than the same model with a Leontief production function. At a total-e↵ect order, the quadruplet
(↵, , , ⌫) is replaced by (↵, , , C)34 which seems coherent with the sensitivity of [Grasselli and
Costa Lima, 2012] be it with a Leontief or a CES production function.

Quantitatively, at first-order the main influent parameter on the wage share is now the nominal
interest rate r around 0.174 (since ⌫ is now endogeneized), but at a total-e↵ect order, this input is
not very influent due to its few joint e↵ect with others inputs.

The main influent parameter on the employment rate is still the labor productivity growth rate ↵
around 0.196 (greater than with a Leontief production function) and the main influent parameter on
the debt ratio is the factor productivity C around 0.127. At a total-e↵ect level, the most impacting
inputs (with all joint e↵ects) are still the labor productivity growth rate ↵, the population growth
rate , the depreciation rate and now the factor productivity C.

Due to the increased number of possible joint e↵ect, the value of the second-order Sobol’ indices
are smaller than with a Leontief production function (66 possible joint e↵ect instead of 45). The
main influent joint e↵ect on the outputs are now ↵ for the wage share, around 0.0383 (instead
of 0.2346 for ↵ ⌫), ↵ for the employment rate, around 0.0397 (instead of 0.1982 for the same
inputs in the model with a Leontief function), ↵ for the debt ratio, around 0.0409 (instead of
0.1897 for ⌫).
⇣ ⌘ 1/⌘
34
Note that: ⌫(t) = C1 1 !(t)b
which means C and ⌫ are closely related. Note e.g. that ⌫(t) ! 1/C when the
elasticity of substitution, ⌘ tends to 1.

177
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

First-order Si Total e↵ect SiT

Input Description !1 1 d1 !1 1 d1
labor
0.135 0.196 0.101 0.822 0.863 0.791
↵ productivity
(0.0788) (0.16477) (0.0806) (0.1971) (0.1470) (0.2102)
growth rate
population 0.096 0.107 0.100 0.754 0.728 0.755
growth rate (0.0923) (0.0705) (0.0664) (0.1876) (0.3113) (0.2464)
depreciation 0.127 0.099 0.124 0.865 0.792 0.868
rate (0.0640) (0.0722) (0.0786) (0.1808) (0.2289) (0.1597)
factor 0.106 0.114 0.127 0.876 0.856 0.876
C
productivity (0.0886) (0.0902) (0.0088) (0.1635) (0.1312) (0.1784)
Share of 0.085 0.115 0.082 0.254 0.288 0.221
b
capital (0.0799) (0.0773) (0.0628) (0.2226) (0.2473) (0.2662)
Controlling
0.087 0.122 0.073 0.282 0.288 0.312
⌘ the elasticity
(0.0707) (0.0934) (0.0535) (0.2978) (0.2763) (0.3624)
of substitution
Philips Curve
0.0091 0.0121 0.0085 0.185 0.445 0.224
0 ( )
(0.0078) (0.0082) (0.0060) (0.1385) (0.2238) (0.3037)
parameter
Philips Curve
0.0089 0.0154 0.0087 0.185 0.547 0.272
1 ( )
(0.0079) (0.0011) (0.0059) (0.1386) (0.5121) (0.6022)
parameter
Investment
0.083 0.137 0.096 0.320 0.320 0.224
k0 function (⇡)
(0.0826) (0.0784) (0.0589) (0.3640) (0.3583) (0.3318)
constant
(⇡) 0.099 0.124 0.093 0.343 0.383 0.308
k1
parameter (0.0913) (0.0998) (0.0644) (0.3907) (0.3354) (0.4289)
(⇡)
0.105 0.141 0.094 0.429 0.445 0.387
k2 exponential
(0.0891) (0.1035) (0.0611) (0.4132) (0.4493) (0.4506)
parameter
nominal 0.174 0.128 0.090 0.310 0.265 0.182
r
interest rate (0.1029) (0.1016) (0.0589) (0.0544) (0.1323) (0.0094)

Table A.12: Sobol’ indices and standard deviations (with 10 bootstrap replicates) of all inputs of
the [Grasselli and Costa Lima, 2012] model with a CES production function on the outputs: the
Good Equilibrium values of wages share !1 , employment rate 1 and debt ratio d1

The [Grasselli and Costa Lima, 2012] model is slightly more robust to its inputs than the same
model with Leontief production function. In order to improve the first-order performance of the
model in terms of robustness, one can be interested in endogeneizing the nominal interest rate r by
introducing a Taylor’s rule, or the labor productivity growth rate (as suggested before in the model
with Leontief production function).

178
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

N = 100
Second-order Sobol’ indices Sij of the [Grasselli and Costa Lima, 2012] model
simulations
Joint-e↵ect Joint-e↵ect
!1 1 d1 !1 1 d1
of inputs of inputs
0.0383 0.0221 0.0409 0.0092 0.0187 0.0166
↵- 1 - k1
(0.0340) (0.0249) (0.0642) (0.0094) (0.0121) (0.0221)
0.0343 0.0318 0.0257 0.0268 0.0222 0.0156
- k0 - k2
(0.0316) (0.0571) (0.0203) (0.0285) (0.0162) (0.0081)
0.0282 0.0353 0.0205 0.0015 0.0012 0.0011
-C k1 - r
(0.0307) (0.0559) (0.0176) (0.0011) (0.0009) (0.0005)
0.0214 0.0118 0.0387 0.0168 0.0266 0.0289
C -b ↵-C
(0.0252) (0.0126) (0.0602) (0.0178) (0.0351) (0.0316)
0.0023 0.0129 0.0009 0.0026 0.0222 0.0317
b-⌘ -b
(0.0027) (0.0122) (0.0006) (0.0021) (0.0236) (0.0435)
0.0035 0.0168 0.0011 0.0187 0.0149 0.0198
⌘- 0 -⌘
(0.0005) (0.0136) (0.0667) (0.0131) (0.0119) (0.0198)
0.0004 0.0138 0.0011 0.0123 0.0138 0.0010
0 - 1 C - 0
(0.0006) (0.009) (0.0006) (0.0156) (0.0085) (0.0006)
0.0008 0.0179 0.0010 0.0199 0.0172 0.0253
1 - k0 b- 1
(0.0005) (0.011) (0.0006) (0.0236) (0.0599) (0.0297)
0.0111 0.0167 0.0025 0.0026 0.0012 0.0015
k0 - k1 ⌘ - k0
(0.0062) (0.0112) (0.0039) (0.0032) (0.0009) (0.0015)
0.0223 0.0162 0.0134 0.0200 0.0175 0.0256
k1 - k2 0 - k1
(0.0282) (0.0107) (0.0052) (0.0197) (0.0129) (0.0253)
0.0012 0.0046 0.0011 0.0200 0.0197 0.0256
k2 - r 1 - k2
(0.0011) (0.0067) (0.0005) (0.0197) (0.0136) (0.0253)
0.0307 0.0397 0.0246 0.0009 0.0016 0.0018
↵- k0 - r
(0.0443) (0.0866) (0.0179) (0.0007) (0.0011) (0.0023)
0.0133 0.0381 0.0262 0.0302 0.0157 0.0256
-C ↵-b
(0.0069) (0.0748) (0.0227) (0.0374) (0.0132) (0.0277)
0.0252 0.0127 0.0217 0.0139 0.0150 0.0205
-b -⌘
(0.0273) (0.0131) (0.0192) (0.0119) (0.0115) (0.0167)
0.0215 0.0131 0.0182 0.0190 0.0145 0.0256
C -⌘ - 0
(0.0209) (0.0105) (0.0165) (0.0176) (0.0103) (0.0261)
0.0025 0.0141 0.0023 0.0174 0.0209 0.0242
b- 0 C - 1
(0.0035) (0.0131) (0.0025) (0.0165) (0.0159) (0.0242)
0.0016 0.0154 0.0016 0.0223 0.0187 0.0029
⌘- 1 b - k0
(0.0018) (0.0118) (0.0016) (0.0188) (0.0151) (0.0029)
0.0046 0.0174 0.0011 0.0257 0.0176 0.0055
0 - k0 ⌘ - k1
(0.0051) (0.0122) (0.0006) (0.0272) (0.0142) (0.0061)

Table A.13: Part 1 / Second-order Sobol’ indices and standard deviations (with 10 bootstrap
replicates and 100 simulations) of all inputs in the [Grasselli and Costa Lima, 2012] model with
CES production function on the outputs: the Good Equilibrium values of wages share !1 and
employment rate 1

179
APPENDIX A. BUILDING A GLOBAL SENSITIVITY ANALYSIS TO QUANTIFY THE
ROBUSTNESS OF MACRO-ECONOMIC MODELS

N = 100
Second-order Sobol’ indices Sij of the [Grasselli and Costa Lima, 2012] model
simulations
Joint-e↵ect Joint-e↵ect
!1 1 d1 !1 1 d1
of inputs of inputs
0.0199 0.0113 0.0256 0.0203 0.0145 0.0157
0 - k2 C - k2
(0.0197) (0.0099) (0.0100) (0.0123) (0.0101) (0.0105)
0.0003 0.0013 0.0011 0.0008 0.0154 0.0011
1 -r b-r
(0.0002) (0.0010) (0.0006) (0.0004) (0.0138) (0.0006)
0.0256 0.0157 0.0203 0.0248 0.0191 0.0322
↵-⌘ ↵ - k0
(0.0355) (0.0119) (0.0221) (0.0283) (0.0125) (0.0312)
0.0193 0.0155 0.0255 0.0324 0.0157 0.0308
- 0 - k1
(0.0188) (0.0081) (0.0249) (0.0466) (0.0074) (0.0511)
0.0190 0.0173 0.0258 0.0165 0.0146 0.0177
- 1 - k2
(0.0176) (0.0137) (0.0261) (0.0108) (0.0111) (0.0174)
0.0275 0.0150 0.0020 0.0446 0.0012 0.0023
C - k0 C -r
(0.0381) (0.0122) (0.0019) (0.0833) (0.0009) (0.0039)
0.0186 0.0154 0.0019 0.0214 0.0147 0.0334
b - k1 ↵ - k1
(0.0153) (0.0127) (0.0018) (0.012) (0.0150) (0.0545)
0.0199 0.0197 0.0114 0.0242 0.0189 0.0243
⌘ - k2 - k2
(0.0217) (0.0138) (0.0141) (0.0224) (0.0115) (0.0335)
0.0007 0.0168 0.0010 0.0238 0.0149 0.0181
0 -r -r
(0.0015) (0.0129) (0.0005) (0.0161) (0.0107) (0.0163)
0.0213 0.0159 0.0271 0.0182 0.0174 0.0134
↵- 0 ↵ - k2
(0.0215) (0.0123) (0.0265) (0.0130) (0.0121) (0.0070)
0.0193 0.0155 0.0255 0.0253 0.0165 0.0254
- 1 -r
(0.0188) (0.0101) (0.0249) (0.0201) (0.0127) (0.0252)
0.0188 0.0259 0.0266 0.0314 0.0176 0.0192
- k0 ↵-r
(0.0159) (0.0363) (0.0306) (0.0249) (0.0133) (0.0180)
0.0323 0.0129 0.0192 0.0247 0.0187 0.0053
C - k1 b - k2
(0.0362) (0.0088) (0.0115) (0.0253) (0.0138) (0.0058)
0.0012 0.0186 0.0010 0.0213 0.0152 0.0271
⌘-r ↵- 1
(0.0006) (0.0009) (0.0006) (0.0798) (0.0406) (0.0489)
0.0015 0.001 0.0017 0.0223 0.0240 0.0286
- k0 - k1
(0.0455) (0.0115) (0.0411) (0.0239) (0.0363) (0.0305)

Table A.14: Part 2 / Second-order Sobol’ indices and standard deviations (10 bootstrap replicates,
100 simulations) of all inputs in the CES [Grasselli and Costa Lima, 2012] model on the outputs

180
Appendix B

Money Velocity in an Imperfectly

Competitive, Stock-Flow Consistent
Dynamics

B.1 Long-run properties of the moneyless model

This section analyzes the local, asymptotic stability of long-run steady states in the vector field
(26).

B.1.1 The reduced dynamical system

Notice first that the behavior of the households’ debt ratio, dh , in (26), can be deduced from the
other variables while the latter are independent from dh . We therefore focus on the 3-dimensional
system (!, , df ), together with the following auxiliary variables:
8
>
> 1 ⇣1 ! ⌘ 1/⌘
>
> ⌫(!) =
>
> A b
>
>
>
> i(!) = ⌘p (µ! 1)
<
⇡e = 1 ! rdf ⌫(!) (1)
>
>
>
> yd = ch (! + rdf + ⇡e ) + (⇡e )
>
> ✓ ◆
>
> (⇡e ) !˙
>
> ˙
: dh = ch (! + rdf + ⇡e ) ! rdf ⇡e dh + i(!) .
⌫ ⌘(1 !)

The associated Jacobian matrix is:

0 @f ⌘
1
1
! ⌘+1 0( ) 0
@!
B C
B C
B @f2 @f2 0 (⇡ C
e) C
J(!, , df ) = B
B @! @ r ⌫(!) C (2)
B C
@ A
@f3 df ! 0( @f3
@! ⌘+1 1 ! ) @df

181
APPENDIX B. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE,
STOCK-FLOW CONSISTENT DYNAMICS

where
@f1 ⌘ h i ⌘!
:= ( ) ↵ (1 )i(!) (1 )⌘p µ
@! ⌘+1 ⌘+1
✓ 0 ✓ ◆ ◆
@f2  (⇡e ) ⌫(!) (⇡e ) ( ) ↵ (1 )i(!) (1 )⌘p µ
:= 1+ + +
@! ⌫(!) ⌘(1 !) ⌘(1 !)⌫(!) (⌘ + 1)(1 !)2 (⌘ + 1)(1 !)
@f2 (⇡e ) ( ) ↵ (1 )i(!) 0 ( )
:= (↵ + n + )
@ ⌫(!) (⌘ + 1)(1 !) (⌘ + 1)(1 !)
✓ 0 ✓ ◆ ◆
@f3  (⇡e ) ⌫(!) (⇡e ) ( ) ↵ (1 )i(!) !(1 )⌘p µ
:= df 1+ + ⌘p µ +
@! ⌫(!) ⌘(1 !) ⌘(1 !)⌫(!) (⌘ + 1)(1 !)2 (⌘ + 1)(1 !)
✓ ◆
⌫(!) @y d (1 ) ⌫(!)
0 (⇡e ) 1 + + (yd ) (1 !) +
⌘(1 !) @! ⌘(1 !)
✓ ◆
@f3 0 @yd (⇡e ) ( ) ↵ (1 )i(!)
:= (1 )r r (⇡e ) (1 !) + i(!) !
@df @df ⌫(!) (⌘ + 1)(1 !)
0
 (⇡e )
+ rdf
⌫(!)
(3)

B.1.2 Existence and stability of long-run equilibria

Let us adopt the following set of rather weak assumptions:1
Assumption A

1. The (short-run Phillips) function 2 is continuously di↵erentiable on [0, 1] and

0
(i) ( ) > 0 on [0; 1]
(ii) (0) < ↵ + (1 )i(0) = ↵ (1 )⌘p (4)
(iii) lim ( ) = +1.
!1

2. The investment function (·) is increasing, continuously di↵erentiable on R. There is some

⇢ > 0 such that, as long as rdf > ⇢:3

(i) 0 (⇡e ) > 0 on [0, 1]

b1/⌘
(ii) lim (!, df ) = 0 (df )  (↵ + n + )⌫(0) = (↵ + n + ) (5)
!!0 A
(iii) lim ⇡ 2 0 (⇡e ) < +1.
⇡e ! 1 e
1
Of course, if one is only concerned with existence, continuous di↵erentiability can be replaced by local Lips-
chitzianity.
2
Its form and parameterization guarantee that 8 t, 0  (t)  1, as in [Desai et al., 2006]. Observe also that (ii)
implies (0) < ↵ because i(0) = ⌘p < 0.
3
These restrictions are sufficient to guarantee the existence of the relevant equilibria alluded to in section 4.3, as
in [Grasselli and Costa Lima, 2012]. The lower bound, ⇢, will be useful when dealing with persistence in section 3.5.2.

182
APPENDIX B. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE,
STOCK-FLOW CONSISTENT DYNAMICS

3. The consumption function ch (·) is strictly increasing and continuously di↵erentiable on R+ ,

and verifies:4
(i) 8 x  0, ch (x) = c > 0
(ii) lim ch (x) = c+ . (6)
x!+1

1
4. Eventually, r + A b ⌘ (0) > 0.5

Our numerical example 4.4.2 complies with this set of assumptions.

The Solovian equilibrium

The Solovian equilibrium corresponds to a desirable long-run steady state with positive wages
and employment, and a finite level of debt (!1 ; 1 ; df1 ). To check its existence, observe that ˙ = 0
in the second equation of (26) whenever
✓ ◆ 1
↵+n+ 1 !1 ⌘
(⇡1 ) = = (↵ + n + )⌫(!1 ). (7)
A b

Equation (15) can be uniquely solved for ⇡1 because of condition (9), which implies that df1 = (1
⇡1 ! 1 ⌫(!1 ))/r is uniquely determined by !1 . Given !1 , we also obtain 1 = 1 (↵ +(1 )i1 ),
where i1 = ⌘p (µ!1 1), which belongs to [0; 1], courtesy of (8) (ii) and provided i1 > 0.

Because dh can be deduced from the other state variables, ch (!1 + rdf1 + ⇡1 ) depends only upon
!1 . The same holds for yd1 = (⇡1 ) + ch (!1 + rdf1 + ⇡1 ). Consequently, in order to guarantee
d˙ = 0 in the third equation of (26), we need: (⇡1 ) (1 !1 )(yd1 ) + (1 ) ⌫(!1 ) df1 [↵ +
n + i1 (1 )r] = 0, which leads to a nonlinear equation in !1 .

Each side of (15) is a continuous function of ! 2 [0; 1]. The right-hand side of the equation is
null if ! = 1, and equals (↵ + n + )⌫(0) > 0 if ! = 0. In the left-hand side, (·) is an increasing
function of !. Consequently, assumption (9) (ii) ensures the existence of at least one solution of
(15). At this steady state, the growth rate of the economy, g = ẎY + ṗp with ẎY = (⇡
⌫(!)
e) !˙
⌘(1 !) ,
becomes: g = ↵ + n + i1 .
The Jacobian matrix becomes:
0 1
Ki K0 0
B C
B C
B K1 K! K1e K rK1 C
B C (8)
B C
B C
@K2 K! + K e + Ky! Kd f rK2 + (1 )r Kyd (↵ + n + ī)A
2

4
Condition (i) expresses the fact that households need a minimum level of subsistence consumption even at negative
income, whereas (ii) says that total normalized consumption can be higher than 1 (and consequently exceed total
output, by sourcing from the stock of inventories) but must have an upper bound. See [Giraud and Grasselli, 2019]
for details.
5
This condition is borrowed from [Bastidas et al., 2019], and turns out to be sufficient for the slavery equilibrium
to exist.

183
APPENDIX B. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE,
STOCK-FLOW CONSISTENT DYNAMICS

with
⌘!1 ⌘!1 0
Ki = (1 )⌘p µ , K0 = ( 1 ),
⌘+1 ⌘+1
0 (⇡1 ) 0( )
1
K1 = 1 , K = 1
⌫(!1 ) (⌘ + 1)(1 !1 )
h (⇡1 ) (1 )⌘p µ i
K1e = 1 ,
⌘(1 !1 )⌫(!1 ) (⌘ + 1)(1 !1 )
✓ ◆
df1 ⌫(!1 ) 0 ⌫(!1 )
K2 =  (⇡1 ) , K! = 1 + (9)
⌫(!1 ) ⌘(1 !1 )
✓ ◆
(⇡1 ) (1 )⌘p µ
K2e = d f1 ⌘ p µ !1
⌘(1 !1 )⌫(!1 ) (⌘ + 1)(1 !1 )
@yd (1 ) ⌫(!1 )
Ky! = yd 1 (1 !1 ) +
@! |(!1 ; 1 ;df1 ) ⌘(1 !1 )
0
d f1 ! 1 ( 1 ) @yd
Kdf = , Kyd = (1 !1 ) .
(⌘ + 1)(1 !1 ) @df1 |(!1 ; 1 ;df )
1

The characteristic polynomial is P (X) = X 3 + a2 X 2 + a1 X + a0 , where

a0 = K0 [rK1 (K2 K! + K2e + Ky! ) b2 (K1 K! + K1e )] Ki (rK1 Kdf b2 K )

a1 = rK1 Kdf b2 K + Ki (b2 K ) + K0 (K1 K! + K1e ) (10)
a2 = K b2 Ki where b2 = rK2 + (1 )r Kyd (↵ + n + ī).

a3 a0 a2 a1
The Routh-Hurwitz Stability criterion requires a3 0, a2 0, a2 0, a0 0, which,
here, means:

K rK2 (1 )r + Kyd + (↵ + n + i1 ) Ki 0, (11)

K0 [rK1 (K2 K! + K2e + Ky! ) b2 (K1 K! + K1e )] Ki (rK1 Kdf b2 K ) 0, (12)
a2 a1 a0 0. (13)

This condition is satisfied for a wide range of parameters and, in particular, by the parameterization
of our numerical example.6

Deflationary equilibria with finite debt

1) Another economically doubtful equilibrium exists, (!, , df ) = (!2 , 0, df2 ), provided some
(!2 , df2 ) solves the nonlinear equations:
8
> (0) ↵
>
<i(!2 ) = 1
h (⇡ ) i (14)
>
> 2
: 0 = (⇡2 ) (1 !2 )(yd2 ) + (1 ) ⌫(!2 ) d f2 (1 )r + i(!2 )
⌫(!2 )

with ⇡2 = 1 !2 rdf2 ⌫(!2 ) and yd2 = (⇡2 ) + ch (!2 + rdf2 + ⇡2 ). Assumption (8) (ii)
implies that i(!2 ) < 0, which motivates the name of this deflationary equilibrium. Positive wages
6
Section 4.4.2 and the final values : (!1 , 1 , d f 1 , ⇡ 1 , i 1 , yd1 ) = (0.791, 0.969, 0.592, 0.161, 0.013, 0.915).

184
APPENDIX B. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE,
STOCK-FLOW CONSISTENT DYNAMICS

being distributed to unemployed workers make however this Cornucopian equilibrium economically
hardly relevant. Its Jacobian matrix is lower triangular. The stability criterion is:
(1 )r r0 (⇡2 ) (1 !2 ) (1 )rc0h (!2 + rdf2 + ⇡2 ) r0 (⇡2 )
✓ ◆
(⇡2 ) 0 (⇡2 ) (15)
+ i(!2 ) + rdf2 < 0,
⌫(!2 ) ⌫(!2 )
which is equivalent to
d !2 ⌫(!2 ) (⇡2 )
(1 )r(1 (1 !2 )c0h (!2 + rdf2 + ⇡2 )) + i(!2 ) + r0 (⇡2 ) f2 ⌫(!2 ) ⌫(!2 ) < 0.
Given that > 0, !2 2 [0, 1], c0h 0 and i(!2 ) < 0, the local stability cannot be ruled out a
priori. Notice, however, that, with the peculiar investment function chosen for our leading numerical
example, 0 (⇡e ) = k1 k2 ek2 ⇡e = k2 ((⇡e ) k0 ) which implies that the stability criterion becomes:
d f2 !2 ⌫(!2 )
(1 )r(1 (1 !2 )c0h (!2 + rdf2 + ⇡2 )) + i(!2 ) rk2 k0
⌫(!2 )
(16)
rk2 (df2 !2 ⌫(!2 )) 1
+ (⇡2 ) < 0.
⌫(!2 )
Zero-profit should presumably lead to zero investment: (0) = k0 + k1 = 0, so that k0 should be
negative. If this is so, and for a wide range of parameters r and k2 , criterion (16) is violated. This
holds, in particular, for our leading parameterization.

2) An economically more relevant, deflationary steady state is given by (!, , df ) = (0, 0, df3 ),
provided df3 solves the nonlinear equation:
h (⇡ ) i
3
(⇡3 ) (1 )yd3 + (1 ) ⌫(0) df3 (1 )r + i(0) = 0, (17)
⌫(0)
where ⇡3 = 1 rdf3 ⌫(0) and yd3 = (⇡3 ) + ch (rdf3 + ⇡3 ). Here again, i(0) = ⌘p < 0, and the
Jacobian matrix is lower triangular. Under Assumption (8) (ii), the sign of its eigenvalues depends
on the chosen parameterization, so that local stability cannot be ruled out a priori. Indeed, the
stability criterion is:7
✓ ◆
0 0 (⇡3 ) 0 (⇡3 )
(1 )r (1 )rch (rdf3 + ⇡3 ) + r (⇡3 ) ⌘p + rdf3 <0 (18)
⌫(0) ⌫(0)
With our specific investment function, (⇡e ) = k0 + k1 ek2 ⇡e , 0 (⇡e ) = k1 k2 ek2 ⇡e = k2 ((⇡e ) k0 ),
so that the stability criterion becomes :
df 3 ⌫(0) rk2 (df3 ⌫(0)) 1
(1 )(1 c0h (rdf3 + ⇡3 ))r + + ⌘p rk2 k0 + (⇡3 ) < 0.
⌫(0) ⌫(0)
(19)
As before, there is a broad range of parameters r, k2 , — including our canonical parameterization
— that does not satisfy this stability condition.On the other hand, (19) does not depend upon .

