Experts' understanding of partial derivatives using the Partial Derivative Machine

Partial derivatives are used in a variety of different ways within physics. Most notably, thermodynamics uses partial derivatives in ways that students often find confusing. As part of a collaboration with mathematics faculty, we are at the beginning of a study of the teaching of partial derivatives, a goal of better aligning the teaching of multivariable calculus with the needs of students in STEM disciplines. As a part of this project, we have performed a pilot study of expert understanding of partial derivatives across three disciplines: physics, engineering and mathematics. Our interviews made use of the Partial Derivative Machine (PDM), which is a mechanical system featuring four observable and controllable properties, of which any two are independent. Using the PDM, we probed expert understanding of partial derivatives in an experimental context in which there is not a known functional form. Through these three interviews, we found that the mathematicians exhibited a striking difference in their understanding of derivatives relative to the other groups. The physicists and engineers were quick to use measurements to find a numeric approximation for a derivative. In contrast, the mathematicians repeatedly returned to speculation as to the functional form, and although they were comfortable drawing qualitative conclusions about the system from measurements, were reluctant to approximate the derivative through measurement. This pilot study led us to further questions. How do fields differ in their experts' concept image of partial derivatives? What representations of partial derivatives are preferred by experts? We plan to address these questions by means of further interviews with a wider range of disciplinary experts.


I. INTRODUCTION
Thermo is hard. In a recent national workshop on the upper-division physics curriculum, approximately 1 /3 of the faculty indicated, in an informal show of hands, that they are uncomfortable enough with the content of thermodynamics that they would be reluctant to teach it. There are a number of reasons why thermodynamics is hard. One reason is the large number of different kinds of partial derivative manipulations that need to be performed to solve many theoretical problems. In previous work, our group has analyzed expert problem solving in this context using a framework of epistemic games. 1,2 Other research groups have identified a variety of other difficulties both mathematical and conceptual. [3][4][5][6] In this paper, we are beginning an investigation of two further reasons why thermodynamics is hard. First, the nature of the variables in thermodynamics. The independent variables are measurable (and changeable) physical variables such as pressure and volume rather than immutable background markers such as space and time. 7 Furthermore, which of these variables are independent and which are dependent varies with the context. In particular, the conjugate pair associated with heating, i.e. temperature and entropy is known to be troublesome for students. 8 Second, many important physical quantities in thermodynamics are actually partial derivatives of other physical quantities. Thermodynamics involves an apparent surfeit of variables in the sense that extensive variables such as volume have intensive conjugate pairings such as pres-sure that have independent operational definitions and are independently measurable, and may seem to be independently controllable. Because of this apparent surfeit of variables, thermodynamics is typically the first time that physics students encounter scenarios in which the quantities held fixed when taking a partial derivative are ambiguous. In mathematics courses, students are taught that when taking partial derivatives, all the independent variables are held fixed. Nevertheless, we have found that most students come into our course with a firm belief that when taking a partial derivative everything else is held fixed.
Two years ago, one of us (DR) developed the partial derivatives machine (PDM), a simple mechanical device of springs and pulleys as a classroom manipulative. (See Section IV A for a more complete description of the PDM.) Classroom activities involving the PDM exhibit many of the same features as experiments and calculations that students encounter in thermodynamics. All the same issues about independent and dependent, extensive and intensive variables arise. And the question of which variables to hold fixed also arises, but in a somewhat simpler context in that the variables involved are concrete and tangible (lengths and forces). Our hope was that students would benefit from classroom experience with the PDM, and, to some extent, this appears to be the case.
However, we have now become aware of a more fundamental underlying problem. In the first experiment with the PDM, students were asked to find a partial derivative from experimental data. Anecdotally, it became clear that many students did not immediately understand that a derivative can be effectively approximated by the ratio of small numerical differences.
We do not believe that the issues we have observed with partial derivatives are limited to students. Indeed, we hypothesize that many of the issues we have observed are due to the ways in which different disciplines use and think about derivatives and partial derivatives. In this study, we conducted small group interviews with experts in several STEM disciplines. By studying experts' thinking about derivatives and partial derivatives, we hope to obtain a better benchmark for comparison in the study of students' thinking about those same ideas. These interviews were most similar to a clinical interview except that the group setting provided a means for participants to listen and respond to each other's ideas, rather than just the interviewers'. The purpose of this paper is explore the research question in what ways do disciplinary experts in physics, engineering and mathematics think about partial derivatives?.
In the course of gaining insight into this question, we describe how the responses were similar and differed across disciplines and consider the role and affordances of the partial derivative machine in the experts' responses.
In the remainder of this paper, we will give a description of the Partial Derivative Machine (PDM), describe the method we used to study our research question and give the results of our analyses from the expert interviews. We close by considering both the pedagogical and the research implications of our results for student thinking and learning in similar contexts.

II. THEORETICAL GROUNDING
A. Concept Images and Concept Definitions Thompson 9 argued that the development of coherent meanings is at the heart of the mathematics that we want teachers to teach and what we want students to learn. He argued that meanings reside in the minds of the person producing it and the person interpreting it. We hypothesized that we could study experts' meanings by studying their images, definitions, and representations for a concept and using these to model their meaning for an idea. We rely on Vinner's 10 language of concept images and concept definitions as an orienting framework. Vinner described the concept image as "the total cognitive structure that is associated with the concept, which includes all the mental pictures and associated properties and processes" (p. 152). A concept definition is a verbal definition that accurately explains the concept in a non-circular way. While we were primarily interested in experts' concept images, as illustrated by our tasks and method, we also considered their concept definitions that in some cases underpin those images. We see both the concept image and concept definition as a means to operationalize and explore the meanings that experts had for derivatives.
Vinner's definition of concept image explicitly allows a particular concept image to involve many properties and many mental pictures. We believe that mathematicians, engineers, and physicists have multifaceted and detailed concepts images for derivative. However, Browne 11 showed that middle-division physics students did not necessarily move spontaneously between various facets even when changing to a different facet might make solving a particular problem easier for these relative novices. Our own classroom experience bears out this observation. Therefore, one of our research question is to explore which facets of the concept image of derivative are cued for different content experts by an open-ended prompt involving numerical data from the PDM.

