betaFeedback

WorksPhysics and Astronomy

Analysis of Stability in Multistage Feedforward Operational Transconductance Amplifiers using Successive One-Pole Approximation

Abstract

This paper presents analysis results of the operational transconductance amplifiers (OTAs) that combine feedforward paths and multistage amplifiers to achieve high-gain wideband operation as well as frequency compensation. To analyze multistage feedforward OTAs and provide an intuitive design method, the successive one-pole approximation (SOPA) is used for each substage of a multistage feedforward OTA. Using SOPA, the stability analysis is carried out from the two-stage feedforward OTA to the four-stage feedforward OTA in this work.

INTRODUCTION

Semiconductor technology has continued to develop, which has significantly advanced electronic devices in terms of power, speed, and size. As CMOS technology has advanced rapidly, computing engines and communication devices with high data rates become available. In these applications that require high data rates, a high-gain wideband operational transconductance amplifier (OTA) is essential for various analog blocks. As the technology node shrinks, digital circuits can be developed to occupy less area, consume lower power, and enable high-speed operation. However, when developing analog circuits such as OTAs using advanced CMOS technologies, they must be carefully designed while considering the performance such as stability and gain.

As CMOS technology has advanced, the minimum length of transistors has decreased, which has improved the speed and power efficiency of digital circuits such as flip-flops and logic gates. However, an intrinsic gain of a single transistor is degraded due to the minimized gate length, causing gain errors in negative feedback systems. To overcome this gain degradation, the OTA can be developed by cascading multiple gain stages [1, 2, 3, 4, 5]. Especially, a two-stage OTA has been widely employed in bias circuits, analog front-ends, analog-to-digital converters, and wireless receivers for applications such as communications [6, 7, 8, 9] and sensor interfaces [10, 11, 12, 13].

As the number of cascaded gain stages increases, additional poles are created, causing phase shift and degrading stability in feedback systems. Therefore, in an OTA used in a feedback system, the technique called Miller compensation has been widely used to compensate for stability by controlling phase shift [4, 5]. Among the diverse compensation techniques to ensure stability in multiple gain stages, including the Miller compensation technique, the feedforward structure can be employed as a frequency compensation solution in a multistage OTA, which improves stability by canceling out poles [1]. This paper provides the analysis of stability in multistage feedforward OTAs which achieve high-gain wideband operation. In particular, the successive one-pole approximation (SOPA) is developed in this work to conduct intuitive stability analysis in multistage feedforward OTAs. By utilizing SOPA, systematic stability analysis is possible in feedforward OTAs with increasing gain stages. SOPA means that the one-pole approximation is successively applied to each substage of the multistage feedforward OTA, which is introduced in the remaining work of this paper. In this paper, by employing SOPA, the stability analysis is carried out from the two-stage feedforward OTA to the four-stage feedforward OTA. Additionally, the design limitation of increasing the number of feedforward stages is described based on SOPA after discussing the four-stage feedforward OTA.

This paper is organized as follows. Section II reviews Miller compensation which is a widely used method to secure stability for feedback systems. Section III provides the stability analysis by employing SOPA from the two-stage feedforward OTA to the four-stage feedforward OTA. Finally, the conclusion is drawn in Section IV. All analysis results are verified using a 65-nm CMOS technology.

General Consideration

Two-stage OTAs are widely used to provide a sufficient gain, but an additional pole contributes to phase shift, ultimately degrading the phase margin (PM). To mitigate the PM degradation, Miller compensation has been widely employed in numerous OTA designs. As shown in Figure 1(a), the Miller capacitor CmC_m is placed between the input and output nodes of the second stage M2M_2, splitting poles and ultimately improving the PM. At the interstage node XX, the capacitance looking into the gate terminal of M2M_2 increases to (1+Av2)Cm(1+A_{v2})C_m, shifting the dominant pole toward the origin in low frequencies [4]. When looking into the output node of M2M_2 at very high frequencies, the resistance decreases from ro2∥ro3r_{o2}\|r_{o3} to approximately Ro1∥(1/gm2)∥ro2∥ro3≈1/gm2R_{o1}\|(1/g_{m2})\|r_{o2}\|r_{o3}\approx1/g_{m2} since CmC_m becomes a low-impedance path at high frequencies [4]. Accordingly, when CmC_m is employed, the pole at node XX moves toward the origin, the pole at the output node moves away from the origin, and two poles are split [4]. Ro1R_{o1} means the output resistance of the first stage. ro2r_{o2} and ro3r_{o3} are the output resistances of M2M_2 and M3M_3, respectively.

Figure 1. (a) Miller compensation using CmC_m. (b) Addition of RzR_z to remove the RHP zero.

However, employing CmC_m also provides the feedforward path from node XX to the output node, contributing to a negative phase shift. In other words, as CmC_m creates the feedforward path, the gain becomes zero at ωz\omega_z called a right-half-plane (RHP) zero. In Figure 1(a), the RHP zero is obtained by solving the condition of VXCmωz=VXgm2V_X C_m \omega_z=V_X g_{m2} which means that the small-signal output becomes zero at ωz\omega_z [4]. Considering the gate-drain parasitic capacitance of M2M_2, Cgd2C_{gd2}, ωz\omega_z can be rewritten as gm2/(Cm+Cgd2)g_{m2}/(C_m+C_{gd2}). The RHP zero contributes to a negative phase shift as with a pole, resulting in the PM degradation. As a solution to remove the RHP zero, the compensation resistor RzR_z can be series-connected with CmC_m as shown in Figure 1(b). After adding RzR_z, ωz\omega_z is obtained as 1/[Cm(gm2−1−Rz)]1/[C_m(g_{m2}^{-1}-R_z)] by solving VXgm2=VX[Rz+(Cmωz)−1]V_Xg_{m2}=V_X[R_z+(C_m\omega_z)^{-1}] [4]. Therefore, the RHP zero can move to the left-half plane by satisfying Rz>gm2−1R_z>g_{m2}^{-1}. The left-half-plane (LHP) zero contributes to a positive phase shift, which can cancel out a pole through appropriate design techniques and eventually improve the phase margin and stability.