Trivial and slavery steady states

As in the literature quoted, there exist as well a trivial and a slavery equilibrium, given respectively
by:
(!, , df ) = (0, 0, 0) and (!, , df ) = (0, 5 , df 5 ) . (20)
7
Whenever (0) < ↵ + (1 )i(0) is replaced with the (stronger) condition (0) < ↵, the first eigenvalue becomes
positive and so this equilibrium is locally unstable.

185
APPENDIX B. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE,
STOCK-FLOW CONSISTENT DYNAMICS

In our di↵erential system (26), the presence of depreciation into the profit equation implies that
(!, , df ) = (0, 0, 0) leads to d˙ = 0 if, and only if, (⇡4 ) (yd4 ) + (1 ) ⌫(0) = 0, where
⇡4 = 1 ⌫(0) and yd4 = (⇡4 ) + ch ( ⇡4 ). This a structurally unstable condition on the model’s
parameters.

In a similar way, for (0, 1 (⌫(0)(↵ +

5 , df 5 ) to be an equilibrium, one needs 1 rdf5 ⌫(0) = 
n + )) where

⌫(0)(↵ + n + ) (yd5 ) + (1 ) ⌫(0)

d f5 = ,
↵+n (1 )r + i(0)
and yd5 = ⌫(0)(↵ + n + ) + ch (rdf5 + ⇡5 ).

Hence, both equilibria are structurally unstable.

Infinite-valued debt-deflationary equilibria

Contrary to [Grasselli and Nguyen-Huu, 2018], a long-run equilibrium with an unbounded debt
ratio, (!, , df ) = (0, 0, +1), exists as a solution of (26). Let us indeed make the change of variable
q = 1/df , as in [Grasselli and Costa Lima, 2012], and observe that (!, , q) = (0, 0, 0) is a zero of
the transformed system:
8 ⌘! h i
>
> !
˙ = ( ) ↵ (1 )i(!)
>
> ⌘+1
>
< h i
˙ = (⇡e ) !˙
(↵ + n + ) (21)
>
> ⌫(!) ⌘!(1 !)
>
> h i h i
>
: q̇ = q g(!, 1/q) (1 )r + i(!) q 2 (⇡e ) (1 !)(yd ) + (1 ) ⌫(!)

The Jacobian matrix becomes:

0 1
@f1 ⌘ 0( )
@! ! ⌘+1 0
B C
B C
B @f2 @f2 0 (⇡e ) C
B r C
B @! @ q2 ⌫(!) C (22)
B C
B ⇤ C
B @f3 q ! 0( @f3⇤ C
@ @! ⌘+1 1 ! ) @q A

where
✓ ✓ ◆ ◆
@f3⇤ 0 (⇡e ) ⌫(!) (⇡e ) ( ) ↵ (1 )i(!) !(1 )⌘p µ
= q 1+ + ⌘p µ +
@! ⌫(!) ⌘(1 !) ⌘(1 !)⌫(!) (⌘ + 1)(1 !)2 (⌘ + 1)(1 !)
 ✓ ◆
⌫(!) @yd (1 ) ⌫(!)
q2 0 (⇡e ) 1 + + yd (1 !) +
⌘(1 !) @! ⌘(1 !)
⇤
✓ ◆ 0
@f3 (⇡e ) ( ) ↵ (1 )i(!) r (⇡e )
= ! (1 )r + i(!) +
@q ⌫(!) (⌘ + 1)(1 !) q⌫(!)
@yd
2q[(⇡e ) (1 !)(yd ) + (1 ) ⌫(!)] r0 (⇡e ) + q 2 (1 !)
@q
(23)

186
APPENDIX B. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE,
STOCK-FLOW CONSISTENT DYNAMICS

Since (⇡e ) ! k0 , 0 (⇡e ) ! 0, i(!) ! ⌘p , Assumption (10) (ii) c(x) ! c+ 8 and

q!0 q!0 !!0 x!+1
Assumption (9) (iii), lim ⇡e2 0 (⇡e ) < +1, the Jacobian matrix becomes:
⇡e ! 1

0 ⌘ 1
⌘+1 ( (0) ↵ + (1 )⌘p ) 0 0
B C
B C
B 1 1/⌘ (0) ↵+(1 )⌘p C
B 0 k0 A b (↵ + n + ) ⌘+1 0 C
B C (24)
B C
B C
@ 1 1/⌘
0 0 k0 A b (1 )r ⌘p A

The (explosive) debt-deflationary steadys state is locally, asymptotically stable if, and only if,:

(0) ↵ + (1 )⌘p  0
✓ ◆1/⌘
1 (0) ↵ + (1 )⌘p
k0 A (↵ + n + ) 0
b ⌘+1 (25)
✓ ◆1/⌘
1
k0 A (1 )r ⌘p  0 ,
b

which is satisfied by a broad range of parameterizations, including our numerical example.

Notice that there exists another debt-deflationary equilibrium, (!; ; df ) = (!6 ; 6 ; +1), where
(!6 ; 6 ) is solution of the following equation:9
8 ✓ ◆ 1
>
< ↵+n+ 1 !6 ⌘
k0 = = (↵ + n + )⌫(!6 ),
A b (26)
>
: 1
6 = (↵ + (1 )i6 )

with i6 = ⌘p (µ!6 1), which, courtesy of (8) (ii), always exist provided i6 > 0.
An educated guess suggests, however, that the first condition is likely to be violated: ↵, n and
should be of the magnitude 10 2 , ⌫(!) ' 3 whereas k0  10 4 (and in our specific investment
function, k0 is even negative).Consequently, this additional debt-deflationary equilibrium is very
unlikely to exist.

Eventually, a deflationary state with an infinite debt ratio, (!, , df ) = (!2 , 0, +1)10 is also a zero
of (21). In the neighborhood of this equilibrium, ⇡8 ! 1, (⇡8 ) ! k0 and yd8 ! k0 + c+ . The
Jacobian matrix can be transformed into a lower triangular matrix whose eigenvalues are easy to
identify. Their sign depends on the selected parameterization, so that the stability of the steady state
cannot be ruled out a priori. With our particular parameterization, the three stability conditions
above are verified. None the less, a steady state with positive wages and zero-employment is hardly
economically meaningful.
8
Notice that ! + rdf + ⇡e ! +1, hence, yd = (⇡) + ch (! + rdf + ⇡e ) ! k0 + c+ > 0. Consequently, on a
path to this explosive debt-deflationry equilibrium, the demand ratio yd can be greater or lower than 1, depending
upon the upper bound c+ . Because production is collapsing anyway, demand Yd = yd Y is following the same path.
9
Remember that df tends to infinity, d˙ ! 0, ⇡e ! 1 and (⇡e ) ! k0 .
10
With ⌘p (µ!2 1) = (0)1
↵
< 0 due to Assumption (8) (ii).

187
APPENDIX B. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE,
STOCK-FLOW CONSISTENT DYNAMICS

B.2 The monetary dynamics

Let us now introduce money and Income velocity as in section 3.4.

B.2.1 The reduced monetary dynamical system

As before, the last variable, dh , of the system (36), can be deduced from (!, , df , m). The same
holds for Income velocity v = yd /m, while the evolution of (!, , df , m) is independent from dh and
v. We therefore focus on the following 4-dimensional system with (!, , df , m) as state variables:

8
> ⌘! h i
>
> !˙ = ( ) ↵ (1 )⌘ p v(µ! 1)
>
> ⌘+1
>
> h (⇡ ) i
< ˙ e !˙
= (↵ + n + ) (27)
> ⌫(!) ⌘!(1 !)
>
> ⇥ ⇤
>
> d˙f = (⇡e ) (1 !)(yd ) + (1 ) ⌫(!) df g(!, , df ) (1 )r + ⌘p v(µ! 1)
>
>
:
ṁ = !(1 yd ) + ⌫(!) m[g(!, , df , m) + ⌘p v(µ! 1)].

New auxiliary variables are:

8
> 1 ⇣1 ! ⌘ 1/⌘
>
> ⌫(!) =
>
> A b
>
>
>
> ⇡ = 1 ! rdf ⌫(!)
>
> e
>
>
>
> yd = ch (! + rdf + ⇡e ) + (⇡e )
>
> ✓ ◆
>
< (⇡e ) !˙
d˙h = ch (! + rdf + ⇡e ) ! rdf ⇡e dh + i(!)
> ⌫ ⌘(1 !)
>
> yd
>
> v=
>
>
>
> m
>
>
>
> i(!, m) = ⌘ p v(µ! 1)
>
>
>
> (⇡e ) !˙
>
:g(!, , df , m) = .
⌫(!) (1 !)⌘
(28)

The Jacobian matrix, J ⇤ (!, , df , m), associated with (27) is:

0 @f ⇤ ⌘ ⌘! (1 )⌘p @yd ⌘!(1 ) ⌘p yd
1
! ⌘+1 0( )
@df (µ! 1) (µ! 1)
1
@! ⌘+1 m ⌘+1 m2
B C
B C
B @f ⇤ @f2⇤ @f2⇤ (1 )⌘p yd (µ! 1) C
B 2 C
B @! @ @df (⌘+1)(1 !)m2 C
B C
B h iC (29)
B @f3⇤ df @f3⇤ yd !(1 ) C
B ! 0( ) df ⌘p m 2 (µ! 1) +1 C
B @! ⌘+1 1 ! @df (1 !)(⌘+1) C
B C
@ ⇤ A
@f4 m ! 0( @f4⇤ @f4⇤
@! ⌘+1 1 ! ) @df @m

188
APPENDIX B. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE,
STOCK-FLOW CONSISTENT DYNAMICS

where
@f1⇤ ⌘ h i ⌘! ⌘p h @yd i
= ( ) ↵ (1 )⌘p v(µ! 1) (1 ) yd µ + (µ! 1)
@! ⌘+1 ⌘+1 m @!
✓ 0 ✓ ◆
@f2⇤  (⇡e ) ⌫(!) (⇡e ) ( ) ↵ (1 )⌘p v(µ! 1)
= 1+ + +
@! ⌫(!) ⌘(1 !) ⌘(1 !)⌫(!) (⌘ + 1)(1 !)2
h i ◆
(1 )⌘p @yd
yd µ + (µ! 1)
(⌘ + 1)(1 !)m @!
✓ 0 ✓ ◆ 
@f3⇤  (⇡e ) ⌫(!) (⇡e ) ⌘p @yd ( ) ↵ (1 )⌘p v(µ! 1)
= df 1+ + yd µ + (µ! 1) +
@! ⌫(!) ⌘(1 !) ⌘(1 !)⌫(!) m @! (⌘ + 1)(1 !)2
 ◆ ✓ ◆
!(1 ) ⌘p @yd ⌫(!) @yd (1 ) ⌫(!)
yd µ + (µ! 1) 0 (⇡e ) 1 + + yd (1 !) +
(⌘ + 1)(1 !) m @! ⌘(1 !) @! ⌘(1 !)
✓ 0 ✓ ◆
@f4⇤ @yd ⌫(!)  (⇡e ) ⌫(!) (⇡e )
= (1 yd ) + ! + +m 1+ +
@! @! ⌘(1 !) ⌫(!) ⌘(1 !) ⌘(1 !)⌫(!)
  ◆
( ) ↵ (1 )⌘p v(µ! 1) !(1 ) ⌘p @yd ⌘p @yd
+ y d µ + (µ! 1) y d µ + (µ! 1)
(⌘ + 1)(1 !)2 (⌘ + 1)(1 !) m @! m @!
(30)

0
@f2 (⇡e ) ( ) ↵ (1 )⌘p v(µ! 1) ( )
= (↵ + n + )
@ ⌫(!) (⌘ + 1)(1 !) (⌘ + 1)(1 !)
⇤ 0
@f2  (⇡e ) (1 )⌘p @yd (µ! 1)
= r
@df ⌫(!) m(⌘ + 1) @df (1 !)
⇤
✓ ◆
@f3 @yd (⇡e ) ( ) ↵ (1 )⌘p v(µ! 1)
= (1 )r r0 (⇡e ) (1 !) + ⌘p v(µ! 1) !
@df @df ⌫(!) (⌘ + 1)(1 !)
 0
r (⇡e ) ⌘p @yd !(1 ) ⌘p @yd
+df (µ! 1) (µ! 1)
⌫(!) m @df (⌘ + 1)(1 !) m @df

@f4⇤ @yd @yd r0 (⇡e ) !(1 ) ⌘p @yd
=! ⌘p (µ! 1) + m (µ! 1)
@df @df @df ⌫(!) (⌘ + 1)(1 !) m @df
@f4⇤ !(1 ) ⌘ p yd
= g(!, , df , m) + (µ! 1).
@m (⌘ + 1)(1 !) m
(31)

B.2.2 Long-run equilibria with money

Here is the list of steady states for (27).

– Finite-debt ratio equilibria

– The monetary Solovian equilibrium, (!1⇤ , ⇤ ⇤ ⇤
1 , df1 , m1 ) 2 R4++ , where m⇤1 is uniquely determined
by !1⇤ :11
1 ⇥ ⇤ ⇤
m⇤1 = !1 (1 yd⇤1 ) + ⌫(!1⇤ ) ⌘p yd⇤1 (µ!1⇤ 1) . (32)
↵+n

– A deflationary state with non-zero wage (!2⇤ , 0, d⇤f2 , m⇤2 ) where (!2⇤ , d⇤f2 , m⇤2 ) solves the nonlinear
11
Because yd⇤1 is also only depending on !1⇤ , see Appendix C.1.4.

189
APPENDIX B. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE,
STOCK-FLOW CONSISTENT DYNAMICS

equations:
8
> y⇤ (0) ↵
>i(!2⇤ , m⇤2 ) = ⌘p d2 (µ!2⇤ 1) =
> < 0 due to technical conditions (8) (ii)
>
> m2⇤ 1
>
>
< h (⇡ ⇤ ) i
0 = (⇡2⇤ ) (1 !2⇤ )(yd⇤2 ) + (1 ) ⌫(!2⇤ ) d⇤f2 2
(1 )r + i(! ⇤
2 , m⇤
2 ) ,
>
> ⌫(!2⇤ )
>
> 
>
> (⇡2⇤ )
>
: 0 = !2⇤ (1 yd⇤2 ) ⌫(!2⇤ ) + ⌘p yd⇤2 (µ!2 1) + m⇤2
⌫(!2⇤ )
(33)
⇤ ⇤ ⇤ ⇤ ⇤ ⇤ ⇤
with ⇡2 = 1 !2 rdf2 ⌫(!2 ) and yd2 = (⇡2 ) + ch (!2 + rdf2 + ⇡2 ).
The characteristic polynomial of the Jacobian matrix at this equilibrium is
✓ ◆
(⇡2⇤ )
(↵ + n + ) X P3 (X),
⌫(!2⇤ )

where P3 (X) = X 3 + a2 X 2 + a1 X + a0 , with

a0 = J31 (J13 J44 J43 J14 ) J11 (J33 J44 J43 J34 ) J41 (J13 J34 J33 J14 )
a1 = J44 J33 J34 J43 + J11 (J44 + J33 ) J31 J13 J14 J41
(34)
@fi⇤
a2 = (J11 + J44 + J33 ) with Jij =
@xj |(!⇤ ;0;d⇤ ⇤
2 f2 ,m2 )

According to the Routh-Hurwitz* criterion, this equilibrium is locally stable if, and only if, a3 0,
a2 0, a2 a1 a3 a0 , a0 0. Whether these conditions are satisfied depends upon the selected
parameterization and, in general, local (un)stability cannot be rule out. Notice that, with the
particular parameterization of our numerical example (4.4.2), the second condition, a2 0, is
not verified, so that the debt-deflationary equilibrium turns out to be unstable.
– A deflationary state with zero-employment-wage (0, 0, d⇤f3 , m⇤3 ),
! ⇤ ; ¯ ⇤ ; d¯⇤ ; +1),
– An equilibrium with infinite money ratio (¯ f

– A Trivial Equilibrium (0, 0, 0, m⇤4 ).

– A Slavery Equilibria (0, ⇤5 , d⇤f5 , m⇤5 ) with ⇤
5 > 0.
– Any equilibrium with m = 0, such as (!6 , 6 , df6 , 0).

– Steady states with infinite debt ratio

– Debt-deflationary equilibria with finite money ratio (0, 0, +1, m⇤7 ) or (!9⇤ ; ⇤ ⇤
9 ; +1, m9 ) > 0,
– A Cornucopian deflationary state with infinite debt ratio (!8⇤ , 0, +1, m⇤8 ),
⇤
– Debt-deflationary equilibria with skyrocketing levels of money (!10 ; ⇤10 ; +1, +1) or (0, 0, +1, +1)

Monetary Solovian equilibrium

At the monetary Solovian equilibrium (!1⇤ ; ⇤ ⇤ ⇤
1 ; df1 , m1 ) defined above, the Jacobian matrix is given by:
0 1
J11 J12 J13 J14
B C
B C
B ⇤
1 (K1 K! + K21 )
⇤
1 K22
⇤
1 (rK1 + K23 ) ⇤ C
1 K24 C
B
B C, (35)
B C
BK2 K! + df1 K31 + Ky! J32 rK2 + (1 )r Kyd (↵ + n + i1 ) K3 J34 C
B C
@ A
J41 J42 J43 J44

190
APPENDIX B. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE,
STOCK-FLOW CONSISTENT DYNAMICS

with
⌘!1⇤ ⌘p h ⇤ @yd i ⌘!1⇤
J11 = (1 ) y µ + (µ!1⇤ 1) , J12 = 0
( ⇤
1 ),
⌘+1 m⇤1 d1 @! |(!1⇤ ; ⇤ ;d⇤ ,m⇤ )
1 f1 1
⌘+1
⌘!1⇤ ⌘p @yd yd⇤1
J13 = (1 ) (µ!1⇤ 1) , i⇤1 = ⌘p (µ!1⇤ 1)
⌘+1 m⇤1 @df |(!⇤ ; ⇤ ;d⇤ ,m⇤ ) m⇤1
1 1 f1 1

0 (⇡1⇤ ) ( ⇤1 ) 0
(1 )i1
K1 = , K22 = , K24 =
⌫(!1⇤ ) (⌘ + 1)(1 ⇤
!1 ) (⌘ + 1)(1 !1⇤ )m⇤1
⇤
(⇡1 ) (1 )⌘p h @yd i
K21 = yd µ + (µ!1⇤ 1) , J14 = ⌘!1⇤ K24 (1 !1⇤ ),
⌘(1 !1⇤ )⌫(!1⇤ ) (⌘ + 1)(1 !1⇤ )m⇤1 @! |(!1⇤ ; ⇤1 ;d⇤f1 )
(⇡1⇤ ) ⌘p h @yd i✓ !1⇤ (1 )
◆
⇤
K31 = y d µ + (µ! 1 1) 1 +
⌘(1 !1⇤ )⌫(!1⇤ ) m⇤1 @! |(!1⇤ ; ⇤1 ;d⇤f1 ,m⇤1 ) (⌘ + 1)(1 !1⇤ )
✓ ◆
⌘p !1⇤ (1 ) @yd (1 )⌘p (µ!1⇤ 1) @yd
K3 = ⇤ 1 + ⇤ , K23 =
m1 (⌘ + 1)(1 !1 ) @df |(!⇤ ; ⇤ ;d⇤ ,m⇤ ) (⌘ + 1)(1 !1⇤ )m⇤1 @df |(!⇤ ; ⇤ ;d⇤ ,m⇤ )
1 1 f1 1 1 1 f1 1

@yd (1 ) ⌫(!1⇤ ) @yd

Ky! = yd⇤1 (1 !1⇤ ) + , Kyd = (1 !1⇤ )
@! |(!1⇤ ; ⇤ ;d⇤ ,m⇤ )
1 f1 1
⌘(1 !1⇤ ) @df |(!⇤ ; ⇤ ;d⇤ ,m⇤ )
1 1 f1 1
✓ ◆
d⇤f1 i1 !1⇤ (1 )
J32 = d⇤f1 !1⇤ K22 , J34 = 1+ , J42 = m⇤1 !1⇤ K22
m⇤1
(⌘ + 1)(1 !1⇤ )
@yd ⌫(!1⇤ )
J41 = 1 + yd⇤1 + !1⇤ + + m⇤1 (K1 K! + K31 ) , K2 = d⇤f1 K1 0 (⇡1⇤ )
@! |(!1⇤ ; ⇤1 ;d⇤f1 ,m⇤1 ) ⌘(1 !1⇤ )
✓ ◆
!1⇤ (1 )⌘p @yd
J43 = m⇤1 rK1 + !1⇤ ⌘p (µ!1⇤ 1) (µ! ⇤
1 1)
(⌘ + 1)(1 !1⇤ ) @df |(!⇤ ; ⇤ ;d⇤ ,m⇤ )
1 1 f1 1

(36)
✓ ◆
!1⇤ (1 )⌘p yd⇤1 ⌫(!1⇤ )
J44 = (↵ + n) + (µ!1⇤ 1) , K! = 1+ (37)
(⌘ + 1)(1 !1⇤ )m⇤1 ⌘(1 !1⇤ )

The characteristic polynomial is P (X) = X 4 + a3 X 3 + a2 X 2 + a1 X + a0 , where

a4 = 1
⇤
a3 = 1 K22 J11 J33 J44
⇤
a2 = J33 J44 J43 J34 + J31 J13 J41 J14 + (J33 + J44 )(J11 1 K22 )
⇤
+ 1 [J32 (rK1 + K23 ) + K24 J42 K22 J11 + J12 (K1 K! + K21 )]
⇤
a1 = 1 [(J13 J32 + J14 J42 (J33 + J44 )J12 )(K1 K! + K21 ) + K22 [J11 (J33 + J44 ) J14 J41 J13 J31 ]
+ K24 (J42 J33 J32 J43 J41 J12 + J11 J42 ) + (J32 J44 J11 J32 + J42 J34 + J31 J12 )(rK1 + K23 )] (38)
⇤
(J11 1 K22 )(J33 J44 J34 J43 ) J31 (J13 J44 J43 J14 ) + J41 (J13 J34 J33 J44 )
⇤
a0 = 1 [K22 (J31 (J13 J44 J43 J14 ) J11 (J33 J44 J43 J34 ) + J41 (J33 J14 J13 J34 ))
+ K24 (J11 (J32 J43 J42 J33 ) + J31 (J42 J13 J12 J43 ) + J41 (J12 J33 J32 J13 ))
+ (K1 K! + K21 )(J12 (J33 J44 J43 J34 ) + J42 (J13 J34 J33 J14 ) J32 (J13 J44 J43 J14 ))
+ (rK1 + K23 )(J31 (J42 J14 J12 J44 ) + J11 (J32 J44 J42 J34 ) + J41 (J12 J34 J32 J14 ))]

with J31 = K2 K! + df1 K31 + Ky! and J33 = rK2 + (1 )r Kyd ) (↵ + n + i1 ) K3

The Routh-Hurwitz Stability criterion leads to the following local, asymptotic stability criterion: a3 0,
a3 a2 a4 a1 0, (a3 a2 a4 a1 )a1 a0 a23 > 0 and a0 0 .

191
APPENDIX B. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE,
STOCK-FLOW CONSISTENT DYNAMICS

The parameterization of our numerical example (4.4.2) yields:

(!1⇤ , ⇤ ⇤ ⇤ ⇤ ⇤ ⇤
1 , df1 , m1 , ⇡1 , i1 , yd1 ) = (0.791, 0.985, 0.059, 0.011, 0.177, 0.078, 0.980) ,
(39)
we obtain (a3 , a2 , a1 , a0 ) = (5.954, 42.862, 44.882, 8.547) ,
which satisfy the stability criterion.