B. A framework for student understanding of derivatives
The framework developed by Zandieh for student understanding of derivative is a valuable tool in this work. 12 This framework is aimed at mapping student concept images for derivative at the level of first-year calculus. It begins by breaking the formal symbolic definition of the derivative into three process-object layers.
These three layers are the ratio layer, in which one finds a ratio of changes, the limit layer in which one takes the limit as the changes become small, and finally the function layer, in which one recognizes that this could be done for any value of x, and thus describes a function. These three layers are each required in a complete understanding of derivative. Moreover, each of these layers can be seen both as a process and as a reified object. As a process, each layer is a procedure that you could use to find a value. But alternatively, one can understand each of these layers as a static object, which exists independently, and can be and is acted upon by other processes. Zandieh identifies an orthogonal dimension of representation (or alternatively context) with four possible representations: graphical, verbal, symbolic, and "paradigmatic physical." Each of these representations exists for each process-object layer. We introduced Zandieh's symbolic representation in the previous paragraph and will here briefly outline the graphical representation of derivative, which is slope. At the ratio layer, the graphical representation is the slope of the secant line to a curve (which itself is the graphical representation of a function). At the limit layer, one has the slope of a tangent line. And finally, at the function layer, one recognizes that the slope of the tangent line is itself a function that could be visualized as a curve.
This framework is valuable because it makes explicit the three process-object layers that exist in the concept of the derivative, and which can be used separately. In Section III B we will introduce our perspective on the different representations of the derivative, which is expanded beyond that considered by Zandieh in order to explicitly include physical representations at a level beyond that treated by Zandieh.

III. BACKGROUND AND LITERATURE
The purpose of this section is threefold. First, we describe in what ways students and experts have been shown to think about the concepts of derivative and partial derivative, and hone in on particular difficulties for students. Second, we articulate various meanings for derivative and partial derivative that come from both research literature and our own experience working with students and colleagues. In this subsection, we also briefly elucidate several language issues that have arisen as we ourselves, from our different disciplinary perspectives, have discussed these various concepts of derivative and we detail the specific language choices that we have made in this paper. Finally, we consider the importance of studying experts' thinking about derivatives and partial derivatives as a means to identify important learning goals for students in physics, mathematics and engineering. The overarching purpose of the section is to demonstrate that while mathematics and physics education research have gained insight into students' thinking about derivatives, they have not fully explored thinking about partial derivatives. Understanding how experts think about these ideas is a natural first step to exploring how we might want students to reason about them.

A. Students' Ways of Thinking about Derivative
A number of researchers have identified difficulties students have in thinking about rate of change of one variable functions. These difficulties range from students thinking about a graph as representing its derivative, confounding average and instantaneous rate of change, 13 conceptualizing rate as the slope or steepness of a graph, 14 using statistical meanings for average to interpret average rate of change, 15 and inattention to how fast quantities are changing with respect to one another. Some students conflate the average rate of change of a function with the average value of a function and compute average rate of change by computing an arithmetic mean and do not distinguish between the graph of the function and the graph of the function's rate of change. 16 Researchers have suggested that some of these difficulties might be attributable to students not conceiving of rate of change as a quotient. For instance, students often discuss the rate of change as a slope but do not speak of slope as a quotient (the change in a function's value being so many times as large as the corresponding change in its argument). Instead, they talk about slope as the function's steepness. 12 As another example, students often use a tangent line and rely on visual judgments to sketch the derivative function. 17 In yet another study, students who were able to correctly rank the slope at points on a graph were less able to find the sign of the derivative at those points. 18 While this approach is not necessarily problematic, thinking about sliding tangent lines does not necessitate images of variation beyond visual judgments of steepness. Zandieh also argued that students can answer many standard calculus questions without needing to think about functions as relationships between variables nor needing to think about the rate of change of one quantity with respect to another. Students often respond to the directive "find the derivative of g(x)" by acting on a symbolic expression using standard rules for differentiation, without thinking about functions or rates of change. For both slope and symbolic expression issues, students might not see the role of ratio and quotient in understanding rate and derivative. Indeed, researchers have documented that these issues keep students from thinking about the derivative as a ratio of changes in quantities, or interpreting that ratio as measuring how many times larger one quantity is than the other over a particular, small interval. 12,19,20 There is only limited literature that addresses how students think about rates of change in the context of functions of two variables. In a math setting, Yerushalmy 21 provides an indication of natural questions that might arise as students conceptualize rate of change in multivariable settings. She illustrated students' struggles with how to think about dependence in a system with three quantities and how to represent multiple quantities and their changes in a single graph. Her students struggled to describe the change in a particular direction of a linear function of two variables. Part of the reason may have been that there are infinitely many directions in which to move from a given point at a constant rate, yet, in general, each direction yields a different slope.
In a physics setting, researchers have investigated mixed partial derivatives and differentials in thermodynamics. 3,22 They address students' ability to translate back and forth between "physical processes" and partial derivatives, and found that students were more able to go from a partial derivative to a physical process than the other way around. Similarly, researchers investigating physical chemistry have found that students need help in interpreting mathematics in a thermodynamics context. 23 It is clear that students do not necessarily reason about derivatives and rates of change as we might hope. Furthermore, very little literature has considered how these issues extend to functions of more than one variable, particularly how students might think about partial derivatives. This is surprising, given that many scenarios in mathematics, physics and engineering require thinking about systems or scenarios in which many variables may be changing simultaneously. Researchers have identified the need for students to be fluent in looking at the concept image of derivative using multiple perspectives. Derivatives are commonly described as slopes, as ratios of small changes, as difference quotients, and as rates of change, among others. However, these phrases are used in different ways in different disciplines, and by different individuals within those disciplines. For instance, the description of slope as "rise over run" could be either numerical or symbolic, as well as graphical. This complexity reflects the multifaceted concept images of the derivative held by experts.
A common framework for multiple representations is the Rule of Four introduced by the Calculus Consortium (see Hughes Hallett et al. 24 ), in which key concepts are presented graphically, numerically, symbolically, and verbally. In recent work, 25 we proposed adding a fifth representation, based on experiment. In order to be clear in our discussion of the different facets of expert concept images of the derivative, we will associate each of the verbal descriptions of derivatives given in the previous paragraph with just one of these representations, as shown in Table I. We have based this association on our interpretation of the technical usage of these common descriptions within the mathematics community. The usage may be unfamiliar to our physics readers, but we hope that this unfamiliarity will help the reader notice the nuanced differences in language. In particular, both the use of rate of change to refer to rates that do not involve time and the technical term difference quotient are largely limited to mathematics. We will use the conventions indicated in Table I throughout this paper, except where explicitly stated otherwise.
We therefore consider five different ways to understand and think about the concept image of derivative, each of which is useful in different scenarios.
1. The slope of the tangent line to a curve. This representation describes the slope as a geometric measure of the steepness or slant of a graph of a function.