As a solution to secure the PM and stability, CmC_m and RzR_z can be employed for performing Miller compensation [4], and stability can be also secured by removing poles by adding active feedforward paths [1]. The following stability analysis covers the compensation technique with LHP zeros generated by feedforward paths implemented using active devices. To conduct a systematical stability analysis in multistage feedforward OTAs, SOPA is applied to each design of multistage feedforward OTAs. For simplicity, the stability analysis is first conducted on the two-stage feedforward structure using the one-pole approximation, and SOPA is explained based on the results of the two-stage analysis. Then, the stability analysis of three- and four-stage structures is carried out using SOPA. The design limitation of increasing the number of feedforward stages is also discussed using SOPA at the end of this work.

Stability Analysis using SOPA

Analysis of Two-Stage Feedforward Structure

For simplicity of multistage stability analysis, a two-stage feedforward OTA is first analyzed using the diagram shown in Figure 2(a). When neglecting the decoupling capacitor CdC_d, the gains of the main path are expressed as Av1=gm1Ro1A_{v1}=g_{m1}R_{o1} and Av2=gm2(Ro2∥RoF1)A_{v2}=g_{m2}(R_{o2}\mathbin{\|}R_{oF1}). The gain of the feedforward path is AvF1=gmF1(Ro2∥RoF1)A_{vF1}=g_{mF1}(R_{o2}\mathbin{\|}R_{oF1}). Then, the overall transfer function of the two-stage feedforward structure is given by

Figure 2. (a) Two-stage feedforward OTA and (b) its Bode plots.

VoutVin(s)∣2-stage=Av1Av2(1+sωp1)(1+sωp2)+AvF1(1+sωp2)(1)\left.\frac{V_{out}}{V_{in}}(s)\right|_{2\text{-stage}}=\frac{A_{v1}A_{v2}}{\left(1+\frac{s}{\omega_{p1}}\right)\left(1+\frac{s}{\omega_{p2}}\right)}+\frac{A_{vF1}}{\left(1+\frac{s}{\omega_{p2}}\right)} \tag*{(1)}
=(Av1Av2+AvF1)[1+AvF1s(Av1Av2+AvF1)ωp1](1+sωp1)(1+sωp2)(2)=(A_{v1}A_{v2}+A_{vF1})\frac{\left[1+\frac{A_{vF1}s}{(A_{v1}A_{v2}+A_{vF1})\omega_{p1}}\right]}{\left(1+\frac{s}{\omega_{p1}}\right)\left(1+\frac{s}{\omega_{p2}}\right)} \tag*{(2)}

The multistage feedforward OTAs used in this work are designed based on [1]. Eqs. (1) and (2) are obtained in the same way as [1]. From Eq. (2), the voltage gain at low frequencies can be obtained as gm1Ro1gm2(Ro2∥RoF1)+gmF1(Ro2∥RoF1)g_{m1}R_{o1}g_{m2}(R_{o2}\mathbin{\|}R_{oF1})+g_{mF1}(R_{o2}\mathbin{\|}R_{oF1}) and approximated as gm1Ro1gm2(Ro2∥RoF1)g_{m1}R_{o1}g_{m2}(R_{o2}\mathbin{\|}R_{oF1}). In Figure 2(a), if CdC_d is neglected, the poles at the first and second output stages are obtained as ωp1=1/(Ro1Co1)\omega_{p1}=1/(R_{o1}C_{o1}) and ωp2=1/(Ro2∥RoF1)(Co2+CoF1)\omega_{p2}=1/(R_{o2}\mathbin{\|}R_{oF1})(C_{o2}+C_{oF1}). From Eq. (2), the LHP zero generated by the feedforward path of gmF1g_{mF1} is expressed as ωz1=ωp1[(Av1Av2+AvF1)/AvF1]=ωp1[(gm1Ro1gm2/gmF1)+1]\omega_{z1}=\omega_{p1}[(A_{v1}A_{v2}+A_{vF1})/A_{vF1}]=\omega_{p1}[(g_{m1}R_{o1}g_{m2}/g_{mF1})+1]. If gmF1≫gm1,gm2g_{mF1}\gg g_{m1},g_{m2}, ωz1\omega_{z1} can be approximated as ωp1\omega_{p1}. Therefore, the first pole ωp1\omega_{p1} can be canceled out by choosing gmF1≫gm1,gm2g_{mF1}\gg g_{m1},g_{m2}. This ultimately improves the stability by converting the two-pole system into the one-pole system, which is called the one-pole approximation. Then, SOPA is an extended version of the one-pole approximation.

Although the PM is improved by employing a feedforward path, the voltage gain is degraded by RoF1R_{oF1}. Therefore, to isolate the output resistance of the feedforward path from the main path, the decoupling capacitor CdC_d can be placed between the feedforward path output and the main path output [2]. Compared to the design that places CdC_d at all inputs and outputs of feedforward paths [2], this work places CdC_d only at the output of the outer feedforward path to save the design area. By placing CdC_d as shown in Figure 2(a), the output impedance of the feedforward path becomes high at low frequencies and the effective output resistance of the OTA is dominated by Ro2R_{o2}. Therefore, the voltage gain changes from gm1Ro1gm2(Ro2∥RoF1)g_{m1}R_{o1}g_{m2}(R_{o2}\mathbin{\|}R_{oF1}) to gm1Ro1gm2Ro2g_{m1}R_{o1}g_{m2}R_{o2}. Also, the dominant pole ωp2\omega_{p2} approximately changes from 1/(Ro2∥RoF1)(Co2+CoF1)1/(R_{o2}\mathbin{\|}R_{oF1})(C_{o2}+C_{oF1}) to 1/(Ro2Co2)1/(R_{o2}C_{o2}) as the output impedance of the feedforward path is isolated from the main path. As the frequency increases, CdC_d operates almost as a short circuit. Accordingly, ωp1\omega_{p1} and ωz1\omega_{z1} located at high frequencies are almost the same as in the case without CdC_d.

Figure 2(b) shows the simulation results of gain and PM based on the two-stage feedforward OTA shown in Figure 2(a): using only the two-stage OTA without the feedforward path significantly degrades the PM while achieving a high gain (red line), adding the feedforward path to the main path improves the PM but the gain is reduced due to the output resistance of the feedforward path (blue line), and combining the feedforward path with CdC_d compensates for the gain while ensuring sufficient PM (green line). As shown in the phase shift of Figure 2(b), the two-pole system (red line) is approximated as the one-pole system (blue and green lines) as the feedforward path is added.