Other finite-debt ratio equilibria

In the same way as in section C.1.4, one can exhibit several other equilibria with finite debt ratio.
1)(!, , df , m) = (0, 0, d⇤f3 , m⇤3 ), with (d⇤f3 , m⇤3 ) solving the nonlinear equations:
h (⇡ ⇤ ) i
(⇡3⇤ ) yd⇤3 + + (1 ) ⌫(0) d⇤f3 3
(1 )r + i(0, m⇤3 ) = 0
⌫(0)
 (40)
(⇡3⇤ )
m⇤3 = ⌫(0) + ⌘p yd⇤3
⌫(0)
where ⇡3⇤ = 1 rd⇤f3 ⌫(0) and yd⇤3 = (⇡3⇤ ) + ch (rd⇤f3 + ⇡3⇤ ).
y⇤
Here again, we have i(0, m⇤3 ) = ⌘p md⇤3 < 0, which motivates the name of this deflationary equilibrium.12
3
The characteristic polynomial of the Jacobian matrix at the equilibrium point (0, 0, d⇤f3 , m⇤3 ) is:
0 2 y⇤
31
✓ ◆ ⇤ (0) ↵ + (1 )⌘p md⇤3
⌘ (⇡ )
X [ (0) ↵ + (1 )i(0, m⇤3 )] @X 4 3
(↵ + n + ) 3 5A
P2 (X)
⌘+1 ⌫(0) ⌘+1
(41)
@fi⇤
where P2 (X) = X 2 (J44 + J33 )X + J44 J33 J34 J43 , with Jij = @xj |(0;0;d⇤ ,m⇤ ) . Under the assumptions
f 33
adopted in section C.1.2, the signs of the roots of this polynomial depend on the selected parameterization
of our economy, so that neither stability, nor instability can be a priori ruled out. With the specific pa-
rameterization of our numerical example, the two roots of P2 (X) are positive: this equilibrium is locally
unstable.

2) The change of variable, n = 1/m, enables to verify that (¯ ! ⇤ ; ¯ ⇤ ; d¯⇤f ; +1) is an equilibrium for the
system, with ¯ =
⇤ 1
¯e and d¯f uniquely defined by !
(↵), ⇡ ⇤ ⇤
¯ ⇤ 13
and where ! ¯ ⇤ solves the nonlinear equation:

⇡e⇤ )
0 = (¯ (1 ¯ ⇤ )(ȳd⇤
! ) + (1 !⇤ )
) ⌫(¯ d¯⇤f [↵ + n (1 )r].
Now, the fourth line of the Jacobian matrix contains the following elements:
✓ 0 ✓ ◆
@fn⇤  (⇡e ) ⌫(!) (⇡e ) ( ) ↵ (1 )⌘p v(µ! 1)
=n 1+ + +
@! ⌫(!) ⌘(1 !) ⌘(1 !)⌫(!) (⌘ + 1)(1 !)2
 ✓ ◆◆
⌘p @yd !(1 )
yd µ + (µ! 1) 1 +
m @! (⌘ + 1)(1 !)
 ✓ ◆
2 @yd ⌫(!) @yd
+ n (1 yd ) ! + ⌘ p yd µ + (µ! 1)
@! ⌘(1 !) @!
(42)
@fn⇤ ! 0( )
= n
@ (n + 1)(1 !)
⇤
 
@fn 2 @yd @yd 0 (⇡e )
=n ! + ⌘p (µ! 1) + n r
@df @df @df ⌫(!)
⇤

@fn !(1 )⌘p yd (µ! 1)
= 2n[!(1 yd ) ⌫(!) + ⌘p yd (µ! 1)] + g(!, , df , 1/n) + n .
@n (⌘ + 1)(1 !)
12
The deflationary nature of this equilibrium holds as long as m⇤3 > 0, i.e. (⇡3 ) > ⌫(0), which means a non-
negative growth rate of the economy fueled by firms’ debts.
⇤
1 ⇡
¯e ¯⇤
! ¯⇤)
⌫(!
13
Indeed (¯ ! ⇤ )(↵ + n + ) and d¯⇤f =
⇡e⇤ ) = ⌫(¯ r
.

192
APPENDIX B. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE,
STOCK-FLOW CONSISTENT DYNAMICS

! ⇤ ; ¯ ⇤ ; d¯⇤f ; 0), with the exception of the last

Each element of this fourth line tends to 0 when (!; ; df ; n) ! (¯
@f ⇤
one: @nn ! ↵ + n > 0. This last element is a positive root of the associated characteristic polynomial, which
implies that this equilibrium is locally unstable.

Trivial and Slavery steady states

The trivial Equilibrium is structurally unstable as before. More precisely, (!, , df , m) = (0, 0, 0, m⇤4 ) with
m⇤4 6= 0 is no longer an equilibrium due to the introduction of depreciation into the profit equation: d˙ = 0
if, and only if, (⇡4⇤ ) (1 )yd⇤4 + (1 ) ⌫(0) = 0, with ⇡4⇤ = 1 ⌫(0) and yd⇤4 = (⇡4⇤ ) + c( ⇡f⇤4 ), a
14
condition that is structurally unstable.

Similarly, as before, the so-called Slavery state (0, ⇤5 , d⇤f5 , m⇤5 ) is a long-run equilibrium only provided
a structurally unstable condition be satisfied.15 Moreover, both the trivial and the slavery equilibria lack
economic relevance, and will not be studied further.

More generally one can show that there is no structurally stable equilibrium with m = 0. Indeed, let us
assume the existence of an equilibrium of the form (!6⇤ , ⇤6 , d⇤f6 , 0) for the di↵erential system which implies
that:
ṁ = 0 , 0 = !6⇤ (1 yd⇤6 ) + ⌫(!6⇤ ) ⌘p yd⇤6 (µ!6⇤ 1)
(43)
, !6⇤ (1 yd⇤6 ) + ⌘p yd⇤6 (µ!6⇤ 1) = ⌫(!6⇤ )
y⇤
which can be uniquely solved by !6⇤ 6= 0. This condition implies that ⇤6 = 1
(↵ + (1 )⌘p md⇤6 (µ!6⇤ 1))
6
so that !˙ = 0. Because m⇤6 ! 0, this means that ⇤6 ! 1 according to hypothesis (8) (iii). Let us take:

⇡6⇤ )
(¯ = (↵ + n + )⌫(!6⇤ ), (44)

so that ˙ = 0 in the second equation of (27). We know that equation (44) can be uniquely solved for ⇡6⇤
because of condition (9), which implies that d⇤f6 = (1 ⇡6⇤ !6⇤ ⌫(!6⇤ ))/r is uniquely determined by
˙
!6 . Consequently, in order to guarantee d = 0 in the third equation of (27), we need: 0 = (⇡6⇤ ) (1
⇤

!6⇤ )(yd⇤6 ) + (1 ) ⌫(!6⇤ ) d⇤f6 [↵ + n (1 )r], which leads to a second nonlinear equation that !6⇤
⇤
must satisfy — this over-determination of !6 makes very such equilibrium (if any) structurally unstable since
any infinitesimal change in the coefficients of, say, the first equation will prevent !6⇤ from solving the second,
and vice versa.

The Debt-deflationary equilibrium

To see that (!, , df , m) = (0, 0, +1, +1) is an equilibrium, let use the same change of variable q := 1/df
as in [Grasselli and Costa Lima, 2012], and n := 1/m. Observe that (!, , q, n) = (0, 0, 0, 0) is an equilibrium
point of the transformed system:
8 ⌘! h i
>
> !
˙ = ( ) ↵ (1 )⌘ p v(µ! 1)
>
> ⌘+1
>
>
>
> h i
< ˙ = (⇡e ) !˙
(↵ + n + )
⌫(!) ⌘!(1 !) (45)
>
> h i h i
>
> 2
>
> q̇ = q g(!, , 1/q, 1/n) (1 )r + +⌘p v(µ! 1) q (⇡e ) (1 !)(yd ) + (1 ) ⌫(!)
>
>
:
ṅ = n2 [!(1 yd ) ⌫(!) + ⌘p yd (µ! 1)] + n[g(!, , 1/q, 1/n)].

The Jacobian matrix of (45) is

14
No equilibrium can satisfy m⇤4 = 0. In that case, indeed, ṁ = ⌫(0) + ⌘p yd⇤4 > 0. The monetary ratio, m, tending
to 0 would imply an inflation rate plummeting to 1, which lacks any economic meaning.
15
Here again, the model parameters must indeed satisfy a binding condition which makes this equilibrium struc-
⌫(0)(↵+n+ ) y ⇤ + +(1 ) ⌫(0) ⌫(0)+⌘p y ⇤
turally unstable: 1 rd⇤f5 ⌫(0) =  1 (⌫(0)(↵+n+ )) where d⇤f5 = ↵+n (1
d5
)r+i(0,m⇤
, m⇤5 = d5
5) ↵+n
⇤ ⇤ ⇤ ⇤
and yd5 = ⌫(0)(↵ + n + ) + ch (rdf5 + ⇡5 ). This implies that df5 must solve simultaneously two distinct equations.

193
APPENDIX B. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE,
STOCK-FLOW CONSISTENT DYNAMICS

0 @f ⇤ 1
@!
1 ⌘
! ⌘+1 0
( ) ⌘!
⌘+1 (1 )⌘p n @y
@q (µ!
d
1) ! ⌘(1 )
⌘+1 ⌘p yd (µ! 1)
B C
B ⇣ ⌘ C
B @f ⇤ @f2⇤ r0 (⇡e ) (1 )⌘p n @yd (µ! 1) (1 )⌘p yd (µ! 1) C
B 2 + C
B @! @ q 2 ⌫(!) (⌘+1) @q (1 !) (⌘+1)(1 !) C
B C
B h i C (46)
B @fq⇤ @fq⇤ C
B q ! 0
( ) q !(1 )
+ 1 ⌘p yd (µ! 1) C
B @! ⌘+1 1 ! @q (1 !)(⌘+1) C
B C
@ ⇤ ⇤ ⇤ ⇤
A
@fn @fn @fn @fn
@! @ @q @n

where
✓ ✓ ◆ 
@fq⇤ 0 (⇡e ) ⌫(!) (⇡e ) ⌘p @yd
= q 1+ + yd µ + (µ! 1)
@! ⌫(!) ⌘(1 !) ⌘(1 !)⌫(!) m @!
 ◆
( ) ↵ (1 )⌘p v(µ! 1) !(1 ) ⌘p @yd
+ y d µ + (µ! 1)
(⌘ + 1)(1 !)2 (⌘ + 1)(1 !) m @!
 ✓ ◆
⌫(!) @yd (1 ) ⌫(!)
q2 0 (⇡e ) 1 + + yd (1 !) +
⌘(1 !) @! ⌘(1 !)
✓ ◆
@fq⇤ (⇡e ) ( ) ↵ (1 )⌘p v(µ! 1) r0 (⇡e ) @yd
= (1 )r + ⌘p v(µ! 1) ! + ⌘p n (µ! 1)
@q ⌫(!) (⌘ + 1)(1 !) q⌫(!) @q
!(1 )⌘p n @yd @yd
(µ! 1) 2q[(⇡e ) (1 !)(yd ) + (1 ) ⌫(!)] r0 (⇡e ) + q 2 (1 !)
(⌘ + 1)(1 !) @q @q
✓ 0 ✓ ◆
@fn⇤  (⇡e ) ⌫(!) (⇡e )
=n 1+ +
@! ⌫(!) ⌘(1 !) ⌘(1 !)⌫(!)
 ✓ ◆◆
( ) ↵ (1 )⌘p v(µ! 1) ⌘p @yd !(1 )
+ y d µ + (µ! 1) 1 +
(⌘ + 1)(1 !)2 m @! (⌘ + 1)(1 !)
 ✓ ◆
@y d ⌫(!) @y d
+ n2 (1 yd ) ! + ⌘ p yd µ + (µ! 1)
@! ⌘(1 !) @!
@fn⇤ ! 0( )
= n
@ (n + 1)(1 !)
 
@fn⇤ 2 @y d @y d r0 (⇡e ) !(1 )⌘p n @yd
=n ! + ⌘p (µ! 1) + n 2
(µ! 1)
@q @q @q q ⌫(!) (⌘ + 1)(1 !) @q

@fn⇤ !(1 )⌘p yd (µ! 1)
= 2n[!(1 yd ) ⌫(!) + ⌘p yd (µ! 1)] + g(!, , df , 1/n) + n
@n (⌘ + 1)(1 !)
(47)
Observe (!, , q, n) = (0, 0, 0, 0) is an equilibrium point of (45), which implies that (⇡e ) ! k0 , 0 (⇡e ) ! 0,
q!0 q!0
i(!) ! 0. Consequently, because of condition (10) (ii) c(x) ! c+ 16 and condition 9 (iii), lim ⇡e2 0 (⇡e ) <
!!0 x!+1 ⇡e ! 1
+1, the matrix (46) becomes:
16
One can notice here that yd = (⇡) + ch (! + rdf + ⇡e ) ! k0 + c+ (because ! + rdf + ⇡e ! +1), a finite
positive value. Consequently, on a path to the infinite-valued equilibrium, the demand ratio yd can be greater or
lower than 1 according to the upper bound c+ and, because production is collapsing, demand Yd = yd Y is following
the same path.

194
APPENDIX B. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE,
STOCK-FLOW CONSISTENT DYNAMICS

0 ⌘ 1
⌘+1 ( (0) ↵) 0 0 0
B C
B C
B 0 k0
(↵ + n + ) (0) ↵
0 0 C
B ⌫(0) ⌘+1 C
B C
B C
B k0 C (48)
B 0 0 ⌫(0) (1 )r 0 C
B C
B C
B C
@ 0 0 0 k0 A
⌫(0)

Local, asymptotic stability obtains if, and only if,:

(0) ↵0,
k0 (0) ↵
(↵ + n + ) 0,
⌫(0) ⌘+1
k0 (49)
(1 )r  0 ,
⌫(0)
k0
0.
⌫(0)

which is satisfied with our parameterization.

Other equilibria with infinite debt ratio

h i
(⇡7⇤ )
1) A second debt-deflationary equilibrium exists: (!, , df , m) = (0, 0, +1, m⇤7 ) with m⇤7 ⌫(0) =
⌫(0) + ⌘p yd⇤7 . Because df ! +1, we have ⇡7⇤ ! 1. Consequently, our specification of the increasing
investment function (⇡7⇤ ) ! k0 < 0 implies m⇤7 < 0 (economically meaningless). The Jacobian matrix with
the change of variable q = 1/df is a lower triangular matrix at this equilibrium point, from which we can
immediately obtain the four eigenvalues:
y
yd k0 (0) ↵ + (1 )⌘p md⇤7
(0) ↵ + (1 )⌘p ⇤7 , (↵ + n + ) 7

m7 ⌫(0) (⌘ + 1) (50)
⇤ k0
J3,3 (!, , q, m)|(0;0;0;m⇤7 ) , +
⌫(0)

With this specific investment function, the last eigenvalue is positive which implies that this equilibrium is
locally unstable.

2) There exists also a deflationary state with infinite-debt level, (!, , df , m) = (!8⇤ , 0, +1, m⇤8 ), where
(!8⇤ , m⇤8 ) must be solution of the following system:
8 yd⇤8
> ⇤ ⇤ (0) ↵
> i(!
< 8 8 , m ) = ⌘ p ⇤ (µ!8⇤ 1) = < 0 due to technical conditions (8) (ii)
m8 1
 (51)
>
> (⇡8⇤ )
: 0 = !8⇤ (1 yd⇤8 ) ⌫(!8⇤ ) + ⌘p yd⇤8 (µ!8⇤ 1) + m⇤8
⌫(!8⇤ )

In the neighborhood of this equilibrium, ⇡8⇤ ! 1, (⇡8⇤ ) ! k0 and yd⇤8 ! k0 + c+ . As for deflationary
states with finite-debt levels, one can transform the Jacobian matrix into a lower triangular matrix leading
to the identification of its eigenvalues. Their sign depends on the selected parameterization and cannot be
determined a priori. With the particular parameterization of our numerical example, the first three stability
condition are still verified, but the last one is not, which means that this equilibrium is locally unstable.

195
APPENDIX B. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE,
STOCK-FLOW CONSISTENT DYNAMICS

3) Due to the introduction of a CES production function, another debt-deflationary equilibrium appears,
namely (!; ; df ; m) = (!9⇤ ; ⇤9 ; +1, m⇤9 ) or (!9⇤ ; ⇤10 ; +1, +1), where (!9⇤ ; ⇤9 ) solves the following equations:
8 ✓ ◆ ⌘1
>
< ↵+n+ 1 !9⇤
k0 = = (↵ + n + )⌫(!9⇤ ),
A b (52)
>
: ⇤ 1
9 = (↵ + (1 )i9 ) ,
y⇤
where df ! 1, d˙ ! 0, ⇡e ! 1 and (⇡e ) ! k0 , and with i⇤9 = ⌘p md⇤9 (µ!9⇤ 1).17
9

As shown above, the ṁ - equation of the system can be solved either with

!9⇤ (1 yd⇤9 ) + ⌫(!9⇤ ) + ⌘p yd⇤9 (µ!9⇤ 1)

m⇤9 =
↵+n
or with m ! +1, which implies i(!, m) ! 0 and consequently, ⇤10 = 1
(↵). The first condition in
(52), however, is likely to be empirically violated because ↵, n and are close to 10 2 , ⌫(!) ' 3 whereas
k0  10 4 .18 Consequently, we will not further consider this equilibrium.

17
Courtesy of (8) (ii), a solution to (52) exists provided i⇤9 0.
18
For the specific investment function of our numerical example, k0 is even negative.

196
APPENDIX B. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE,
STOCK-FLOW CONSISTENT DYNAMICS

SUPPLEMENTARY MATERIALS

B.3 Normalization of the Philips’ Curve

In this section, we just give a quick taste of the impacts that an other definition of the Philips’ Curve could
have on the original model [Grasselli and Costa Lima, 2012]. In order to avoid ! > 1, one can choose e.g. a
normalized Phillips curve to constraint ! 2 [0, 1].

In [Grasselli and Costa Lima, 2012], the Philips’ Curve, ( ), is defined by:

1
( )= 0 +
(1 )2
.
We here propose a normalization such as the di↵erential system becomes

!˙ = ! ((1 !) ( ) ↵)
✓ ◆
1 ! (53)
˙ = ↵ .
⌫

and we obtain the following dissipative phenomenon:

Figure B.1: Goodwin with Normalized Philips Curve !0 = 0.8 and 0 = 0.9

B.3.1 Properties
Using the Lyapunov function previously associated to the system,

Z
(s) ! 1
H(!, ) = ds ↵ ln( ) + ( ↵ ) ln(!) , (54)
0
s ⌫ ⌫

197
APPENDIX B. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE,
STOCK-FLOW CONSISTENT DYNAMICS

and calculating, with the new system (!˙ ; ˙ ), dH

dt ˙ ˙ ) leads to:
= rH · (!,
dH @H @H ˙
= · !˙ + ·
dt @! @

✓ ◆
1 1 1
= ( (↵ + + )) · ! ((1 !) ( ) ↵)
⌫ ⌫ !

✓ ◆ ✓ ◆
( ) ↵ 1 !
+ · (↵ + + )
⌫
(55)
✓ ◆
! 1
= ( (↵ + + )) ( !) ( )
⌫ ⌫

˙
= ! ( )

✓ ◆
1 !
= ↵ ! ( ),
⌫
which means that the conservative behaviour of the system depends on the +/- sign of this last equation.
We first study the sign of ( )
1
( )= 0 + (56)
(1 )2
0.04 0.043
with 0 = 1 0.042 and 1 = 1 0.042

Consequently, for each 0.96 we have ( ) 0

Assuming 0.96, we study the function f (defined for x 2 [0; 1]):

✓ ◆
1 x
f (x) = ↵ x ( ) (57)
⌫
which is a polynomial with two roots :
⇢
x1 = 0
(58)
x2 = 1 ⌫(↵ + + )⇠
= 0.835
We then study the sign of the derivative of f
✓ ◆
1 x x
f˙(x) = (↵ + + ) ( ) (59)
⌫ ⌫
1 ⌫(↵ + + )
) f˙(x) 0,x (60)
2
Thus, it implies the following evolution of f :
This means that the derivative of the potential function H(!, ) (Lyapunov function associated to the sys-
tem) is negative for all !p 2]x2 ; 1] and positive otherwise, when 0.96.

Assuming now  0.96, which implies ( )  0, as a consequence :

✓ ◆
1 x x
f˙(x) = (↵ + + ) ( ) (61)
⌫ ⌫
1 ⌫(↵ + + )
) f˙(x) 0,x (62)
2

198
APPENDIX B. MONEY VELOCITY IN AN IMPERFECTLY COMPETITIVE,
STOCK-FLOW CONSISTENT DYNAMICS

Figure B.2: Evolution of the function f for = 0.98

Thus, it implies the following evolution of f :

Figure B.3: Evolution of the function f for = 0.95

This means that the derivative of the potential function H(!, ) (Lyapunov function associated to the system)
is negative for all !p 2 [0; x2 [ and positive otherwise, when  0.96. Thus, the system has a dissipative
behavior in general, except for the value of !p 2 [x2 ; 1] when  0.96 and value of ! 2 [0; x1 [ when 0.96,
which seems strange (we do not observe expansive behavior on the graph between these values) What could
be sumed up in the following figure representing the di↵erent behaviors of the system according to value of
(!; ) :

Figure B.4: Evolution of the function f for = 0.95

[Grasselli and Nguyen Huu, 2016]

199
Appendix C

Directed Technical Change in an

inventory Stock-flow Consistent
Dynamics with Government: from
dirty to clean Capital

C.1 Stability of equilibria in the transition model with government

This section analyzes the local, asymptotic stability of long-run steady states of the dynamical
system (33).

C.1.1 The reduced dynamical system

Notice first that the behavior of the households’ debt ratio, dh , in (33), can be deduced from the
other variables while the latter are independent from dh .

Based on the observation of actual french government’s interest rates, we consider rg close to 0,
in order to simplify the model. This implies that consumption only depends now on wage share and
firm’s debt ratio c(!, rdf ), and that dg becomes an auxiliary variable of the di↵erential system, as
no other variable depends on it.

As in [Costa Lima and Grasselli, 2014], “as long as the dynamics for government expenditures
does not depend explicitly on the level of government debt, equation (48) can be solved separately
first and then used to solve equation (49), and government debt ratio converges exponentially fast
with rate ” (⇡
⌫✓
e)
+ i rg , to the equilibrium value:
8
>
> ḡe + ḡ ⌧¯ (⇡e )
>
< if A = + i rg > 0
A ⌫✓
d¯g = (1)
>
> +1 if A < 0 or A = 0 and ḡe + g1 ⌧1 > 0
>
:
0 if A = 0 and ḡe + g1 ⌧1 < 0.

200
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Moreover, with the same method as in [Costa Lima and Grasselli, 2014], using the di↵erential
system and that: ⇡e = 1 ! rdf + g ⌧ E ⌫1 ", we can write:

⇡˙ e = !˙ rd˙f + ġ ⌧˙ pE ⌫1 "˙
✓ ◆
"˙ ⌫
= ! ( ) (1 )i ↵
µ"
 ✓ ◆
(⇡e ) (2)
r (⇡e ) (1 !)yd + E ⌫1 " g + ⌧ + df r + i + ( )
⌫✓
✓ ◆ ✓ ◆
(⇡e ) (⇡e )
g +i ⇥(⇡f ) + ⌧ +i pE ⌫1 "˙
⌫✓ ⌫✓

✓ ◆
(⇡e )
) ⇡˙ e = !( ( ) (1 )i ↵) + ( ) ⇥(⇡f ) + [1 ! ⇡e E ⌫1 "] +i
⌫✓
✓ ◆ (3)
⌫
r [(⇡e ) ⇡e + (1 !)(1 yd )] + "˙ ! pE ⌫ 1 .
µ"

Thus, the previous system reduces to the following five-dimensionnal system1 in which we consider
ge quickly convergent to its equilibrium value ḡe :
8
> !˙ ⌫
>
> = ( ) (1 )i ↵ "˙
>
> ! µ "
>
>
>
> ˙
> = (⇡e ) ↵ n
>
>
⌫
"˙
>
> ⌫✓ µ"
>
> ✓ ◆
>
>
>
<⇡˙ e = !( ( ) (1 (⇡e )
)i ↵) + ( ) ⇥(⇡f ) + [1 ! ⇡e E ⌫1 "] +i
⌫✓ (4)
>
> ✓ ◆
>
> ⌫
>
> r [(⇡e ) ⇡e + (1 !)(1 yd )] + "˙ ! pE ⌫ 1
>
> µ"
>
> ✓ ◆
>
> (⇡e ) µ" 1
>
> "˙ =
>
> ⌫ ⌫✓ ⌫1 ⌫2
>
>
>
: ˙✓ = ✓(1 ✓)( E ⌫),
1
It will be enough to characterize the equilibria in which the economy either prospers or collapses. But we have
to notice that when working with the reduced system, we can not recover df , g and ⌧ separately, but rather the
combination rdf g + ⌧ .