2.
A numerical ratio of small changes, by which we mean an explicit numerical quotient, involving actual values of the "rise" and the "run". By "small changes" we mean small enough that the quotient represents a reasonable estimate of the derivative (within the physical context of the problem).
3. The result of algebraic manipulation of a symbolic expression. Formally, these manipulations involve the difference quotient f (x+∆x)−f (x) ∆x , but in practice a memorized set of derivative rules is used instead. (For the purposes of this paper, we will conflate these two algebraic manipulations.) This meaning for derivative is a process applied to an algebraic object, but does not include an image of that process as measuring how one quantity changes with respect to another.
4. The rate of change of one quantity with respect to another, which is here used to mean a description in words of that change. (Outside of mathematics, the word "rate" implies that the second quantity is assumed to be time, but we will use this term more generically.) In this representation, the derivative measures covariation, i.e. how one physical quantity changes with respect to another.
5. Finally, as introduced in our previous work, 25 particular derivatives can be associated with particular experiments, such as measuring the change of volume in a piston of gas as weights are added to the top of the piston. Determining which experiment might correspond to which derivative provides a representation of the derivative, which we call name the experiment.
In the second case, it is most natural to think of the derivative as a number. One picks a point at which to take the derivative and computes a numerical ratio. While that number will be different at other pointsmaking the derivative actually a function-this aspect of the derivative may often be ignored. When considering the slope of the tangent line to a curve, it is clearer that the derivative is a function, but it is also natural to think of the derivative as a number, the slope at a single point on the curve. When using the symbolic approach, the derivative is inherently a function, and while that function could be evaluated at a point, its value cannot be determined until after its functional form is known.
Most of these aspects of the concept image of derivative (namely, 1, 3, 4, 5) find approximate analogues in Zandieh's framework. 12 Interestingly, the second aspect, which is not present in Zandieh's framework, turned out to be the most important one in the analysis of our interviews and is a major theme throughout this paper.
2. What is a partial derivative? Partial derivatives differ from ordinary derivatives in important ways. How we understand this difference can vary with how we understand derivatives.
1. A tangent line turns into a tangent plane in three dimensions, and a partial derivative becomes the slope of the plane in a given direction in the domain, at a given point.
2. When considering a ratio of small changes, a partial derivative requires that we specify not only which quantities are changing, but also which quantities to hold fixed.
3. The algebraic procedure to find a partial derivative of a symbolic expression is identical to that for an ordinary derivative, provided there are no interdependencies among the variables in the expression.
4. The verbal description of derivatives as rates of change must explicitly mention the independent variable(s) in order to describe a partial derivative. e.g. the derivative of volume with respect to pressure would be "the rate of change of volume as pressure is changed, with either temperature or entropy held constant." 5. The representation of derivatives in terms of experiments is designed precisely to take into account which physical quantities are controlled, and which are not. As such, it is particularly well suited to descriptions of partial derivatives.
In thermal physics, and other areas of mathematics, the quantities that are being held fixed are contextdependent. In general, one has a set of interrelated variables, of which a few may be fixed. The number of independent variables is itself context-dependent, and in physical situations we are seldom provided with symbolic equations connecting the set of interdependent variables. More often we rely on physical intuition and argumentation to establish how many variables may be controlled independently. "If I fix the pressure, temperature and number of molecules, I could measure the volume and the mass, therefore I believe I have three independent degrees of freedom." How we respond to the ambiguity provided by abundant physical variables depends deeply on our concept of a derivative. However, if students' concept of derivative is not rooted in an image of measuring how fast one quantity changes with respect to one or more other quantities, then it is unlikely they will understand derivative in the ways we intend. In the previous two sections, we have made the case that there are a variety of ways to think about derivatives and partial derivatives, but students have difficulty thinking in the ways we might intend because of their inability to think about a derivative as measuring the ratio of small numerical changes between quantities. This is the case even for functions of a single variable. Earlier, we also argued that most real-world scenarios involve reasoning about multiple quantities and the relationships between them, a goal that seems especially problematic for students who have difficulty reasoning even about simple systems. Our reasons for studying the thinking of a variety of experts across mathematics, engineering and physics were a) we believed they would be accustomed to working with situations involving multiple quantities and relationships, b) we believed their experience would allow us to observe sophisticated reasoning patterns that we could only hope to observe in extremely advanced students, c) we anticipated their ways of thinking about partial derivatives could help us identify "end goals" for how we want students to think and d) we expected their thinking would vary across disciplinary areas, allowing us to better understand how students in these fields might need to reason in different ways about derivatives and partial derivatives.
As mentioned earlier, the overarching purpose of the section was to demonstrate that while mathematics and physics education research have gained insight into students' thinking about derivatives, they have not fully explored thinking about partial derivatives. Furthermore, understanding how experts think about these ideas is a natural first step to exploring how we might want students to reason about them. In the subsequent sections, we describe how we studied experts' thinking about these ideas.

A. The Partial Derivative Machine
We have developed and used two versions of the Partial Derivative Machine (PDM). The first version of this device is documented in reference 26, and features a central system that is attached to four strings. The simplified version of this device-which will be discussed in this paper-is shown in Fig. 1, and consists of an anchored elastic system, which is constructed of springs and strings. In both versions, the elastic system may be manipulated using two strings independently. Each of these two strings has a scalar position that can be measured with a measuring tape and a tension that can be adjusted by adding to or removing weights from a hanger. Detailed instructions for constructing a Partial Derivative Machine, including a parts list and photographs of ad- The Partial Derivative Machine. The machine consists of a system (inside the dotted quadrilateral) which may be manipulated by pulling on two strings. Each string features a flag, which may be used to measure its position, and a weight that can fix the tension in the string. In addition, thumb nuts may be used to fix the position of each string independently.
ditional central systems, are available on our Paradigms website. 27 The usefulness of the PDM emerges because it is an exact mechanical analogue for a thermodynamic system. The system contains a potential energy U (analogous to the internal energy) that cannot be directly measured. The system has four directly measurable-and controllable-state properties: two positions x and y and two tensions F x and F y . These four state properties play roles analogous to volume, entropy, pressure and temperature in a thermodynamic system.
Although the PDM has four measurable and manipulable properties, like its analogous thermal system, it only has two degrees of freedom. One cannot independently control the tension and position of a single string, unless one uses the other string to do so. In the PDM, the two degrees of freedom are physically manifest: each corresponds to one string. We can choose to manipulate that string either by changing its position, or by changing its tension.
As in thermodynamics, the choice of which properties to treat as independent variables is context-dependent. While it is experimentally easiest to control the two weights as independent variables while measuring positions, it is sometimes theoretically more convenient to view the positions as the independent variables. Most notably, when using work to determine the potential en-FIG. 2. The system we used in the PDM for for the expert interviews. The system contains a single spring, attached by string to an off-center post. The "observable" strings are routed around two horizontal pulleys towards the front of the machine where they are measured, and are attached to weights. ergy, the positions are the "natural" variables, as seen in the total differential that is analogous to the thermodynamic identity: We can relate this total differential to the mathematical expression By equating coefficients of dx and dy, we can find expressions for the two tensions as partial derivatives of the potential energy.
This enables us to clarify the interdependence of the four directly observable quantities. We use the PDM to teach a mathematical introduction to thermodynamics prior to our junior-level course in thermal physics, Energy and Entropy. This introduction uses seven contact hours, and covers the range of mathematical topics generally taught in undergraduate thermodynamics. We begin with total differentials and integration along a path, discuss partial derivatives and chain rules, mixed partial derivatives and Maxwell relations, and end with Legendre transformations. Throughout both this mathematical introduction and Energy and Entropy, a focus is placed on connecting the mathematical expressions with tangible reality. 28