Figure 2(b) is obtained using the schematic shown in Figure 3 by setting Cd=2C_d = 2 pF, VDD=1.2V_{DD} = 1.2 V, Vcm=0.6V_{cm} = 0.6 V, and Iref=400I_{ref} = 400 μA. The output load impedance is modeled as 10 MΩ∥\parallel2 pF on Vout−V_{out-} and Vout+V_{out+}. For simplicity of the analysis throughout this work, the equation derivation is conducted using the single-ended schematic (e.g., Figure 2(a)), and the simulation results are obtained using the fully-differential schematic (e.g., Figure 3).

Figure 3. (a) Fully differential two-stage feedforward OTA, (b) its bias block, and (c) the common-mode feedback used in the output stage.

Figure 3 shows the overall schematics of the fully differential two-stage feedforward OTA. The input stage gm1g_{m1} is implemented by employing the cross-coupled structure M2M_2 and the diode-connected load M3M_3, which improves the voltage gain by maximizing the output resistance of the first stage [3]. The output stages of gm2g_{m2} and gmF1g_{mF1} are implemented as the current reuse (or inverter) structure. The common-mode voltage of each output stage is set using the common-mode feedbacks CM1CM_1 and CM2CM_2. gm1g_{m1} consumes 273 μA. gm2g_{m2} and gmF1g_{mF1} consume 1.48 mA and 3 mA, respectively. Each CMn=1,2CM_{n=1,2} consumes 104 μA. Throughout this paper, the multistage feedforward OTAs are implemented by employing the bias block and the common-mode feedback, as shown in Figure 3(b) and (Figure 3c). Note that the values of transconductances such as gm1g_{m1}, gm2g_{m2}, and gmF1g_{mF1} are summarized in the table at the end of this work.

As discussed above using the schematic of Figure 2(a), the two-pole system can be converted into the one-pole system by choosing gmF1g_{mF1} sufficiently large as shown in Figure 4(a). To carry out the systematical stability analysis of multistage feedforward OTAs, the one-pole approximation shown in Figure 4(a) is successively applied to the multistage feedforward structures, defined as the successive one-pole approximation (SOPA) in this work. Figure 4(a) describes the one-pole approximation depending on CdC_d, which is possible by canceling out the pole using the LHP zero through the feedforward path as [1, 2, 3, 7].

Figure 4. (a) Approximation from the two-pole system to the one-pole system without (or with) CdC_d. (b) Successive one-pole approximation (SOPA).

Figure 4(b) describes the concept of SOPA that approximates the four-stage feedforward OTA into the one-pole system. SOPA is conducted by assuming that each feedforward path of each substage is designed to have a sufficiently large transconductance than that of the main path in the substage. Based on SOPA, the stability analysis of the three- and four-stage feedforward structures is carried out in the remaining work. Note that the substages in Figure 4(b) are distinguished using different color boxes.

Analysis of Three-Stage Feedforward Structure

Extending Eqs. (1) and (2) to a three-stage structure, the OTA produces three poles and two zeros in Figure 5(a). However, obtaining poles and zeros becomes complex as the number of stages extends. Therefore, the three-stage feedforward OTA is approximated by replacing the internal two-stage feedforward structure with the one-pole system as shown in Figure 5(b). As discussed in the two-stage feedforward OTA shown in Figure 2(a), ωp1\omega_{p1} can be canceled out by ωz1\omega_{z1} through a feedforward path. Accordingly, a two-stage feedforward OTA can be approximated as a one-pole system with a pole of 1/(Ro2∥RoF1)(Co2+CoF1)1/(R_{o2}\mathbin{\|}R_{oF1})(C_{o2}+C_{oF1}) and a gain of gm1Ro1gm2(Ro2∥RoF1)g_{m1}R_{o1}g_{m2}(R_{o2}\mathbin{\|}R_{oF1}) as shown in Figure 5(b).

Figure 5. (a) Three-stage feedforward OTA and (b) its approximated schematic.

The Miller capacitors, Cm1C_{m1} and Cm2C_{m2}, are employed in Figure 5(a) and (Figure 5b). However, to simplify the circuit analysis, Cm1C_{m1} and Cm2C_{m2} are first neglected. Extending Eq. (1) to the three-stage feedforward structure while neglecting CdC_d, the overall transfer function of Figure 5(b) is given by

VoutVin(s)∣3−stage=A2−stageAv3(1+sωp1′)(1+sωp2′)+AvF2(1+sωp2′)(3)\left.\frac{V_{\mathrm{out}}}{V_{\mathrm{in}}}(s)\right|_{\mathrm{3-stage}} = \frac{A_{\mathrm{2-stage}}A_{\mathrm{v3}}} {\left(1+\frac{s}{\omega'_{p1}}\right)\left(1+\frac{s}{\omega'_{p2}}\right)} + \frac{A_{\mathrm{vF2}}}{\left(1+\frac{s}{\omega'_{p2}}\right)} \tag*{(3)}

where A2−stageA_{\mathrm{2-stage}} is an approximated voltage gain of the internal two-stage structure, Av3A_{\mathrm{v3}} is the output stage gain of the main path, AvF2A_{\mathrm{vF2}} is the output stage gain of the feedforward path, ωp1′\omega'_{p1} is a pole at the output of the internal two-stage structure, and ωp2′\omega'_{p2} is an output pole. Neglecting CdC_d for simplicity, each stage has a gain as follows: A2−stage=gm1Ro1gm2(Ro2∥RoF1)A_{\mathrm{2-stage}}=g_{m1}R_{o1}g_{m2}(R_{o2}\mathbin{\|}R_{oF1}), Av3=gm3(Ro3∥RoF2)A_{\mathrm{v3}}=g_{m3}(R_{o3}\mathbin{\|}R_{oF2}), and AvF2=gmF2(Ro3∥RoF2)A_{\mathrm{vF2}}=g_{mF2}(R_{o3}\mathbin{\|}R_{oF2}). To obtain poles and zero in Figure 5(b), Eq. (3) can be rewritten as

(A2-stageAv3+AvF2)[1+AvF2s(A2-stageAv3+AvF2)ωp1′](1+sωp1′)(1+sωp2′)(4)\frac{\left(A_{2\text{-stage}}A_{v3}+A_{vF2}\right)\left[1+\frac{A_{vF2}s}{\left(A_{2\text{-stage}}A_{v3}+A_{vF2}\right)\omega'_{p1}}\right]}{\left(1+\frac{s}{\omega'_{p1}}\right)\left(1+\frac{s}{\omega'_{p2}}\right)} \tag*{(4)}