201
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

with the following auxiliary variables:

8
>
> rdf g + ⌧ = 1 ! ⇡e E ⌫1 "
>
>
>
> 1 1 ✓ ✓
>
> = +
>
> ⌫✓ ⌫1 ⌫2
>
>
>
> 1 " "
>
> µ" = +
>
> ⌫1 ⌫2
>
>
>
> 1 1
>
< ⌫=
⌫1 ⌫2 (5)
>
> yd = c(!, df ) + (⇡e ) + ge
>
>
>
> ✓ ◆
>
> (⇡e )
>
>
> d˙h = c(!, df ) ! rdf dh +i
>
> ⌫✓
>
> ✓ ◆
>
> (⇡e )
>
> d˙g = ge + g ⌧ + dg rg + i
>
> ⌫✓
>
>
:
⇡e ⇡f = (1 !)(1 yd ),
⇣ ⌘
E (!, ,⇡,g,⌧,GE ,Y ) (⇡e )
while assuming that g˙e = Y ge ⌫✓ quickly converges to a finite equilibrium
value ḡe .

Considering ge at its equilibrium value, the associated Jacobian matrix is:

0 ⇣ ⌘ 1
̇(⇡e ) µ"
@f1
@! ! ˙( ) ! µ" ⌫✓
1
⌫ 1 ⌫2 ! (⇡ e) ⌫
µ2" ⌫1 ⌫2
! ⌫(⇡e )
B C
B C
B C
B 0 @f2 ̇(⇡e ) (⇡e ) ⌫
0 C
B @ µ" ⌫ 1 ⌫ 2 µ2" ⌫1 ⌫2 C
B C
B C
B @f3 ˙ C
B @!
B ( ) ! ˙( ) @f3
@⇡e
@f3
@"
@f3
@✓
C
C (6)
B C
B ⇣ ⌘ C
B ̇(⇡e ) µ" 1 (⇡e ) C
B 0 0 ⌫ ⌫✓ ⌫1 ⌫2 ⌫✓ (⇡e )µ" C
B C
B C
B C
@ 0 0 0 0 (1 2✓)( E ⌫)A

202
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Where:
@f1 "˙ ⌫
= ( ) ↵ (1 )i !(1 )⌘p m
@! µ"
@f2 (⇡e ) ⌫
= ↵ n "˙
@ ⌫✓ µ"

✓ ◆
@f3 "˙ ⌫ @⇡f ˙
= ( ) ↵ (1 )i + !⌘p (1 )m ⇥(⇡f )
@! µ" @!
✓ ◆ ✓ ◆
(⇡e ) @yd
+ (1 ! ⇡e E ⌫1 ")⌘p m + i + r 1 yd + (1 !)
⌫✓ @!
✓ ◆ 
@f3 @⇡f ˙ ̇(⇡e ) (⇡e ) @yd
= ⇥(⇡f ) + (1 ! ⇡e E ⌫1 ") +i r ̇(⇡e ) 1 (1 !)
@⇡e @⇡e ⌫✓ ⌫✓ @⇡e
✓ ◆
! ̇(⇡e ) µ" 1
+
µ" ⌫✓ ⌫1 ⌫2
✓ ◆ ✓ ◆ ✓ ◆
@f3 (⇡e ) ⌫ 2 ! ⌫ (⇡e )
= pE ⌫ 1 + i + "!˙ pE ⌫ 1
@" ⌫✓ µ" µ" ⌫✓
✓ ◆
@f3 ! ⌫
= pE ⌫1 (⇡e )µ" (1 ! ⇡e E ⌫1 ")(⇡e ) ⌫
@✓ µ"
(7)

C.1.2 Existence of long-run equilibria

Let us adopt the following set of weak assumptions:
– The (Phillips) function 2 is continuously di↵erentiable on [0, 1] and
0
(i) ( ) > 0 on [0; 1]
(ii) (0) < ↵ + (1 )i(0) (8)
(iii) lim ( ) = +1,
!1

– The investment function (·) is increasing, continuously di↵erentiable on R, and satisfies:3

(i) 0 (⇡e ) > 0 on [0; 1]
(ii) lim (⇡e ) = 0 (9)
⇡e !0
(iii) lim ⇡ 2 0 (⇡e ) < +1
⇡e ! 1 e

– The consumption function c(·) is strictly increasing and continuously di↵erentiable on R+ ,

and verifies4 :
(i) 8 x  0, c(x) = c > 0
(10)
(ii) lim c(x) = c+ ,
x!+1
2
Its form and parameterization guarantee that 8 t, 0  (t)  1, as in [Desai et al., 2006]. Observe also that (ii)
implies (0) < ↵ because i(0) = ⌘p < 0.
3
These restrictions are sufficient to guarantee the existence of the relevant equilibria alluded to in section 4.3, as
in [Grasselli and Costa Lima, 2012].
4
Condition (i) expresses the fact that households need a minimum level of subsistence consumption even at negative
income, whereas (ii) says that total normalized consumption can be higher than 1 (and consequently exceed total
output, by sourcing from the stock of inventories) but must have an upper bound. See [Giraud and Grasselli, 2019]
for details.

203
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

– The tax function ⇥(·) is increasing, continuously di↵erentiable on R and satisfies:

(i) 8 x 0 ⇥0 (x) > 0

(ii) lim ⇥(x) = 0 (11)
x!0
2 0
(iii) lim x ⇥ (x) < +1 ,
x! 1

– the subsidies function (·) is decreasing, continuously di↵erentiable on [0; 1] and satisfies:
0
(i) 8 x 0 (x) < 0
(ii) lim (x) = 0 >0 (12)
x!0
(iii) lim (x) = 0.
x!1

Our Leading Example 4.3 complies with this set of assumptions.

C.1.3 Analysis of the di↵erent equilibria

The last equation of the system (33) admits only two solutions:✓ = 0 (corresponding to a society
where there is no more investment in green capital) or ✓ = 1 (corresponding to a society investing
only in green capital).
1. ✓0 = 0, solution for the fifth equation: ✓˙ = ✓(1 ✓)( E ⌫), implies ⌫✓ = ⌫1 while
1
"˙ = 0 , µ" = ⌫2 ) " = 1. In this case, we have to study the so-called brown equilibria of the
following system:
8
> !˙
>
> ! = ( ) ↵ (1 )i
>
>
>
>
> ˙
< = (⇡e ) ↵ n
>
⌫1
> ✓ ◆
>
> (⇡e )
>
> ⇡˙ e = !( ( ) (1 )i ↵) + ( ) ⇥(⇡f ) + [1 ! ⇡e pE ⌫1 ] +i
>
> ⌫1
>
>
:
r [(⇡e ) ⇡e + (1 !)(1 yd )] .
(13)
The di↵erential system is a Goodwin-Keen system with government, inventories and prices
and a net profit which is at equilibrium values ⇡ ¯ rd¯f + ḡ ⌧¯ pE ⌫1 .5
¯e = 1 !
2. ✓1 = 1, solution for the fifth equation: ✓˙ = ✓(1 ✓)( E ⌫), implies ⌫✓ = ⌫2 while
"˙ = 0 , µ" = ⌫11 ) " = 0. In this case, the di↵erential system is now a Goodwin-Keen model
with government, inventories and prices, with ⇡e = 1 ! rdf + g ⌧ . So we have to study
the so-called green equilibria of the system:
8
>
> !˙ = !( ( ) ↵ (1 )i)
>
> ✓ ◆
>
> (⇡ )
> e
< ˙ =
>
⌫2
↵ n
✓ ◆ (14)
>
> (⇡e )
>
> ⇡˙e = !( ( ) (1 )i ↵) + ( ) ⇥(⇡f ) + [1 ! ⇡e ] +i
>
> ⌫2
>
>
:
r[(⇡e ) ⇡e + (1 !)(1 yd )].
5
in that case: ⌫✓ = ⌫1 .

204
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

In this case, the di↵erential system is a Goodwin-Keen system with government, inventories
and prices and a net profit which is at equilibrium values ⇡ ¯ rd¯f + ḡ ⌧¯.
¯e = 1 !

Both di↵erential system (13) and (14) admits 7 equilibria:

1. a Solovian-like equilibrium (!1 , 1 , ⇡e1 ) which is locally stable with our Leading Example
2. a debt-deflationary equilibrium (0, 0, 1) which is locally unstable with our Leading Example
3. a slavery equilibrium (0, 3 , ⇡e3 ) – structurally unstable
4. Deflationary finite-valued equilibria (!4 , 0, ⇡e4 ) and (0, 0, ⇡e5 ) – both structurally unstable
5. Other infinite-valued equilibria (0, 0, +1) and (!4 , 0, ±1) – both structurally unstable

C.1.4 Stability study of equilibria with brown or green capital

The Solovian-like equilibrium
The Solovian-like equilibrium corresponds to a desirable long-run steady state with positive wages
and employment, and a finite level of debt (!1 , 1 , ⇡e1 , ). To check its existence, observe that ˙ = 0
in the second equation of (33) whenever
(⇡e1 ) = (↵ + n + )⌫✓ . (15)

Equation (15) can be uniquely solved for ⇡e1 because of condition (9). Given !1 , we also obtain
1 (↵ + (1
1 = )i1 ), where i1 = ⌘p (m!1 1), which belongs to [0; 1], courtesy of (8) (ii) and
provided i1 > 0.

Because of assumption (10), c(·) is uniquely determined by ! and df and so does yd , and conse-
quently ⇡f . Assuming !1 6= 1 6 implies that df has to converge to d¯f , be it finite or not, in order to
have ⇡˙ e = 0. This leads to a finite value of c(!1 + rd¯f ) = c1 2 [c ; c+ ]. Thus yd1 = (⇡e1 ) + c1 + ḡe
and ⇡f1 = ⇡e1 (1 !1 )(1 yd1 ).

Consequently, in order to guarantee ⇡˙ e = 0 in the third equation of (33), we need !1 such that:
( 1 ) ⇥(⇡f1 ) + [1 !1 ⇡e1 pe ⌫1 "¯](↵ + n + i1 ) r[(⇡e1 ) ⇡f1 )] = 0. By the way, (!1 , 1 , ⇡e1 )
as defined above, is an equilibrium for (33).

The Jacobian matrix becomes:

0 1
!1 (1 )⌘p m K0 0 ! 1 K" !1 ⌫(⇡e1 )
B C
B C
B ̇(⇡e1 ) C
B 0 0 1 µ" ⌫ 1 ⌫ 2 1 K" 0 C
B C
B C
B C
B K31 ˙ ( 1 ) K0 K33 K34 K35 C
J1 = B
B
C
C (16)
B C
B (⇡e1 ) C
B 0 0 0 ⌫✓ (⇡e1 )µ" C
B C
B C
B C
@ 0 0 0 0 (1 ¯
2✓)( E ⌫)A

6
Notice that !1 can not be equal to 1, otherwise ⇡˙e 6= 0.

205
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Where:

K0 = ! 1 ˙ ( 1 )
(⇡e1 ) ⌫
K" =
µ2" ⌫1 ⌫2
✓ ◆
K31 = !1 ⌘p m(1 ) ˙ f ) @⇡f
⇥(⇡ + Kd ⌘ p m (↵ + n + i1 ) + r 1 yd1 + (1 !1 )
@yd
@! |(!1 , 1 ,⇡e1 ) @! |(!1 , 1 ,⇡e1 )

Kd = 1 !1 ⇡e1 pE ⌫1 "¯
" #
K33 ˙ f ) @⇡f
= ⇥(⇡ + Kd
̇(⇡e1 )
(↵ + n + i1 ) r ̇(⇡e ) 1 (1 !1 )
@yd
1
@⇡e |(!1 , 1 ,⇡e1 )
⌫✓ @⇡e |(!1 , 1 ,⇡e1 )
✓ ◆
!1 ⌫ !1 ⌫
K34 = pE ⌫1 ( i1 ) (↵ + n + ) K35 = pE ⌫1 (⇡e1 )µ" Kd ⌫(⇡e1 )
µ" µ"
⇣ ⌘ (17)
¯ (⇡e1 )
The characteristic polynomial of this matrix is: P (X) = (X (1 2✓)( E ⌫)) X + ⌫✓ P3 (X)
where P3 (X) = a3 X 3 + a2 X 2 + a1 X + a0 with:

a3 = 1
a2 = K33 + ⌘p m(1 )!1
1 ̇(⇡e1 )
a1 = K33 !1 ⌘p m(1 ) (˙( 1) K0 ) (18)
µ" ⌫ 1 ⌫ 2
̇(⇡e1 ) ̇(⇡e1 )
a0 = (˙( 1) K0 ) 1 ⌘p m(1 )!1 K0 K31 1 .
µ" ⌫ 1 ⌫ 2 µ" ⌫ 1 ⌫ 2
The stability of the equilibrium is guaranteed by the Routh-Hurwitz’s criterion if and only if:

a3 0
a2 0
a0 0
a2 a1a3 a0 0 (19)
(⇡e1 )
>0
⌫✓
(1 ¯ E
2✓)( ⌫) < 0.

These conditions are satisfied for a wide range of parameters and, in particular, by the parameteri-
zation of our Leading Example.

Debt-deflationary Equilibrium
To study the debt-deflationary Equilibrium (0; 0; 1), let us make the change of variable ⇡e = 1q .
The new system is the same as before except for the ⇡e -equation:
 ✓ ◆✓ ◆
2 1 (1/q)
q̇ = q !( ( ) (1 )i ↵) + ( ) ⇥(⇡f ) + 1 ! E ⌫1 " +i
q ⌫✓
✓ ◆ ✓ ◆
1 ⌫
r (1/q) + (1 !)(1 yd ) + "˙ ! pE ⌫ 1 .
q µ"
(20)

206
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

The Jacobian matrix associated to this new system is:

0 ⇣ ⌘ 1
! ̇(1/q)
@f1
! ˙( ) 2
1 µ"
! µ2⌫ (1/q) ! ⌫(1/q)
B @! q µ" ⌫1 ⌫ 2 ⌫✓ " ⌫1 ⌫ 2 C
B C
B C
B 0 @f2 ̇(1/q) ⌫ (1/q)
0 C
B @ q 2 µ" ⌫ 1 ⌫ 2 µ2" ⌫1 ⌫2 C
B C
B C
B @f3 C
B @! q 2 ( ˙ ( ) ! ˙ ( )) @f3 @f3 @f3 C
(21)
B @q @" @✓ C
B C
B ⇣ ⌘ C
B ̇(1/q) µ" 1 (1/q) C
B 0 0 q 2 ⌫ ⌫✓ ⌫1 ⌫ 2 ⌫✓ (1/q)µ" C
B C
B C
B C
@ 0 0 0 0 (1 2✓)( E ⌫)A

Where:
@f1 "˙ ⌫
= ( ) ↵ (1 )i !(1 )⌘p m
@! µ"
@f2 (⇡e ) ⌫
= ↵ n "˙
@ ⌫✓ µ"
 ✓ ◆ ✓ ◆
@f3 "˙ ⌫ @⇡f ˙ 1
= q2 ( ) ↵ (1 )i + !⌘p (1 )m ⇥(⇡f ) + 1 ! ⌫
E 1 " ⌘p m
@! µ" @! q
✓ ◆ ✓ ◆
(1/q) @yd
+ i + r 1 yd + (1 !)
⌫✓ @!
 ✓ ◆ ✓ ◆
@f3 @⇡f ˙ 1 ̇(1/q) 1 (1/q)
= q2 ⇥(⇡f ) + 1 ! ⌫
E 1 " + + i
@q @q q q 2 ⌫✓ q2 ⌫✓
✓ ◆
̇(1/q) 1 @yd
r 2
+ 2 (1 !)
q q @q
  ✓ ◆
1 (1/q)
2q !( ( ) (1 )i ↵) + ( ) ⇥(⇡f ) + 1 ! pE ⌫ 1 " +i
q ⌫✓
 ✓ ◆
1 ⌫
r (1/q) + (1 !)(1 yd ) + "˙ ! pE ⌫ 1
q µ"
" ✓ ◆ ✓ ◆ ✓ ◆ #
@f3 2 (1/q) ⌫ 2 ! ⌫ (1/q)
= q pE ⌫ 1 + i + "! ˙ pE ⌫ 1
@" ⌫✓ µ" µ" ⌫✓
 ✓ ◆ ✓ ◆
@f3 ! ⌫ 1
= q2 pE ⌫1 (1/q)µ" 1 ! E ⌫1 " (1/q) ⌫
@✓ µ" q
(22)

207
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Consequently, at the equlibrium (!, , q) = (0, 0, 0), the Jacobian Matrix becomes:
0 1
(0) ↵ + (1 )⌘p 0 0 0 0
B C
B C
B 0 k0
↵ n 0 0 0 C
B ⌫✓ C
B C
B C
B k0 C
B 0 0 ⌫✓ ⌘p r 0 0 C
B C (23)
B C
B k0 C
B 0 0 0 k0 µ" C
B ⌫✓ C
B C
B C
@ 0 0 0 0 (1 ¯
2✓)( E ⌫)A

Which is stable if and only if:

(0) ↵ + (1 )⌘p < 0
k0
↵ n < 0
⌫✓
k0
⌘p r< 0 (24)
⌫✓
k0
< 0
⌫✓
(1 ¯
2✓)( E ⌫) < 0
The fourth condition, k0 > 0, is likely to be violated under any reasonable specifications of the
investment function: it is hard, indeed, to imagine a productive sector with zero profit that would
still spend a non-negligible amount of money on investment. In particular with our aggregate
investment function satisfying assumptions (9) (ii), k0 < 0.7

Whence, the debt-deflationary equilibrium admits a one-dimensional repulsive submanifold of the

phase space, and is therefore locally unstable. Numerical simulations suggest that the inertia of
aggregate investment plays a role in this lack of asymptotic stability. Consider, indeed, an economy
trapped in the neighborhood of a debt-deflationary steady state: the excess of private debt induced
by the cost of investing in either type of capital has a strong recessionary e↵ect on investment, so
that output is shrinking to zero, as well as employment, profits and wages. Despite the overall
depression, the inertia embodied in (4.2.4) implies that output, Y , is still produced out of K2 and
K1 , so that " = Y1 /Y need not converge to 1 (resp. 0) as it should were the economy to converge
towards the brown (resp. green) deflationary equilibrium.

In order to display the trajectories leading to these debt-deflationary equilibria, one can make
other assumptions (e.g k0 > 0) corresponding to firms investing even if there is no expected profit
(which is not economically meaningful) so that the debt-deflationary Equilibrium will become locally
stable, as in Figure C.1b. This assumption (k0 = 0.05) does not impact the local stability of the
other equilibria, but does not meet our general assumptions.
Here are some illustrations of the paths leading to these two locally stable equilibria (see Table [4.2]
for the parameterization).

7
Indeed, assuming lim (⇡e ) = 0 implies that k0 < 0 with our specific aggregate investment function.
⇡e !0

208
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

(a) Trajectory from initial values: ! = 0.6, = 0.9, (b) Trajectory (with k0 = 0.05) from initial val-
d = 1, " = 0.9, ✓ = 0.3, g = 0.003, ⌧ = 0.08 to Solovian- ues: ! = 0.53, = 0.9, d = 1, " = 0.9, ✓ = 0.3,
like equilibrium with final values: !1 = 0.5546858 , g = 0.09, ⌧ = 0.08 to debt-deflationary equilibrium
1 = 0.9733608 , d1 = 1.478804, " ¯ = 0.9999996, ✓¯ = with final values: !2 = 0 , 2 = 0 , d2 = 351.99,
1.829661e 07, g1 = 0.003236408, ⌧1 = 0.006333863, "¯ = 1, ✓¯ = 0, g2 = 26, 99, ⌧2 = 0.67
⇡1 = 0.1378527

Figure C.1: Trajectories to the brown locally stable equilibria of (33)

(a) Trajectory from initial values: ! = 0.6, = 0.9, (b) Trajectory (with k0 = 0.05) from initial val-
d = 1, " = 0.9, ✓ = 0.3, g = 0.09, ⌧ = 0.08 to Solovian- ues: ! = 0.5, = 0.9, d = 1, " = 0.9, ✓ = 0.3,
like equilibrium with final values: !1 = 0.8226136 , g = 0.09, ⌧ = 0.08 to debt-deflationary equi-
1 = 0.9713984 , d1 = 0.1589752, " ¯ = 0.0217836, librium with final values: !2 = 0 , 2 = 0
✓¯ = 0.9888603, g1 = 0.0021762718, ⌧1 = 0.005458815, , d2 = 1.1 E12, "¯ = 0, ✓¯ = 1, g2 = 0.12,
⇡1 = 0.1614925 ⌧2 = 6.8 E10

Figure C.2: Trajectories to the green locally stable equilibria of (33)

To reach the green equilibria, one can notice the initial positive value of g ⌧ needed to ensure the
convergence of these trajectories to green equilibria. These values will decrease as the transition is
e↵ective (" ! 0). The private debt is also falling in the long-run because profits are not impinged
by fossil energy prices.

209
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Slavery Equilibrium
As in the literature quoted, there exist a slavery equilibrium (0; 3 ; ⇡e 3 ; "
¯ be it brown or green:
¯; ✓),

⇡e3 = ⇡e1 =  1 (⌫✓ (↵ + n + ))

 ✓ ◆
1 (⇡e3 )
and 3 = ⇥(⇡ f3 ) [1 ⇡e3 p ⌫
E 1 "
¯ ] ⌘p + r [(⇡e3 ) ⇡e3 + (1 yd3 )] .
⌫✓
(25)
Consequently the Jacobian Matrix becomes:
0 1
(0) ↵ + (1 )⌘p 0 0 0 0
B C
B C
B ̇(⇡e3 ) (⇡e3 ) ⌫ C
B 0 0 3 µ" ⌫ 1 ⌫ 2 3 µ2 ⌫ 1 ⌫ 2 0 C
B " C
B C
B C
B K31 ˙ ( 3) K33 K34 K35 C
B C (26)
B C
B C
B (⇡e3 ) C
B 0 0 0 ⌫✓ µ" (⇡e3 ) C
B C
B C
B C
@ 0 0 0 0 (1 ¯
2✓)( E ⌫)A

Where:
Kd = 1 ⇡e3 pE ⌫1 "¯

K31 = ( (0) ↵ + (1 )⌘p ) ˙ f ) @⇡f
⇥(⇡ + Kd ⌘ p m (↵ + n ⌘p ) + r 1 yd 3 +
@yd3
3
@! |(0, 3 ,⇡e3 ) @! |(0, 3 ,⇡e3 )
" #
K33 = ˙ f ) @⇡f
⇥(⇡ + Kd
̇(⇡e3 )
(↵ + n ⌘p ) r ̇(⇡e3 ) 1
@yd3
3
@⇡e |(0, 3 ,⇡e3 )
⌫✓ @⇡e |(0, 3 ,⇡e3 )

K34 = pE ⌫1 ( + ⌘p ) K35 = pE ⌫1 µ" (⇡e3 ) Kd ⌫(⇡e3 )

(27)
The characteristic
⇣ polynomial⌘ of this matrix is: P (X) = (X ( (0) ↵ + (1 )⌘p ))(X (1
¯ E (⇡ )
2✓)( ⌫)) X + ⌫1 3 P2 (X) where P2 (X) = a2 X 2 + a1 X + a0 with:
e

a2 = 1
a1 = K33
(28)
˙( 3 )̇(⇡e3 )
a0 = 3 .
µ" ⌫ 1 ⌫ 2
The stability of the equilibrium is guaranteed by the Routh-Hurwitz’s criterion if and only if:
a2 0
a1 0
a0 0
(⇡e3 ) (29)
>0
⌫✓
(0) ↵ + (1 )⌘p < 0
(1 ¯ E
2✓)( ⌫) < 0.

210
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

which is not satisfied by a broad range of parameterizations,8 including our Leading Example.
Whence, the slavery equilibrium is not structurally stable and will not be further studied.