B. Design and Conduct of the Expert interviews
To gain insight into our research question, we performed three expert interviews, each of which lasted around an hour. We interviewed three groups of experts: physicists, engineers and mathematicians. For the physicists we interviewed one associate professor and one full professor. Both use computational methods in their research, which is in astrophysics and optics. We interviewed three engineers: one chemical engineer who is a full professor with considerable research and teaching expertise in thermodynamics, and two assistant professors who study student thinking and epistemology in engineering. Finally, we interviewed two mathematicians who are both assistant professors and whose research is in mathematics education at the collegiate level. All of those we interviewed have doctoral degrees in their disciplinary areas and work in physics, engineering, or mathematics departments.
These interviews were mostly closely aligned with the notion of a clinical interview in which one focuses on gaining insight into another's thinking through systematic questioning and open ended prompts. At the same time, it is atypical that clinical interviews occur in group format because of the difficulty of ascertaining an individual's thinking. Our use of groups was purposeful. We believed that the conversations between the experts and their questioning of each others' responses would be just as important as the questioning and prompts from the interview team. Indeed, we saw that some of the most interesting data we collected came from the experts debating each others' responses.

C. Prompts and Purpose of Task
Our overall research question is to explore in what ways do disciplinary experts in physics, engineering, and mathematics think about partial derivatives. Important subquestions included how the experts use notation, how they think about derivatives and partial derivatives, how they think about the various variables and which, if any, should be held fixed, and how they related those ways of thinking to the PDM. We chose to exploit a pedagogical design feature of the PDM to explore our experts' understandings of partial derivatives outside the world of symbolic manipulation.
We began by introducing the experts to the PDM, and showing them how to manipulate the machine. We then gave the experts the following prompt: We asked experts to "find" a partial derivative for which they are given no functional form. The PDM allows for measurement of changes in the positions and tensions at discrete data points and the prompt was designed in such a way that a ratio of changes in quantities would be the only easy response. However, we deliberately asked the prompt in an open ended way so that it would not cue a particular aspect of the concept of derivatives. All parts of the system were chosen to be visible in order to encourage discussion of the possibility of finding an analytic expression by which one could determine the derivative, although the actual determination of such an analytical expression would have been prohibitively difficult. When we provided this prompt, we did not define either x or F x , but rather let the interviewees discuss what these quantities might mean. After they had discussed the meanings of these terms, and we agreed that their meaning was sufficiently clear to us, we clarified if necessary that x was the position of one flag (i.e. one string) and that F x was the tension in that string, which was determined by the weight. We note that this aspect of the task came at the cost of not explicitly exploring graphical representations of tabular experimental data with all groups of participants.

D. Analytical Method
The analysis of the data collected for this study relied on systematically creating descriptions of experts' thinking about rate of change (i.e. their concept images and definitions). These hypotheses resulted from in-the-moment observations and short reflections by the researchers between interviews. We also created descriptions of experts' thinking about derivatives and partial derivatives from the interviews using retrospective analysis. The retrospective analyses involved making interpretations and hypotheses about participants' thinking by analyzing videos of the interviews after they were completed. Thus, we worked from two sets of observations: those formed during the interviews and those formed from analysis of the videos as a whole. These analyses helped us to think about the categories of concept images and definitions we present in the analyses. Thus, analysis of the data led to two things: a log of the hypotheses developed about experts' thinking in the interviews and descriptions of the experts' thinking made by the researchers after the interviews were completed. This data corpus provided a means to describe patterns in experts' thinking, which in turn helped us focus on various concept images and definitions they appeared to have for derivative.

V. RESULTS
During the course of the three interviews, a number of themes emerged, each of which provided insight into the experts' concept images and definitions for derivatives and partial derivatives. These themes were a combination of issues we noticed as we did the interviews and issues that emerged as we conducted the data analysis described in the analytical method section. In the sections below we describe these themes and articulate how we saw each group of interviewees in the context of that theme. Where it is possible, we provide transcript excerpts from the interviews to support the claims we make.

A. Identifying x and Fx
Since the invention of the first partial derivative machine, we have enjoyed sharing this tangible experimental system with colleagues. During these informal "interviews," we had noticed that different individuals interpreted the symbols in the algebraic expression ∂x ∂F x differently and began to expect that there might be disciplinary reasons for these differences. Therefore, we chose, in these interviews, not to tell the interviewees our own meaning for the symbols (at first) but rather to let them explore in their groups what these symbols might mean. All three groups spent significant time thinking about and debating over how to identify the quantities x and F x . We expected that the physicists and probably the engineers would share our understanding that x was the position of the pointer on the x measuring tape and that F x was the tension in the x labeled spring. We expected that the mathematicians would try to invoke the mathematics convention that a subscript indicates a partial derivative. We were surprised by several other unexpected types of confusion that our notation caused.
The physicists initially identified x with the elongation of the spring and then wanted F x to represent a force in the same direction as the spring. Their mental focus was clearly on the internal mechanics of the system and they (at first) ignored the measuring tapes on the PDM as a potential method for measuring a position x. When they recognized that their interpretation of the symbols would lead to a "total" derivative rather than a partial derivative, they then switched to an interpretation in which they wanted the two strings to be perpendicular to each other, even going so far as to manipulate the PDM to make this true. Presumably, they were invoking a physics convention that x (and y) are independent rectangular coordinates and F x (and F y ) are perpendicular rectangular components of a single, net force vector. We might have anticipated some of this interpretation since one of the interviewers has such a strong connection of the symbol F x to the x component of a force that she even referred to F x in this way in the interview with the mathematicians.
The engineers immediately showed a preference for having x represent the horizontal coordinate (the strings were vertical) but acknowledged, with laughter, that this must not be the case because it would have trivialized the problem. Unlike the physicists, they rapidly make use of the left and right measuring tapes and are happy to invent their own notation, calling the position of one pointer x L and the position of the other pointer x R . They do not have any difficulty identifying F x as the weight on one of the strings.
The mathematicians were not only puzzled by the meaning of the subscript, as we expected, but also expected the capital symbol F to represent a function rather than a force.