In Eq. (4), the gain at low frequencies is gm1Ro1gm2(Ro2∥RoF1)gm3(Ro3∥RoF2)+gmF2(Ro3∥RoF2)g_{m1}R_{o1}g_{m2}(R_{o2}\mathbin{\|}R_{oF1})g_{m3}(R_{o3}\mathbin{\|}R_{oF2})+g_{mF2}(R_{o3}\mathbin{\|}R_{oF2}). This gain is approximated by gm1Ro1gm2(Ro2∥RoF1)gm3(Ro3∥RoF2)g_{m1}R_{o1}g_{m2}(R_{o2}\mathbin{\|}R_{oF1})g_{m3}(R_{o3}\mathbin{\|}R_{oF2}). Two poles are obtained as ωp1′=1/(Ro2∥RoF1)(Co2+CoF1)\omega'_{p1}=1/(R_{o2}\mathbin{\|}R_{oF1})(C_{o2}+C_{oF1}) and ωp2′=1/(Ro3∥RoF2)(Co3+CoF2)\omega'_{p2}=1/(R_{o3}\mathbin{\|}R_{oF2})(C_{o3}+C_{oF2}). The zero by the feedforward path of gmF2g_{mF2} is expressed as

ωz1′=ωp1′(A2-stageAv3AvF2+1)=ωp1′[gm1Ro1gm2(Ro2∥RoF1)gm3gmF2+1](5)\begin{aligned} \omega'_{z1} &= \omega'_{p1}\left(\frac{A_{2\text{-stage}}A_{v3}}{A_{vF2}}+1\right) \\ &= \omega'_{p1}\left[\frac{g_{m1}R_{o1}g_{m2}(R_{o2}\mathbin{\|}R_{oF1})g_{m3}}{g_{mF2}}+1\right] \tag*{(5)} \end{aligned}

In Figure 5(b), considering CdC_d at low frequencies, the output impedance of the feedforward path becomes high. At low frequencies, the output impedance is therefore dominated by Ro3R_{o3} and Co3C_{o3} while ignoring RoF2R_{oF2} and CoF2C_{oF2}. Accordingly, the voltage gain from Eq. (4) can be expressed as gm1Ro1gm2(Ro2∥RoF1)gm3Ro3g_{m1}R_{o1}g_{m2}(R_{o2}\mathbin{\|}R_{oF1})g_{m3}R_{o3}. Also, the dominant pole ωp2′\omega'_{p2}, which is in a relatively lower frequency range than ωp1′\omega'_{p1} and ωz1′\omega'_{z1}, changes from 1/(Ro3∥RoF2)(Co3+CoF2)1/(R_{o3}\mathbin{\|}R_{oF2})(C_{o3}+C_{oF2}) to 1/(Ro3Co3)1/(R_{o3}C_{o3}). As the frequency increases, CdC_d becomes almost a short circuit. Therefore, ωp1′\omega'_{p1} and ωz1′\omega'_{z1}, which are in the relatively higher frequency range, remain similar to the state in the absence of CdC_d.

Recall that ωz1=ωp1[(gm1Ro1gm2/gmF1)+1]\omega_{z1}=\omega_{p1}[(g_{m1}R_{o1}g_{m2}/g_{mF1})+1] from the two-stage feedforward OTA. If gmF1g_{mF1} is large enough, ωz1\omega_{z1} can be designed as ωp1\omega_{p1}, and ωp1\omega_{p1} can be canceled out by ωz1\omega_{z1}. Similar to this operation, in Eq. (5), ωz1′\omega'_{z1} can be designed as ωp1′\omega'_{p1} if gmF2g_{mF2} is large enough. However, in practical design, there is a limit to increasing gmF2g_{mF2} and gm1Ro1gm2(Ro2∥RoF1)g_{m1}R_{o1}g_{m2}(R_{o2}\mathbin{\|}R_{oF1}) is also large, so ωz1′\omega'_{z1} is generally designed to be larger than ωp1′\omega'_{p1}. Figure 6(a) shows the phase shift of the three-stage feedforward OTA and the phase peaking is generated due to the difference between ωp1′\omega'_{p1} and ωz1′\omega'_{z1}.

Figure 6. (a) Phase shift of the three-stage feedforward OTA, (b) Bode plots considering feedforward path and CdC_d, and (c) Bode plots considering CmC_m.

Figure 6(b) presents the simulated Bode plots of the three-stage feedforward OTA according to gmF2g_{mF2} and CdC_d. By including the feedforward amplifier gmF2g_{mF2} (blue line), the PM is improved but the difference between ωp1′\omega'_{p1} and ωz1′\omega'_{z1} accompanies the phase peaking. Also, employing CdC_d improves the voltage gain at low frequencies (green line). To further improve the PM, the Miller capacitors, Cm1C_{m1} and Cm2C_{m2} shown in Figure 5, are employed. As shown in Figure 6(c), the PM is improved through Miller compensation [4], which exchanges the positions of ωp1′\omega'_{p1} and ωp2′\omega'_{p2}, but the bandwidth is sacrificed. ωz1′\omega'_{z1} also moves slightly toward the origin as ωp1′\omega'_{p1} approaches the origin. In this three-stage OTA design, CdC_d is set to 2 pF, and Cm1C_{m1} and Cm2C_{m2} are set to 50 fF. This means that CdC_d makes the low-impedance path than Cm1C_{m1} and Cm2C_{m2}. Therefore, the OTA is still dominated by the feedforward path than the Miller compensation path at high frequencies, which means that Eq. (5) is still valid to describe ωz1′\omega'_{z1}. Comparing the three-stage feedforward OTA and the two-stage feedforward OTA, the voltage gain improves as the number of stages increases. However, the three-stage structure shows a rapid phase change because ωz1′\omega'_{z1} is not completely designed as ωp1′\omega'_{p1} in an actual circuit design.