Deflationary finite-valued equilibria

¯ exists if there exists
1) Another economically doubtful equilibrium (!, , ⇡e , ", ✓) = (!4 , 0, ⇡e4 , "¯, ✓)
(!4 , ⇡e4 ) solving the nonlinear equations:
8
> (0) ↵
>i(!4 ) =
< 1
✓ ◆
>
> (⇡e4 )
: 0 = (0) ⇥(⇡f4 ) + [1 !4 ⇡e4 pE ⌫1 "¯] + i(!4 ) r [(⇡e4 ) ⇡e4 + (1 !4 )(1 yd4 )] .
⌫✓
(30)
Assumption (8) (ii) implies that i(!2 ) < 0, which motivates the name of this deflationary equi-
librium. Positive wages being distributed to unemployed workers make however this Cornucopian
equilibrium economically hardly relevant.
Consequently the Jacobian Matrix becomes:
0 1
!4 (1 )⌘p m K0 0 ! 4 K" !4 ⌫(⇡e4 )
B C
B C
B 0 K22 0 0 0 C
B C
B C
B C
B K31 ˙ (0) K0 K33 K34 K35 C
B C
B C (31)
B C
B (⇡e4 ) C
B 0 0 0 (⇡e4 )µ" C
B ⌫✓ C
B C
B C
@ 0 0 0 0 ¯
(1 2✓)( E ⌫)A


̇(⇡e3 ) @yd3
8
Indeed K33 = ˙ f3 ) @⇡f
⇥(⇡ + Kd (↵ + n ⌘p ) r ̇(⇡e3 ) 1 with
@⇡f
= @yd
.
@⇡e |(0, 3 ,⇡e3 ) ⌫✓ @⇡e |(0, 3 ,⇡e ) @⇡e @⇡e
3
˙ f3 ) > 0. For instance, in our leading, r is around 10 and ⇥(⇡
For a broad range of parameters r ⇥(⇡ ˙ f3 ) is around 2
@⇡f @yd3
3 ˙
10 . Thus ⇥(⇡f3 ) @⇡e + r @⇡e > 0. By the way, ↵ + n ⌘p r ⇠ 10 2
and negative for a
|(0, 3 ,⇡e3 ) |(0, 3 ,⇡e3 )
broad range of values (and so it is in our Leading Example). Lastly Kd /⌫✓ ⇠ 10 1 for a broad range of values and
so Kd /⌫✓ > r. Consequently, there is a large range of parameters for which K33 is positive, which implies that the
Routh-Hurwitz’s criterion is not satisfied.

211
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Where:

K0 = !4 ˙ (0)
(⇡e4 ) ⌫
K" =
µ2" ⌫1 ⌫2
✓ ◆
K31 = !4 ⌘p m(1 ˙ f ) @⇡f
) ⇥(⇡ + Kd ⌘ p m
(⇡e4 )
+ i(!4 )
4
@! |(!4 ,0,⇡e4 ) ⌫✓
✓ ◆
@yd
+ r 1 yd4 + (1 !4 )
@! |(!4 ,0,⇡e4 )
Kd = 1 ! 4 ⇡e4 pE ⌫1 "¯
(⇡e4 )
K22 = ↵ n
⌫✓
✓ ◆ " #
˙ @⇡f ̇(⇡e4 ) (⇡e4 ) @yd
K33 = ⇥(⇡f4 ) + Kd + i(!4 ) r ̇(⇡e4 ) 1 (1 !4 )
@⇡e |(!4 ,0,⇡e4 ) ⌫✓ ⌫✓ @⇡e |(!4 ,0,⇡e4 )
✓ ◆ ✓ ◆
(⇡e4 ) !4 ⌫ (⇡e4 )
K34 = pE ⌫ 1 + i(!4 ) pE ⌫ 1
⌫✓ µ" ⌫✓
✓ ◆
!4 ⌫
K35 = pE ⌫1 (⇡e4 )µ" Kd ⌫(⇡e4 )
µ"
(32)
This matrix can be transformed into an upper triangular matrix of which stability criterion are the
(⇡e4 ) ¯ E
signs of the eigenvalues: K33 , ⌘p m(1 )!4 , K22 , ⌫✓ and (1 2✓)( ⌫). According
to our assumptions, all these eigenvalues have negative parts but K33 , which is positive for a broad
range of parameterization.9 Consequently this equilibrium is structurally unstable.

¯
2) An economically more relevant, deflationary steady state is given by (!, , ⇡e , ", ✓) = (0, 0, ⇡e6 , "¯, ✓),
provided ⇡e5 solving the nonlinear equation:
✓ ◆
(⇡e6 )
0 = (0) ⇥(⇡f5 ) + [1 ⇡e6 pE ⌫1 "¯] ⌘p r [(⇡e6 ) ⇡e6 + (1 yd5 )] . (33)
⌫✓

Here again, i(0) = ⌘p < 0, and the Jacobian matrix can be transformed into an upper triangular
matrix.
9
Indeed, for the same reasons as in the slavery case, there exists a broad range of parameters values that makes
(⇡ ) (⇡ )
K33 > 0. If K22 = ⌫✓e4 ↵ n < 0 is negative, there is a broad range of value that makes also ⌫✓e4 ⌘p r <0
@yd
because r, ⌘p ↵ and n are of the same order of magnitude. Kd /⌫✓ is still greater than r. Lastly r(1 !4 ) @⇡ e |(!4 ,0,⇡e4 )
@⇡
is still greater than ⇥(⇡˙ f4 ) @⇡fe for a broad range of values.
|(!4 ,0,⇡e4 )

212
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Consequently the Jacobian Matrix becomes:

0 1
(0) ↵ + (1 )⌘p 0 0 0 0
B C
B C
B 0 K22 0 0 0 C
B C
B C
B C
B K31 (0) K33 K34 K35 C
B C
B C (34)
B C
B (⇡e6 ) µ" C
B 0 0 0 ⌫(⇡ e1 C
)
B ⌫1 ⌫ C
B C
B C
@ 0 0 0 0 ( E ⌫) A

Where:
(⇡e6 ) ⌫ (⇡e6 )
⌫(⇡e1 ) = Kd = 1 ⇡e6 pE ⌫1 "¯ K22 = ↵ n
⌫12 ⌫✓
✓ ◆
K31 = ( (0) ↵ + (1 )⌘p ) ˙ f ) @⇡f
⇥(⇡ + Kd ⌘ p m
(⇡e6 )
⌘p
5
@⇡e |(0,0,⇡e5 ) ⌫✓
✓ ◆
@yd
+r 1 yd 5 +
@! |(0,0,⇡e5 )

✓ ◆ " #
K33 = ˙ f ) @⇡f
⇥(⇡ + Kd
̇(⇡e6 ) (⇡e6 )
⌘p r ̇(⇡e6 ) 1
@yd
5
@⇡e |(0,0,⇡e5 ) ⌫✓ ⌫✓ @⇡e |(0,0,⇡e5 )
(⇡e6 )µ"
K34 = pE ⌫1 ( + ⌘p ) K35 = pE + Kd ⌫(⇡e5 )
⌫1
(35)
For the same reasons as before, either K22 or K33 is positive for a broad range of value which means
that all eigenvalues have not negative parts. Consequently this equilibrium is structurally unstable.

Other infinite-valued equilibria

¯ Indeed ⇡e ! +1
Another infinite-valued equilibrium is given by (!, , ⇡e , ", ✓) = (0, 0, +1, "¯, ✓).
1
implies q = ⇡e ! 0 that solves q̇-equation (46) of the new system. The Jacobian matrix becomes
an upper triangular matrix. One of its eigenvalues is (1/q)
⌫✓ ↵ n which can never be negative
because  is assumed to be strickly increasing. So this equilibrium is structurally unstable.

Eventually, a deflationary state with an infinite profit ratio, be it positive or negative, (!, , df ) =
(!4 , 0, ±1)10 is also a zero of (4). In the neighborhood of this equilibrium, two cases can be studied:
• If ⇡e ! 1, due to an explosive level of private debt for example, (⇡e ) ! k0 and yd !
k0 + c+ + ḡe . The Jacobian matrix can be transformed into an upper triangular matrix whose
eigenvalues are easy to identify. Their sign depends on the selected parameterization, so that
the stability of the steady state cannot be ruled out a priori. As in the case of the Bad equi-
librium, the stability condition k⌫✓0 < 0 is not verified with our particular parameterization.
Other values of parameterization might satisfy this condition. Nonetheless, a steady state
with positive wages, negative profits and zero-employment is hardly economically meaningful.
10 (0) ↵
With ⌘p (µ!4 1) = 1
< 0 due to Assumption (8) (ii).

213
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

• If ⇡e ! +1, due to an explosive level of subsidies for example, the characteristic polynomial
associated to the Jacobian Matrix admits (1/q)
⌫✓ ↵ n as root, that is not negative
because  is assumed to be strickly increasing. So this equilibrium is structurally unstable.

C.2 Stability of equilibria with subsidies and carbon tax

This section analyzes the local, asymptotic stability of long-run steady states of the dynamical
system (55).

C.2.1 The reduced dynamical system

Notice as before that the behavior of the households’ debt ratio, dh , in (55), can still be deduced
from the other variables while the latter are independent from dh .

Regarding again rg close to 0, consumption only depends now on wage share and firm’s debt ratio
c(!, rdf ), and consequently dg is still an auxiliary variable of the di↵erential system, as no other
variable depends on it.

With the same consideration about ge as before, we assume that its equation dynamics can be
solved separetely.

Eventually, using the same method as [Costa Lima and Grasselli, 2014] and because ⇡e = 1 !
rdf + g ⌧ ⌧c E ⌫1 ", we can write:

⇡˙ e = !˙ rd˙f + ġ ⌧˙ ( E + c )⌫1 "˙

✓ ✓ ◆◆
"˙ ⌫ "
= ! ( ) (1 )i ↵ ⌫ ⇢
ˆ 1
µ" ⌫ 2 µ"
 ✓ ◆
(⇡e )
r (⇡e ) (1 !)yd + E ⌫1 " g + ⌧ + df r + ⌫⇢
ˆ(1 ") i + ( )
⌫✓
✓ ◆ ✓ ◆
(⇡e ) (⇡e )
g + ⌫ ⇢ˆ(1 ") + i ⇥(⇡f ) + ⌧ + ⌫⇢
ˆ(1 ") + i ( E + c )⌫1 "˙
⌫✓ ⌫✓
(36)
✓ ✓ ◆◆
"
) ⇡˙ e = ! ( ) (1 )i ↵ ⌫⇢
ˆ ⇥(⇡f )1 + ( )
⌫ 2 µ"
✓ ◆
(⇡e )
+ [1 ! ⇡e ( E + c )⌫1 "] + ⌫ ⇢ˆ(1 ") + i (37)
⌫✓
✓ ◆
⌫
r [(⇡e ) ⇡e + (1 !)(1 yd )] + "˙ ! ( E + c )⌫1 .
µ"
Thus, the previous system reduces to the following five-dimensionnal system11 in which we consider
11
It will be enough to characterize the equilibria in which the economy either prospers or collapses. But we have
to notice that when working with the reduced system, we can not recover df , g and ⌧ separately, but rather the
combination rdf g + ⌧ .

214
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

ge quickly convergent to its equilibrium value ḡe :

8 ✓ ◆
>
> !˙ "˙ ⌫ "
>
> = ( ) (1 )i ↵ ⌫⇢
ˆ 1
>
> ! µ" ⌫ 2 µ"
>
> ✓ ◆
>
> ˙ (⇡e ) "˙ ⌫ 1
>
> = ↵ n
>
> ⌫⇢ˆ" 1
>
> ⌫✓ µ" ⌫ 2 µ"
>
> ✓ ✓ ◆◆ ✓ ◆
>
> " ⌫
>
> ⇡˙ e = ! ( ) (1 )i ↵ ⌫⇢ˆ 1 + ( ) ⇥(⇡f ) + "˙ ! ( E + c )⌫1
>
> ⌫ 2 µ" µ"
< ✓ ◆
(⇡e )
>
> + [1 ! ⇡e ( E + c )⌫1 "] + ⌫ ⇢ˆ(1 ") + i r [(⇡e ) ⇡e + (1 !)(1 yd )]
>
> ⌫✓
>
> ✓ ◆
>
> (⇡e ) µ" 1
>
> "˙ =
> ⌫⇢ˆ(1 ")"
>
> ⌫ ⌫✓ ⌫1 ⌫2
>
>
>
> ✓˙ = ✓ (1 ✓ )( E + c
> ⇤ ⇤ ⇤
> ⌫)
>
>
>
> ⇤
(✓RD ⇢)✓ (⇡e ) + RD ⌧c
>
: ⇢˙ = .
⌫2 (1 ")
(38)
with the following auxiliary variables:
8
>
> rdf g + ⌧ + ⌧c = 1 ! ⇡e E ⌫1 "
>
>
>
> ⌫ˆ2 = ⌫⇢
ˆ
>
>
>
>
>
> ⌧c = c ⌫ 1 "
>
>
>
> i = ⌘p (m! 1).
>
>
>
>
>
> 1 1 ✓ ✓
>
< = +
⌫✓ ⌫1 ⌫2
1 " " (39)
>
> µ = +
>
> "
>
> ⌫1 ⌫2
>
> 1 1
>
>
>
> ⌫=
>
> ⌫ 1 ⌫ 2
>
>
>
> Yd
>
> yd := = c(! + rdf + rg dg ) + (⇡e ) + ge
>
> Y
>
:
⇡e ⇡f = (1 !)(1 yd )
⇣ ⌘
E (!, ,⇡,g,⌧,GE ,Y ) (⇡e )
while assuming that g˙e = Y ge ⌫✓ quickly converges to a finite equilibrium
value ḡe .

215
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Considering ge at its equilibrium value, the associated Jacobian matrix is:

0 ⇣ ⌘ 1
@f1 ˙( ) ̇(⇡e ) µ" 1 @f1 @f1
! ! ! ⌫(⇡e )
B @! µ" ⌫✓ ⌫1 ⌫ 2 @" @⇢
C
B C
B C
B 0 @f2 ̇(⇡e ) @f2
0 @f2 C
B @ µ" ⌫ 1 ⌫ 2 @" @⇢ C
B C
B C
B @f3 ˙ C
B @! ( ) ! ˙( ) @f3 @f3 @f3 @f3 C
B @⇡e @" @✓ @⇢ C
B C
B ⇣ ⌘ C (40)
B ̇(⇡e ) µ" 1 @f4 @f4 @f4 C
B 0 0 ⌫ ⌫✓ ⌫1 ⌫2 @" @✓ ⇤ @⇢ C
B C
B C
B C
B 0 0 0 0 (1 2✓⇤ )( E + c ⌫) ✓⇤ (1 ✓⇤ ) ⌫2⌫⇢ C
B C
B C
B C
B (✓RD ⇢)✓ ⇤ ̇(⇡e ) @f6 @f6 @f6 C
@ 0 0 ⌫2 (1 ") @" @✓ ⇤ @⇢
A

216
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Where:
✓ ◆
@f1 "˙ ⌫ "
= ( ) ↵ (1 )i ⌫⇢ˆ 1 !(1 )⌘p m
@! µ" ⌫ 2 µ"
 ✓ ◆ ✓ ◆
@f1 (⇡e ) ⌫ ⌫(1 2") ⌫ 2 1 ⌫"
=! + ⌫ ⇢ˆK⇢1 where K⇢1 = + (1 ")" + +
@" µ2" ⌫1 ⌫2 µ" µ" ⌫2 µ" ⌫2 µ2"
✓ ◆ 2 ˆ"
✓ ◆ ✓ ◆
@f1 ⇢ˆ " ⌫⇢ 1 ⌫ @f4 @ ⌫
=! ⌫ 1 +! 1 ! ! "˙
@⇢ ⇢ ⌫ 2 µ" ⇢⌫2 µ" µ" ⌫ 2 µ" @⇢ @⇢ µ"
✓ ◆
@f2 (⇡e ) "˙ ⌫ 1
= ↵ n ⌫⇢ˆ" 1
@ ⌫✓ µ" ⌫ 2 µ"
 ✓ ◆
@f2 (⇡e ) ⌫ µ" (1 2") + "(1 ") ⌫ 1 " ⌫
= + ⌫ ⌫ ⇢ˆ ⌫⇢
ˆ 1
@" µ2" ⌫1 ⌫2 µ2" ⌫ 2 µ" ⌫ 2 µ"
✓ ◆ ✓ ◆ 2 ˆ"
✓ ◆
@f2 ✓ ⌫ @ ⌫ ⌫ @f4 ⇢ˆ 1 ⌫⇢ "
= (⇡e ) "˙ + ⌫" 1 + 1
@⇢ ⌫2 ⇢ @⇢ µ" µ" @⇢ ⇢ ⌫ 2 µ" µ" ⌫ 2 ⇢ µ" ⌫ 2
✓ ✓ ◆◆ 
@f3 " ˙ f ) 1 yd + (1 !) @yd
= ( ) ↵ (1 )i ⌫⇢ ˆ 1 + !⌘p (1 )m ⇥(⇡
@! ⌫ 2 µ" @!
✓ ◆ ✓ ◆
⌫ (⇡e ) @yd
+ "˙ + ⌘p m(1 ! ⇡e ( E + c )⌫1 ") + ⌫ ⇢ˆ(1 ") + i + r 1 yd + (1 !)
µ" ⌫✓ @!
✓ ◆
@f3 @⇡f ˙ ̇(⇡e ) (⇡e )
= ⇥(⇡f ) + [1 ! ⇡e ( E + c )⌫1 "] + ⌫ ⇢ˆ(1 ") + i
@⇡e @⇡e ⌫✓ ⌫✓
 ✓ ◆ ✓ ◆
@yd ! ⌫ ̇(⇡e ) µ" 1
r ̇(⇡e ) 1 (1 !) + ( e + c )⌫1
@⇡e µ" ⌫ ⌫✓ ⌫1 ⌫2
✓ ◆2 ✓ ◆✓ ◆
@f3 µ" + ⌫" ⌫ ! ⌫ (⇡e )
= ! ⌫ ⇢ˆ + "!
˙ + ( E + c )⌫1 ⌫⇢ˆ(1 2")
@" ⌫2 µ2" µ" µ" ⌫✓

(⇡e )
( E + c )⌫1 + ⌫ ⇢ˆ(1 ") + i ⌫⇢ˆ[1 ! ⇡e ( E + c )⌫1 "]
⌫✓
✓ ◆
@f3 ! ⌫ " @f6
⇤
= ( E + c )⌫1 (⇡e )µ" ⌫ (1 ")
@✓ µ" ⇢ @✓⇤

⌫ (1 ") @f6
(1 ! ⇡e ( E + c )⌫1 ") (⇡e ) ⌫
⇢ @✓⇤
✓ ◆ ✓ ◆ ✓ ◆ 
@f3 ⇢ˆ " @ 1 ⌫" ! ⌫ ⌫ @f4
= ! ⌫ 1 ! ⌫ ⇢ˆ" + "˙ ! ⌫ 2
+ ! ( E + c )⌫1
@⇢ ⇢ ⌫ 2 µ" @⇢ ⌫2 µ" ⌫2 ⇢µ" µ" ⌫ 2 ⇢ µ" @⇢
✓ ◆
⌫✓ ⇢ˆ @f6 (✓RD ⇢)(⇡e )
+ [1 ! ⇡e ( E + c )⌫1 "] (⇡e ) ⌫ (1 ") ; =
⌫2 ⇢ ⇢ @✓⇤ ⌫2 (1 ")
@f4 (⇡e ) @f4 " @f6
= ⌫⇢ˆ(1 2") ; ⇤
= (⇡e )µ" ⌫ (1 ")
@" ⌫✓ @✓ ⇢ @✓⇤
✓ ◆ ✓ ◆✓ ◆
@f4 (⇡e ) 1 @µ" µ" ⌫ ✓ ⌫ (⇡e ) ⌫ µ" 1 ⇢ˆ
= + + + ⌫ (1 ")"
@⇢ ⌫ ⌫✓ @⇢ ⌫2 ⇢ ⌫1 ⌫2 ⇢ ⌫2 ⌫2 ⇢ ⌫✓ ⌫1 ⌫2 ⇢
⇤ ⇤
✓ ◆
@f6 (✓RD ⇢)✓ (⇡e ) + RD c ⌫1 @f6 ✓ (⇡e ) ✓RD
= 2
; = 1 ⌫ 1
@" ⌫2 (1 ") @⇢ ⌫2 (1 ") ⇢
@µ" " ⌫ @1/⌫✓ ✓ ⌫ @⌫2 @⌫2 ⌫2 ⇢ @t ⌫ˆ2 ⌫2 ⌫2
= ; = ; = = = ⌫
@⇢ ⌫2 ⇢ @⇢ ⌫2 ⇢ @⇢ @t ⌫2 ⇢ @⇢ ⇢ˆ ⇢ ⇢
(41)

217
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

C.2.2 Analysis of the di↵erent equilibria

With the same set of assumptions as before and with the parameter’s value of our Leading
Example, we study the stability of the two locally stable equilibria of the previous model: the
Solovian-like and the Debt-deflationary Equilibrium. We let to an other study, the possibility to
analyse other meaningful equilibria.

The Solovian-like Equilibrium

This Solovian-like equilibrium corresponds to a desirable long-run steady state with positive wages
and employment, and a finite level of debt (!1 , 1 , ⇡e1 ). This equilibrium may be green or brown
according to the asymptotic value of (", ✓).

At the Green Solovian-like Equilibrium, ✓ = 1 and " = 0: ⇢ has to be equal to ⇢1 := ✓RD in order
to ensure ⇢˙ = 0. At the Brown Solovian-like Equilibrium (✓ = 0 and " = 1), one can use the change
of variable ⇢ = 1/ to show that ⇢ ! +1. We study now the stability of these two equilibria.

At the Green Solovian-like Equilibrium, the Jacobian matrix becomes is the same as the 5x5
matrix of the appendix C.1.4, J1 , with the following sixth line and column:
0 1
0
B C
B C
B (⇡e1 ) ⌫ C
B 1 ⌫¯2 ⇢1 C
B C
B C
B (⇡ ) C
B J1 (1 !1 ⇡e1 ) ⌫¯2 ⇢1 C e 1 ⌫
B C
B C
B C (42)
B 0 C
B C
B C
B C
B C
B 0 C
B C
B C
B (⇡e1 ) C
@0 0 0 ⌫1
RD c ⌫¯2 0
A
⌫¯2

which can be transformed into an upper triangular matrix.

(⇡e1 )
The characteristic polynomial of this matrix admits a new negative root, ⌫¯2 , which means
that the Green Solovian-like Equlibrium is still locally asymptotically stable.

At the Brown Solovian-like Equilibrium, (!, , ⇡e , ", ✓⇤ , ) = (!1 , 1 , ⇡e1 , 1, 0, 0), with the change
(✓RD 1)✓ ⇤ (⇡e )+ 2 RD ⌧c
of variable = 1/⇢, the last equation of the system becomes: ˙ = ⌫2 (1 ") .

218
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Consequently the Jacobian matrix for the modified system becomes:

0 @f ⇤ ⇣ ⌘ 1
˙( ) ̇(⇡e ) µ" 1 @f1⇤ @f1⇤
@!
1
! ! µ" ⌫✓ ⌫1 ⌫2 @" ! ⌫(⇡e ) @
B C
B C
B @f2⇤ @f2⇤ @f2⇤
C
B 0 ̇(⇡e )
0 C
B @ µ " ⌫ 1 ⌫2 @" @ C
B C
B C
B @f3⇤ ˙ ˙( ) @f3⇤ @f3⇤ @f3⇤ @f3⇤ C
B @! ( ) ! C
B @⇡ e @" @✓ @ C
B C
B ⇣ ⌘ C (43)
B ̇(⇡e ) µ" 1 @f4⇤ @f4⇤ @f4⇤ C
B 0 0 ⌫ ⌫ ⌫ ⌫ @" @✓ ⇤ @ C
B ✓ 1 2 C
B C
B C
B 0 0 0 0 (1 2✓⇤ )( E + c ⌫) ✓⇤ (1 ✓⇤ ) ⌫2⌫ C
B C
B C
B C
B (✓RD 1)✓ ⇤ ̇(⇡e ) 2
RD c ⌫1 (✓RD 1)(⇡e ) @f6⇤ C
@ 0 0 ⌫2 (1 ") ⌫2 (1 ")2 ⌫2 (1 ") @
A

In order to analyse the local stability of this brown Solovian like equilibrium (where ⇢ tends to
+1), let us endorse Assumption C: ⇢˙ remains bounded from above as t ! +1. This implies that:

1 ⇢˙
= . (44)
⌫2 (1 ") (✓RD ⇢)✓⇤ (⇡ e) + RD ⌧c

Consequently, each term of the last line of the Jacobian matrix (43) tends to 0. Indeed:12

(✓RD 1)✓⇤ ̇(⇡e ) (✓RD 1)✓⇤ ̇(⇡e )⇢˙

= !0
⌫2 (1 ") (✓RD ⇢)✓⇤ (⇡e ) + RD ⌧c
2 2 2
RD c ⌫1 RD c ⌫1 ⇢˙
= !0
⌫2 (1 ")2 [(✓RD ⇢)✓⇤ (⇡e ) + RD ⌧c ]2
(✓RD 1)(⇡e ) (✓RD 1)(⇡e )⇢˙
= ⇤
!0 (45)
⌫2 (1 ") (✓RD ⇢)✓ (⇡e ) + RD ⌧c
@f6⇤ [(2 ⌫ ) ✓RD (1 ⇤
⌫ )]✓ (⇡e ) + (2 ⌫) RD ⌧c
=
@ ⌫2 (1 ")
[[(2 ⇤
⌫ ) ✓RD (1 ⌫ )]✓ (⇡e ) + (2 ⌫) RD ⌧c ]⇢˙
= !0
(✓RD ⇢)✓⇤ (⇡e ) + RD ⌧c

This means that the Jacobian matrix (43) can be turned into an upper triangular matrix of which
the characteristic polynomial admits a non-negative root. Consequently, the dirty Solovian-like
equilibrium is locally asymptotically unstable. Our numerical Example illustrates how the intro-
duction of green R&D destabilizes the brown Solovian-like steady state (see Figure 4.3b: green
transition does occur — with high private debt levels and much more slowly but still).
12
Note that RD ⌧c tends to a constant: RD c ⌫1 .