Physicists
The physicists began the task thinking that the position x might be the elongation of the spring and F x a force in the direction of the spring. They rapidly corrected this one-dimensional interpretation when they realized that they were asked to find a partial derivative. Excerpt 1. At 7:46, the interviewer clarified the meanings of x, y, F x , and F y . From this stage on, the physicists had no difficulty using this notation.

Excerpt
At the end of this portion of the interview, after the interviewers interpretation of the symbols has been explained, P1 volunteers the interpretation that, at the beginning of the interview he was trying to measure the spring constant, which explains why he was looking at the internal mechanics of the system.

Excerpt 3
8:37 P1 Yeah, I was more thinking we wanted to measure the spring constant.

Engineers
The engineers began with the professional but humorous recognition that there is an interpretation of the symbols that would make the problem trivial: They then spent 2-3 minutes discussing how they expected the system to behave as they added weights, based on the geometry of the system, especially the angle of the spring. During this analysis, they discussed whether these symbols should refer to one side of the PDM or both: From the initial moments of the interview, the two mathematicians puzzled over the meaning of the subscript on F x , noting that "we have not seen this type of notation before" and "it looks like a derivative but we are unsure what the symbols are." While they attended to the position of the strings, they did not interpret F x as the force related to x until the interviewer explicitly suggested this to them part way through the interview.
In Excerpt 9, the mathematicians explain their confusion with the notation and in doing so, reveal that they often associate a derivative with the process of differentiating an explicitly defined function. This was a theme we observed throughout the interview with the mathematicians, yet only observed briefly in the other groups. A few minutes later, the mathematicians began to focus on the meaning of the subscripts and revealed how problematic the notation was for them. In particular, we think their uncertainty reflects possible disciplinary differences in notation, an issue we anticipated prior to the interviews. At 5:34 the interviewer interrupted to clarify that F means "force" and that the x subscript indicates the component of the force in the x direction. The mathematicians immediately grasped that there could then be a y and an F y , although-like the engineers-they did not guess that y would be the position of the other string and F y the tension on that string.

Excerpt
B. What is a (partial) derivative?
The prompt to find a partial derivative yielded an interesting picture of how our experts understood the concept of derivative. In the end, all of our experts found an approach to measure the derivative experimentallywhich practically requires application of the ratio layer of Zandieh 12 (see Section II B). Each group immediately proceeded to explore whether their changes were sufficiently small. This reflects a strong recognition of the limit layer of Zandieh, 12 and a recognition that a measurement of the derivative must account for this. We saw a large difference between disciplines in the prominence of the function layer of Zandieh, 12 which we recognize as an exploration of the dependence of the derivative on the forces F x and F y . And finally, we saw a large difference in the degree of comfort with a numerical approximation for the derivative.
The physicists and engineers were very comfortable with the derivative as a number, and quickly computed this number as a ratio of small changes. In both cases, they took an additional measurement in order to verify that their changes were sufficiently small that they were within the linear regime. Their comfort with a single numerical answer for the derivative suggests to us that the physicists and engineers were satisfied with a derivative that omits the function layer of Zandieh. 12 Both groups did acknowledge and discuss that there is a functional dependence of the derivative on the force.
Our mathematicians, in contrast, saw the derivative as a function, and expressed concern about numerical approximation. Interestingly, although the mathematicians were persistent in seeking a symbolic functional form for the derivative, in the process they were quite comfortable and creative with drawing conclusions about the derivative through experimentation, and specifically investigated the functional behavior of the derivative in how it changes when different parameters are modified.
Unlike the physicists, the engineers went on to mention other representations of the derivative in their discussions, such as the slope of a graph, and a symbolic expression derived from statics. The mathematicians made use of essentially every representation for derivative except for the ratio of small changes.

Physicists
After some technical difficulties, at 13:56 the physicists had collected their first data, which was sufficient to find a simple ratio of changes. In Excerpt 11, they discussed whether two values each for x and F x are sufficient. In particular, they expressed the idea that "because it is a derivative" it may need smaller increments.

Excerpt 11
14:29 P2: The other question is, because it is a derivative, does it need to have smaller increments? P1: Umm... P2: The difference is... P1: Yeah... unless the system is linear or not. P2: So maybe we should try... P1: Try it with fifty.
The went ahead and took an additional data point to verify that their ∆F x value was small enough that they were working in the linear regime, which they did in Excerpt 12. They computed the derivative as a ratio of their ∆x and ∆F x , confirmed that the two sizes of ∆F x gave similar answers, and concluded that they had made a measurement of the derivative. They proceeded to discuss the possibility that this derivative might not be a constant, i.e. might depend on the value of F x , acknowledging that that the derivative is a function, although they did not feel that this it was necessary to explore this dependence in order to answer the prompt.

Engineers
We begin with the engineers at the same stage as we began with the physicists, after they had taken two measurements each of x and y (x left and x right in their notation) with different values of F x and the same F y , and thus had enough data to compute an estimate of the derivative as a ratio of changes. In Excerpt 13 the engineers discussed whether to go on taking more data.

Excerpt 13
14:17 E1: So shall we put another hundred grams on to see if it's linear? E2: I think so. Yeah. E3: I'm feeling like we should be writing some huge equation to describe this and not have to mess with this, but I'm unwilling to start that procedure.
Also, in Excerpt 13, E3 suggested that perhaps they ought to be performing a symbolic calculation, but was reluctant to do so, and the subject was dropped. On another occasion, the engineers spent some time discussing the possibility of using statics to find a symbolic solution, and also dropped the idea. They clearly recognized the possibility of a symbolic expression, but were unwilling to pursue it. At this point the engineers computed values for the derivative for two values of F x using a ratio of changes, and found somewhat different numerical values. In Excerpt 14 they discussed how to interpret these differing ratios.