Figure 6(b) and (Figure 6c) are obtained using the fully differential three-stage feedforward OTA shown in Figure 7. The bias block and common-mode feedback are the same as those shown in Figure 3(b) and (Figure 3c). The output load impedance is modeled as 10 MΩ∥\Omega\mathbin{\|}2 pF on Vout−V_{out-} and Vout+V_{out+}. The supply voltage VDDV_{DD} is set to 1.2 V. The common-mode voltages of Vout+/−V_{out+/-} and Vin+/−V_{in+/-} are set to 0.6 V. The current consumption of each stage is as follows: gm1g_{m1} (274 μ\muA), gm2g_{m2} (794 μ\muA), gmF1g_{mF1} (821 μ\muA), gm3g_{m3} (1.64 mA), gmF2g_{mF2} (3.28 mA), and CMn=1,2,3CM_{n=1,2,3} (104 μ\muA). Note that each CMn=1,2,3CM_{n=1,2,3} sets the common-mode voltage using VcmV_{cm} (= 0.6 V) as shown in Figure 3(c).

Figure 7. Fully differential three-stage feedforward OTA.

Analysis of Four-Stage Feedforward Structure

Recall that ωz1=ωp1[(gm1Ro1gm2/gmF1)+1]\omega_{z1} = \omega_{p1} [(g_{m1}R_{o1}g_{m2}/g_{mF1})+1] in the two-stage OTA, then the two-stage feedforward OTA can be approximated as a one-pole system if ωz1=ωp1\omega_{z1} = \omega_{p1} by making gmF1g_{mF1} large enough. By employing the approximated model of the two-stage feedforward OTA, the three-stage design also can be approximated as shown in Figure 5(b).

Then, the zero of ωz1′\omega'_{z1} is obtained as ωp1′[(gm1Ro1gm2(Ro2∥RoF1)gm3/gmF2)+1]\omega'_{p1} [(g_{m1}R_{o1}g_{m2}(R_{o2}\mathbin{\|}R_{oF1})g_{m3}/g_{mF2})+1] as shown in Eq. (5). Similar to ωz1\omega_{z1} of the two-stage design, if gmF2g_{mF2} is large enough, ωz1′\omega'_{z1} becomes ωp1′\omega'_{p1}, and ωp1′\omega'_{p1} can be canceled out by ωz1′\omega'_{z1}. Therefore, theoretically, the three-stage OTA can be simplified into the two-pole system by making gmF1g_{mF1} large and further simplified into the one-pole system by making gmF2g_{mF2} sufficiently large, which is described through SOPA as shown in Figure 4(b).

Figure 8 shows the four-stage feedforward OTA including the approximated circuit model of the preceding stages through SOPA. Note that the Miller capacitors Cm1,2,3C_{m1,2,3} are employed in the four-stage design, but will be first neglected in the analysis for simplicity and discussed in the Bode plot results. In the red box of Figure 8, if gmF1g_{mF1} is large enough, the two-stage design can be simplified as the one-pole system employing Gm′=(gm1Ro1)gm2G'_m=(g_{m1}R_{o1})g_{m2}, Ro′=Ro2∥RoF1R'_o=R_{o2}\mathbin{\|}R_{oF1}, and Co′=Co2+CoF1C'_o=C_{o2}+C_{oF1}. In the blue box of Figure 8, if gmF1g_{mF1} and gmF2g_{mF2} are large enough, the three-stage design can be simplified into the two-pole system and further simplified into the one-pole system by employing Gm′′G''_m, Ro′′R''_o, and Co′′C''_o. In Eq. (4), Gm′′G''_m is approximately obtained as gm1Ro1gm2(Ro2∥RoF1)gm3g_{m1}R_{o1}g_{m2}(R_{o2}\mathbin{\|}R_{oF1})g_{m3}. Ro′′R''_o and Co′′C''_o are obtained as Ro3∥RoF2R_{o3}\mathbin{\|}R_{oF2} and Co3+CoF2C_{o3}+C_{oF2}, respectively.

Figure 8. Four-stage feedforward OTA with SOPA.

Leveraging approximated one-pole models of the two- and three-stage OTAs while neglecting CdC_d in Figure 8, the transfer function of the four-stage OTA can be given by

VoutVin(s)∣4-stage=A3-stageAv4(1+sωp1′′)(1+sωp2′′)+AvF3(1+sωp2′′)(6)\left.\frac{V_{out}}{V_{in}}(s)\right|_{4\text{-stage}}= \frac{A_{3\text{-stage}}A_{v4}}{\left(1+\frac{s}{\omega''_{p1}}\right)\left(1+\frac{s}{\omega''_{p2}}\right)} +\frac{A_{vF3}}{\left(1+\frac{s}{\omega''_{p2}}\right)} \tag*{(6)}
=(A3-stageAv4+AvF3)[1+AvF3s(A3-stageAv4+AvF3)ωp1′′](1+sωp1′′)(1+sωp2′′)(7)=\left(A_{3\text{-stage}}A_{v4}+A_{vF3}\right) \frac{\left[1+\frac{A_{vF3}s}{\left(A_{3\text{-stage}}A_{v4}+A_{vF3}\right)\omega''_{p1}}\right]} {\left(1+\frac{s}{\omega''_{p1}}\right)\left(1+\frac{s}{\omega''_{p2}}\right)} \tag*{(7)}

where A3-stage=Gm′′Ro′′A_{3\text{-stage}}=G''_mR''_o, Av4=gm4(Ro4∥RoF3)A_{v4}=g_{m4}(R_{o4}\mathbin{\|}R_{oF3}), and AvF3=gmF3(Ro4∥RoF3)A_{vF3}=g_{mF3}(R_{o4}\mathbin{\|}R_{oF3}). The poles of the three- and four-stage outputs are expressed as ωp1′′=1/(Ro3∥RoF2)(Co3+CoF2)\omega''_{p1}=1/(R_{o3}\mathbin{\|}R_{oF2})(C_{o3}+C_{oF2}) and ωp2′′=1/(Ro4∥RoF3)(Co4+CoF3)\omega''_{p2}=1/(R_{o4}\mathbin{\|}R_{oF3})(C_{o4}+C_{oF3}), respectively. In Eq. (7) obtained using SOPA, the zero of the four-stage feedforward OTA is given by

ωz1′′=ωp1′′(A3-stageAv4AvF3+1)=ωp1′′[ ]\begin{aligned} \omega''_{z1}=\omega''_{p1}\left(\frac{A_{3\text{-stage}}A_{v4}}{A_{vF3}}+1\right) \\ &=\omega''_{p1}\left[\ \right] \end{aligned}
⇓\Downarrow
[gm1Ro1gm2(Ro2∥RoF1)gm3(Ro3∥RoF2)gm4gmF3+1](8)\left[ \frac{g_{m1}R_{o1}g_{m2}(R_{o2}\mathbin{\|}R_{oF1})g_{m3}(R_{o3}\mathbin{\|}R_{oF2})g_{m4}}{g_{mF3}}+1 \right] \tag*{(8)}