219
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Debt-deflationary Equilibrium
As before, to study the debt-deflationary Equilibrium (0; 0; 1), let us make the change of
variable ⇡e = 1q . The new system is the same as before except for the ⇡e -equation:
 ✓ ✓ ◆◆ ✓ ◆
2 " ⌫
q̇ = q ! ( ) (1 )i ↵ ⌫⇢ˆ 1 + ( ) ⇥(⇡f ) + "˙ ! ( E + c )⌫1
⌫ 2 µ" µ"
✓ ◆✓ ◆ ✓ ◆
1 (1/q) 1
+ 1 ! ( E + c )⌫1 " + ⌫ ⇢ˆ(1 ") + i r (1/q) + (1 !)(1 yd ) .
q ⌫✓ q
(46)
The Jacobian matrix associated to this new system is:
0 @f ⇣ ⌘ 1
1 ˙( ) ! ̇(1/q) 1 µ" ⌫ (1/q) @f1
! 2 ! 2 ! ⌫(1/q)
B @! q µ" ⌫1 ⌫ 2 ⌫✓ µ" ⌫ 1 ⌫ 2 @⇢
C
B C
B C
B 0 @f 2 ̇(1/q) ⌫ (1/q)
0 @f 2C
B @ 2
q µ " ⌫1 ⌫2 2
µ" ⌫ 1 ⌫ 2 @⇢ C
B C
B C
B @f3 @f3 C
B @! q 2 ( ˙ ( ) ! ˙ ( )) @f3 @f3 @f3
C
B @q @" @✓ @⇢ C
B C
B ⇣ ⌘ C (47)
B ̇(1/q) µ" 1 (1/q) @f4 C
B 0 0 2
q ⌫ ⌫✓ ⌫1 ⌫2 ⌫✓ (1/q)µ " @⇢ C
B C
B C
B C
B 0 0 0 0 (1 2✓⇤ )( E + c ⌫) @f 5C
B @⇢ C
B C
B C
B ⇤
(✓RD ⇢)✓ ̇(⇡e ) @f6 @f6 @f6 C
@ 0 0 ⌫2 (1 ") @" @✓ ⇤ @⇢
A

Where:
 ✓ ✓ ◆◆
@f3 2 "˙ ⌫ " @⇡f ˙
= q ( ) ↵ (1 )i ⌫⇢ˆ 1 + !⌘p (1 )m ⇥(⇡f )
@! µ" ⌫ 2 µ" @!
✓ ◆ ✓ ◆ ✓ ◆
1 (1/q) @yd
+ 1 ! ( E + c )⌫1 " ⌘p m + ⌫ ⇢ˆ(1 ") + i + r 1 yd + (1 !)
q ⌫✓ @!
 ✓ ◆ ✓ ◆
@f3 2 @⇡ f ˙ 1 ̇(1/q) 1 (1/q)
= q ⇥(⇡f ) + 1 ! ( E + c )⌫1 " + 2 + ⌫ ⇢ˆ(1 ") + i
@q @q q q 2 ⌫✓ q ⌫✓
✓ ◆
̇(1/q) 1 @yd
r + 2 (1 !) 2q [ !( ( ) (1 )i ↵) + ( ) ⇥(⇡f )
q2 q @q
 ✓ ◆  ✓ ◆
1 (1/q) 1 ⌫
+ 1 ! pE ⌫ 1 " +i r (1/q) + (1 !)(1 yd ) + "˙ ! pE ⌫ 1
q ⌫✓ q µ"
" ✓ ◆ ✓ ◆2 ✓ ◆ #
@f3 (1/q) ⌫ ! ⌫ (1/q)
= q2 pE ⌫ 1 + i + "! ˙ pE ⌫ 1
@" ⌫✓ µ" µ" ⌫✓
 ✓ ◆ ✓ ◆
@f3 ! ⌫ 1
= q2 pE ⌫1 (1/q)µ" 1 ! E ⌫1 " (1/q) ⌫
@✓ µ" q
 ✓ ◆ ✓ ◆ ✓ ◆ 
@f3 2 ⇢ˆ " @ 1 ⌫" ! ⌫ ⌫ @f4
= q ! ⌫ 1 ! ⌫ ⇢ˆ" + "˙ ! ⌫ + ! ( E + c )⌫1
@⇢ ⇢ ⌫ 2 µ" @⇢ ⌫2 µ" ⌫2 ⇢µ2" µ" ⌫ 2 ⇢ µ" @⇢
✓ ◆
1 ⌫✓ ⇢ˆ
+[1 ! ( E + c )⌫1 "] (1/q) ⌫ (1 ")
q ⌫2 ⇢ ⇢
(48)

220
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Consequently, at the equlibrium (!, , q) = (0, 0, 0), the Jacobian Matrix becomes:
0 1
(0) ↵ + (1 )⌘p 0 0 0 0 0
B C
B C
B 0 k0
↵ n 0 0 0 0 C
B ⌫✓ C
B C
B C
B 0 0 k0
⌘p r 0 0 0 C
B ⌫✓ C
B C
B C
B k0 @f4 @f4 C
B 0 0 0 ¯ ⇢) C (49)
B ⌫✓ @✓ |(0,0,0,¯
",✓,¯ @⇢ |(0,0,0,¯ ¯ ⇢) C
",✓,¯
B C
B C
B @f5 @f5 C
B 0 0 0 0 ¯ ⇢) C
B @✓ |(0,0,0,¯
",✓,¯ @⇢ |(0,0,0,¯ ¯ ⇢) C
",✓,¯
B C
B C
B @f6 (✓RD ⇢¯)k0 @f6 C
@ 0 0 0 @" |(0,0,0,¯ ¯ ⇢)
",✓,¯ ⌫¯2 (1 "¯) @⇢ ¯ A
|(0,0,0,¯
",✓,¯
⇢)

with
@f4 "¯ (✓RD ⇢¯)k0
= k0 µ" ⌫ (1 "¯)
@✓ |(0,0,0,¯",✓,¯
¯ ⇢) ⇢¯ ⌫¯2 (1 "¯)
✓ ◆ ✓ ◆✓ ◆
@f4 k0 "¯ ⌫ µ" ⌫ ✓¯ ⌫ k0 ⌫ µ" 1
= ¯ + + 2
@⇢ |(0,0,0,¯",✓,¯
¯ ⇢) ⌫ ⌫✓ ⌫¯2 ⇢¯ ⌫¯2 ⇢¯ ⌫1 ⌫¯2 ⇢¯ ¯⌫ ⌫¯2 ⇢¯ ⌫✓ ⌫1 ⌫¯2
@f5 ¯
= (1 2✓)( E + c ⌫)
@✓ |(0,0,0,¯",✓,¯
¯ ⇢)
@f5 (50)
⌫
= ✓¯⇤ (1 ✓¯⇤ )
@⇢ |(0,0,0,¯",✓,¯
¯ ⇢) ⌫¯2 ⇢¯
@f6 ¯⇤
(✓RD ⇢¯)✓ k0 + RD c ⌫1
=
@" |(0,0,0,¯",✓,¯
¯ ⇢) ⌫¯2 (1 "¯)2
✓ ◆
@f6 ✓¯⇤ k0 ✓RD
= 1 ⌫ 1
@⇢ |(0,0,0,¯",✓,¯
¯ ⇢) ⌫¯2 (1 "¯) ⇢¯
where, whether " ! 1 and ✓ ! 0 (brown equilibrium), or " ! 0 and ✓ ! 1 (green):

@f4
!0
@⇢ |(0,0,0,¯",✓,¯
¯ ⇢)
(51)
@f5
!0
@⇢ |(0,0,0,¯",✓,¯
¯ ⇢)

Consequently, the Jacobian Matrix of the modified system can be turned into an upper triangular

221
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

matrix which is stable if and only if:

(0) ↵ + (1 )⌘p < 0
k0
↵ n < 0
⌫✓
k0
⌘p r< 0
⌫✓
k0 (52)
< 0
⌫✓
(1 ¯
2✓)( E + c ⌫) < 0
@f6
< 0
@⇢ |(0,0,0,¯",✓,¯
¯ ⇢)

The condition k⌫10 < 0 is not verified with our aggregate investment function satisfying assumptions
(9) (ii) and thus, the debt-deflationary equilibria, be it brown or green, is still locally unstable. 13

Hybrid Equilibrium
Due to the introduction of a variable ⌫2 , we can know study another type of equilibrium where ⌫2
is a solution of ✓˙ = 0. Indeed, ⌫¯12 = ⌫11 E
˙
c ) ✓ = 0 and so ⌫
¯2 is solution of the fifth equation
of (38).

Because ⌫ˆ2 = ⌫⇢
ˆ, if ⌫2 tends to ⌫¯2 , consequently ⇢ tends to ⇢3 , which has no reason to be equal
to ⇢1 = ✓RD (like it was the case in the Solovian-like Equilibrium).

As a consequence, if (", ✓) = ("3 , ✓3 ) such that:

8
< µ"3 = 1
⌫ ✓3 ⌫1 ⌫¯2 (53)
:
(✓RD ⇢3 )✓3 (⇡e3 ) + RD c ⌫1 "3 =0
with ⇡e3 =  1 (⌫✓3 (↵ + n + )), 3 = 1 (↵ + (1 )i3 ), and !3 , solution of the third equation of
(38), consequently (!3 , 3 , ⇡e3 , "3 , ✓3 , ⇢3 ) is an equilibrium for (38).

The Jacobian matrix becomes:

0 1
J11 J12 0 !3 ⌫(⇡e3 ) J16
B C
B C
B 0 0 J23 J24 0 J26 C
B C
B C
B C
BJ J36 C
B 31 J32 J33 J34 J35 C
B C
B C
J3 = B C (54)
B 0 0 0 J44 J45 J46 C
B C
B C
B C
B 0 0 0 0 0 J56 C
B C
B C
B C
@ 0 0 J63 J64 J65 J66 A

13
Indeed, assuming lim (⇡e ) = 0 implies that k0 < 0.
⇡e !0

222
APPENDIX C. DIRECTED TECHNICAL CHANGE IN AN INVENTORY STOCK-FLOW
CONSISTENT DYNAMICS WITH GOVERNMENT: FROM DIRTY TO CLEAN CAPITAL

Where:
(⇡e3 ) ⌫ ⌫
J11 = !3 (1 )⌘p m J12 = !3 ˙ ( 3) J14 = !3 K" with K" = J16 = !3 J46
µ2"3 ⌫1 ⌫¯2 µ"3
̇(⇡e3 ) ✓3 ⌫ ⌫
J23 = 3 J24 = 3 K" J26 = 3 (⇡e3 ) 3 J46
µ"3 ⌫1 ⌫¯2 ⌫¯2 ⇢3 µ"3

⌫ (⇡e3 ) "3 1
with J46 = + µ "3 ✓3
⌫ ⌫¯2 ⇢3 ⌫✓3 ⌫1

⌫ ⌫ ✓3 ⌫
J36 = !3 ( E + c )⌫1 J46 + (1 !3 ⇡e3 ( E + c )⌫1 "3 )(⇡e3 ) J56 = ✓3 (1 ✓3 )
µ"3 ⌫¯2 ⇢3 ⌫¯2 ⇢3
˙ f ) f @⇡
J31 = !3 ⌘p m(1 ) ⇥(⇡ 3 + Kd ⌘p m (↵ + n + i3 )
@! |(!3 , 3 ,⇡e3 ,"3 ,✓3 ,⇢3 )
✓ ◆
@yd
+ r 1 yd3 + (1 !3 )
@! |(!3 , 3 ,⇡e3 ,"3 ,✓3 ,⇢3 )
✓ ◆
˙ ✓3 (⇡e3 ) ✓RD
Kd = 1 !3 ⇡e3 ⌫
E 1 3 " J 32 = ( 3 ) J 12 J 66 = 1 ⌫ 1
⌫¯2 (1 "3 ) ⇢3
" #
˙ f ) @⇡f ̇(⇡ e 3 ) @yd
J33 = ⇥(⇡ 3 + Kd (↵ + n + i3 ) r ̇(⇡e3 ) 1 (1 !3 )
@⇡e |(!3 , 3 ,⇡e3 ,"3 ,✓3 ,⇢3 ) ⌫ ✓3 @⇡e |(!3 , 3 ,⇡e3 ,"3 ,✓3 ,⇢3 )
✓ ◆ ✓ ◆
!3 ⌫ !3 ⌫
J34 = ( E + c )⌫1 ( i3 ) (↵ + n + ) J35 = ( E + c )⌫1 J45
µ "3 µ"3
"3
J44 = (↵ + n + ) J45 = (⇡e3 )µ"3 + ⌫ (1 "3 ) J65
⇢3
(✓RD ⇢3 )✓3⇤ ̇(⇡e3 ) RD c 1 ⌫ (✓RD ⇢3 )(⇡e3 )
J63 = J64 = J65 =
⌫¯2 (1 "3 ) ⌫¯2 (1 "3 ) ⌫¯2 (1 "3 )
(55)
The characteristic polynomial of this matrix is: P (X) = a6 X 6 + a5 X 5 + a4 X 4 + a3 X 3 + a2 X 2 +
a1 X + a0 with:

a6 = 1
a5 = (J11 + J33 + J44 + J66 )
a4 = J11 (J33 + J44 + J66 ) + J33 (J44 + J66 ) J56 J65 + J44 J66 J46 J64 J23 J32
a3 = J11 [J33 (J44 + J66 ) J56 J65 + J44 J66 J46 J64 ] + J44 J56 J65 J64 J45 J56 J31 J12 J23 J16 J31 J63
+ J33 [J56 J65 J44 J66 + J64 J46 ] J32 J63 J26 + J32 J11 J23 + J32 J23 (J44 + J66 )
a2 = J11 [J44 J56 J65 J64 J45 J56 + J33 (J56 J65 J44 J66 + J64 J46 )] J33 (J44 J56 J65 + J64 J45 J56 )
J32 J11 [J23 (J44 + J66 ) J63 J26 ] + J32 [J23 (J56 J65 J66 J44 + J46 J64 ) + J63 J26 J44 ]
+ J31 J12 J23 (J44 + J66 ) + J24 J31 J12 J63 + J31 J63 (J14 J46 + J44 J16 + !3 ⌫(⇡e3 )J56 )
a1 = J11 J33 (J44 J56 J65 + J64 J45 J56 ) J32 J11 [J23 (J56 J65 J44 J66 + J64 J46 ) J63 J24 J46 + J65 J26 J44 ]
J32 J23 J44 J56 J65 + J32 J23 J64 J45 J56 + J32 J63 J24 J45 J56 + J31 J12 J23 (J56 J65 J44 J66 + J46 J64 )
J31 J12 J63 (J24 J46 + J26 J44 ) + J31 J63 (J14 J45 J56 + J44 J56 !3 ⌫(⇡e3 )
a0 = J32 J11 [ J23 J44 J56 J65 + J23 J64 J45 J56 + J63 J24 J45 J56 ] + J31 J12 J23 ( J44 J56 J65 + J64 J45 J56 )
J31 J12 J63 J24 J45 J56 .
(56)

223
The stability of the equilibrium is guaranteed by the Routh-Hurwitz’s criterion if and only if:

a6 0
a5 0
a5 a4 a3 a6
b1 = 0
a5 (57)
b1 a 3 a 5 b2 a3 a2 a1 a4
c1 = 0 where b2 =
b1 a3
c 1 b2 c 2 b1 b2 a 1 a0 a3
d1 = >0 where c2 = .
c1 b2
These conditions are satisfied for a wide range of parameters and, in particular, by the parameteri-
zation of our Leading Example.
Bibliography

[Abel, 2013] Abel, W. (2013). Agricultural Fluctuations in Europe: From the Thirteenth to twentieth
centuries. Routledge.

[Acemoglu et al., 2012] Acemoglu, D., Aghion, P., Bursztyn, L., and Hemous, D. (2012). The
environment and directed technical change. American Economic Review, 102(1):131–66.

[Adjemian et al., 2011] Adjemian, S., Bastani, H., Juillard, M., Mihoubi, F., Perendia, G., Ratto,
M., and Villemot, S. (2011). Dynare: Reference manual, version 4.

[Akerlof and Stiglitz, 1969] Akerlof, G. and Stiglitz, J. E. (1969). Capital, Wages and Structural
Unemployment. Economic Journal, 1969, vol. 79, issue 314, 269-81.

[Alex Mara and Rakoto Joseph, 2008] Alex Mara, T. and Rakoto Joseph, O. (2008). Comparison
of some efficient methods to evaluate the main e↵ect of computer model factors. Journal of
Statistical Computation and Simulation, 78(2):167–178.

[An and Schorfheide, 2007] An, S. and Schorfheide, F. (2007). Bayesian analysis of dsge models.
Econometric Reviews, 26(2-4):113–172.

[Anderson et al., 2014] Anderson, B., Borgonovo, E., Galeotti, M., and Roson, R. (2014). Uncer-
tainty in climate change modeling: Can global sensitivity analysis be of help? Risk Analysis,
Vol. 34, No. 2, DOI: 10.1111/risa.12117.

[Annicchiarico, 2010] Annicchiarico, B. (2010). Microfoundations of dsge models: Lecture.

[Arrow et al., 2004] Arrow, K., Dasgupta, P., Goulder, L., Daily, G., Ehrlich, P., Heal, G., Levin,
S., Mäler, K.-G., Schneider, S., Starrett, D., et al. (2004). Are we consuming too much? Journal
of Economic Perspectives, 18(3):147–172.

[Arrow, 1999] Arrow, K. J. (1999). Amartya k. sen’s contributions to the study of social welfare.
The Scandinavian Journal of Economics, 101(2):163–172.

[Aruoba et al., 2006] Aruoba, S. B. B., Fernández-Villaverde, J., and Rubio-Ramı́rez, J. F. (2006).
Comparing solution methods for dynamic equilibrium economies. Journal of Economic Dynamics
and Control 30 (12), 2477-2508.

[Athanassoglou, 2014] Athanassoglou, S. (2014). Multidimensional welfare rankings under weight

imprecision: a social choice perspective. Social Choice and Welfare 44, 719-744.

[Auder et al., 2010] Auder, B., Iooss, B., and Marquès, M. (2010). Computer experiments with
functional outputs: Global sensitivity analysis and metamodeling. Procedia - Social and Behav-
ioral Sciences, 2(6):7583 – 7584.

225
BIBLIOGRAPHY

[Augier, 2018] Augier, S. (2018). Finding basin of attraction: Di↵erent algorithms. Work in
progress, Agence Française de développement.

[Ayres and Warr, 2005] Ayres, R. U. and Warr, B. (2005). Accounting for growth: the role of
physical work. Structural Change and Economic Dynamics, 16(2):181–209.

[Bank of England, 2020] Bank of England, . (2020). Monetary policy report. Technical report,
Bank of England.

[Bardaji et al., 2017] Bardaji, J., Campagne, B., Khder, M.-B., La↵éter, Q., Simon, O., Dufer-
nez, A. S., Elezaar, C., Leblanc, P., Masson, E., and Partouche, H. (2017). The MESANGE
macroeconometric model: re-estimation and innovations. Technical report, Institut National de
la Statistique et des Etudes Economiques.

[Barnosky et al., 2012] Barnosky, A. D., Hadly, E. A., Bascompte, J., Berlow, E. L., Brown, J. H.,
Fortelius, M., Getz, W. M., Harte, J., Hastings, A., Marquet, P. A., et al. (2012). Approaching
a state shift in earth’s biosphere. Nature, 486(7401):52.

[Bastidas et al., 2019] Bastidas, D., Fabre, A., and Mc Isaac, F. (2019). Minskyan classical growth
cycles: stability analysis of a stock-flow consistent macrodynamic model. Mathematics and Fi-
nancial Economics, 13(3):359–391.

[Bastidas et al., 2017] Bastidas, D., Giraud, G., and McIsaac, F. (2017). Growth, underemploy-
ment, capital and international trades in brazil: a multisectorial, monetary and stock-flow con-
sistent analysis. Workingpaper, AFD.

[Beckerman et al., 2007] Beckerman, W., Hepburn, C., et al. (2007). Ethics of the discount rate
in the stern review on the economics of climate change. World Economics-Henley on Thames-,
8(1):187.

[Bell, 1988] Bell, D. E. (1988). Decision making: Descriptive, normative, and prescriptive interac-
tions. cambridge university Press.

[Bellman, 1957] Bellman, R. E. (1957). Dynamic programming. Rand Corporation. Princeton

University Press. ISBN 978-0-691-07951-6.

[Berveiller, 2005] Berveiller, M. (2005). Stochastic finite element analysis: intrusive and non-
intrusive approaches for reliability analysis. Theses, Université Blaise Pascal - Clermont-Ferrand
II.

[Beven, 1993] Beven, K. (1993). Prophecy, reality and uncertainty in distributed hydrological mod-
elling. Advances in water resources, 16(1):41–51.

[Beven and Beven, 2001] Beven, K. J. and Beven, K. (2001). Rainfall-runo↵ modelling. Wiley
Online Library.

[Bjornland, 2000] Bjornland, H. C. (2000). Var models in macroeconomic research. Statistics Nor-
way, 14:1–29.

[Blanchard, 2009] Blanchard, O. (2009). The state of macro. Annu. Rev. Econ., 1(1):209–228.

[Blanchard, 2016] Blanchard, O. (2016). Do dsge models have a future? Revista de Economı́a
Institucional, 18(35):39–46.

226
BIBLIOGRAPHY

[Blatman and Sudret, 2010] Blatman, G. and Sudret, B. (2010). Efficient computation of global
sensitivity indices using sparse polynomial chaos expansions. Reliability Engineering & System
Safety, 95:1216–1229.

[Borgonovo et al., 2011] Borgonovo, E., Castaings, W., and Tarantola, S. (2011). Moment indepen-
dent importance measures: New results and analytical test cases. Risk Analysis, 31(3):404–428.

[Borgonovo and Plischke, 2016] Borgonovo, E. and Plischke, E. (2016). Sensitivity analysis: A
review of recent advances. European Journal of Operational Research, 248(3):869–887.

[Bouleau, 2008] Bouleau, N. (2008). Error calculus for finance and physics. the lan-
guage of dirichlet forms. Berlin, Boston: De Gruyter. Retrieved 14 Jun. 2019, from
https://www.degruyter.com/view/product/36729.

[Bouleau, 2019] Bouleau, N. (2019). Théorie des erreurs. Spartacus-Cassini.

[Bouleau, 2022] Bouleau, N. (2022). The mathematics of errors. Springer International Publishing.

[Bovari et al., 2020] Bovari, E., Giraud, G., and McIsaac, F. (2020). Financial impacts of climate
change mitigation policies and their macroeconomic implications: a stock-flow consistent ap-
proach. Climate Policy, 20(2):179–198.

[Bovari et al., 2018] Bovari, E., McIsaac, F., and Giraud, G. (2018). Coping with Collapse : A
Stock-Flow Consistent Monetary Macrodynamics of Global Warming. Ecological Economics Vol-
ume 147, pages 383–398.

[Box, 1976] Box, G. E. (1976). Science and statistics. Journal of the American Statistical Associa-
tion, 71(356):791–799.

[Box and Behnken, 1960] Box, G. E. P. and Behnken, D. W. (1960). Some new three level designs
for the study of quantitative variables. Technometrics, 2(4):455–475.

[Boyer, 2012] Boyer, R. (2012). La discipline économique des années 1930 à nos jours. Le Débat,
2:148–166.

[BP, 2019] BP (2019). 2019 bp statistical review of world energy. Technical Report 2019 - 68th,
BP Report 2019.

[Bratley et al., 1992] Bratley, P., Fox, B. L., and Niederreiter, H. (1992). Implementation and tests
of low-discrepancy sequences. ACM Trans. Model. Comput. Simul., 2(3):195–213.

[Brenner, 1976] Brenner, R. (1976). Agrarian class structure and economic development in pre-
industrial europe. Past & present, 70(1):30–75.

[Burke et al., 2015] Burke, M., Hsiang, S. M., and Miguel, E. (2015). Global non-linear e↵ect of
temperature on economic production. Nature, 527(7577):235.

[Burman and Plackett, 1946] Burman, J. P. and Plackett, R. L. (1946). The Design of Optimum
Multifactorial Experiments. Biometrika, 33(4):305–325.