Excerpt 14
18:49 E3: Are those the same? E1: Well so, I would recommend, let's let's crank it a bunch and see if we come up with the same number. E3: Yeah. E1: And then if it's not, we could either, we could either plot it and take slopes, or we could say, hey is that good enough. E3: Right.
After some further discussion they concluded that the two slopes were the same within their experimental error, and that they had a good measurement of the derivative. In checking that the response is linear they addressed the limit layer in a way that could be surprising: they increased the size of their change in order to show that it was sufficiently small. This expert behavior reflected a recognition that experimental uncertainty would make smaller changes harder to measure.
The engineers returned to the idea of graphing and finding the slope much later in the interview. After being prompted with how they would perform an analysis if they had sufficient experimental data, the engineers returned in Excerpt 15 to the idea of the slope of a graph to explain how they could find the F x dependence of the partial derivative. In this discussion the engineers gave a clear picture of how they could graphically obtain the derivative as a function of both F x and y (which they called L 2 . The engineers continued to further discus how they could obtain the slope from the graphical data by performing a curve fit.

Mathematicians
The mathematicians spent much longer than the physicists or engineers before finding an answer to the prompt that they were satisfied with. In the process, they used physical manipulation of the machine to reach several conclusions regarding properties of x.
Very early on, the mathematicians grappled with identifying the arguments of the function x. This was in strong contrast to the physicists and engineers, who did not talk about x as a function until after having experimentally found the partial derivative. In Excerpt 16, the mathematicians used physical reasoning to conclude that x is a function of F x and F y . This discussion followed considerable manipulation of the machine. The mathematicians have concluded that by fixing y, x becomes independent of F y . This reasoning probably reflects an interpretation of F y as the mass on the hanger, rather than the tension in the string-an interpretation that is entirely consistent with the information they were given, although this was not the interpretation we intended.
A few minutes later, in Excerpt 17, the mathematicians discussed the derivative as "rate of change" and addressed how their understanding of x as a function of the two forces relates to the partial derivative they were asked to find. , the derivative of f with respect of x, that partial derivative is like the rate of change with respect to x. But now x is gonna be dependent on these two forces and so it's like we're finding out the rate of change in x with regard to this particular force [points to ∂x ∂Fx ]. So originally that notation is like why, 'these physicists are doing it all wrong, they're putting it all in the wrong place!' But now I understand that, now I understand the notation. Does that make sense? M1: Uh huh. M2: So if x is dependent on these two forces [pulls on F x and F y ], then, right, which we're figuring that out, um, then this notation makes sense. M1: Mhmm. Mhmm.
After this, the mathematicians began discussing how to address the prompt. This led to a discussion of what would happen if they fixed the y string, which appears in Excerpt 28 in the following section.
In the middle of the interview, the mathematicians began discussing what it is that they are being asked. In Excerpt 18 the interviewer responded by explicitly asking what they think their task is, and clarified that they were not actually asked for a symbolic expression for the partial derivative.  Throughout the interview, the mathematicians returned to speculation as to the functional form of x. In some cases, they used hypothetical symbolic expressions to reason about possible behavior of x. As the end of the time allocated for the interview approached, the interviewer began pressing the mathematicians for an explicit answer to the prompt. In Excerpt 20 one mathematician replied jokingly with a guess as to a symbolic expression. When pressed by the interviewer, the mathematicians ended up describing a process by which they could find a numeric answer from the slope of a graph of their data. The mathematicians proceeded to write ordered pairs of numbers on their board (using data that they had previously collected), rather than drawing the graph they described. Like the engineers and physicists, soon after they recognized the existence of a numerical solution, they began considering whether those changes were small enough. In Excerpt 21, the mathematicians discussed the accuracy of their approximation, and concluded that they could improve it by adding smaller weights.

Excerpt 21
47:30 M1: Okay so we have those differences, and so these are, we can think of them as just being points on our function like whatever that function is. M2: Actually a really rough approximation.. After this, they spent some time adding small weights and measuring changes in x-although they never did actually perform a division or write down a ratio.
C. What should be held fixed, and does it matter?
A defining feature of a partial derivative is that some other variables must be held fixed. In thermodynamics in particular, this choice of what to hold fixed has important implications, and we were very interested to see how our experts treated this question. We believed that the prompt to

Find
∂x ∂F x was ambiguous, and were interested to see how our experts responded to this ambiguity. We expected that at least some experts would point out the ambiguity and ask us which property to hold fixed. Our interview had other results: all our experts assumed that since we were taking a partial derivative with respect to one force, then the other force must be held constant. In this respect, the PDM differs from thermodynamics, in which a derivative such as

∂p ∂V
does not have any inherent information about whether entropy or temperature should be held fixed.
In further discussion, each group addressed the question of what would happen if they fixed y instead of F y . Our experts felt that clamping the y string would change either the system or the function x, and thus the derivative would change. The mathematicians measured how x changed with and without y fixed, and concluded that it moved less when y was fixed, although they did not discuss this result in terms of derivatives.
Also related to this question is whether the derivative depends on the value of F y . This is a further level of the concept of function than we explored in the previous section. Ordinary derivatives are a function of the single independent variable, but partial derivatives are functions of multiple independent variables.

Physicists
One physicist decided that F y must be held fixed when measuring ∂x/∂F x , because they were taking a derivative with respect to F x . Thus if they took the inverse derivative ∂F x /∂x, then y would have been held fixed. He stated in Excerpt 24 that this is "what they taught me." When he considered fixing the thumb nut, he stated that this (physical act?) would change x to be a different function, a function of F x and y.
In Excerpt 22, before starting to take data the physicists began discussing the meaning of the partial derivative, and what to hold fixed. Thus the physicists concluded that they should fix F y in order to measure the partial derivative with which they were prompted. Later in the interview, the interviewer prompted the physicists by asking them to measure the derivative ∂x ∂F x y .
In Excerpt 24, the physicists correctly explained how to measure this derivative by clamping the other string.
They also explained about why they had assumed that F y should be held fixed, and went on to discuss how a partial derivative is like a derivative along a path in a multidimensional space. They end by suggesting that clamping the y string changes the system itself, creating a different function.

Engineers
At a point in the middle of the interview, E2 pointed out that they had been instructed that they could fix y using the clamp, leading to a discussion of what that would mean and how it would affect their results.

Excerpt 25
27:00 E2: They said we could stop [inaudible, points at the thumb nut for fixing y.]. Does that help us at all? E3: That would be crazy! [jokingly] After this the engineers spent a couple minutes discussing what they would find by performing the same derivative measurement while the thumb screw holds y fixed. After a bit of discussion involving how to perform the measurements, E1 concluded in Excerpt 26 that they are now holding something different constant.

Excerpt 26
29:06 E1: We're holding something different constant, and we could think about, if, once we look at the numbers and see how they define that. E3: But it's sort of like you're saying, certainly a different problem holding that [points to y string] constant. E1: Yep.
At this point the engineers proceeded to take some data with y fixed. Then the engineers discussed how to notate what they just measured with y fixed. E1 (who teaches thermodynamics) suggests that they use the subscript convention to notate which quantity is held fixed. E1: That's awesome.
The engineers at this stage were using the symbol L 2 for the distance we call y. Thus this notation is in complete accord with our understanding of the problem. We find this unsurprising, given that E1 is entirely familiar with thermodynamics. E3, who was writing on the board, seemed considerably less comfortable with this notation.