Table 1 summarizes the zeros according to the number of stages. Recall that ωz1′\omega'_{z1} and ωz1′′\omega''_{z1} are obtained assuming that gmF1g_{mF1} and gmF2g_{mF2} are large enough. Similar to two- and three-stage OTAs, for improving the PM in the four-stage feedforward OTA, gmF3g_{mF3} must also be large enough to cancel out ωp1′′\omega''_{p1}. From the equations of ωz1\omega_{z1}, ωz1′\omega'_{z1}, and ωz1′′\omega''_{z1} as summarized in Table 1, it can be concluded that the transconductance of the feedforward path must be sufficiently large to secure the PM. Assuming that gmF1g_{mF1} and gmF2g_{mF2} are large enough, Eq. (7) can be rewritten as

VoutVin(s)∣4-stage=A4-stage(1+sωz1)(1+sωz1′)(1+sωp1)(1+sωp1′)(1+sωz1′′)(1+sωp1′′)(1+sωp2′′)(9)\left.\frac{V_{out}}{V_{in}}(s)\right|_{4\text{-stage}}=\frac{A_{4\text{-stage}}\left(1+\frac{s}{\omega_{z1}}\right)\left(1+\frac{s}{\omega'_{z1}}\right)}{\left(1+\frac{s}{\omega_{p1}}\right)\left(1+\frac{s}{\omega'_{p1}}\right)} \frac{\left(1+\frac{s}{\omega''_{z1}}\right)}{\left(1+\frac{s}{\omega''_{p1}}\right)\left(1+\frac{s}{\omega''_{p2}}\right)} \tag*{(9)}

where A4-stageA_{4\text{-stage}} can be approximated by Gm′′Ro′′gm4(Ro4∥RoF3)G''_{m}R''_{o}g_{m4}(R_{o4}\mathbin{\|}R_{oF3}). The four-stage OTA shown in Figure 8 consists of four poles and three zeros in Eq. (9). Ideally, by making gmF1g_{mF1} large enough, [1+(s/ωp1)]\left[1+\left(s/\omega_{p1}\right)\right] in the denominator can be canceled out by [1+(s/ωz1)]\left[1+\left(s/\omega_{z1}\right)\right] in the numerator. Similarly, [1+(s/ωp1′)]\left[1+\left(s/\omega'_{p1}\right)\right] in the denominator can be canceled out by [1+(s/ωz1′)]\left[1+\left(s/\omega'_{z1}\right)\right] in the numerator as gmF2g_{mF2} becomes large enough. Eventually, the four-stage feedforward OTA can be converted into the one-pole system assuming that [1+(s/ωp1′′)]\left[1+\left(s/\omega''_{p1}\right)\right] in the denominator is canceled out by [1+(s/ωz1′′)]\left[1+\left(s/\omega''_{z1}\right)\right] in the numerator as gmF3g_{mF3} becomes sufficiently large. The approximation procedure of Eq. (9) described above is illustrated by SOPA in Figure 4(b). As mentioned above, to satisfy the condition that the poles are canceled out by the zeros, all feedforward stages need to be designed to have sufficiently large transconductances while sustaining gmF1≪gmF2≪gmF3g_{mF1}\ll g_{mF2}\ll g_{mF3} as shown in Table 1. However, as the number of stages increases, the burden on the transconductance of the feedforward path increases for canceling out poles, which is associated with design area and power consumption.

OrderZero
2-stageωz1=ωp1(gm1Ro1gm2gmF1+1)\omega_{z1}=\omega_{p1}\left(\frac{g_{m1}R_{o1}g_{m2}}{g_{mF1}}+1\right)
3-stageωz1′=ωp1′[gm1Ro1gm2(Ro2∣RoF1)gm3gmF2+1]\omega'_{z1}=\omega'_{p1}\left[\frac{g_{m1}R_{o1}g_{m2}(R_{o2}\mathbin{|}R_{oF1})g_{m3}}{g_{mF2}}+1\right]
4-stageωz1′′=ωp1′′[gm1Ro1gm2(Ro2∣RoF1)gm3(Ro3∣RoF2)gm4gmF3+1]\omega''_{z1}=\omega''_{p1}\left[\frac{g_{m1}R_{o1}g_{m2}(R_{o2}\mathbin{|}R_{oF1})g_{m3}(R_{o3}\mathbin{|}R_{oF2})g_{m4}}{g_{mF3}}+1\right]

Table 1. Zeros according to the number of stages

In other words, to cancel out a pole by a zero in each stage, the transconductance ratio between the output stages of each order must be greater than the gain of the previous stage. Therefore, the following conditions must be satisfied:

(gmF1gm2)≫gm1Ro1(10)\left(\frac{g_{mF1}}{g_{m2}}\right)\gg g_{m1}R_{o1} \tag*{(10)}
(gmF2gm3)≫gm1Ro1gm2(Ro2∥RoF1)(11)\left(\frac{g_{mF2}}{g_{m3}}\right)\gg g_{m1}R_{o1}g_{m2}(R_{o2}\mathbin{\|}R_{oF1}) \tag*{(11)}
(gmF3gm4)≫gm1Ro1gm2(Ro2∥RoF1)gm3(Ro3∥RoF2)(12)\left(\frac{g_{mF3}}{g_{m4}}\right)\gg g_{m1}R_{o1}g_{m2}(R_{o2}\mathbin{\|}R_{oF1})g_{m3}(R_{o3}\mathbin{\|}R_{oF2}) \tag*{(12)}

However, considering DC bias conditions, power consumption, and design area, it becomes more difficult to satisfy the above conditions as the number of stages increases. For these reasons, it becomes difficult to cancel out a pole that is generated as the number of stages increases, ultimately leading to the degradation of stability.