[Butler et al., 2014] Butler, M. P., Reed, P. M., Fisher-Vanden, K., Keller, K., and Wagener, T.
(2014). Identifying parametric controls and dependencies in integrated assessment models using
global sensitivity analysis. Environmental Modelling & Software 59 (2014) 10-29.

227
BIBLIOGRAPHY

[Cameron and Martin, 1947] Cameron, R. H. and Martin, W. T. (1947). The orthogonal develop-
ment of non-linear functionals in series of fourier-hermite functionals. Annals of Mathematics,
48(2):385–392.

[Campbell et al., 2006] Campbell, K., McKay, M. D., and Williams, B. J. (2006). Sensitivity anal-
ysis when model outputs are functions. Reliability Engineering and System Safety, 91(10):1468 –
1472.

[Campolongo, 2005] Campolongo, F. (2005). Screening methods in sensitivity analysis: The morris
method and its applications. JRC-Ispra (Joint research centre, European commission).

[Campolongo et al., 2006] Campolongo, F., Cariboni, J., and Schoutens, W. (2006). The impor-
tance of jumps in pricing european options. Reliability Engineering & System Safety, 91:1148–
1154.

[Campolongo et al., 1999] Campolongo, F., Saltelli, A., Jensen, N., Wilson, J., and Hjorth, J.
(1999). The role of multiphase chemistry in the oxidation of dimethylsulphide (dms). a lati-
tude dependent analysis. Journal of Atmospheric Chemistry, 32(3):327–356.

[Canfin et al., 2016] Canfin, P., GrandJean, A., and Mestrallet, G. (2016). Rapport de la mission,
propositions pour des prix du carbone alignés avec l’accord de paris.

[Canova, 1995] Canova, F. (1995). Sensitivity analysis and model evaluation in simulated dynamic
general equilibrium economies. International Economic Review, 36(2):477–501.

[Cariboni et al., 2007] Cariboni, J., Gatelli, D., Liska, R., and Saltelli, A. (2007). The role of
sensitivity analysis in ecological modelling. Ecological modelling, 203(1-2):167–182.

[Carney, 2015] Carney, M. (2015). Breaking the tragedy of the horizon - climate change and financial
stability - speech by mark carney. https://www.bankofengland.co.uk/speech/2015/breaking-the-
tragedy-of-the-horizon-climate-change-and-financial-stability. 2015-09-29.

[Carney, 2019] Carney, M. (2019). A new report from the network for greening the financial sys-
tem. https://www.theguardian.com/commentisfree/2019/apr/17/the-financial-sector-must-be-
at-the-heart-of-tackling-climate-change. 2019.

[Cass and Stiglitz, 1969] Cass, D. and Stiglitz, J. (1969). The implications of alternative saving
and expectations hypotheses for choices of technique and patterns of growth. Journal of Political
Economy.

[Catton, 1982] Catton, W. R. (1982). Overshoot: The ecological basis of revolutionary change.
University of Illinois Press.

[Chastaing et al., 2013] Chastaing, G., Prieur, C., and Gamboa, F. (2013). Generalized sobol sen-
sitivity indices for dependent variables: numerical methods. Journal of Statistical Computation
and Simulation, 85.

[Chongvilaivan, 2012] Chongvilaivan, A. (2012). Thailand’s 2011 flooding: Its impact on direct
exports and global supply chains. Technical report, ARTNeT working paper series.

[Christiaens, 2018] Christiaens, T. (2018). Can we avoid another financial crisis? (steve keen).
Marx & Philosophy Review of Books.

228
BIBLIOGRAPHY

[Christiano et al., 2010] Christiano, L. J., Trabandt, M., and Walentin, K. (2010). Dsge models
for monetary policy analysis. In Handbook of monetary economics, volume 3, pages 285–367.
Elsevier.

[Cleveland et al., 2015] Cleveland, S., Schuwerk, R., and Weber, C. (2015). Carbon asset risk: From
rhetoric to action.

[Cobb and Douglas, 1928] Cobb, C. W. and Douglas, P. H. (1928). A theory of production. The
American Economic Review, Vol. 18, No. 1, Supplement, Papers and Proceedings of the Fortieth
Annual Meeting of the American Economic Association (Mar., 1928), pp. 139- 165.

[Collard, 2012] Collard, D. (2012). Generations of economists. Routledge.

[Corson, 1992] Corson, G. (1992). Input-output sensitivity of building energy simulations. ASHRAE
Winter Meeting, 98(1), pp. 618-626.

[Costa Lima and Grasselli, 2014] Costa Lima, B. and Grasselli, M. R. (2014). Destabilizing a sta-
ble crisis: employment persistence and government intervention in macroeconomics. Structural
Change and Economic Dynamics, Volume 30, Pages 30-51.

[Crosetto and Tarantola, 2001] Crosetto, M. and Tarantola, S. (2001). Uncertainty and sensitiv-
ity analysis: tools for gis-based model implementation. International Journal of Geographical
Information Science, 15(5):415–437.

[Cukier et al., 1973] Cukier, R., Fortuin, C., Shuler, K., Petschek, A., and Schaibly, J. (1973).
Study of the sensitivity of coupled reaction systems to uncertainties in rate coefficients: I theory.
The Journal of chemical physics, 59(8), pp. 3873.

[Cukier et al., 1978] Cukier, R., Levine, H., and Shuler, K. (1978). Nonlinear sensitivity analysis
of multiparameter model systems. Journal of Computational Physics, 26(1):1 – 42.

[Cukier et al., 1975] Cukier, R. I., Schaibly, J. H., and Shuler, K. E. (1975). Study of the sensitivity
of coupled reaction systems to uncertainties in rate coefficients. iii. analysis of the approximations.
The Journal of Chemical Physics, 63(3):1140–1149.

[De Rocquigny, 2008] De Rocquigny, E. (2008). Uncertainty in Industrial Practice. John Wiley &
Sons, Ltd, 1 edition.

[De Wit, 2001] De Wit, M. (2001). Uncertainty in predictions of thermal comfort in buildings. Delft
University, PhD thesis.

[Del Negro and Schofheide, 2008] Del Negro, M. and Schofheide, F. (2008). Forming priors for dsge
models (and how it a↵ects the assessment of nominal rigidities). Journal of Monetary Economics,
55(7):1191 – 1208.

[Den Haan et al., 2011] Den Haan, W. J., Judd, K. L., and Juillard, M. (2011). Computational
suite of models with heterogeneous agents ii: Multi-country real business cycle models. Journal
of Economic Dynamics and Control 35 (2), 175-177.

[Desai et al., 2006] Desai, M., Henry, B., and Mosley, A.and Pemberton, M. (2006). Clarification
of the goodwin model of the growth cycle. J. Econ. Dyn. Control 30(12), 2661-2670.

229
BIBLIOGRAPHY

[Diamond and Van Dijk, 2008] Diamond, J. M. and Van Dijk, M. (2008). Collapse: How societies
choose to fail or succeed. Anthropos: International Review of Anthropology and Linguistics,
103(2):587.

[Dietz and Stern, 2015] Dietz, S. and Stern, N. (2015). Endogenous growth, convexity of damage
and climate risk: How nordhaus’ framework supports deep cuts in carbon emissions. Economic
Journal, 0(583):574-620.

[Dossetto and Giraud, 2020] Dossetto, E. and Giraud, G. (2020). Money velocity in an imperfectly
competitive, stock-flow consistent dynamics. forthcoming.

[Dubus et al., 2003] Dubus, I. G., Brown, C. D., and Beulke, S. (2003). Sensitivity analyses for
four pesticide leaching models. Pest Management Science, 59(9):962–982.

[Eddington, 2012] Eddington, A. (2012). The nature of the physical world: Gi↵ord lectures (1927).
Cambridge University Press.

[EPA, 2003] EPA (2003). Crem draft guidance for environmentalmodels. council for regulatory
environmental modelling. Technical report, CREM.

[Erb and Michaels, 1998] Erb, R. S. and Michaels, G. S. (1998). Sensitivity of biological models
to errors in parameter estimates. Pacific Symposium on Biocomputing. Pacific Symposium on
Biocomputing, pages 53–64.

[Erceg et al., 2006] Erceg, C., Guerrieri, L., and Gust, C. (2006). Sigma: A new open economy
model for policy analysis. International Journal of Central Banking, 2(1).

[European Central Bank, 2017] European Central Bank, . (2017). Economic bulletin issue 3.
https://www.ecb.europa.eu/pub/pdf/other/ebbox2017030 3.en.pdf.

[European Central Bank, 2021] European Central Bank, . (2021). The ecb‘s negative interest
rate. https://www.ecb.europa.eu/ecb/educational/explainers/tell-me-more/html/why-negative-
interest-rate.en.html.

[European Commision, 2009] European Commision, . (2009). Global sensitivity analysis for macroe-
conomic models. European Commission : Joint Research Centre - Institute for the Protection
and Security of the Citizen (IPSC). Available at: http://ipsc.jrc.ec.europa.eu/events.php?idx=79.

[Fajraoui, 2014] Fajraoui, N. (2014). Analyse de sensibilite globale et polynomes du chaos pour
l’estimation des parametres : application aux transferts en milieu poreux. PhD thesis Ecole
Doctorale Sciences de la Terre et Environnement Strasbourg.

[Federal Reserve, 2021a] Federal Reserve, . (2021a). The liquid-

ity trap: An alternative explanation for today’s low inflation.
https://www.federalreserve.gov/aboutthefed/files/newyorkfinstmt2020.pdf.

[Federal Reserve, 2021b] Federal Reserve, . (2021b). Open market operations. Technical report,
Board of Governors of the Federal Reserve System.

[Felipe and McCombie, 2013] Felipe, J. and McCombie, J. S. (2013). The Aggregate Production
Function and the Measurement of Technical Change: ŠNot Even Wrong. Edward Elgar Publish-
ing.

230
BIBLIOGRAPHY

[Fenton and Murchison, 2006] Fenton, P. and Murchison, S. (2006). Totem: The bank of canada’s
new projection and policy-analysis model. Bank of Canada Review, 2006:5–18.

[Ferraton, 2008] Ferraton, C. (2008). Les valeurs guident et accompagnent notre recherche,
l’institutionnalisme de Myrdal. ENS éditions.

[Ferraton, 2011] Ferraton, C. (2011). L’objectivité en économie politique selon gunnar myrdal.
Revue de philosophie économique, 12(2):29–52.

[Fisher, 1912] Fisher, I. (1912). “the equation of exchange” for 1911, and forecast. The American
Economic Review, 2(2):302–319.

[Fisher, 1933] Fisher, I. (1933). The debt-deflation theory of great depressions. Econometrica:
Journal of the Econometric Society, pages 337–357.

[Friedman M., 1963] Friedman M., S. A. (1963). A monetary history of the united states.

[Fruth et al., 2014] Fruth, J., Roustant, O., and Kuhnt, S. (2014). Total interaction index: A
variance-based sensitivity index for second-order interaction screening. Journal of Statistical
Planning and Inference, 147:212–223.

[Galbraith, 2009] Galbraith, J. K. (2009). Pourrait-on enfin avancer sur la réforme de l’enseignement
de l’économie ? Revue du MAUSS, pages 261–265.

[Galiánd Monacelli, 2005] Gali,́ J. and Monacelli, T. (2005). Monetary policy and exchange rate
volatility in a small open economy. Review of Economic Studies 72.

[Georgescu-Roegen, 1993] Georgescu-Roegen, N. (1993). The entropy law and the economic prob-
lem. Valuing the Earth: Economics, ecology, ethics, pages 75–88.

[Ghanem and Spanos, 1991] Ghanem, R. and Spanos, P. (1991). Stochastic finite elements : a
spectral approach. Berlin journal Springer.

[Ghersi, 2020] Ghersi, F. (2020). IMACLIM-3ME. Centre international de recherche sur

l’environnement et le développement.

[Gilquin et al., 2016] Gilquin, L., Capelle, T., Arnaud, E., and Prieur, C. (2016). Sensitivity analy-
sis and optimisation of a land use and transport integrated model. Journal de la Société Française
de Statistique hal-01291774.

[Giraud, 2014] Giraud, G. (2014). Illusion financière (3e édition revue et augmentée). Editions de
l’Atelier.

[Giraud and Grasselli, 2017] Giraud, G. and Grasselli, M. (2017). The macrodynamics of household
debt, growth, and inequality. Technical report, mimeo: Centre déconomie de la Sorbonne, Paris.

[Giraud and Grasselli, 2019] Giraud, G. and Grasselli, M. (2019). Household debt: The missing
link between inequality and secular stagnation. Journal of Economic Behavior & Organization.

[Giraud and Kahraman, 2014] Giraud, G. and Kahraman, Z. (2014). How dependent is growth
from primary energy? the dependency ratio of energy in 33 countries (1970-2011). Documents
de Travail du Centre d’Economie de la Sorbonne.

231
BIBLIOGRAPHY

[Giraud and Pottier, 2016] Giraud, G. and Pottier, A. (2016). Debt-deflation versus the liquidity
trap: the dilemma of nonconventional monetary policy. Economic Theory, 62(1-2):383–408.

[Godard, 2007] Godard, O. (2007). Le rapport stern sur l’économie du changement climatique
était-il une manipulation grossière de la méthodologie économique ? Revue d’économie politique,
117(4):475–506.

[Godley and Lavoie, 2006] Godley, W. and Lavoie, M. (2006). Monetary economics: an integrated
approach to credit, money, income, production and wealth. Springer.

[Goldstone, 1991] Goldstone, J. A. (1991). Revolution and rebellion in the early modern world. Univ
of California Press.

[Goodwin, 1967] Goodwin, R. M. (1967). A growth cycle, in: Feinstein, c. h. (ed.), socialism,
capitalism and economic growth . Cambridge University Press, London, pp. 54-58.

[Gordon, 2011] Gordon, R. J. (2011). The history of the phillips curve: Consensus and bifurcation.
Economica, 78(309):10–50.

[Gordon, 2013] Gordon, R. J. (2013). The phillips curve is alive and well: Inflation and the nairu
during the slow recovery. Technical report, National Bureau of Economic Research.

[Gourieroux et al., 2000] Gourieroux, C., Laurent, J. P., and Scaillet, O. (2000). Sensitivity analysis
of values at risk. Journal of Empirical Finance, 7(3-4):225–245.

[Grandjean and Giraud, 2017] Grandjean, A. and Giraud, G. (2017). Comparaison des modèles
météorologiques, climatiques et économiques : quelles capacités, quelles limites, quels usages ?
WP.

[Grasselli and Costa Lima, 2012] Grasselli, M. R. and Costa Lima, B. (2012). An Analysis of the
Keen Model for Credit Expansion, Asset Price Bubbles and Financial Fragility. Mathematical
and Financial Economics - Volume 6, pages 191–210.

[Grasselli and Maheshwari, 2017] Grasselli, M. R. and Maheshwari, A. (2017). A comment on test-
ing goodwin: growth cycles in ten oecd countries? Cambridge Journal of Economics, 41(6):1761–
1766.

[Grasselli and Nguyen Huu, 2016] Grasselli, M. R. and Nguyen Huu, A. (2016). Inventory growth
cycles with debt-financed investment. Working paper.

[Grasselli and Nguyen-Huu, 2018] Grasselli, M. R. and Nguyen-Huu, A. (2018). Inventory growth
cycles with debt-financed investment. Structural Change and Economic Dynamics, 44:1–13.

[Guerrien and Gun, 2014] Guerrien, B. and Gun, O. (2014). En finir, pour toujours, avec la fonction
de production agrégée. jesus felipe and john sl mccombie, the aggregate production function and
the measurement of technical change: Not even wrong.

[Guesnerie, 2005] Guesnerie, R. (2005). Assessing Rational Expectations 2: eductive stability in

economics. MIT press.

[Guesnerie, 2013] Guesnerie, R. (2013). L’économie saisie par le doute. 2013-11-13.

232
BIBLIOGRAPHY

[Gurley and Shaw, 1960] Gurley, J. G. and Shaw, E. S. (1960). Money in a theory of finance (wash-
ington, dc: Brookings institution). Transactions Demand for Cash. The Review of Economics
and Statistics, 38(3).

[Hamby, 1994] Hamby, D. (1994). A review of techniques for parameter sensitivity analysis of
environmental models. Environmental Monitoring and Assessment, 32, pp. 135-154.

[Harenberg et al., 2017] Harenberg, D., Marelli, S., Sudret, B., and Winschel, V. (2017). Uncer-
tainty quantification and global sensitivity analysis for economic models. Economics Work-
ing Paper Series, No. 17/265, ETH Zurich, CER-ETH - Center of Economic Research,
Zurich,http://dx.doi.org/10.3929/ethz-a-010820135.

[Harenberg et al., 2019] Harenberg, D., Marelli, S., Sudret, B., and Winschel, V. (2019). Uncer-
tainty quantification and global sensitivity analysis for economic models. Quantitative Economics,
10(1):1–41.

[Harrison et al., 2005] Harrison, R., Nikolov, K., Quinn, M., Ramsay, G., Scott, A., and Thomas, R.
(2005). The Bank of England Quarterly Model. Technical report, Bank of England, Threadneedle
Street, London, EC2R 8AHs.

[Heal and Millner, 2014] Heal, G. and Millner, A. (2014). Reflections: Uncertainty and decision
making in climate change economics. Review of Environmental Economics and Policy 8, 120-
137.

[Helton et al., 2006] Helton, J., Johnson, J., Sallaberry, C., and Storlie, C. (2006). Survey of
sampling- based methods for uncertainty and sensitivity analysis. Reliability Engineering and
System Safety, 91(10), pp. 1175-120.

[Hodgson, 2012] Hodgson, G. M. (2012). From pleasure machines to moral communities: an evolu-
tionary economics without homo economicus. University of Chicago Press.

[Hoes and De Vann, 2005] Hoes, P. and De Vann, C. (2005). De gevoeligheidsanalyse van gebouw
simulations. Masterproject 3. Technische Universiteit Eindhoven, Unit Building Physics and
Systems (BPS).

[Hooker, 2007] Hooker, G. (2007). Generalized functional anova diagnostics for high-dimensional
functions of dependent variables. Journal of Computational and Graphical Statistics, 16(3):709–
732.

[Hopfe et al., 2007] Hopfe, C., Hensen, J., and Plokker, W. (2007). Uncertainty and sensitivity
analysis for detailed design support. Proceedings of the 10th International IBPSA Building Sim-
ulation Conference 2007.

[Hoppensteadt, 2006] Hoppensteadt, F. (2006). Predator-prey model. Scholarpedia, 1(10):1563.

[Hurwitz, 1963] Hurwitz, A. (1963). Ueber die bedingungen, unter welchen eine gleichung nur
wurzeln mit negativen reellen teilen besitzt. In Mathematische Werke, pages 533–545. Springer.

[Iman, 1999] Iman, R. (1999). Appendix a: Latin hypercube sampling. Encyclopedia of Statistical
Sciences, Update, 3, pp. 408-411.

[Iooss, 2011] Iooss, B. (2011). Revue sur l’analyse de sensibilité globale de modèles numériques.
Journal de la Societe Française de Statistique, 152(1):1–23.

233
BIBLIOGRAPHY

[Iooss and Lemaı̂tre, 2015] Iooss, B. and Lemaı̂tre, P. (2015). A review on global sensitivity anal-
ysis methods. In Meloni, C. and Dellino, G., editors, Uncertainty management in Simulation-
Optimization of Complex Systems: Algorithms and Applications. Springer.

[Iooss and Ribatet, 2009] Iooss, B. and Ribatet, M. (2009). Global sensitivity analysis of computer
models with functional inputs. Reliability Engineering and System Safety, 94(7):1194 – 1204.

[IPCC, 2014] IPCC (2014). Climate change 2014 - working group iii contribution to the fifth
assessment report of the intergovernmental panel on climate change. IPCC.

[IPCC, 2018] IPCC (2018). Global Warming of 1.5 C: An IPCC Special Report on the Impacts
of Global Warming of 1.5 C Above Pre-industrial Levels and Related Global Greenhouse Gas
Emission Pathways, in the Context of Strengthening the Global Response to the Threat of Climate
Change, Sustainable Development, and E↵orts to Eradicate Poverty. Intergovernmental Panel on
Climate Change.

[IPCC, 2021] IPCC (2021). AR6 Climate Change 2021: The Physical Science Basis. Intergovern-
mental Panel on Climate Change - WG I.

[Iskrev, 2010] Iskrev, N. (2010). Local identification in dsge models. Journal of Monetary Eco-
nomics, 57(2):189 – 202.

[Jacobi et al., 2019] Jacobi, L., Joshi, M., and Zhu, D. (2019). Automated sensitivity computations
for bayesian markov chain monte carlo inference: A new approach for prior robustness and
convergence analysis. Working progress.

[Jacques, 2011] Jacques, J. (2011). Pratique de l’analyse de sensibilité : comment évaluer l’impact
des entrées aléatoires sur la sortie d’un modèle mathématique. PUB. IRMA, LILLE 2011, Vol.
71, No III, p 9.

[Janon et al., 2014] Janon, A., Klein, T., Lagnoux, A., Nodet, M., and Prieur, C. (2014). Asymp-
totic normality and efficiency of two sobol index estimators. ESAIM: Probability and Statistics,
18:342–364.

[Jansen, 1999] Jansen, M. J. (1999). Analysis of variance designs for model output. Computer
Physics Communications, 117(1-2):35–43.

[Jany-Catrice and Méda, 2015] Jany-Catrice, F. and Méda, D. (2015). Faut-il attendre la croissance
? La Documentation française.

[Joe and Kuo, 2008] Joe, S. and Kuo, F. (2008). Constructing sobol sequences with better two-
dimensional projections. SIAM Journal on Scientific Computing, 30(5):2635–2654.

[Johansen, 1959] Johansen, L. (1959). Substitution versus fixed production coefficients

in the theory of economic growth : A synthesis. Econometrica, 27(2), 157-176.
http://www.jstor.org/stable/1909440.

[Kammen et al., 1994] Kammen, D. M., Smith, K. R., Rambo, A. T., and Khalil, M. (1994). Prein-
dustrial human environmental impacts: are there lessons for global change science and policy?
Chemosphere.

[Karna et al., 2012] Karna, D. S., D Ph, K., and P Scholar, A. (2012). An overview on taguchi
method. International Journal of Engineering and Mathematical Sciences, Volume 1, pp.11-18.

234
BIBLIOGRAPHY

[Kaya, 2014] Kaya (2014). Equation de kaya par jm jancovici. https://jancovici.com/changement-

climatique/economie/quest-ce-que-lequation-de-kaya/.

[Keen, 1995] Keen, S. (1995). Finance and economic breakdown: Modeling minsky’s financial in-
stability hypothesis. Journal of Post-Keynesian Economics, 1995, vol. 17, issue 4, 607-635.

[Keen, 2011] Keen, S. (2011). Debunking economics: The naked emperor dethroned? Zed Books
Ltd.

[Keen, 2013] Keen, S. (2013). A monetary minsky model of the great moderation and the great
recession. Journal of Economic Behavior & Organization, 86:221–235.

[Keynes, 1936] Keynes, J. M. (1936). The general theory of employment, interest, and money - ed
2018. Springer.

[Kioutsioukis et al., 2005] Kioutsioukis, I., Melas, D., Zerefos, C., and Ziomas, I. (2005). Efficient
sensitivity computations in 3d air quality models. Computer Physics Communications, 167(1):23–
33.

[Krugman and Eggertsson, 2013] Krugman, P. and Eggertsson, G. (2013). Debt, deleveraging, and
the liquidity trap: A fisher-minsky-koo approach. The Quaterly Journal of Economics.

[Kucherenko, 2010] Kucherenko, S. . (2010). The importance of being global. application of global
sensitivity analysis in monte carlo option pricing. Convergence (London, England).ISSN: 1354-
8565. Volume: 1 Issue: 2 Page: 2.

[Kucherenko and Shah, 2007] Kucherenko, S. and Shah, N. (2007). The importance of being global
application of global sensitivity analysis in monte carlo option pricing. Imperial College London,
SW7 2AZ, UK,.

[Kucherenko et al., 2012] Kucherenko, S., Tarantola, S., and Annoni, P. (2012). Estimation of
global sensitivity indices for models with dependent variables. Computer Physics Communica-
tions, 183:937–946.

[Kuroda, 2016] Kuroda, H. (2016). The battle against deflation: The evolution of monetary policy
and japan’s experience. Speech given at Columbia University, New York, April, 13.

[Ladurie, 1987] Ladurie, E. L. R. (1987). The French Peasantry, 1450-1660. Univ of California
Press.

[La↵argue, 2012] La↵argue, J.-P. (2012). Une comparaison des modèles macro-économétriques
et dsge dans l’évaluation des politiques économiques: une comparaison basée sur les modèles
mésange et egée. Economie et statistique, 451(1):45–68.

[Lam and Hui, 1995] Lam, J. and Hui, S. (1995). Sensitivity analysis of energy performance of
office buildings. Building and Environment, 31(1), pp. 27-39.