Mathematicians
We saw in Excerpt 16 a mathematician said that fixing y would change the system by making x no longer a function of F y . This would make x a function of only one variable, since they did not talk about the value of y as a variable. A few minutes later, in Excerpt 28, the mathematicians explored what happens if they fix y by clamping its string. They observed a smaller change in x, and decided that this made sense, since the clamp impedes the motion. After this, they spent some time discussing what value of F x to use for their derivative, not being sure what to focus on. We believe based on the above quote that they understood this to mean that the derivative itself was smaller if y were fixed, although they did not explicitly state this.
D. How many independent variables are there?
As discussed in Section III B, the number of independent variables-or the number of degrees of freedom-is a critical property of a physical system. It determines the number of parameters we must control in order to determine the state of that system, and at the same time limits the number of parameters we can fix when finding a partial derivative. The PDM has two degrees of freedom, as discussed in Section IV A.
The existence of two degrees of freedom means that the derivative ∂x/∂F x (and the variable x itself) should be viewed as a function of two variables. This is an expansion of the layer of function into multiple dimensions. In this section, we discuss how our experts interpreted the number of degrees of freedom of the system.
The question of how many independent variables were present arose in each interview. Both the physicists and engineers treated this question explicitly (the physicists, at the prompting of the interviewer), and went through a stage of talking of x being a function of the remaining three variables. They then concluded that one could eliminate one of those three, and that only two were independent. The mathematicians were not asked this question explicitly, but addressed the question during their discussion of the meaning of the partial derivative.
All of our experts were able to discern the number of degrees of freedom present in the system, but we were surprised at how long it took the physicists and engineers to agree upon and express the fact of the interdependence among the controllable quantities.

Physicists
When asked how many independent variables there were in Excerpt 29 below, the physicists recognized that only two variables were independent, because they could independently control the two forces. After some more discussion, they agreed that the derivative must be a function of the two independent variables F x and F y . A bit later in Excerpt 30, the physicists pondered how to understand the possibility of directly controlling y, suggesting that x depends on three variables each of which they could control, but they could not control all three independently. Based on their gestures leading up to this and the discussion in Excerpt 29 above, we believe they arrived at this conclusion based on "physical" reasoning.

Excerpt 30
27:05 P1: Can we say that x is actually a function of y, F x , and F y , but then these three are not independent.

Engineers
Towards the end of their interview (without interviewer prompting), the engineers discussed their concept image of a partial derivative. In Excerpt 31 below, the senior engineer E1 decided to ask his two partners how they understand the concept of a partial derivative. This led to a discussion of how many independent variables were present in the system. At this stage we believe E1 knew the answers to his questions, and had taken on the role of interviewer of his younger colleagues, who were less certain as to how to treat this multidimensional system. Discussion after the interview confirmed that E1 had been thinking already about how to employ the PDM in the context of a class in thermodynamics.
The engineers (like the physicists) concluded by discussing the quantity x as a function of the three other variables, and recognized that those three variables are not themselves independent.

Mathematicians
As we described above, the mathematicians were very quick (once they knew what the variables meant) to recognize that x depends on F x and F y (see Excerpt 16), indicating that they recognized two independent degrees of freedom. Moreover, they found in Excerpt 28 that the derivative ∂x/∂F x was smaller if they fixed y rather than fixing F y , indicating that fixing the third quantity y had some effect. They did not, however, count the independent degrees of freedom, nor did we prompt them to do so.
The mathematicians began discussing how to treat the second side of the system in Excerpt 32. M2 expressed a tension between a "mathy" understanding and their physical experimentation which led them to believe that the partial derivative depended on what was done on the y side of the PDM. After agreeing to explore the dependence on the y side of the system, in Excerpt 33, the mathematicians concluded based on experimentation that x is indeed affected by the value of the other force F y , confirming that the derivative is a function of two variables. We note that their language is at times confusing, as they often use x and y to refer to the two independent variables (which we call F x and F y ), and had written f (x, y) on the board to refer to a generic function of two dimensions. Thus the mathematicians concluded that the partial derivative was a function of the two forces.

VI. CONCLUSIONS: CONCEPT IMAGES AND CONCEPT DEFINITIONS
In this section, we discuss the experts' concept images and definitions for partial derivatives. The commonalities in the concept images are as follows (1) a partial derivative is a function; (2) to find a partial derivative, something must be held fixed; (3) the notion of slope; (4) the derivative is a ratio of small changes. There are, however, some differences: the mathematicians and the engineers were the only ones to talk about the derivative as a rate of change and as a slope, and the engineers and mathematicians were the only groups to consider writing an explicitly defined function. We also note that having more concept images is not necessarily "better." Indeed, given the prompt provided to the experts, measuring and then determining a ratio of small changes was sufficient to complete the task. The experts' use of multiple approaches and images suggests that their concept images are complex yet they may not have immediately known which image would provide a solution to the given prompt.

Physicists
The physicists thought of a partial derivative as a ratio of small changes and as a function. However, at no point did they consider writing an explicit function definition; rather, they attended to determining how many independent variables were in the system, which of them to fix, and collecting data.

Engineers
Like the physicists, the engineers thought of the partial derivative as something that could be approximated experimentally, and spent a lot of time collecting data so that they could represent the partial derivative as a ratio of small changes. They also talked about the derivative as a slope, and that they could graph their data to determine an explicitly-defined function for the derivative by making use of the notion of derivative as slope. They further noted that their experimentally-determined function should match some sort of theoretical equation. The engineers were the only ones to state a concept definition, albeit incomplete and bordering on inaccurate. They defined a partial derivative as "how one dimension in a multidimensional system changes when the other dimensions are fixed."

Mathematicians
The mathematicians repeatedly expressed their interest in determining an algebraic expression for ∂x/∂F x ; they seemed to interpret the task statement as directing them to find an explicit function definition. Once they had collected data, they talked about the derivative as a slope, as a ratio of small changes, and as a rate of change. Like the physicists, they recognized the need to fix one quantity, which for them was connected with the idea of holding a variable constant when using symbolic differentiation rules for a multivariable function.

VII. CONCLUSIONS
In this section, we summarize the experts' concept images and definitions for partial derivatives, propose the need for an extended theoretical framework, describe differences in the concept of limit between mathematics/theory and experiment, and discuss the limitations of our study and the prospects for future work.