Figure 9(a) shows the simulated Bode plots obtained using the four-stage feedforward OTA without Miller compensation. Employing the feedforward path with gmF3g_{mF3} and without CdC_d alleviates the phase shift at high frequencies compared to the design without the feedforward path (blue and red lines in Figure 9(a)). However, the zero is located at a higher frequency than the pole, and the zero does not completely cancel out the pole, leading to the PM degradation. As shown in the green line in Figure 9(a), by isolating the output feedforward path using CdC_d, the voltage gain is improved from Gm′′Ro′′gm4(Ro4∥RoF3)G^{\prime\prime}_{m}R^{\prime\prime}_{o}g_{m4}(R_{o4}\mathbin{\|}R_{oF3}) to Gm′′Ro′′gm4Ro4G^{\prime\prime}_{m}R^{\prime\prime}_{o}g_{m4}R_{o4} at low frequencies. Also, the dominant pole ωp2′′\omega^{\prime\prime}_{p2} changes from 1/(Ro4∥RoF3)(Co4+CoF3)1/(R_{o4}\mathbin{\|}R_{oF3})(C_{o4}+C_{oF3}) to 1/(Ro4Co4)1/(R_{o4}C_{o4}) as CdC_d is employed. The other poles and zeros located at high frequencies are almost unaffected by CdC_d since CdC_d becomes almost a short circuit.

Bode plots of the four-stage feedforward OTA without and with Miller capacitors

Figure 9. (a) Bode plots of the four-stage feedforward OTA without the Miller capacitors and (b) Bode plots employing the Miller capacitors.

As shown in Figure 9(b), the PM can be further improved by employing the Miller capacitors Cm1,2,3C_{m1,2,3} shown in Figure 8. Similar to Miller compensation in the two-stage design, the poles and zeros can be located closely together through pole splitting [4]. Therefore, the zero can effectively cancel out the pole that shifts toward high frequencies by pole splitting, resulting in an improved PM. However, the pole that moves toward the origin through pole splitting reduces the bandwidth. In this analysis, Cm1C_{m1} and Cm3C_{m3} are set to 10 fF, and Cm2C_{m2} is set to 200 fF. Also, the load impedance is set to 10 MΩ∥\Omega\mathbin{\|}2 pF on VoutV_{out} for this analysis. Note that the output load impedance is set to 10 MΩ∥\Omega\mathbin{\|}2 pF in analyzing all multistage feedforward OTAs throughout this work, and all output stages are designed using large device size to achieve a large transconductance. Accordingly, when analyzing the OTAs without Miller capacitors, it is assumed that ωp2\omega_{p2}, ωp2′\omega^{\prime}_{p2}, and ωp2′′\omega^{\prime\prime}_{p2} are the dominant poles.

Fig. 10 shows the fully differential feedforward OTA used to obtain the results shown in Figure 9. In Figure 10, the bias block and common-mode feedbacks CMn=1,2,3\mathrm{CM}_{n=1,2,3} are the same as the circuits used in two- and three-stage OTAs. The common-mode voltages of Vin+/−V_{in+/-} and Vout+/−V_{out+/-} are set to 0.6 V. CdC_d is set to 2 pF. The current consumption of the four-stage feedforward OTA is as follows: gm1g_{m1} (273 μ\muA), gm2g_{m2} (273 μ\muA), gmF1g_{mF1} (820 μ\muA), gm3g_{m3} (1.61 mA), gmF2g_{mF2} (1.64 mA), gm4g_{m4} (1.63 mA), gmF3g_{mF3} (3.28 mA), and CMn=1,2,3\mathrm{CM}_{n=1,2,3} (104 μ\muA). The OTA is driven using a 1.2 V supply voltage.

Figure 10. Fully differential four-stage feedforward OTA.

Table 2 summarizes the overall transconductances used in this analysis according to the number of stages. Note that the transconductance ratio is defined as gmF(k−1)/gm(k)g_{mF(k-1)}/g_{m(k)}, k=2,3,4k=2,3,4 in this analysis. As discussed above, gmF(k−1)/gm(k)g_{mF(k-1)}/g_{m(k)} must be designed as large as possible to secure sufficient phase margin in a multistage feedforward OTA. However, as the OTA order increases, achieving sufficient gmF(k−1)/gm(k)g_{mF(k-1)}/g_{m(k)} is a challenging task and becomes even more difficult when considering DC bias conditions, power consumption, and design area. Therefore, the stability analysis is carried out by setting gmF(k−1)/gm(k)g_{mF(k-1)}/g_{m(k)} to be larger than 2 in this work, as summarized in Table 2.

OrderTransconductance
2-stagegm1g_{m1} (1.99 mS)gm2g_{m2} (22.12 mS)gmF1g_{mF1} (44.97 mS)
3-stagegm1g_{m1} (1.99 mS)gm2g_{m2} (2.27 mS)gmF1g_{mF1} (5.97 mS)
3-stagegm3g_{m3} (28.04 mS)gmF2g_{mF2} (56.29 mS)
4-stagegm1g_{m1} (1.99 mS)gm2g_{m2} (0.79 mS)gmF1g_{mF1} (5.96 mS)
4-stagegm3g_{m3} (4.76 mS)gmF2g_{mF2} (11.95 mS)
4-stagegm4g_{m4} (28.05 mS)gmF3g_{mF3} (56.28 mS)

Table 2. Overall transconductances

Conclusion

In this work, the stability analysis of multistage feedforward OTAs is conducted by employing SOPA. In each multistage feedforward OTA, SOPA is conducted as follows: (1) the first two-stage feedforward structure is modeled as the one-pole system by assuming that the first feedforward transconductance is large enough, (2) the next stage is again approximated identically by leveraging the previous one-pole system and making the second feedforward transconductance sufficiently large, and (3) the approximation is successively conducted by employing the previous one-pole system and making the next feedforward transconductance large enough.

To ensure stability, each feedforward structure must satisfy Eq. (13).

gmF(k−1)gm(k)≫gm1Ro1×[gm2(Ro2∥RoF1)]×⋯⋯×[gm(k−2)(Ro(k−2)∥RoF(k−3))]×[gm(k−1)(Ro(k−1)∥RoF(k−2))](13)\frac{g_{mF(k-1)}}{g_{m(k)}} \gg g_{m1}R_{o1} \times\left[g_{m2}\left(R_{o2}\mathbin{\|}R_{oF1}\right)\right] \times\cdots \cdots\times\left[g_{m(k-2)}\left(R_{o(k-2)}\mathbin{\|}R_{oF(k-3)}\right)\right] \times\left[g_{m(k-1)}\left(R_{o(k-1)}\mathbin{\|}R_{oF(k-2)}\right)\right] \tag*{(13)}

where kk (= 2, 3, 4, ...) means the order of a multistage feedforward OTA, and RoF(0)R_{oF(0)} is an infinite value. Eq. (13) is obtained by generalizing Eqs. (10), (11), and (12). Then, the generalized transfer function of a multistage feedforward OTA can be expressed as Eqs. (14) and (15).