[Lamboni et al., 2013] Lamboni, M., Iooss, B., Popelin, A.-L., and Gamboa, F. (2013). Derivative-
based global sensitivity measures: General links with sobol’ indices and numerical tests. Mathe-
matics and Computers in Simulation.

[Laxton et al., 2010] Laxton, D., Mursula, S., Kumhof, M., and Muir, D. (2010). The Global
Integrated Monetary and Fiscal Model (GIMF): Theoretical Structure. Technical Report 10/34,
International Monetary Fund.

235
BIBLIOGRAPHY

[Le Gratiet et al., 2018] Le Gratiet, L., Cannamela, C., and Iooss, B. (2018). A bayesian approach
for global sensitivity analysis of (multifidelity) computer codes. SIAM/ASA Journal on Uncer-
tainty Quantification, 2(1):336–363.

[Le Gratiet and Garnier, 2014] Le Gratiet, L. and Garnier, J. (2014). Recursive co-kriging model
for design of computer experiments with multiple levels of fidelity. International Journal for
Uncertainty Quantification, 4(5).

[Le Maı̂tre et al., 2002] Le Maı̂tre, O. P., Reagan, M. T., Najm, H. N., Ghanem, R. G., and Knio,
O. M. (2002). A Stochastic Projection Method for Fluid Flow. II. Random Process. Journal of
Computational Physics, 181:9–44.

[Lefebvre et al., 2010] Lefebvre, S., Roblin, A., Varet, S., and Durand, G. (2010). A methodological
approach for statistical evaluation of aircraft infrared signature. Reliability Engineering & System
Safety, 95(5):484–493.

[Leontief, 1951] Leontief, W. W. (1951). The structure of american economy, 1919-1939 : an em-
pirical application of equilibrium analysis. New York :. Oxford University Press,. 1951.. 264 p. :
Il., tablas. (Es). Edición ; 2 ed.

[Li et al., 2010] Li, G., Rabitz, H., Yelvington, P. E., Oluwole, O. O., Bacon, F., Kolb, C. E., and
Schoendorf, J. (2010). Global sensitivity analysis for systems with independent and/or correlated
inputs. The Journal of Physical Chemistry A, 114(19):6022–6032.

[Li and Zhang, 2007] Li, H. and Zhang, D. (2007). Probabilistic collocation method for flow in
porous media: Comparisons with other stochastic methods. Water Resources Research, 43(9).

[Liu and Owen, 2006] Liu, R. and Owen, A. B. (2006). Estimating mean dimensionality of analysis
of variance decompositions. Journal of the American Statistical Association, 101(474):712–721.

[Lomas and Eppel, 1992] Lomas, K. and Eppel, H. (1992). Sensitivity analysis techniques for build-
ing thermal simulation programs. Energy and Buildings, 19(1), pp. 21-44.

[Lotka, 1920] Lotka, A. J. (1920). Analytical note on certain rhythmic relations in organic systems.
Proceedings of the National Academy of Sciences, 6(7):410–415.

[Lubik and Schorfheide, 2005] Lubik, T. and Schorfheide, F. (2005). A bayesian look at new open
economy macroeconomics. NBER macroeconomics annual, 20:313–366.

[Lubik and Schorfheide, 2007] Lubik, T. A. and Schorfheide, F. (2007). Do central banks respond
to exchange rate movements? A structural investigation. Journal of Monetary Economics,
54(4):1069–1087.

[M. Sobol, 1990] M. Sobol, I. (1990). Sensitivity estimates for nonlinear mathematical models.
Matematicheskoe Modelirovanie, 2.

[MacDonald, 2002] MacDonald, I. (2002). Quantifying the e↵ects of uncertainty in building simu-
lation. Department of Mechanical Engineering. University of Strathclyde, PhD thesis.

[Macdonald et al., 1999] Macdonald, I., Clarke, J., and Strachan, P. (1999). Assessing uncertainty
in building simulation. Proceedings 6th Building Simulation 1999.

236
BIBLIOGRAPHY

[Makowski et al., 2006] Makowski, D., Naud, C., Jeu↵roy, M.-H., Barbottin, A., and Monod, H.
(2006). Global sensitivity analysis for calculating the contribution of genetic parameters to the
variance of crop model prediction. Reliability Engineering & System Safety, 91(10-11):1142–1147.

[Manganelli, 2004] Manganelli, S. (2004). Asset allocation by variance sensitivity analysis. Journal
of Financial Econometrics, 2:370–389.

[Mankiw, 2001] Mankiw, N. G. (2001). The inexorable and mysterious tradeo↵ between inflation
and unemployment. The Economic Journal, 111(471):45–61.

[Mankiw, 2016] Mankiw, N. G. (2016). Principles of economics. Cengage Learning.

[Mankiw and Romer, 1991] Mankiw, N. G. and Romer, D. (1991). New Keynesian Economics-Vol.
1: Imperfect Competition and Sticky Prices, volume 1. The MIT Press.

[Mara and Tarantola, 2011] Mara, T. and Tarantola, S. (2011). Variance-based sensitivity indices
for models with dependent inputs. Reliability Engineering and System Safety - RELIAB ENG
SYST SAFETY, 107.

[Martinez, 2011] Martinez, J. (2011). Analyse de sensibilité globale par décomposition de la vari-
ance. Presentation in Journée des GdR Ondes & Mascot, 13.

[Matthies and Keese, 2003] Matthies, H. G. and Keese, A. (2003). Galerkin methods for linear and
nonlinear elliptic stochastic partial di↵erential equations. Informatik-Berichte der Technischen
Universität Braunschweig, 2003-08.

[Mazzucato, 2011] Mazzucato, M. (2011). The entrepreneurial state. Soundings, 49(49):131–142.

[McCandless and Gustafson, 2017] McCandless, L. C. and Gustafson, P. (2017). A comparison of

bayesian and monte carlo sensitivity analysis for unmeasured confounding. Statistics in Medicine,
36(18):2887–2901.

[McIsaac, 2016] McIsaac, F. (2016). Energy and Money in New Frameworks for Macro-dynamics.
PhD thesis, Université Paris 1, Panthéon - Sorbonne.

[McIsaac, 2021] McIsaac, F. (2021). Testing goodwin with a stochastic di↵erential approach? the
united states (1948–2019). Metroeconomica, 72(4):696–730.

[McKay et al., 2000] McKay, M. D., Beckman, R. J., and Conover, W. J. (2000). A comparison of
three methods for selecting values of input variables in the analysis of output from a computer
code. Technometrics, 42.

[McMahon et al., 2001] McMahon, A., Heathcote, J., Carey, M., Erskine, A., and Barker, J. (2001).
Guidance on assigning values to uncertain parameters in subsurface contaminant fate and trans-
port modelling. National Groundwater and Contaminated Land Centre report NC/99/38/3.

[Meadows and Randers, 2012] Meadows, D. and Randers, J. (2012). The limits to growth: the
30-year update. Routledge.

[Meadows et al., 1972] Meadows, D. H., Meadows, D. H., Randers, J., and Behrens III, W. W.
(1972). The limits to growth: a report to the club of rome (1972). Google Scholar.

237
BIBLIOGRAPHY

[Meadows et al., 1974] Meadows, D. L., Behrens, W. W., Meadows, D. H., Naill, R. F., Randers,
J., and Zahn, E. (1974). Dynamics of growth in a finite world. Wright-Allen Press Cambridge,
MA.

[Mertens and Rubinchik, 2012] Mertens, J.-F. and Rubinchik, A. (2012). Intergenerational equity
and the discount rate for policy analysis. Macroeconomic Dynamics, 16(1):61–93.

[Miftakhova, 2019] Miftakhova, A. (2019). Global sensitivity analysis for integrated assessment
modeling. Working paper, University of Zurich.

[Miftakhova, 2021] Miftakhova, A. (2021). Global sensitivity analysis for optimal climate policies:
Finding what truly matters. Economic Modelling, 105:105653.

[Minsky, 1982] Minsky, H. P. (1982). Can “it” happen again? essays on instability and finance, me
sharpe. Inc.: Armonk, NY.

[Monod et al., 2006] Monod, H., Naud, C., and Makowski, D. (2006). Uncertainty and sensitivity
analysis for crop models. Working with dynamic crop models: Evaluation, analysis, parameteri-
zation, and applications, 4:55–100.

[Morris, 1991] Morris, M. D. (1991). Factorial sampling plans for preliminary computational ex-
periments. Communications in Statistics - Theory and Methods 33 (2), 161174.

[Motesharrei et al., 2014] Motesharrei, S., Rivas, J., and Kalnay, E. (2014). Human and nature
dynamics (handy): Modeling inequality and use of resources in the collapse or sustainability of
societies. Ecological Economics, 101:90–102.

[Mussa, 1986] Mussa, M. (1986). Nominal exchange rate regimes and the behavior of real exchange
rates: Evidence and implications. In Carnegie-Rochester Conference series on public policy,
volume 25, pages 117–214. Elsevier.

[Nikulin, 2001] Nikulin, M. (2001). Hazewinkel, michiel, encyclopaedia of mathematics: an updated

and annotated translation of the soviet’ mathematical encyclopaedia. Reidel Sold and distributed
in the USA and Canada by Kluwer Academic Publishers.

[Nordhaus, 2008] Nordhaus, W. (2008). A question of balance: Weighing the options on global
warming policies. Yale University Press.

[Nordhaus, 2014] Nordhaus, W. (2014). Estimates of the social cost of carbon: Concepts and
results from the dice-2013r model and alternative approaches. Journal of the Association of
Environmental and Resource Economists, Vol. 1, No. 1/2 (Spring/Summer 2014), pp. 273-312.

[Oakley and O’Hagan, 2004] Oakley, J. E. and O’Hagan, A. (2004). Probabilistic sensitivity anal-
ysis of complex models: A bayesian approach. Journal of the Royal Statistical Society. Series B
(Statistical Methodology), 66(3):751–769.

[OECD, 1996] OECD (1996). OECD employment outlook: July 1996. OECD, Paris, France.

[on Climate Change 1/CP.26, 2021] on Climate Change 1/CP.26, U. N. F. C. (2021). Draft cop
decision proposed by the president.

[Owen, 2013] Owen, A. B. (2013). Better estimation of small sobol’sensitivity indices. ACM Trans-
actions on Modeling and Computer Simulation (TOMACS), 23(2):1–17.

238
BIBLIOGRAPHY

[Piketty, 2009] Piketty, T. (2009). Fini le pib, retournons au revenu national. 2009-10-06.
[Pindyck, 2017] Pindyck, R. S. (2017). The use and misuse of models for climate policy. Review of
Environmental Economics and Policy, 11(1):100–114.
[Ponting, 1993] Ponting, C. (1993). A green history of the world. The environment and the collapse
of great civilizations. New York New York Penguin Books 1993.
[Popper, 1963] Popper, K. (1963). Conjectures and refutations: The growth of scientific knowledge
- ed 2014. routledge.
[Popper, 1973] Popper, K. (1973). La logique de la découverte scientifique (1934), trad. Nicole
Thyssen-Rutten et Philippe Devaux, Payot, coll.Bibliotheque scientifique, Paris.
[Pottier, 2014] Pottier, A. (2014). Léconomie dans l’impasse climatique: développement matériel,
théorie immatérielle et utopie auto-stabilisatrice. PhD thesis, Paris, EHESS.
[Pottier and Nguyen-Huu, 2017] Pottier, A. and Nguyen-Huu, A. (2017). Debt and investment in
the keen model: a reappraisal of modelling minsky. Review of Keynesian Economics, 5(4):631–
647.
[Pujol et al., 2015] Pujol, G., Iooss, B., Iooss, M. B., and DiceDesign, S. (2015). Package “sensitiv-
ity”.
[Pytlarczyk, 2005] Pytlarczyk, E. (2005). An estimated dsge model for the german economy within
the euro area. Discussion Paper Series 1: Economic Studies 2005,33, Deutsche Bundesbank.
[Ratto, 2008] Ratto, M. (2008). Analysing DSGE Models with Global Sensitivity Analysis. Com-
putational Economics, 31(2):115–139.
[Ratto and Pagano, 2010] Ratto, M. and Pagano, A. (2010). Using recursive algorithms for the
efficient identification of smoothing spline anova models. AStA Advances in Statistical Analysis,
94(4):367–388.
[Redman, 2004] Redman, C. L. (2004). The archaeology of global change: the impact of humans on
their environment. Smithsonian Books.
[Rizvi, 2006] Rizvi, S. A. T. (2006). The sonnenschein-mantel-debreu results after thirty years.
History of Political Economy, 38(Suppl 1):228–245.
[Roemer, 2011] Roemer, J. E. (2011). The ethics of intertemporal distribution in a warming planet.
Environmental and resource economics, 48(3):363–390.
[Romer, 2016] Romer, P. (2016). The trouble with macroeconomics. The American Economist,
20:1–20.
[Rosen, 1991] Rosen (1991). A comprehensive inquiry into the nature, origin, and fabrication of
life.
[Routh, 1877] Routh, E. J. (1877). A treatise on the stability of a given state of motion: particularly
steady motion. Macmillan and Company.
[Saint-Geours et al., 2011] Saint-Geours, N., Lavergne, C., Bailly, J.-S., and Grelot, F. (2011).
Analyse de sensibilité de sobol d’un modèle spatialisé pour l’évaluation économique du risque
d’inondation. Journal de la Société Française de Statistique, 152:24-46.

239
BIBLIOGRAPHY

[Saltelli, 2002] Saltelli, A. (2002). Making best use of model evaluations to compute sensitivity
indices. Applied Statistics, Institute for the Protection and the Safety of the Citizen (IPSC), The
European Commission,Joint Research Centre, TP 361, 21020 Ispra (VA), Italy.

[Saltelli and Bolado, 1998] Saltelli, A. and Bolado, R. (1998). An alternative way to compute fourier
amplitude sensitivity test (fast). Computational Statistics and Data Analysis, 26(4):445 – 460.

[Saltelli et al., 2008a] Saltelli, A., Chan, K., and Scott, E. (2008a). Sensitivity analysis. Chichester:
Wiley.

[Saltelli et al., 2000] Saltelli, A., Chan, K., Scott, M., et al. (2000). Sensitivity analysis. probability
and statistics series.

[Saltelli et al., 2008b] Saltelli, A., Ratto, M., Andres, T., Campolongo, F., Cariboni, J., Gatelli, D.,
Saisana, M., and Tarantola, S. (2008b). Global Sensitivity Analysis The Primer. John Wiley &
Sons Ltd, The Atrium, Southern Gate, Chichester, West Sussex PO19 8SQ, England.

[Saltelli et al., 2004] Saltelli, A., Tarantola, S., Campolongo, F., and Ratto, M. (2004). Sensitivity
analysis in practice: a guide to assessing scientific models. Chichester, England.

[Saltelli et al., 1999] Saltelli, A., Tarantola, S., and Chan, K. P.-S. (1999). A quantitative model-
independent method for global sensitivity analysis of model output. Technometrics, 41(1):39–56.

[Samuelson, 1966] Samuelson, P. A. (1966). A summing up. The quarterly journal of economics,
80(4):568–583.

[Saporta, 1990] Saporta, G. (1990). Probabilités, analyse des données et statistique. editor Technip
report.

[Sathaye and Shukla, 2013] Sathaye, J. and Shukla, P. (2013). Methods and models for costing
carbon mitigation. Annual Review of Environment and Resources, 38:137–168.

[Sbordone et al., 2010] Sbordone, A., Tambalotti, A., Rao, K., and Walsh, K. (2010). Policy anal-
ysis using dsge models: An introduction. Economic Policy Review, pages 23–43.

[Schaibly and Shuler, 1973] Schaibly, J. H. and Shuler, K. E. (1973). Study of the sensitivity of
coupled reaction systems to uncertainties in rate coefficients. ii applications. The Journal of
Chemical Physics, 59(8):3879–3888.

[Schorfheide, 2000] Schorfheide, F. (2000). Loss function-based evaluation of dsge models. Journal
of Applied Econometrics, 15(6):645–670.

[Schumpeter and Nichol, 1934] Schumpeter, J. A. and Nichol, A. J. (1934). Robinson’s economics
of imperfect competition. Journal of political economy, 42(2):249–259.

[Shannon, 1949] Shannon, C. E. (1949). Communication in the presence of noise. Proceedings of

the Institute of Radio Engineers. 37 (1): Reprint as classic paper in: Proc. IEEE, Vol. 86, No.
2, (Feb 1998).

[Sheng, 2013] Sheng, A. (2013). The end of financialization.

https://www.ineteconomics.org/perspectives/blog/the-end-of-financialization. 2013-09-18.

240
BIBLIOGRAPHY

[Smets and Wouters, 2003] Smets, F. and Wouters, R. (2003). An estimated dynamic stochastic
general equilibrium model of the euro area. Journal of the European Economic Association,
1(5):1123–1175.

[Smets and Wouters, 2007] Smets, F. and Wouters, R. (2007). Shocks and frictions in us business
cycles: A bayesian dsge approach. American economic review, 97(3):586–606.

[Smith and Thieme, 2011] Smith, H. L. and Thieme, H. R. (2011). Dynamical systems and popu-
lation persistence, volume 118. American Mathematical Soc.

[Sobol, 1967] Sobol, I. (1967). On the distribution of points in a cube and the approximate evalu-
ation of integrals. USSR Computational Mathematics and Mathematical Physics, 7(4):86 – 112.

[Sobol, 1998] Sobol, I. (1998). On quasi-monte carlo integrations. Mathematics and Computers in
Simulation, 47(2):103 – 112.

[Sobol, 1993] Sobol, I. M. (1993). Sensitivity analysis for non-linear mathematical models. Mathe-
matical modelling and computational experiment, 1:407–414.

[Sobol’ and Kucherenko, 2009] Sobol’, I. and Kucherenko, S. (2009). Derivative based global sen-
sitivity measures and their link with global sensitivity indices. Mathematics and Computers in
Simulation, 79(10):3009 – 3017.

[Solow, 1956] Solow, R. (1956). A contribution to the theory of economic growth. The Quarterly
Journal of Economics. 70 (1): 65-94. doi:10.2307/1884513. JSTOR 1884513.

[Stansbury and Summers, 2018] Stansbury, A. M. and Summers, L. H. (2018). Productivity and
pay: Is the link broken? Technical report, National Bureau of Economic Research.

[Stiglitz, 1974a] Stiglitz, J. (1974a). Growth with exhaustible natural resources: efficient and opti-
mal growth paths. The review of economic studies, 41:123–137.

[Stiglitz, 2015] Stiglitz, J. (2015). The great divide. Penguin UK.

[Stiglitz, 1974b] Stiglitz, J. E. (1974b). The cambridge-cambridge controversy in the theory of

capital; a view from new haven: A review article. Journal of Political Economy, 82(4):893–903.

[Stiglitz, 2018] Stiglitz, J. E. (2018). Where Modern Macroeconomics Went Wrong. Oxford Review
of Economic Policy, Volume 34, Issue 1-2, pages 70–106.

[Stiglitz et al., 2017] Stiglitz, J. E., Stern, N., Duan, M., Edenhofer, O., Giraud, G., Heal, G. M.,
la Rovere, E. L., Morris, A., Moyer, E., Pangestu, M., et al. (2017). Report of the high-level
commission on carbon prices.

[Sudret, 2008] Sudret, B. (2008). Global sensitivity analysis using polynomial chaos expansions.
Reliability Engineering & System Safety, 93(7):964 – 979. Bayesian Networks in Dependability.

[Tarantola et al., 2007] Tarantola, S., Gatelli, D., Kucherenko, S., Mauntz, W., et al. (2007). Es-
timating the approximation error when fixing unessential factors in global sensitivity analysis.
Reliability engineering & system safety, 92(7):957–960.

[Tarantola and Mara, 2017] Tarantola, S. and Mara, T. A. (2017). Variance-based sensitivity in-
dices of computer models with dependent inputs: The Fourier Amplitude Sensitivity Test. Inter-
national Journal for Uncertainty Quantification, 7(6):511–523.

241
BIBLIOGRAPHY

[Taylor, 1993] Taylor, J. B. (1993). Discretion versus policy rules in practice. In Carnegie-Rochester
conference series on public policy, volume 39, pages 195–214. Elsevier.
[Taylor, 2009] Taylor, L. (2009). Reconstructing macroeconomics: Structuralist proposals and cri-
tiques of the mainstream. Harvard University Press.
[Tian, 2013] Tian, W. (2013). A review of sensitivity analysis methods in building energy analysis.
Renewable and Sustainable Energy Reviews, 20, pp. 411-419.
[Tirole and Gollier, 2015] Tirole, J. and Gollier, C. (2015). The end of financialization. 2013-06-03.
[Tissot and Prieur, 2015] Tissot, J.-Y. and Prieur, C. (2015). A randomized orthogonal array-
based procedure for the estimation of first-and second-order sobol’indices. Journal of Statistical
Computation and Simulation, 85(7):1358–1381.
[Tol, 1965] Tol, R. (1965). The damage costs of climate change toward more comprehensive calcu-
lations. Environmental & Resource Economics 5, 353-374.
[Turchin and Nefedov, 2009] Turchin, P. and Nefedov, S. A. (2009). Secular cycles. Princeton
University Press.
[Turner, 2014] Turner, G. (2014). Is global collapse imminent? University of Melbourne, Melbourne
Sustainable Society Institute.
[Turner, 2008] Turner, G. M. (2008). A comparison of the limits to growth with 30 years of reality.
Global environmental change, 18(3):397–411.
[Turner, 2012] Turner, G. M. (2012). On the cusp of global collapse? updated comparison of the
limits to growth with historical data. GAIA-Ecological Perspectives for Science and Society,
21(2):116–124.
[UNEP, 2021] UNEP, . (2021). Emissions Gap Report 2021. United Nations Environment Program.
[UNFCCC, 2021] UNFCCC, . (2021). Nationally determined contributions under the Paris Agree-
ment Synthesis report by the secretariat. United Nations Framework Convention on Climate
Change Secretariat.
[Van der Ploeg, 1985] Van der Ploeg, F. (1985). Classical growth cycles. Metroeconomica 37 (2),
221-230.
[Volterra, 1928] Volterra, V. (1928). Variations and fluctuations of the number of individuals in
animal species living together. ICES Journal of Marine Science, 3(1):3–51.
[Wang, 2014] Wang, M. (2014). Sensitivity analysis and evolutionary optimization for building
design. PhD thesis Loughborough University Institutional Repository.
[Weitzman, 2012] Weitzman, M. L. (2012). Ghg targets as insurance against catastrophic climate
damages. Journal of Public Economic Theory, 14(2):221–244.
[Werner, 2014a] Werner, R. A. (2014a). Can banks individually create money out of nothing? the
theories and the empirical evidence. International Review of Financial Analysis, 36:1–19.
[Werner, 2014b] Werner, R. A. (2014b). Enhanced debt management: Solving the eurozone crisis
by linking debt management with fiscal and monetary policy. Journal of International Money
and Finance, 49:443–469.

242
[Werner, 2014c] Werner, R. A. (2014c). How do banks create money, and why can other firms not
do the same? an explanation for the coexistence of lending and deposit-taking. International
Review of Financial Analysis, 36:71–77.

[Werner, 2016] Werner, R. A. (2016). A lost century in economics: Three theories of banking and
the conclusive evidence. International Review of Financial Analysis, 46:361–379.

[Wiener, 1938] Wiener, N. (1938). The homogeneous chaos. American Journal of Mathematics,
60(4):897–936.

[Wood, 1998] Wood, J. W. (1998). A theory of preindustrial population dynamics demography,

economy, and well-being in malthusian systems. Current Anthropology, 39(1):99–135.

[Wright, 2004] Wright, R. (2004). A short history of progress: House of anansi press.

[Xu and Gertner, 2008] Xu, C. and Gertner, G. Z. (2008). Uncertainty and sensitivity analysis for
models with correlated parameters. Reliability Engineering and System Safety, 93(10):1563 –
1573.

[Zádor and Turány, 2006] Zádor, J. Z. I. and Turány, T. (2006). Local and global uncertainty anal-
ysis of complex chemical kinetic systems. Reliability Engineering & System Safety, 91(10):1232
– 1240. The Fourth International Conference on Sensitivity Analysis of Model Output (SAMO
2004).

[Zaldivar et al., 2000] Zaldivar, J., Campolongo, F., et al. (2000). An application of sensitivity
analysis to fish population dynamics. Saltelli et al., editors. Sensitivity analysis. New York:
Wiley, pages 367–83.

[Zaldıvar et al., 1998] Zaldıvar, J., Kourti, N., Villacastin, C., Strozzi, F., and Campolongo, F.
(1998). Analysing dynamics of complex ecologies from natural recordings: an application to fish
population models. JRC-ISIS Technical Note I, 98.