A. Concept images of derivative
All three groups of experts assumed that when taking a partial derivative with respect to F x , F y can be assumed to be held fixed. This finding was unexpected, particularly with regard to those experts who work with thermodynamics in their research: both physicists and one engineer. This question merits further study, particularly to probe when and how experts approach problems in which there is ambiguity in the choice of "independent" degrees of freedom.

Physicists
At first, the physicists thought that they were being asked to find (the inverse of) the spring constant. Once they recognized that the spring constant would be represented by a total derivative rather than a partial derivative, they spent some time exploring other possible meanings for the interview prompt. After they were told how to interpret the prompt, they moved immediately to describing the partial derivative as a ratio of small numerical changes, collecting data, and calculating a number. They spent some time establishing that their ratio was accurate enough and attended to determining how many independent variables were in the system and which of them to fix. While they acknowledged that the derivative was a function, they did not try to evaluate this function in any way. They did not mention slope at all, nor did they try to express the relationship between position and tension analytically so that they could take a symbolic derivative.

Engineers
Like the physicists, the engineers thought of the partial derivative as something that could be approximated experimentally, and spent a lot of time collecting data so that they could represent the partial derivative as a ratio of small numerical changes. They mentioned that the derivative could be found as a slope and that they could therefore determine the derivative by graphing their data, although they did not pursue this approach. They further noted that their experimentally-determined function should match some sort of theoretical equation. The engineers were the only ones to state a concept definition, albeit incomplete and bordering on inaccurate. They defined a partial derivative as "how one dimension in a multidimensional system changes when the other dimensions are fixed."

B. Need for an Extended Framework
In section II B, we gave a brief summary of Zandieh's framework for the concept image of derivative. In section III B, we gave our own list of five different ways to understand and think about the concept of derivative, which has substantial overlap with Zandieh's framework, but included a specific numerical category. Throughout our analysis, our interdisciplinary team found it useful to use elements of both of these descriptions. In particular, we found ourselves frequently exploiting Zandieh's choice of process-object pairs. However, we had deliberately made the choice to embed our interviews in an task in which the interviewees had easy access to numerical data, but not to an analytic expression for the relationship between the physical quantities. This is, after all, the environment in which most experimentalists typically find themselves. For this reason, our analysis focused on a numerical representation within the concept image of derivative.
We were surprised that Zandieh's framework did not include a numerical representation, even though her introduction made it clear that she was aware of this possibility. A careful reading of her paper shows several examples of students using numbers and these discussions are included in her analysis. This suggests that she considers numerical evaluation as a later step in each of her representations rather than as a representation in its own right. In our own list of five categories, we are now finding the possible need to distinguish between two numerical representations: exact numerical work (such as would happen when plugging particular values for the independent variable into an explicit analytical model for a physical process) and experimental numerical work (such as data from an experiment, including its experimental uncertainties). We are currently working on a paper which further elucidates these issues.
It will also be necessary to extend any resulting theoretical framework for the concept image of derivative to elements of the concept of partial derivative that are distinct from derivative. In particular, we see the need to add a layer for what is held fixed when finding a particular partial derivative.

C. Limits and the Real World
There were striking differences between the mathematicians and the physicists and engineers. The physicists and engineers were happy and relatively quick to find the derivative at a point as the numerical ratio of small changes. In this process, they were so casual with approximations that they were willing to give this ratio as their final answer to the interview prompt asking for a partial derivative without qualifying their answer with the word approximation. They did signal that they understood that their answer was an approximation by checking that the ratio they calculated was "good enough" by either decreasing (physicists) or increasing (engineers) the change in the independent variable. They also demonstrated an understanding of the function layer but couldn't be bothered to find the value of the derivative at more than one point. On the other hand, the mathematicians were firmly embedded in the function layer, pursuing the desire to find an analytic expression until they were explicitly asked by the interviewer "What else could you do?" Subsequently, when they were discussing finding a slope from their numerical data, they continued to doubt the reasonableness of the resulting approximation (see Excerpt 21).
We note that, in the idealized world of pure mathematics (and theoretical physics), the need to approximate rarely arises, whereas, for experimentalists, a ratio of small numerical changes is often the best answer that they have for a derivative, particularly in the absence of a theoretical model for the process they are studying. So the shorthand of calling this ratio "the derivative" rather than the more cumbersome "an approximation to the derivative" makes cultural sense. Experimentalists can always hope to design a better measuring apparatus or at least imagine a gedanken experiment to improve their approximation.
In the extreme case, physicists, in particular, are likely aware that continuity itself and therefore the ability to take formal limits in derivatives, are properties of the continuous models that they use for the physical world and not properties of the real world itself on atomic scales and below. The real world imposes a fundamental limitation on the concept of limit.

D. Limitations of this Work
Obviously, we cannot use a single interview with two or three content experts to make conclusions about every expert in a given field, nor can we come to definite conclusions about what these experts may or may not have been able to do in other settings or contexts. We plan to interview different types of mathematicians (especially computational and applied mathematicians who may have more experience with approximation), different types of physicists (including experimentalists and non-computational theorists), and different types of engineers. We would also like to interview experts in such fields as economics or oceanography, whose mathematical cultures may differ substantially from those already considered (and from each other). In particular, we hypothesize that thermodynamics experts from a variety of fields will behave more similarly on this task than nonthermodynamics experts within a single field.

E. Other Future Directions
It would be interesting to ask interviewees how they would report their data for the partial derivative (as-suming they were to take lots of data) if they were to publish this result in a paper. We are interested to see (a) if they address both dimensions of Zandieh's representation and (b) how they would represent the result: tables, single or multiple one-dimensional plots, threedimensional surfaces, contours, etc.
We would also be interested to explicitly prompt experts to give us a concept definition, to see how their concept image (as determined from their approach to this novel task) relates to their concept definition.
Every group spent some time exploring the physical system of the PDM. In future, it would be interesting to explore this aspect of the interviews and to think about the pedagogical and research implications of this type of play as an aspect of expert reasoning. In particular, the mathematicians spent more time than other groups exploring the covariation of the physical quantities by pulling on various strings and making expert-like observations of how the system responded.
Finally, we are interested in pursuing the pedagogical consequences of this study for classroom learning trajec-tories. For example, physics students have traditionally been exposed to the discreteness of data and the necessity of addressing experimental uncertainty in their lower division laboratories. But this exposure is disappearing with the near ubiquitous use of motion sensors and computer interfacing that blur the distinction between discrete data and continuous plots. What other physics experiences (research, advanced laboratory courses, etc.) will be needed to reinforce these aspects of the concept image of derivative?