Vout(s)Vin∣(k)-stage=A(k)-stage[(1+sωz1)(1+sωz1′)×⋯×(1+sωz1′′⋯′)][(1+sωp1)(1+sωp1′)×⋯×(1+sωp1′′⋯′)](1+sωp2′′⋯′)(14)\left.\frac{V_{out}(s)}{V_{in}}\right|_{(k)\text{-stage}}= \frac{A_{(k)\text{-stage}}\left[\left(1+\frac{s}{\omega_{z1}}\right)\left(1+\frac{s}{\omega'_{z1}}\right)\times\cdots\times\left(1+\frac{s}{\omega^{\prime\prime\cdots\prime}_{z1}}\right)\right]} {\left[\left(1+\frac{s}{\omega_{p1}}\right)\left(1+\frac{s}{\omega'_{p1}}\right)\times\cdots\times\left(1+\frac{s}{\omega^{\prime\prime\cdots\prime}_{p1}}\right)\right]\left(1+\frac{s}{\omega^{\prime\prime\cdots\prime}_{p2}}\right)} \tag*{(14)}
=A(k)-stage[(1+sωz1)×∏#=1′′′′(1+sω#z1)][(1+sωp1)×∏#=1′′′′(1+sω#p1)](1+sωp2′′⋯′)(15)=\frac{A_{(k)\text{-stage}}\left[\left(1+\frac{s}{\omega_{z1}}\right)\times\prod_{\#=1}^{\prime\prime\prime\prime}\left(1+\frac{s}{\omega_{\#z1}}\right)\right]} {\left[\left(1+\frac{s}{\omega_{p1}}\right)\times\prod_{\#=1}^{\prime\prime\prime\prime}\left(1+\frac{s}{\omega_{\#p1}}\right)\right]\left(1+\frac{s}{\omega^{\prime\prime\cdots\prime}_{p2}}\right)} \tag*{(15)}

where A(k)-stageA_{(k)\text{-stage}} means the voltage gain of the kk-stage feedforward OTA at low frequencies. Eq. (14) is extended from the four-stage transfer function shown in Eq. (9) and further generalized as shown in Eq. (15). Also, Eqs. (14) and (15) describe the behavior of SOPA shown in Figure 4(b). Although designing a reliable OTA becomes difficult as the number of stages increases in actual circuit design, SOPA provides an intuitive way to analyze the stability of a multistage feedforward OTA.

References

  1. [1]B. K. Thandri and J. Silva-Martinez, “A robust feedforward compensation scheme for multistage operational transconductance amplifiers with no Miller capacitors,” IEEE J. Solid-State Circuits, vol. 38, no. 2, pp. 237–243, Feb. 2003.DOI
  2. [2]H. Jung, D. R. Utomo, S.-K. Han, J. Kim, and S.-G. Lee, “An 80 MHz bandwidth and 26.8 dBm OOB IIP3 transimpedance amplifier with improved nested feedforward compensation and multi-order filtering,” IEEE Trans. Circuits Syst. I: Regul. Pap., vol. 67, no. 10, pp. 3410–3421, Oct. 2020.DOI
  3. [3]X. Yang and H.-S. Lee, “Design of a 4th-order multi-stage feedforward operational amplifier for continuous-time bandpass delta sigma modulators,” in Proc. IEEE Int. Symp. Circuits Syst., May 2016, pp. 1058–1061.
  4. [4]B. Razavi, Design of Analog CMOS Integrated Circuits, 2nd ed. (New York: McGraw-Hill, 2016).
  5. [5]R. G. H. Eschauzier, L. P. P. T. Kerklaan, and J. H. Huijsing, “A 100-MHz 100-dB operational amplifier with multistage nested Miller compensation structure,” IEEE J. Solid-State Circuits, vol. 27, no. 12, pp. 1709–1717, Dec. 1992.
  6. [6]J. Lee and S. H. Cho, “A 1.4-μW 24.9-ppm/°C current reference with process-insensitive temperature compensation in 0.18-μm CMOS,” IEEE J. Solid-State Circuits, vol. 47, no. 10, pp. 2527–2533, Oct. 2012.DOI
  7. [7]S. Yun et al., “A 2.4/5 GHz dual-band low-noise and highly linear receiver with a new power-efficient feedforward OPAMP for WiFi-6 applications,” IEEE Access, vol. 11, pp. 137264–137273, Dec. 2023.DOI
  8. [8]M. S. Kappes, “A 2.2-mW CMOS bandpass continuous-time multibit Δ–Σ ADC with 68 dB of dynamic range and 1-MHz bandwidth for wireless applications,” IEEE J. Solid-State Circuits, vol. 38, no. 7, pp. 1098–1104, Jul. 2003.DOI
  9. [9]K. J. de Langen and J. H. Huijsing, “1-GHz operational amplifier with multipath nested Miller compensation,” in Proc. IEEE Int. Symp. Circuits Syst., May 1994, pp. 517–520.
  10. [10]A. K. George, J. Lee, Z. H. Kong, and M. Je, “A 0.8 V supply- and temperature-insensitive capacitance-to-digital converter in 0.18-μm CMOS,” IEEE Sens. J., vol. 16, no. 13, pp. 5354–5364, Jul. 2016.DOI
  11. [11]X. Zou et al., “A 100-channel 1-mW implantable neural recording IC,” IEEE Trans. Circuits Syst. I: Regul. Pap., vol. 60, no. 10, pp. 2584–2596, Oct. 2013.
  12. [12]S.-J. Kim et al., “A 0.5-V sub-µW/channel neural recording IC with delta-modulation-based spike detection,” in Proc. IEEE Asian Solid-State Circuits Conf., Nov. 2014, pp. 189–192.DOI
  13. [13]T. Lee, M. K. Kim, H. J. Lee, and M. Je, “A multimodal neural-recording IC with reconfigurable analog front-ends for improved availability and usability for recording channels,” IEEE Trans. Biomed. Circuits Syst., vol. 16, no. 2, pp. 185–199, Apr. 2022.DOI

Paper details

vlnc:W_4ne9jh5v
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents